commit 3953762847cd8f9413e4b37bd69763ed0cd10187 Author: Renewable Trading Docs Date: Sat Jul 11 00:32:27 2026 +0000 Initial commit: bilingual (EN/ZH) research & design doc set - pjm/: market primer (new, written for Chinese readers), product design (Models A-C, newsvendor, BORD, glossary), US vendor due diligence - china/: market & vendor DD (Doc 136/394, MLT-spot deep dive, six vendors, glossary), product design (Model D MLT position optimizer with math) - All documents in matched EN/ZH pairs; README with document map and conventions (incl. CJK bold-spacing rule) diff --git a/README.md b/README.md new file mode 100644 index 0000000..1358fd3 --- /dev/null +++ b/README.md @@ -0,0 +1,74 @@ +# Renewable Power Trading — Research & Design Docs +# 新能源电力交易——研究与设计文档 + +A bilingual (EN/中文) document set for a venture helping renewable generators +trade in power markets: bid optimization in **PJM** (US) and trading +optimization under the post-136 market reform in **China**. + +一套双语文档，服务于一个帮助新能源发电商参与电力市场交易的创业项目：美国 +**PJM** 市场的竞价优化，以及中国 136号文改革后的交易优化。 + +**Status 状态:** working drafts, v0.x — living documents for internal +discussion. Regulatory and vendor facts date from July 2026 and require +re-verification before external use. +工作草稿（v0.x），供内部讨论持续迭代。监管与供应商信息截至 2026 年 7 月， +对外使用前须重新核实。 + +--- + +## Document map 文档地图 + +Every document exists in both languages (`_en` / `_zh`), with matching +section structure so the two versions can be read side by side. +每份文档均有英文与中文两个版本（`_en` / `_zh`），章节结构一一对应，可对照阅读。 + +### `pjm/` — US market 美国市场 + +| Doc 文档 | EN | 中文 | Contents 内容 | +|---|---|---|---| +| Market primer 市场入门 | [market_primer_en.md](pjm/market_primer_en.md) | [market_primer_zh.md](pjm/market_primer_zh.md) | PJM explained for readers familiar with China's market — institutions, DA/RT, LMP, virtuals/FTR, RPM, BORD, tax credits, with a PJM–China comparison table. 面向熟悉中国市场读者的 PJM 入门，含中美对照表。 | +| Product design 产品设计 | [design_models_en.md](pjm/design_models_en.md) | [design_models_zh.md](pjm/design_models_zh.md) | Models A–C (price regime, probabilistic production, DA–RT spread), the newsvendor bid loop, BORD, data/licensing, build order, glossary. 模型 A–C、报童竞价闭环、BORD、数据与许可、建设顺序、术语表。 | +| Vendor due diligence 供应商尽调 | [vendor_dd_en.md](pjm/vendor_dd_en.md) | [vendor_dd_zh.md](pjm/vendor_dd_zh.md) | Six US autobidder/optimization vendors profiled (Fluence Mosaic, Ascend SmartBidder, Tyba, Gridmatic, Dexter, enspired), comparative matrix, whitespace synthesis. 六家美国竞价优化服务商档案、对比矩阵与空白研判。 | + +### `china/` — China market 中国市场 + +| Doc 文档 | EN | 中文 | Contents 内容 | +|---|---|---|---| +| Market & vendor DD 市场与供应商尽调 | [market_vendor_dd_en.md](china/market_vendor_dd_en.md) | [market_vendor_dd_zh.md](china/market_vendor_dd_zh.md) | China market primer (Doc 136/394, mechanism price, MLT–spot deep dive §1.4–1.6, bargaining dynamics), six Chinese vendors (国能日新, 远景, 清能互联, 清鹏, 兰木达, 心知能源), synthesis, glossary. 中国市场入门（含中长期深度解析）、六家中国服务商、综合研判、术语表。 | +| Product design 产品设计 | [product_design_en.md](china/product_design_en.md) | [product_design_zh.md](china/product_design_zh.md) | Chinese adaptations of Models A–C and the full Model D spec (MLT position optimizer) with the hedge-ratio mathematics, build order, open questions. 模型 A–C 中国化调整与模型 D（中长期头寸优化器）完整规格及数学推导。 | + +## Suggested reading order 建议阅读顺序 + +1. **New to PJM 不熟悉 PJM** → `pjm/market_primer_*` → `pjm/design_models_*` +2. **New to China's market 不熟悉中国市场** → `china/market_vendor_dd_*` §1 → `china/product_design_*` +3. **Business/competitive view 商业与竞争视角** → the two vendor DD docs 两份供应商尽调 → synthesis sections 综合研判章节 +4. **Modeling 建模视角** → `pjm/design_models_*` (A–C + newsvendor) → `china/product_design_*` (Model D math) + +## Conventions 体例约定 + +- **Bilingual parity 双语对应**: EN and ZH versions carry identical section + numbering; edits should land in both. 中英文版本章节编号一致；修改须同步两版。 +- **Chinese markdown bold 中文粗体**: `**` markers are padded with ASCII + spaces (`文字 **粗体** 文字`) to render correctly next to full-width CJK + punctuation in strict CommonMark renderers. 中文粗体两侧加半角空格，以规避 + CJK 标点导致的渲染问题。 +- **Terms of art 术语**: each report carries its own glossary appendix; + cross-market contrasts (e.g. BORD vs 两个细则, mechanism price vs PTC/ITC) + are defined where first used. 各报告自带术语表附录。 +- **Claims hygiene 数据口径**: vendor uplift figures are marketing claims on + non-comparable baselines; private-company data is third-party-sourced. + Treat accordingly. 服务商"提升"数字为不可比基准上的营销口径，非上市公司数据 + 转引自第三方，使用须谨慎。 +- **Point-in-time principle 时点原则**: the documents repeatedly rely on + as-of-date information sets (both for modeling and for market facts); + when updating, note the as-of date. 文档反复依赖"以当时为准"的信息集； + 更新时请注明数据时点。 + +## Provenance 来源说明 + +Documents originated from an extended design discussion (July 2026) covering +PJM public data, renewable DA bidding theory (newsvendor + spread), vendor +landscapes in both markets, China's Document-136 reform and MLT–spot +mechanics. 文档源自 2026 年 7 月的一次系列设计讨论，涵盖 PJM 公开数据、 +新能源日前竞价理论（报童模型与价差）、中美两地服务商格局、中国 136号文改革 +与中长期–现货机制。 diff --git a/china/market_vendor_dd_en.md b/china/market_vendor_dd_en.md new file mode 100644 index 0000000..50ef407 --- /dev/null +++ b/china/market_vendor_dd_en.md @@ -0,0 +1,558 @@ +# Market & Vendor Due Diligence Report: Renewable Power Trading in China + +**Prepared for:** [Renewable DA bidding venture — working name TBD] +**Date:** July 10, 2026 +**Scope:** (1) A primer on China's power market structure as it affects renewable +bidding — because it differs from PJM in ways that reshape the entire product; +(2) profiles of six representative vendors in China's renewable +forecasting / price-forecasting / AI-trading / trading-agency space; (3) a +comparative matrix and strategic synthesis. +**Status:** Working draft from public sources (policy documents, state and trade +media, broker research, company sites) retrieved July 2026, supplemented by +market knowledge. Chinese private-company data is even less standardized than US +data; funding and performance figures are as reported and must be re-verified. +Regulatory details change quarterly — verify against the latest NDRC/NEA and +provincial notices before acting. + +--- + +## 1. Market primer: how China's market differs from PJM + +Understanding the vendor landscape requires the market context, because China's +structure changes what "bid optimization" even means. + +### 1.1 The reform timeline that created this market + +- **Document 136 (发改价格〔2025〕136号, Feb 2025)** — the watershed. NDRC/NEA + ended the decades-old "guaranteed volume, guaranteed price" regime: in + principle **all wind and solar output now enters the power market**, with + prices formed by trading. All 31 provinces plus Xinjiang Corps issued + implementing schemes by end-2025 (Guangxi last, Dec 2025; Tibet the remaining + exception on some counts). By end-2025, wind+solar reached ~1.84 TW of + capacity — 47.3% of national installed capacity, for the first time exceeding + thermal. +- **The "mechanism price" (机制电价)** — 136's stabilizer: a CfD-like + out-of-market settlement. Each province sets an annual volume that qualifies; + for new ("incremental") projects the mechanism price is set by **competitive + auction** (lowest bids win; the marginal winning bid sets the price, capped). + When market prices fall below the mechanism price the grid company pays the + difference; above it, the difference is clawed back. 2026 auction results + diverge sharply by province — Shanghai cleared at ¥0.4155/kWh (near its ¥0.42 + cap), Chongqing near its cap, while Shandong solar cleared near its floor — + reflecting local supply-demand and policy stances. Mechanism-volume ratios also + diverge (e.g., Qinghai capped at 40%, Hubei 50% for utility-scale, Shandong + wind 70%, ~20 provinces at 80–90%). +- **Document 394 (发改办体改〔2025〕394号, Apr 2025)** — spot-market + acceleration: essentially nationwide spot market coverage by end-2025, with a + province-by-province timetable (Hubei and Zhejiang designated to enter formal + operation by mid-2025; Anhui and Shaanxi targeting formal operation by + mid-2026). Ancillary services (frequency regulation, etc.) to be co-cleared + with spot energy. As of early 2026, provincial spot markets in **formal + operation** include Shanxi, Guangdong, Shandong, Gansu, and Hubei, with the + rest in trial/continuous-settlement operation. +- **Feb 2026 State Council implementation opinion** on completing the national + unified power market system — pushing inter-provincial spot expansion and + breaking provincial barriers. + +### 1.2 Structural differences from PJM that reshape the product + +1. **Provincial markets, not one market.** Each province has its own rules, + clearing design, price caps/floors, and settlement details ("one province, + one policy"). A vendor must re-implement per province — the Chinese analog of + our "constraint library" is a **provincial rules library**. Inter-provincial + spot exists but is a separate, thinner layer. +2. **Mid-to-long-term (MLT) contracts dominate; spot is the residual.** Most + volume settles through annual/monthly/multi-day bilateral or listed MLT + contracts with time-of-use curves; the spot market prices the deviation. The + core renewable optimization problem is therefore **contract-curve design + + contract/spot position splitting + day-ahead quantity-price declaration**, not + pure DA-vs-RT newsvendor. The DA–RT spread logic still applies (day-ahead vs. + real-time settlement of deviations) but sits inside a heavier MLT wrapper. +3. **Renewables can "report quantity & price" or be price-takers.** 136 allows + both. As of early 2026, roughly 8 provinces (Shanxi, Shandong, etc.) let + renewables voluntarily submit DA quantity-price bids, with ~16 more committed + to follow — meaning the addressable market for true bidding optimization is + opening province by province right now. +4. **Deviation assessment ("两个细则") instead of BORD-style charges.** (BORD = + PJM's Balancing Operating Reserve Deviation charges: ex-post, cost-causation + allocation of real-time balancing uplift to participants who deviated from + their DA positions, at a daily $/MWh rate scaled to actual uplift incurred.) + China instead regulates forecast accuracy administratively: stations face + penalties for power-forecast errors and schedule deviations under grid + company rules, largely independent of what balancing actually cost the + system that day, with proceeds redistributed among generators. PJM prices + the externality; China grades the homework — in PJM you optimize deviations + against a forecastable *price*, in China you engineer forecasts against a + fixed *rulebook*. This created the compliance-driven forecasting industry + (every utility-scale station must procure a forecasting system) — a dynamic + with no US equivalent, and the reason China's forecasting vendors have + thousands of station clients. +5. **Zonal/unified provincial pricing more than nodal.** Most provinces clear at + a unified provincial price or coarse zones (some, like Guangdong/Shanxi, use + nodal prices on the generation side). Congestion alpha is smaller and + different; weather/supply-demand regime forecasting matters relatively more. +6. **Price caps are tight and floors can be negative.** Caps reference + industrial peak retail rates (commonly ~¥1–1.5/kWh); floors in several + provinces are ¥0 or slightly negative (Shandong famously printed sustained + negative prices; Shanxi's rain-driven spike in July 2025 took prices from a + few jiao to over ¥1/kWh within hours). Volatility exists but the tails are + administratively compressed relative to ERCOT/PJM scarcity pricing. +7. **Data access is thinner and fragmented.** No Data Miner equivalent. Each + provincial exchange/dispatch publishes different disclosures behind member + portals; third-party data aggregation is itself a business (and a moat). + Point-in-time archives are even more valuable — and harder — than in PJM. +8. **The buyer landscape is SOE-heavy.** The "Big Five/Six" genco groups and + provincial energy SOEs own most utility-scale renewables; they increasingly + build in-house trading teams and prefer procurement-by-tender. Private IPPs, + distributed solar aggregators, and industrial owners form the long tail that + outsources. After 136, demand for trading consulting/managed trading + exploded — one service provider reported consulting inquiries up 4x and + closed deals up 2x year-on-year (China Securities Journal reporting). + +### 1.3 What "our three models" become in China + +| Our PJM design | Chinese translation | +|---|---| +| Model A: price level/regime | Provincial spot price forecasting (DA & RT), supply-demand regime, negative-price/cap-hit classifiers — per province | +| Model B: production quantiles | Same core, but doubly monetized: deviation-assessment ("两个细则") penalty avoidance + market bidding; mandatory-procurement channel already exists | +| Model C: DA–RT spread | DA vs RT deviation settlement spread + **MLT contract vs spot basis**; plus contract curve decomposition optimization | +| *(no PJM equivalent)* | **Model D: MLT position optimizer** — forward-curve views by time segment, monthly-horizon basis forecasting, shape-risk quantification, rolling rebalancing (see §1.4–1.6) | +| Newsvendor bid optimizer | Declaration strategy (报量报价) + contract position management + mechanism-price auction strategy | +| PJM redistribution license | Provincial exchange membership, data compliance, and grid-company relationships | + +### 1.4 Deep dive: how MLT and spot interact + +**Architecture.** China's design philosophy is "MLT as ballast, spot as +corrector" (中长期为主、现货为补充) — the inverse of spot-native PJM. Policy +expects most volume (often 80%+ for thermal, lower for renewables) locked in +MLT contracts before the operating day. Critically, **in spot provinces MLT +contracts are financial, not physical**: they are contracts-for-differences +settled against spot. Dispatch is determined entirely by the spot market; the +contracts are a settlement overlay. + +**The three-layer settlement stack.** For a generator, each hour settles +approximately as: + +``` +Revenue ≈ Q_da·P_da + (Q_rt − Q_da)·P_rt + (P_c − P_da)·Q_c + [DA layer] [RT deviation layer] [MLT CfD layer] +``` + +where Q_c/P_c are contract volume/price, Q_da/P_da the day-ahead cleared +quantity/price, Q_rt/P_rt the metered output and real-time price (the CfD +reference price is usually DA, province-dependent). Two spreads therefore +drive P&L, not one: the familiar **DA–RT spread** (same newsvendor logic as +PJM for the DA declaration) and the **MLT–spot basis** — the larger driver, +because Q_c is the majority of volume. + +**The curve requirement — where renewables get hurt.** MLT contracts must +carry time-segment curves (带曲线签约; peak/flat/valley at minimum). This +creates a shape-matching problem under production uncertainty: a solar plant +contracting evening-peak segments is implicitly short hours it cannot produce +(buying them back at spot — a levered short if evening spot spikes); a solar +plant contracting the midday belly hedges exactly the hours where spot +crashes, at the worst forward prices. MLT contracting for renewables is a +portfolio optimization, not a simple hedge ratio: total volume, curve shape, +prices, and continuous adjustment are all decision variables. + +**Rolling adjustment machinery.** Annual → monthly → intra-month/multi-day +rolling windows (many provinces run D-2/D-3 windows) via bilateral, listing +(挂牌), and centralized matching (集中竞价). Sophisticated renewable traders +manage MLT like a slow futures book: conservative annual base, monthly and +intra-month true-ups as production forecasts sharpen. The real workflow spans +five or six nested decision timescales (annual auction, monthly adjustment, +multi-day rolling, DA declaration, RT) versus essentially two in PJM — the +"multi-window high-frequency trading" pain point vendors cite post-136. + +**Mechanism-price layering.** Volume inside the 机制电价 mechanism settles +through a fourth, out-of-market CfD at the auction-determined price, and +generally must not be double-hedged with MLT. A plant's expected output thus +decomposes into: mechanism volume (job: the annual auction bid), MLT-contracted +market volume (job: volume/curve/timing optimization), and open spot exposure +(job: DA declaration + spread management). + +### 1.5 Why the MLT–spot basis persists (unlike PJM forwards) + +PJM-linked forwards are pinned to expectations by an arbitrage enforcement +mechanism requiring four ingredients: **open participation** (financial +traders, no physical asset needed), **two-sidedness** (shorting as easy as +buying), **cash convergence** (futures settle against realized spot, so +systematic error is realized as P&L and bad forecasters exit), and a +**continuous term structure** (new information reprices the curve in minutes). +What survives arbitrage is a modest, insurance-like risk premium — hard to +beat, cheap to trust. + +China's MLT market lacks all four: + +1. **Participation is physical-only.** Registered members with physical + positions trade; there is no financial-participant category and no liquid + power futures covering these provinces. No outside capital can attack a + mispricing. +2. **Two-sidedness is crippled.** Position rules tied to physical scale, + deviation assessments, and compliance norms preclude meaningful speculative + shorts; corrective flow in both directions is structurally limited. +3. **Participation is mandated, not voluntary.** Coverage-ratio requirements + make generators price-inelastic sellers — calendar-driven flow (everyone + selling in the annual window) creates predictable pressure no arbitrageur + leans against. +4. **The administrative anchor.** Thermal MLT prices operate within the + coal-benchmark ±20% band; since coal units dominate MLT selling, the band + anchors the whole provincial forward curve. But the forward is thus tethered + to *administered coal economics* while spot is increasingly driven by + *renewable physics* (midday solar floods, negative prices) — two diverging + drivers with no convergence trade linking them. Price discovery happens + partly through policy revision; the basis is not just persistent but + *predictable*, with correction dates that are announcement dates. + +Amplifiers: information asymmetry is unarbitraged (superior spot modeling +converts directly into better bilateral terms, deal after deal), and coarse +granularity (annual/monthly windows, peak-flat-valley segments) leaves the +curve stale between windows. + +**The entrant's opportunity — three product surfaces:** +- **Basis forecasting as advisory alpha**: time-weighted spot forecasts per + delivery month and segment vs. prevailing MLT levels → how much to contract, + in which window, at what limit prices. The edge is not arbitraged away when + exploited; the client simply wins the negotiation. +- **Curve-shape arbitrage within contracts**: segment definitions misprice + specific hours (a solar-crushed midday inside a "flat"/"peak" segment); + decomposing contract curves against forecast hourly spot distributions finds + systematically cheap/expensive segments. +- **Timing mandated flows**: window- and calendar-driven price patterns + (annual-auction pressure, month-end true-ups) exploitable by choosing *when* + within compliance obligations to transact. + +Caveats: position limits cap how hard a client can lean into a known +mispricing; policy can reset the game overnight (band adjustments); and the +inefficiency shrinks as financial participation expands (futures pilots, +sophisticating retailers). This is a window, not a permanent feature — arguing +for building basis-modeling capability early, while counterparties are still +spreadsheets. + +### 1.6 Bargaining power in MLT negotiations + +**Cyclical layer — the balance has flipped twice in a decade:** +- *2016–2020, buyer's market:* reform phase one effectively let large users + and new retail companies extract discounts from surplus-capacity generators. +- *2021–2023, violent seller's market:* coal price explosion + Doc 1439 + (band widened to ±20%, all industrial/commercial users pushed into market) → + annual contracts cleared at or near the +20% ceiling; retail companies + caught between fixed retail commitments and soaring procurement went + bankrupt in waves; several provinces allowed contract renegotiation. +- *2024–present, back to the buyer:* massive renewable additions, softer coal, + slower demand growth; discounts returned, spot sagged in high-renewable + provinces, thermal leans on the new capacity payment (容量电价, from 2024) + rather than energy margins; 136 then pushed a wall of renewable volume into + a well-supplied market. Absent a fuel shock or demand surprise, the buyer + side holds the upper hand. + +**Structural layer — tilts independent of the cycle:** mandated contracting +makes generators price-inelastic sellers (buyer's card); generation +concentration in a few SOE groups vs. a fragmented buy side converts into real +seller pricing power in tight years (why 2021–22 flipped so violently); +government window guidance sands down both extremes of the band (policy +leverage is part of negotiating leverage); and information asymmetry currently +favors genco marketing desks over the median retailer/industrial — the +unarbitraged edge sits with sellers, partly offsetting the buyer's cyclical +advantage. + +**Renewables negotiate from the weakest chair regardless:** their outside +option is bad and known (walking away means selling into a midday spot +distribution solar itself is crushing → the "green discount" to +thermal-anchored levels, partially offset by green certificate value); their +product is defective from the buyer's view (cannot reliably deliver +evening-peak segments — either the curve excludes the valuable segments or the +renewable takes shape risk it cannot physically cover); and post-136 they must +transact. Thermal, cushioned by capacity payments and scarcity-hour leverage, +negotiates from strength or neutrality in the same window. + +**Business implications.** Prospective clients sit on the structurally weak +side of bilateral negotiations where the strong side has better desks — near- +ideal conditions for decision-support sales. The pitch is not "beat the +market" but "stop losing the negotiation": basis forecasts, segment-level +fair-value marks, walk-away prices per window. In a buyer's market, contract +*structure* (adjustment rights, curve granularity, deviation terms) is worth +more than headline price — flexibility clauses are cheap now, valuable when +the cycle turns. And track the cycle explicitly: when the balance tightens, +clients' optimal posture flips from "minimize forced selling at discounts" to +"lock the ceiling early" — a service that calls that turn a season early pays +for itself many times over. + +--- + +## 2. Vendor profiles + +Archetypes map loosely onto the US report: forecasting incumbent (Sprixin ≈ +UL/DNV+), platform major (Envision ≈ Fluence), market-design software house +(Tsintergy ≈ PCI/structural modeling), AI-agent startups (TsingRoc, Lambda ≈ +Gridmatic/Tyba), and AI managed-trading (Seniverse Energy ≈ enspired/Dexter). + +### 2.1 国能日新 Sprixin (SZSE: 301162) — the forecasting incumbent turning to trading + +**What it is.** China's renewable power-forecasting leader. Founded 2008 +(Beijing), listed on ChiNext April 2022. Core: station-side wind/solar power +forecasting systems and services (hardware + SaaS-like annual service), +grid-connection control, grid-side renewable management; newer lines in power +trading decision support, storage EMS, and virtual power plants (VPP). + +**Scale.** 2019 market share ~22% (solar) / ~19% (wind) per Frost & Sullivan — +industry #1; by 2024, ~4,345 stations served (+21% YoY) and estimated share +above 30%. Serves 400+ GW of renewable clients; customer base spans State +Grid/Southern Grid, the Big Five/Six gencos, major developers and OEMs. +Retention >95%. Forecasting service typically priced ~¥50–60k/station/year — +a useful anchor for Chinese price points (an order of magnitude below US +forecasting fees). + +**Trading push.** Post-136, Sprixin reports sharply higher inbound demand for +trading products; it offers a trading decision-support platform, trading data +services, and **managed trading (交易托管)** across provinces, iterating +per-province rule engines. Its self-developed "旷冥" weather/power large model +(v2.0 in 2025; trained on 45 years of ERA5 reanalysis plus ~6,000 stations of +measured data) feeds both forecasting and spot-price prediction with +quantity-price declaration strategy. Its VPP subsidiary holds aggregator +qualifications in Shaanxi, Gansu, Ningxia, Xinjiang, Qinghai, Zhejiang, Jiangsu, +North China, Hubei. + +**Strengths.** Distribution moat: thousands of station relationships and the +compliance-mandated forecasting channel to upsell trading; listed-company +credibility with SOEs; national service network; data from ~6,000 stations. +**Weaknesses / openings.** Trading decision science is younger than its +forecasting core; product DNA is compliance software (deviation-assessment +driven) rather than P&L-accountable trading; per-province depth varies. +**US analog.** A forecasting incumbent (UL/DNV-like) with Sprixin-scale +distribution attempting the Ascend move. + +### 2.2 远景智能 Envision (EnOS / 格林威治 / 孔明) — the platform major + +**What it is.** Envision Group's digital arm: EnOS IoT/energy-management +platform claimed to manage hundreds of GW of energy assets globally, wind OEM +data advantages, "孔明" (Kongming) weather & power forecasting AI, Greenwich +wind-farm design tools, plus storage (Envision AESC) and green-power solutions. +Post-136 it markets an integrated "forecast → trade → optimize" stack for its +OEM and asset-management clients, including trading decision support and +aggregation services. + +**Strengths.** OEM lock-in (turbine SCADA at source), group balance sheet, +international footprint, strong AI/weather team, credibility for large SOE +tenders. **Weaknesses / openings.** Software is an enabler for the hardware and +asset businesses rather than a neutral product; potential conflict with +non-Envision fleets; trading services are one line among many. +**US analog.** Fluence — hardware major whose bidding software leverages the +installed base. (金风慧能 Goldwind Huineng plays the same role for Goldwind — +winner of Southern Grid's 2024 forecasting competition — and should be tracked +in the same bucket.) + +### 2.3 清能互联 Tsintergy — the market-design software house + +**What it is.** Beijing Tsintergy Technology (清华系, founded mid-2010s from +Tsinghua EE's power-market group). Builds electricity market technical-support +and simulation systems — clearing engines, market simulation, settlement — used +by exchanges/dispatch institutions, and trading decision-support systems for +gencos and retailers. Its position on the "market operator side" of several +provincial spot builds gives it unmatched rules fidelity: it effectively wrote +or shadow-built pieces of the machinery others must reverse-engineer. + +**Strengths.** Deepest rules/clearing-engine fidelity (the "mimic the SCUC" +capability, which in China's less-transparent data regime is worth more than in +PJM); academic authority; exchange-side relationships. **Weaknesses / +openings.** B2G/B2B project-software culture (customized delivery, long cycles) +rather than SaaS; genco-side products compete with in-house teams; conflict +management between operator-side and participant-side work. +**US analog.** A hybrid of PCI (market machinery) and a structural-modeling +consultancy. *[Caveat: company specifics are from industry knowledge; verify +current product lines and references directly.]* + +### 2.4 清鹏智能 TsingRoc — the AI trading-agent challenger + +**What it is.** Beijing TsingRoc (Tsinghua-affiliated startup) bet on power- +trading AI agents in 2022. Public proof points: its AI agent placed 15th of 124 +teams (beating ~90% of human traders from retail companies) in the first +"Insurance Cup" AI power-trading competition using its *conservative* agent +(the aggressive variant backtested cheaper procurement, ¥236 vs ¥307/MWh); +championship results in other AI trading competitions; reported live uplift in +Shanxi of ~¥0.02/kWh for wind and ¥0.005/kWh for solar stations; managed-trading +cooperation with leading retail companies; reported strategy-subscription +coverage of ~30 energy companies. Combines incomplete-information game theory +with deep learning — building a "virtual market" simulator to infer supply- +demand and competitor bidding patterns from cleared prices. + +**Strengths.** Purest algorithm-first player; competition wins are effective +marketing in a market that lacks standardized performance verification; asset- +light strategy subscription scales. **Weaknesses / openings.** Young; live +track record short and concentrated (Shanxi); competition results ≠ live P&L; +faces the same "market-manipulation/algorithmic collusion" regulatory scrutiny +now emerging in Chinese policy discussion. +**US analog.** Early Gridmatic / a quant signal shop. + +### 2.5 兰木达 Lambda — the spot-market trading services specialist + +**What it is.** Shanghai-based power spot-market trading service provider +(quant-finance DNA), providing trading guidance and services to market +participants with a stated mission of improving transparency and liquidity in +high-renewable spot markets. Known in the industry for early, deep Shandong +spot-market analytics (the province with China's most extreme price shapes, +including sustained negative midday prices) and price-forecast-driven strategy +services for renewables, storage, and retailers. + +**Strengths.** Quant credibility in the hardest provincial markets; independent +(no hardware/SOE conflicts); early mover in exactly the negative-price/duck- +curve conditions that most punish naive renewable bidding. **Weaknesses / +openings.** Boutique scale; services business is people-intensive; +province-by-province expansion cost. **US analog.** A specialist quant advisory +— closest to our own starting profile of anyone on this list. *[Public +disclosure on Lambda is thin; treat this profile as directional and verify via +primary diligence.]* + +### 2.6 心知能源 Seniverse Energy — AI managed trading ("trading as a service") + +**What it is.** The energy arm of Seniverse (心知科技, weather-data company), +offering **full-process managed trading**: strategy formulation, trade +execution/monitoring, settlement analysis, and even price guarantees +(电价担保) for gencos, retailers, users, VPPs, and independent storage; plus +trading software and data services on an "AIaaS" platform stack. Marketing +emphasizes the post-136 pain points: multi-window high-frequency trading +(annual/monthly/intra-month/multi-day/spot), scarce trading talent, and the +appeal of risk-sharing with a service provider. Recognized in a 2025 global +energy-tech ranking (self-reported, sole Asian company listed). + +**Strengths.** Weather-data heritage feeding the forecast layer; the +risk-sharing/price-guarantee structure is a genuine differentiator that +resembles European route-to-market firming; TaaS matches what the post-136 +long tail actually wants to buy. **Weaknesses / openings.** Guarantee +structures consume balance sheet and create tail risk (our portfolio-correlation +concern applies with force); brand is young in power circles. +**US/EU analog.** enspired/Dexter hybrid — TaaS with firming. + +### 2.7 Secondary tier (track, not profiled) + +东润环能 (forecasting #2-tier incumbent), 中科伏瑞, 南瑞继保/国电南瑞 (grid-side +giant; also builds exchange systems), 百炼智能 "百炼智电" (spot trading & +decision platform for gencos/retailers; claims 20+ clients and material monthly +retail-side profit uplift), 晶科慧能 (Jinko's energy-services arm aggregating +third-party stations), genco captive traders (Big Five/Six marketing companies), +Huawei-Huadian AI forecasting white-paper alliance, and thousands of retail +companies (售电公司) offering agency trading — the true "Model 3" incumbents. + +--- + +## 3. Comparative matrix + +| | Sprixin 国能日新 | Envision 远景 | Tsintergy 清能互联 | TsingRoc 清鹏智能 | Lambda 兰木达 | Seniverse 心知能源 | +|---|---|---|---|---|---|---| +| **Archetype** | Forecasting incumbent → trading | Platform/OEM major | Market-machinery software | AI trading agents | Quant trading services | AI managed trading (TaaS) | +| **Ownership** | Listed (SZSE 301162) | Envision Group | Private (Tsinghua-affiliated) | Startup (Tsinghua-affiliated) | Private boutique | Private (Seniverse) | +| **Core asset** | 4,000+ station relationships; 旷冥 model | EnOS + OEM data + balance sheet | Rules/clearing-engine fidelity | Game-theoretic AI agents; competition wins | Provincial spot expertise (Shandong etc.) | Weather data + risk-sharing TaaS | +| **Renewable DA bidding relevance** | Direct (报量报价 strategy products, 托管) | Direct for its ecosystem | Indirect (decision-support systems) | Direct (agents, subscriptions, 托管) | Direct (strategy services) | Direct (full 托管 + guarantees) | +| **Provincial coverage** | National service network | National + global | Multiple provinces (operator side) | Shanxi proof; expanding | Selected spot provinces | Multi-segment, expanding | +| **Business model** | Product + annual service + 托管 | Platform + services + hardware pull | Project software + license | Subscription + 托管 share | Advisory/服务费 | 托管 fee + guarantee spread | +| **Key vendor claim (unverified)** | >30% share; 400+ GW served | 100s of GW on EnOS | Operator-grade fidelity | Beat 90% of traders; +¥0.02/kWh wind uplift | Liquidity/transparency leadership | Full-process TaaS with price guarantee | + +--- + +## 4. Synthesis: China vs. the US opportunity + +**4.1 The demand shock is real and current.** 136 turned trading capability from +optional to existential for every renewable owner in one policy stroke — +consulting volume up 4x at some providers, SOEs tendering for services, and +2025 renewable-developer net profits visibly dented by market exposure. This is +a faster, more brutal demand catalyst than anything in the US timeline. + +**4.2 But unit economics are compressed.** The compliance-forecasting anchor +(~¥50–60k/station/year) and SOE procurement culture cap willingness-to-pay far +below US levels; price caps compress the volatility that funds optimization +fees. Winning models therefore skew to (a) volume (thousands of stations × +low fee — Sprixin's game), (b) risk-sharing/托管 with performance economics +(Seniverse/TsingRoc's game), or (c) SOE enterprise deals. A US-style +$1–3k/site/month SaaS translates poorly. + +**4.3 The moats are different.** In PJM our defensible layer was nodal +congestion intelligence plus a point-in-time archive. In China it is: the +**provincial rules library** (31 markets, quarterly rule churn), **data +acquisition itself** (no Data Miner; exchange portals, member-only disclosures), +**SOE relationships and tender qualifications**, and mechanism-price **auction +strategy** (a whole product line with no US equivalent). Weather-to-power +science transfers; market machinery does not. + +**4.4 Model 2 vs Model 3 in Chinese terms.** "Model 3" already exists at +massive scale — thousands of retail companies and genco marketing arms do agency +trading — but mostly with thin analytics. The near-term winning wedge visible in +this research is **algorithms sold into/through the 托管 layer** (TsingRoc's +subscriptions to retail companies; Sprixin's 托管; Seniverse's guarantees) — +i.e., B2B2B rather than direct station SaaS. That mirrors our PJM conclusion +(rent the execution rail) with Chinese characteristics: the rail is the 售电/ +agency ecosystem rather than a QSE. + +**4.5 Regulatory watch items.** The Sep 2025 NDRC/NEA "AI+ Energy" opinion +explicitly promotes AI in trading — a tailwind — while academic/regulatory +discussion of algorithmic collusion and market-power gaming is intensifying +(Jiangsu's 2024 episode of AI systems jointly spiking evening prices to +¥4,500/MWh is the cited cautionary tale). Expect algorithm filing/audit +requirements to emerge; compliance-by-design is a future differentiator. + +**4.6 For a foreign-linked entrant specifically.** Data localization, exchange +membership, grid-company relationships, and SOE procurement rules make a +pure-import play unrealistic; realistic entries are (a) technology licensing to +a domestic partner, (b) a domestic JV with local data/ops, or (c) serving +Chinese developers' *overseas* portfolios first (Sprixin itself is exporting +along Belt-and-Road projects — the reverse flow). Note also that several +Chinese vendors are eyeing outbound expansion, so today's study subjects are +tomorrow's global competitors. + +--- + +## 5. Recommended follow-ups + +1. Province prioritization study: rank Shanxi, Shandong, Guangdong, Gansu, + Hubei (formal-operation spot provinces) by renewable bidding openness + (报量报价 status), volatility, private-IPP density, and data accessibility. +2. Primary diligence on Lambda and TsingRoc (thin public disclosure): client + references, live P&L verification methodology, team backgrounds. +3. Mechanism-price auction analytics as a wedge product: every province now + runs annual auctions; bid-strategy advisory is a low-infrastructure, + high-urgency entry point. +4. Data-supply mapping: per-province disclosure inventory (what the exchange + publishes, at what latency, under what membership) — the Chinese Data-Miner + gap analysis. +5. Partnership scan: retail companies / genco marketing arms seeking algorithm + suppliers; weather-data providers; and the competitive implications of + Sprixin's distribution if it opens a partner program. +6. Legal review: exchange membership requirements per province, algorithm + accountability trends, data-export restrictions, and JV structures for any + foreign-linked entity. +7. Model D scoping (MLT position optimizer): per priority province, inventory + the contracting windows (annual/monthly/rolling cadence and mechanics), + curve-segment definitions, coverage-ratio mandates, and renegotiation/ + adjustment rights — the inputs a contract-curve optimization product must + encode. Prototype the MLT–spot basis forecast (time-weighted spot per + delivery month and segment vs. prevailing contract levels) as the first + deliverable; annual-window advisory is board-level pain today and requires + no real-time infrastructure. +8. Bargaining-cycle monitor: a standing supply-demand balance indicator per + province (capacity additions vs. demand growth, coal price, capacity-payment + policy) to time clients' contracting posture — "minimize forced selling" + in a buyer's market vs. "lock the ceiling early" when the cycle turns. + +--- + +## Appendix: Glossary + +| Term | Definition | +|---|---| +| Document 136 (136号文) | Feb 2025 NDRC/NEA notice pushing essentially all wind/solar output into markets; the demand catalyst for this entire space | +| Document 394 (394号文) | Apr 2025 notice accelerating spot-market coverage nationwide with a provincial timetable | +| Mechanism price (机制电价) | CfD-like out-of-market settlement stabilizing part of renewable revenue; price set by annual provincial auctions for new projects | +| MLT (中长期) | Mid-to-long-term contracts (annual/monthly/rolling) — financial CfDs against spot in spot provinces; the majority of settled volume | +| MLT–spot basis | Gap between contracted price and realized time-weighted spot; the dominant P&L driver and a persistently inefficient price (§1.5) | +| Curve contracting (带曲线签约) | Requirement that MLT contracts specify time-segment (peak/flat/valley or finer) volume curves — the source of renewable shape risk | +| 报量报价 | "Report quantity and price" — renewables submitting genuine DA quantity-price bids rather than being price-takers | +| Two Detailed Rules (两个细则) | Grid-company administrative assessments penalizing power-forecast errors and schedule deviations; created the compliance forecasting market | +| BORD (contrast) | PJM's Balancing Operating Reserve Deviation charges — cost-causation allocation of balancing uplift to DA-vs-RT deviations; China's 两个细则 is the administrative counterpart | +| Coal benchmark ±20% | Administered band around provincial coal benchmark tariffs bounding thermal MLT prices — the anchor of the provincial forward curve | +| Capacity payment (容量电价) | From 2024, availability-based payment to coal (expanding scope) — reduces thermal's dependence on energy margins in negotiations | +| 挂牌 / 集中竞价 | Listing (one side posts, others accept) / centralized matching auction — the two exchange-run MLT trading formats besides bilateral | +| Retail company (售电公司) | Electricity retailer buying wholesale (MLT + spot) and selling to end users; the fragmented buy side and a key algorithm-distribution channel | +| Trading agency / 托管 | Managed trading — a service provider formulates strategy and executes trades on the asset owner's behalf; China's Model-3 analog | +| Green power / GEC (绿电/绿证) | Green power trading bundles certificates with energy at a premium; a parallel MLT channel | +| Provincial exchange (交易中心) | Province-level power trading center operating MLT windows, membership, and disclosures (e.g., 广州/北京电力交易中心 for inter-provincial) | +| Spot formal operation (正式运行) | Regulatory status after extended settlement trial operation; Shanxi, Guangdong, Shandong, Gansu, Hubei as of early 2026 | +| CfD | Contract for differences — settlement of (contract price − reference market price) × volume, no physical delivery | diff --git a/china/market_vendor_dd_zh.md b/china/market_vendor_dd_zh.md new file mode 100644 index 0000000..969acba --- /dev/null +++ b/china/market_vendor_dd_zh.md @@ -0,0 +1,219 @@ +# 市场与供应商尽职调查报告：中国新能源电力交易 + +**编制对象：** [可再生能源日前竞价创业项目 —— 名称待定] +**日期：** 2026年7月10日 +**范围：** （1）中国电力市场结构入门——聚焦其影响新能源报价的方面，因为中国市场与 PJM 的差异足以重塑整个产品形态；（2）中国新能源功率预测 / 电价预测 / AI 交易 / 交易代理领域六家代表性服务商的档案；（3）对比矩阵与战略研判。 +**状态：** 工作草稿，基于 2026 年 7 月检索的公开资料（政策文件、官方与行业媒体、券商研报、公司官网），并辅以行业常识。中国非上市公司数据的标准化程度低于美国，融资与业绩数字均为转引，须重新核实。监管细节按季度演变——行动前请对照国家发改委/国家能源局及各省最新文件核验。 + +--- + +## 一、市场入门：中国市场与 PJM 的差异 + +理解供应商格局必须先理解市场语境——中国的市场结构改变了"竞价优化"本身的含义。 + +### 1.1 塑造当前市场的改革时间线 + +- **136号文（发改价格〔2025〕136号，2025年2月）** ——分水岭。国家发改委、国家能源局终结了延续数十年的"保量保价"模式：原则上 **风电、光伏上网电量全部进入电力市场** ，电价通过交易形成。至 2025 年底，全国 31 省（区、市）及新疆兵团均已出台承接方案（广西 12 月收官；西藏在部分口径下为例外）。截至 2025 年底，风光累计装机历史性突破 18.4 亿千瓦，占全国总装机 47.3%，首次超过火电。 +- **机制电价** ——136号文的稳定器：一种类差价合约（CfD）的市场外差价结算。各省每年确定纳入机制的电量规模；增量项目的机制电价通过 **竞价** 形成（报价由低到高入选，原则上按入选项目最高报价确定，且不得高于竞价上限）。市场价低于机制电价时电网企业补差价，高于时扣回差价。2026 年各省竞价结果分化显著——上海出清 0.4155 元/千瓦时（上限 0.42 元）、重庆接近上限，山东光伏则接近下限——直接反映各地供需与政策取向。机制电量比例同样分化（如青海上限 40%、湖北集中式 50%、山东风电 70%，约 20 个省设 80–90%）。 +- **394号文（发改办体改〔2025〕394号，2025年4月）** ——现货提速：要求 2025 年底前基本实现电力现货市场全国全覆盖，并首次按省划定 20 个地区的运行时间表（湖北、浙江为先行示范区，2025 年 6 月底前转正式运行；安徽、陕西力争 2026 年 6 月底前转正）。调峰、调频等辅助服务与现货联合出清。截至 2026 年初，山西、广东、山东、甘肃、湖北等省级现货已转入 **正式运行** ，其余省份处于连续结算试运行等阶段。 +- **2026年2月国务院办公厅《关于完善全国统一电力市场体系的实施意见》** ——推动省间现货扩容、打破省间壁垒。 + +### 1.2 重塑产品形态的结构性差异（对照 PJM） + +1. **省级市场，而非单一市场。** 各省有各自的规则、出清设计、限价与结算细节（"一省一策"）。服务商必须逐省落地——我们在 PJM 的"约束库"在中国的对应物是 **省级规则库** 。省间现货存在，但是独立且更薄的一层。 +2. **中长期为主，现货结算偏差。** 大部分电量通过年度/月度/多日的双边或挂牌中长期合同（带分时曲线）结算，现货为偏差定价。因此中国新能源的核心优化问题是 **合同曲线设计 + 中长期/现货头寸切分 + 日前报量报价** ，而非纯粹的"日前对实时"报童模型。DA–RT 价差逻辑仍然成立（日前与实时对偏差的结算），但被套在更厚重的中长期外壳之内。 +3. **新能源可"报量报价"或作为价格接受者。** 136号文允许两者。截至 2026 年初，约 8 省（山西、山东等）已允许新能源自愿参与日前报量报价，另有约 16 省明确提出加快推动——这意味着真正的竞价优化的可服务市场正在逐省打开。 +4. **"两个细则"偏差考核，而非 BORD 类费用。** （BORD 即 PJM 的平衡运行备用偏差费用：将实时平衡产生的补偿成本按"成本因果"原则事后分摊给偏离日前头寸的市场主体，费率按日浮动、与当日实际发生的补偿成本挂钩。）中国则以行政方式监管预测精度：场站因功率预测偏差与计划偏差受到并网运行考核，考核力度与当日系统实际平衡成本基本无关，考核资金在发电侧内部再分配。PJM 给外部性定价，中国给作业打分——在 PJM 是对一个可预测的随机 *价格* 优化偏差，在中国是对一套固定的 *规则手册* 优化预测。这催生了合规驱动的功率预测产业（所有集中式场站都必须采购预测系统）——美国没有对应机制，这也是中国预测厂商拥有数千场站客户的原因。 +5. **以省内统一/分区价格为主，节点电价为辅。** 多数省份按全省统一价或粗分区出清（广东、山西等在发电侧采用节点电价）。阻塞阿尔法更小、形态不同；天气与供需格局预测的相对重要性更高。 +6. **限价偏紧，下限可为负。** 申报上限参考工商业尖峰电价（常见约 1–1.5 元/千瓦时）；多省下限为 0 或略为负值（山东午间持续负电价广为人知；2025 年 7 月山西一场大雨使现货价格数小时内从几毛飙升至一元以上）。波动性存在，但尾部相对 ERCOT/PJM 的稀缺定价被行政性压缩。 +7. **数据获取更薄、更碎片化。** 没有 Data Miner 的对应物。各省交易中心/调度机构的披露内容、口径与门户各不相同，且多在会员端口之内；第三方数据聚合本身即是一门生意（也是护城河）。"以报价时点为准"的数据存档在中国比在 PJM 更有价值——也更难。 +8. **买方以国企为主。** 五大六小发电集团与省级能源国企持有大部分集中式新能源；它们日益自建交易团队并偏好招标采购。民营 IPP、分布式光伏聚合商与工商业业主构成外包需求的长尾。136号文后，交易咨询/托管需求爆发——有服务商报告咨询量同比增 4 倍、成单量增 2 倍（中国证券报调研）。 + +### 1.3 "我们的三个模型"在中国的形态 + +| PJM 设计 | 中国对应 | +|---|---| +| 模型 A：价格水平/市场状态 | 分省现货电价预测（日前与实时）、供需格局、负电价/触顶分类器——逐省建设 | +| 模型 B：出力分位数 | 内核相同，但双重变现：规避"两个细则"考核 + 市场竞价；强制采购渠道已然存在 | +| 模型 C：DA–RT 价差 | 日前–实时偏差结算价差 + **中长期合同与现货的基差** ；叠加合同曲线分解优化 | +| *（PJM 无对应物）* | **模型 D：中长期头寸优化器** ——分时段远期曲线观点、月度尺度基差预测、曲线形状风险量化、滚动再平衡（详见 1.4–1.6 节） | +| 报童竞价优化器 | 报量报价策略 + 合同头寸管理 + 机制电价竞价策略 | +| PJM 数据再分发许可 | 各省交易中心会员资格、数据合规与电网企业关系 | + +### 1.4 深度解析：中长期与现货如何联动 + +**总体架构。** 中国的设计哲学是 **"中长期为主、现货为补充"** ——与"现货原生"的 PJM 正好相反。政策预期大部分电量（火电常在 80% 以上，新能源比例较低）在运行日之前锁定于中长期合同。关键在于： **在现货省份，中长期合同是金融性的而非物理性的** ——它们已被转化为对现货结算的差价合约（CfD）。调度完全由现货市场决定（日前申报与实时出清），合同只是叠加其上的结算层。 + +**三层结算结构。** 对发电企业而言，每小时的结算大致为： + +``` +收入 ≈ Q_da·P_da + (Q_rt − Q_da)·P_rt + (P_c − P_da)·Q_c + [日前层] [实时偏差层] [中长期差价层] +``` + +其中 Q_c/P_c 为合同电量/电价，Q_da/P_da 为日前中标电量/电价，Q_rt/P_rt 为实际计量电量与实时电价（差价参考价通常为日前价，因省而异）。因此驱动盈亏的价差有两个而非一个：熟悉的 **日前–实时价差** （日前申报仍适用与 PJM 相同的报童逻辑），以及 **中长期–现货基差** ——后者是更大的盈亏驱动，因为 Q_c 占电量的大头。 + +**曲线要求——新能源的痛点所在。** 中长期合同必须 **带曲线签约** （至少峰平谷分段）。这在出力不确定性下构成一个"形状匹配"问题：光伏电站若签晚高峰段合同，等于在自己无法发电的小时持有隐性空头（须按现货价买回——晚高峰现货飙升时，"套保"变成带杠杆的空头头寸）；若签午间腹部段合同，则恰好在现货最容易崩向零/负值的小时、以最差的远期价格完成了套保。对新能源而言，中长期签约是不确定出力下的组合优化问题——总电量、曲线形状、价格与持续调整都是决策变量——而非简单的套保比例。 + +**滚动调整机制。** 年度 → 月度 → 月内/多日滚动窗口（多省设 D-2/D-3 窗口），通过双边协商、挂牌与集中竞价进行。成熟的新能源交易者把中长期当作一本缓慢管理的期货账簿：年度签保守的基础电量，随出力预测精化按月、按旬调仓。真实工作流横跨五六个嵌套的决策时间尺度（年度竞价、月度调整、多日滚动、日前申报、实时）——而 PJM 设计中基本只有两个。这正是 136号文后服务商反复强调的"多窗口高频交易"痛点。 + +**机制电价的叠加。** 纳入 **机制电价** 的电量通过第四层——由电网企业按竞价形成的机制价执行的市场外差价结算——且一般不得与中长期重复套保。因此电站预期出力实际分解为三块：机制电量（任务：年度竞价报价）、中长期已签市场电量（任务：电量/曲线/时机优化）、现货敞口（任务：日前报量报价 + 价差管理）。 + +### 1.5 为什么中长期–现货基差会持续存在（而 PJM 远期不会） + +PJM 关联的远期价格被一套套利执行机制钉在预期附近，该机制需要四个要素： **开放参与** （金融交易商无需持有实物资产即可入场）、 **双向交易** （做空与做多同样容易）、 **现金收敛** （期货对已实现现货结算，系统性错误会变成交易盈亏，预测差的资金被淘汰出局）、 **连续期限结构** （新信息在几分钟内重新定价整条曲线）。套利之后残留的只是温和的、保险性质的风险溢价——难以战胜，但可以放心信赖。 + +中国中长期市场四个要素俱缺： + +1. **参与者仅限实物主体。** 只有持实物头寸的注册会员可交易；没有金融参与者类别，也没有覆盖这些省份的流动性电力期货。外部资本无法攻击错误定价。 +2. **双向性残缺。** 与实物规模挂钩的头寸规则、偏差考核与合规惯例使有意义的投机性做空基本不可行；双向的纠偏资金流在结构上受限。 +3. **参与是强制而非自愿。** 签约比例要求使发电企业成为价格无弹性的卖方——日历驱动的资金流（所有人都在年度窗口卖出）形成可预测的价格压力，而没有套利者对其逆向操作。 +4. **行政锚。** 火电中长期价格在燃煤基准价 ±20% 区间内运行；煤电是中长期的主力卖方，该区间因此锚定了全省远期曲线。但远期由此被拴在 **行政化的煤电经济性** 上，而现货日益由 **新能源物理特性** 驱动（午间光伏洪峰、负电价）——两个渐行渐远的驱动因素之间没有任何收敛交易将其连接。价格发现部分通过政策修订完成；基差不仅持续，而且 **可预测** ——其修正日期就是文件发布日期。 + +放大因素：信息不对称无人套利（更强的现货建模能力直接转化为一单又一单更优的双边条款），且颗粒度粗糙（年/月窗口、峰平谷分段）使曲线在窗口之间长期陈旧。 + +**进入者的机会——三个产品切面：** +- **基差预测即咨询阿尔法** ：按交割月、分时段的加权现货预测对照当前中长期价格水平 → 回答"签多少、在哪个窗口签、限价多少"。这一优势被利用时不会被套利消灭——客户只是赢下了谈判。 +- **合同内的曲线形状套利** ：分段定义对特定小时存在系统性错误定价（被光伏压垮的午间小时被归入"平段"甚至"峰段"）；将合同曲线对照逐小时现货预测分布做分解，可找出系统性便宜/昂贵的时段。 +- **择时应对强制性资金流** ：窗口与日历驱动的价格规律（年度竞价压力、月末调仓效应），客户可在合规义务范围内选择 *何时* 成交加以利用。 + +注意事项：头寸限制约束了客户对已知错误定价的下注力度；政策可能一夜之间重置游戏（区间调整即刻重定价整条曲线）；且随着金融参与扩大（期货试点、售电公司专业化），低效率会逐步收窄。这是一个窗口期而非永久特征——恰恰支持尽早建设基差建模能力，趁对手方还在用电子表格。 + +### 1.6 中长期谈判中的议价力量 + +**周期层——十年内天平翻转两次：** +- *2016–2020，买方市场* ：改革第一阶段实际上成为大用户与新生售电公司向过剩产能发电企业索取折价的机制。 +- *2021–2023，剧烈反转为卖方市场* ：煤价暴涨叠加 1439号文（区间放宽至 ±20%、工商业用户全部入市）→ 多省年度合同贴着 +20% 上限出清；夹在固定零售承诺与飙升采购成本之间的售电公司成批出险；多省被迫允许合同重新协商。 +- *2024 至今，重回买方* ：新能源大规模投产、煤价回落、需求增速放缓；折价重现、高新能源省份现货疲软、火电转而依靠 2024 年起的容量电价而非电能量利润；136号文又将海量新能源电量推入供给宽松的市场。若无燃料冲击或需求意外， **买方占据上风** 。 + +**结构层——独立于周期的倾斜：** 强制签约使发电企业成为价格无弹性卖方（买方的底牌）；发电侧集中于少数央企/省属集团、买方高度分散——在真正紧张的年份该集中度会迅速转化为定价权（2021–22 反转如此剧烈的原因之一）；政府窗口指导对区间两端都会"打磨"（政策杠杆本身就是谈判杠杆的一部分）；信息不对称目前有利于发电集团的营销交易团队而非中位数的售电公司/工商业用户——未被套利的优势在卖方一侧，部分抵消买方的周期性优势。 + +**新能源无论如何都坐在最弱的一把椅子上：** 其外部选择差且尽人皆知（放弃中长期意味着卖进被光伏自己压垮的午间现货分布 → 相对火电锚定价的"绿色折价"，部分由绿证价值弥补）；其产品在买方眼中有"缺陷"（无法可靠交付买方最想套保的晚高峰时段——要么曲线剔除高价值时段，要么新能源承担自己无法物理覆盖的形状风险）；且 136号文后它 *必须* 成交。而火电有容量电价托底、握有稀缺时段筹码，在同一窗口中以强势或中性姿态谈判。 + +**对本业务的启示。** 潜在客户处于双边谈判中结构性弱势的一侧，而强势一侧拥有更好的交易团队——这是销售决策支持服务近乎理想的环境。卖点不是"战胜市场"，而是 **"不再输掉谈判"** ：基差预测、分时段公允价值标尺、逐窗口的离场价格。在买方市场中，合同 *结构* （调整权、曲线颗粒度、偏差条款）比表面价格更值钱——灵活性条款现在要价便宜、周期反转时价值巨大。并且要显式跟踪周期：一旦供需重新收紧，客户的最优姿态将从"减少被迫折价卖出"翻转为"尽早锁定上限"——能提前一个季度预判拐点的服务，其价值是自身费用的许多倍。 + +类型与美国报告松散对应：预测在位者（国能日新 ≈ UL/DNV 加强版）、平台巨头（远景 ≈ Fluence）、市场机制软件商（清能互联 ≈ PCI/结构建模）、AI 智能体创业公司（清鹏、兰木达 ≈ Gridmatic/Tyba）、AI 交易托管（心知能源 ≈ enspired/Dexter）。 + +### 2.1 国能日新（深交所：301162）——转向交易的预测在位者 + +**定位。** 中国新能源功率预测龙头。2008 年成立于北京，2022 年 4 月创业板上市。核心：场站侧风光功率预测系统与服务（硬件 + 类 SaaS 年度服务）、并网智能控制、电网侧新能源管理；新兴业务包括电力交易辅助决策、储能 EMS 与虚拟电厂。 + +**规模。** 据沙利文报告，2019 年光伏/风电功率预测市占率约 22.1%/18.8%，行业第一；至 2024 年服务场站约 4,345 个（同比 +21%），测算市占率超 30%。为超过 400 GW 新能源客户提供数智化服务；客户覆盖国家电网、南方电网、五大六小及主流开发商与整机厂。客户留存率超 95%。功率预测服务均价约 5–6 万元/场站/年——这是中国价位的重要锚点（比美国预测服务费低一个数量级）。 + +**交易业务推进。** 136号文后，公司报告交易类产品咨询量显著提升；提供电力交易辅助决策平台、交易数据服务与 **交易托管** 三类产品，并按各省规则持续迭代。自研"旷冥"气象/功率大模型（2025 年迭代至 2.0；融合 45 年 ERA5 再分析资料与近 6,000 家场站实测数据），同时支撑功率预测与现货电价预测及报量报价策略。旗下国能日新智慧能源已获陕西、甘肃、宁夏、新疆、青海、浙江、江苏、华北、湖北等电网的（虚拟电厂）聚合商准入资格。 + +**优势。** 分发护城河：数千场站客户关系与合规必采的预测渠道可向上销售交易产品；上市公司身份利于服务国企；全国服务网络；约 6,000 场站的数据积累。 +**弱点 / 我方机会。** 交易决策科学的沉淀晚于预测主业；产品基因是合规软件（考核驱动）而非对盈亏负责的交易；各省深度不均。 +**美国类比。** 拥有 Sprixin 级分发能力的预测在位者（类 UL/DNV），正在尝试 Ascend 式转型。 + +### 2.2 远景智能（EnOS / 格林威治 / 孔明）——平台巨头 + +**定位。** 远景集团的数字化板块：EnOS 物联网/能源管理平台（宣称管理全球数百 GW 能源资产）、风机整机厂的数据禀赋、"孔明"气象与功率预测 AI、格林威治风电场设计工具，叠加储能（远景动力）与绿电解决方案。136号文后，其面向整机与资产管理客户营销"预测 → 交易 → 优化"的一体化组合，含交易辅助决策与聚合服务。 + +**优势。** 整机厂锁定效应（风机 SCADA 数据源头）、集团资产负债表、国际化布局、强 AI/气象团队、大型国企招标中的公信力。 +**弱点 / 我方机会。** 软件是硬件与资产业务的赋能件而非中立产品；对非远景机组存在潜在冲突；交易服务只是众多业务线之一。 +**美国类比。** Fluence——依托装机基础推软件的硬件巨头。（金风慧能对金风科技扮演同样角色——南方电网 2024 赛季功率预测竞赛一等奖——应纳入同一象限跟踪。） + +### 2.3 清能互联（Tsintergy）——市场机制软件商 + +**定位。** 北京清能互联科技（清华系，2010 年代中期由清华电机系电力市场团队孵化）。建设电力市场技术支持与仿真系统——出清引擎、市场仿真、结算——服务交易中心/调度机构；并为发电与售电企业提供交易辅助决策系统。其在多省现货建设中处于"市场运营方一侧"的位置，赋予其无可比拟的规则保真度：别人需要逆向工程的市场机器，它实际参与搭建了其中的部件。 + +**优势。** 最深的规则/出清引擎保真度（即"复刻 SCUC"的能力——在中国数据透明度更低的环境中，这一能力比在 PJM 更值钱）；学术权威；交易中心侧关系。 +**弱点 / 我方机会。** B2G/B2B 项目制软件文化（定制交付、周期长）而非 SaaS；发电侧产品与客户自建团队竞争；运营方与参与方两侧业务的利益冲突管理。 +**美国类比。** PCI（市场机器）与结构建模咨询的混合体。*[提示：公司细节部分来自行业常识，产品线与案例请直接向公司核实。]* + +### 2.4 清鹏智能（TsingRoc）——AI 交易智能体挑战者 + +**定位。** 北京清鹏智能（清华背景创业公司），2022 年押注电力交易 AI 智能体。公开验证点：首届"保险杯"AI 电力交易大赛中，其 AI 智能体在 124 支售电公司队伍中列第 15 名（超越约 90% 的人类交易员），且使用的是*保守型*智能体（激进型回测采购成本更低：236 对 307 元/MWh）；在其他 AI 交易赛事中夺冠；据报道在山西市场为风、光场站分别带来约 0.02 元与 0.005 元/千瓦时的度电收益提升；正与头部售电公司开展交易托管合作；策略订阅服务据称覆盖约 30 家能源企业。技术路线为不完全信息博弈论与深度学习结合——构建"虚拟市场"模拟器，从出清价格反推供需与对手报价模式。 + +**优势。** 最纯粹的算法先行者；在缺乏标准化业绩验证的市场中，大赛成绩是有效营销；轻资产的策略订阅可规模化。 +**弱点 / 我方机会。** 年轻；实盘记录短且集中（山西）；比赛成绩 ≠ 实盘盈亏；将直面中国政策讨论中正在浮现的"算法合谋/市场操纵"监管审视。 +**美国类比。** 早期 Gridmatic / 量化信号商。 + +### 2.5 兰木达（Lambda）——现货交易服务专家 + +**定位。** 总部上海的电力现货市场交易服务商（量化金融基因），为电力市场各方提供交易指导与服务，宣称使命为提升市场透明度、增加交易流动性，支撑高比例新能源电力系统的安全经济运行。业内以对山东现货（中国价格形态最极端的省份，含持续性午间负电价）的早期深度分析著称，为新能源、储能与售电公司提供以电价预测驱动的策略服务。 + +**优势。** 在最难的省级市场中建立的量化公信力；独立（无硬件/国企利益冲突）；在最惩罚"朴素新能源报价"的负电价/鸭子曲线环境中的先发优势。 +**弱点 / 我方机会。** 精品店规模；服务业务人力密集；逐省扩张成本高。 +**美国类比。** 专业量化咨询——是本列表中与我们自身起步画像最接近的一家。*[兰木达公开披露有限；本档案仅供方向性参考，须经一手尽调核实。]* + +### 2.6 心知能源（Seniverse Energy）——AI 交易托管（"交易即服务"） + +**定位。** 心知科技（气象数据公司）的能源板块，提供 **全流程交易托管** ：为发电企业、售电公司、电力用户、虚拟电厂、独立储能提供策略制定、交易操盘、结算分析乃至电价担保；并以"AIaaS"平台提供交易软件与数据服务。营销直击 136号文后的痛点：年度/月度/月内/多日/现货多窗口高频交易、交易人才稀缺，以及与服务商共担风险的吸引力。自称入选 2025 年某全球能源科技 15 强（亚洲唯一，自述口径）。 + +**优势。** 气象数据基因反哺预测层；"风险共担/电价担保"结构是真正的差异化，类似欧洲 route-to-market 的收益托底；TaaS 契合 136号文后长尾客户真正想买的东西。 +**弱点 / 我方机会。** 担保结构占用资产负债表并产生尾部风险（我们关于组合相关性的担忧在此尤为适用）；在电力圈的品牌尚年轻。 +**美欧类比。** enspired 与 Dexter 的混合——带收益托底的 TaaS。 + +### 2.7 第二梯队（跟踪、暂不立档） + +东润环能（预测第二梯队在位者）、中科伏瑞、南瑞继保/国电南瑞（电网侧巨头，亦建设交易中心系统）、百炼智能"百炼智电"（发售电现货交易与决策平台；宣称 20+ 客户、售电侧月均利润提升可观）、晶科慧能（晶科旗下能源服务商，聚合代理第三方场站）、发电集团自有交易公司（五大六小营销公司）、华为–华电 AI 预测白皮书联盟，以及数千家提供代理交易的售电公司——它们才是中国真正的"模式三"在位者。 + +--- + +## 三、对比矩阵 + +| | 国能日新 | 远景智能 | 清能互联 | 清鹏智能 | 兰木达 | 心知能源 | +|---|---|---|---|---|---|---| +| **类型** | 预测在位者→交易 | 平台/整机巨头 | 市场机制软件 | AI 交易智能体 | 量化交易服务 | AI 交易托管（TaaS） | +| **股权** | 上市（301162） | 远景集团 | 民营（清华系） | 创业公司（清华系） | 民营精品店 | 民营（心知科技） | +| **核心资产** | 4000+ 场站关系；旷冥大模型 | EnOS + 整机数据 + 资产负债表 | 规则/出清引擎保真度 | 博弈论 AI 智能体；大赛战绩 | 省级现货专长（山东等） | 气象数据 + 风险共担 TaaS | +| **与新能源日前竞价的相关性** | 直接（报量报价策略产品、托管） | 对自有生态直接 | 间接（辅助决策系统） | 直接（智能体、订阅、托管） | 直接（策略服务） | 直接（全流程托管+担保） | +| **省域覆盖** | 全国服务网络 | 全国 + 全球 | 多省（运营方侧） | 山西实证；扩张中 | 若干现货省份 | 多主体、扩张中 | +| **商业模式** | 产品 + 年度服务 + 托管 | 平台 + 服务 + 硬件带动 | 项目软件 + 授权 | 订阅 + 托管分成 | 咨询/服务费 | 托管费 + 担保价差 | +| **核心宣称（未经验证）** | 市占率超 30%；服务 400+ GW | EnOS 管理数百 GW | 运营方级保真度 | 胜过 90% 交易员；风电度电 +0.02 元 | 流动性/透明度引领 | 全流程 TaaS 含电价担保 | + +--- + +## 四、综合研判：中国与美国机会的对照 + +**4.1 需求冲击真实且正在发生。** 136号文以一纸政策把交易能力从"可选项"变成每个新能源业主的"生存项"——有服务商咨询量增 4 倍、国企开始招标采购服务、2025 年部分新能源企业净利润因市场敞口而阶段性下滑。这是比美国任何时点都更快、更猛烈的需求催化。 + +**4.2 但单位经济被压缩。** 合规预测的价格锚（约 5–6 万元/场站/年）与国企招标文化把支付意愿压到远低于美国的水平；限价机制压缩了本可支撑优化费用的波动率。因此胜出的模式偏向：（a）走量（数千场站 × 低费率——国能日新的打法）；（b）风险共担/托管并绑定业绩分成（心知/清鹏的打法）；（c）国企级企业大单。美式"每站每月 1–3 千美元"的 SaaS 在中国难以照搬。 + +**4.3 护城河不同。** 在 PJM，我们的壁垒是节点阻塞情报与"以报价时点为准"的数据存档。在中国，壁垒是： **省级规则库** （31 个市场、规则季度级变动）、 **数据获取本身** （无 Data Miner；交易中心门户、会员制披露）、 **国企关系与投标资质** ，以及机制电价 **竞价策略** （一条在美国不存在对应物的完整产品线）。"气象到功率"的科学可以迁移；市场机器无法迁移。 + +**4.4 中国语境下的模式二与模式三。** "模式三"在中国已经大规模存在——数千家售电公司与发电集团营销公司在做代理交易——只是大多分析能力薄弱。本次研究可见的近期制胜切口是 **把算法卖进/卖穿托管层** （清鹏向售电公司的订阅、国能日新的托管、心知的担保）——即 B2B2B 而非直接面向场站的 SaaS。这与我们在 PJM 的结论（租用执行轨道）互为镜像，只是"轨道"换成了售电/代理生态而非 QSE。 + +**4.5 监管观察项。** 2025 年 9 月发改委/能源局《关于推进"人工智能+"能源高质量发展的实施意见》明确鼓励 AI 在交易中的应用——顺风；同时学界与监管层对算法合谋与市场力操纵的讨论正在升温（2024 年夏江苏多个 AI 系统同时抬高晚高峰报价、现货瞬时冲至 4,500 元/MWh 的事件是被引用的警示案例）。可预期算法备案/审计要求将逐步出现；"合规内建"（compliance-by-design）将成为未来的差异化能力。 + +**4.6 对具有外资背景的进入者。** 数据本地化、交易中心会员资格、电网企业关系与国企采购规则使"纯进口"打法不现实；现实路径为：（a）向本土伙伴进行技术授权；（b）设立由本土方掌握数据与运营的合资公司；（c）先服务中国开发商的*海外*资产组合（国能日新自身正随"一带一路"项目出海——反向印证了流动方向）。另须注意：多家中国服务商正谋划出海，今天的研究对象就是明天的全球竞争者。 + +--- + +## 五、后续行动建议 + +1. 省份优先级研究：对山西、山东、广东、甘肃、湖北（现货正式运行省份）按新能源报量报价开放度、波动率、民营 IPP 密度与数据可得性排序。 +2. 对兰木达与清鹏智能的一手尽调（公开披露薄弱）：客户访谈、实盘业绩验证方法、团队背景。 +3. 以机制电价竞价分析作为楔子产品：各省每年组织竞价；竞价策略咨询是基础设施要求低、紧迫度高的切入点。 +4. 数据供给测绘：逐省披露清单（交易中心公开什么、延迟多少、需何种会员资格）——中国版"Data Miner 缺口分析"。 +5. 合作伙伴扫描：寻求算法供应商的售电公司/发电集团营销公司；气象数据提供商；以及若国能日新开放伙伴计划，其分发能力对竞争格局的影响。 +6. 法务审查：各省交易中心入市要求、算法问责趋势、数据出境限制，以及外资关联主体的合资架构。 +7. 模型 D 立项调研（中长期头寸优化器）：按优先省份盘点签约窗口（年度/月度/滚动的节奏与机制）、曲线分段定义、签约比例要求与重新协商/调整权——这些是合同曲线优化产品必须编码的输入。以中长期–现货基差预测（按交割月、分时段的加权现货预测对照当前合同价格水平）作为首个交付物原型；年度窗口咨询是当下董事会级别的痛点，且几乎不需要实时基础设施。 +8. 议价周期监测：建立分省供需平衡常设指标（装机增量对需求增速、煤价、容量电价政策），用于把握客户签约姿态的切换时点——买方市场中"减少被迫折价卖出"，周期反转时"尽早锁定上限"。 + +--- + +## 附录：术语表 + +| 术语 | 释义 | +|---|---| +| 136号文 | 2025 年 2 月发改委/能源局文件，推动风光上网电量原则上全部入市；本领域需求爆发的总催化剂 | +| 394号文 | 2025 年 4 月文件，加快现货市场全国全覆盖并给出分省时间表 | +| 机制电价 | 类差价合约的市场外结算机制，稳定新能源部分收入；增量项目价格由各省年度竞价形成 | +| 中长期（MLT） | 年度/月度/滚动合同——在现货省份为对现货结算的金融差价合约；占结算电量的大头 | +| 中长期–现货基差 | 合同价与实际加权现货价之差；最大的盈亏驱动，且是持续低效的价格（见 1.5 节） | +| 带曲线签约 | 中长期合同须约定分时段（峰平谷或更细）电量曲线的要求——新能源形状风险的来源 | +| 报量报价 | 新能源提交真实的日前量价申报，而非作为价格接受者 | +| 两个细则 | 电网企业对功率预测偏差与计划偏差的行政考核；催生了合规驱动的功率预测市场 | +| BORD（对照概念） | PJM 的平衡运行备用偏差费用——按成本因果将平衡补偿成本分摊给日前–实时偏差；"两个细则"是其行政化的对应物 | +| 燃煤基准价 ±20% | 围绕各省燃煤基准电价的行政区间，约束火电中长期价格——全省远期曲线的锚 | +| 容量电价 | 2024 年起对煤电（范围渐扩）按可用容量支付的电价——降低火电在谈判中对电能量利润的依赖 | +| 挂牌 / 集中竞价 | 挂牌（一方挂出、他方摘牌）/ 集中撮合竞价——双边协商之外交易中心组织的两种中长期交易形式 | +| 售电公司 | 从批发侧（中长期+现货）购电并零售给终端用户的主体；分散的买方阵营，也是算法分发的关键渠道 | +| 交易托管 | 服务商代资产业主制定策略并执行交易的托管服务；中国版"模式三" | +| 绿电 / 绿证 | 绿电交易将证书与电能量捆绑并溢价成交；与常规中长期并行的合同通道 | +| 交易中心 | 省级电力交易机构，组织中长期交易窗口、会员管理与信息披露（省间交易由北京/广州电力交易中心组织） | +| （现货）正式运行 | 长周期结算试运行之后的监管状态；截至 2026 年初含山西、广东、山东、甘肃、湖北 | +| 差价合约（CfD） | 按（合同价 − 参考市场价）× 电量结算、不涉及物理交割的合约 | diff --git a/china/product_design_en.md b/china/product_design_en.md new file mode 100644 index 0000000..6c09de6 --- /dev/null +++ b/china/product_design_en.md @@ -0,0 +1,287 @@ +# China Product Design: Forecasting & Optimization Models for Renewable Power Trading + +**Working design document — v0.1 (July 2026)** +**Companion to:** `../pjm/design_models_en.md` (PJM design, Models A–C) and +`market_vendor_dd_en.md` (market & vendor context, esp. §1.4–1.6). + +Scope: the China product line. Models A–C are imported by reference with their +Chinese adaptations summarized in §1; Model D — the MLT position optimizer, +which has no PJM equivalent — receives its full first-class specification here, +including the mathematical formulation of the contract-position problem. + +--- + +## 0. Architecture: four models, six decision timescales + +``` + ┌──────────────────────────┐ + │ Model B: Production │── daily quantiles ──────────┐ + │ (adds MONTHLY-horizon │── monthly distributions ─┐ │ + │ distributions for D) │ │ │ + └──────────────────────────┘ ▼ ▼ +┌──────────────────────┐ regimes ┌──────────────┐ ┌──────────────┐ +│ Model A: Spot price │───────────────────────────▶│ Model D: │ │ DA Declaration│ +│ level/regime (per │── monthly seg. forecasts ─▶│ MLT Position│ │ Optimizer │ +│ province, DA & RT) │ │ Optimizer │ │ (newsvendor + │ +└──────────────────────┘ │ (annual/ │ │ covariance + │ + ┌──────────────────────────┐ │ monthly/ │ │ 两个细则 terms)│ + │ Model C: DA–RT deviation │ │ rolling) │ └──────────────┘ + │ spread (per province) │─▶└──────────────┘ ▲ + └──────────────────────────┘ │ residual spot │ + │ exposure Q_open │ + └────────────────────┘ +``` + +Decision timescales: (1) mechanism-price annual auction → (2) annual MLT +window → (3) monthly adjustments → (4) intra-month/multi-day rolling → (5) DA +declaration (报量报价) → (6) RT. Model D owns (1)–(4); the DA optimizer owns +(5)–(6) with Models B/C exactly as in the PJM design. + +--- + +## 1. Models A–C: Chinese adaptations (imported by reference) + +| Model | PJM spec | Chinese delta | +|---|---|---| +| A — price level/regime | `../pjm/design_models_en.md` §1 | Per-province spot forecasting (DA & RT); negative-price/floor-hit and cap-hit classifiers; supply-demand regime. **New job: monthly-horizon time-weighted spot forecasts per contract segment (D's key input).** Structural-lite features adapt: provincial supply stack, coal price, inter-provincial flows, mechanism/MLT disclosure data | +| B — production quantiles | §2 | Same two-stage pipeline and calibration discipline. **New requirement: monthly production distributions** (sum of hourly, with intra-month correlation preserved — ensemble/scenario route preferred over pure quantile regression). Dual monetization: 两个细则 assessment avoidance + bidding | +| C — DA–RT spread | §3 | Same regime × conditional-distribution × covariance architecture on the DA-vs-RT deviation settlement. Congestion layer largely replaced by provincial supply-demand and renewable-surplus regimes (zonal pricing); 两个细则 assessment terms enter the optimizer alongside the spread (the "BORD-style" analog — see design doc §4.1 contrast) | + +Point-in-time discipline (§5.1 of the PJM doc) applies unchanged, with the +declaration deadline per province replacing the 10:30 AM ET close, and — harder +than PJM — the as-of archive must include **MLT window disclosures** (listed +prices, cleared volumes) which provincial exchanges publish irregularly and +revise. + +--- + +## 2. Model D — MLT Position Optimizer (full specification) + +### 2.1 Decision problem + +For a plant (or portfolio) in province p, at each contracting window w +(annual, monthly, rolling), choose: + +- **x_s** — contract volume to hold in each time segment s (peak/flat/valley + or finer, as the province defines), across each delivery month m; +- **limit prices** for each window's bilateral/listing/auction participation; +- **timing** — which windows to transact in, within mandate constraints; + +given: Model B's monthly production distributions, Model A's segment-level +spot forecasts, the current contract book, mechanism-price volume, and the +province's rulebook (coverage mandates, curve granularity, adjustment rights, +position limits). + +### 2.2 Objective function: the settlement identity + +Per delivery hour t (suppressing the plant index), with contract book giving +hourly contracted volume Q_c(t) = Σ_s x_s·1[t ∈ s] at volume-weighted price +P_c(t), mechanism volume Q_m(t) at mechanism price P_m: + +``` +R(t) = Q_da(t)·P_da(t) + [Q_rt(t) − Q_da(t)]·P_rt(t) (spot layers) + + [P_c(t) − P_ref(t)]·Q_c(t) (MLT CfD; P_ref usually = P_da) + + [P_m − P_ref_m(t)]·Q_m(t) (mechanism CfD) + − A(t) (两个细则 assessments) +``` + +Model D optimizes the **contract terms** {x_s, P_c} against the distribution +of the spot-dependent quantities; the DA optimizer later optimizes Q_da given +the book. Note the separability: for fixed Q_c, the DA/RT problem is exactly +the PJM newsvendor (plus assessment terms). Model D's problem is the outer +portfolio problem. + +Monthly aggregation: let month m have segment hours H_s(m). Define per +segment: + +- **W_s** = Σ_{t∈H_s} P_ref(t) / |H_s| — realized time-weighted spot for the + segment ("segment index"), a random variable at contracting time; +- **G_s** = Σ_{t∈H_s} Q_rt(t) — plant production in the segment, random, + correlated with W_s (negatively for solar in midday segments: high fleet + output depresses the index exactly when own output is high). + +Then the MLT layer's monthly P&L is Σ_s x_s·(P_c,s − W_s)·|H_s| (per-MWh +convention), and the plant's unhedged market revenue is Σ_s G_s·(segment +average capture price). The contract position's economics reduce to a +**portfolio of segment-level basis positions**: long basis (P_c,s − W_s) in +size x_s. + +### 2.3 The core optimization + +Risk-adjusted formulation (per month, extendable across months): + +``` +max_{x} Σ_s x_s·( P_c,s − E[W_s] )·|H_s| (expected basis P&L) + − λ · Risk( Σ_s [ G_s·Ŵ_s + x_s·(P_c,s − W_s)·|H_s| ] ) +s.t. L_p ≤ Σ_s x_s·|H_s| / E[Σ_s G_s] ≤ U_p (coverage mandate band) + x_s ≥ 0 segment-wise or as province allows (position/short limits) + x_s ≤ κ_s·(physical basis) (physical-scale rules) +``` + +where Risk(·) is variance or CVaR of total revenue, λ the client's risk +aversion, and Ŵ_s shorthand for the plant's capture price in segment s. Two +classical results emerge in the mean–variance case with a single segment: + +**Optimal hedge ratio.** With production G (mean μ_G), segment index W +(variance σ_W²), production-index covariance σ_GW, and contract price P_c: + +``` +x* = [ β·μ_G ] + [ (P_c − E[W]) / (2λ·σ_W²·|H|) ] + minimum- speculative tilt + variance (the basis view) + hedge +where β = −Cov(G·W-revenue, W)/Var(W)·(1/|H|) ≈ μ_G + Cov(G,W)/σ_W² +``` + +Interpretation, and the three China-specific corrections that make this more +than textbook hedging: + +1. **The minimum-variance hedge is NOT 100% of expected production.** The + correction term Cov(G,W)/σ_W² is *negative* for solar in midday segments + (own output high ⇔ index low), pushing the pure-hedge position *below* + expected production — the plant is already "short the index" through its + revenue exposure less than naively assumed, because its volumes concentrate + in low-index realizations. This is the contract-market mirror of Model C's + covariance correction, and skipping it systematically over-hedges solar. +2. **The speculative tilt is where Model A's basis forecast enters.** In PJM + we set this term to ~0 (forwards ≈ efficient). In China, per DD report + §1.5, (P_c − E[W]) is persistently non-zero and forecastable — the tilt is + a legitimate, bounded position, capped in practice by the mandate band + [L_p, U_p] and position rules rather than by the optimizer. +3. **Segments the plant cannot produce in are pure basis bets.** For s where + G_s ≈ 0 (solar in evening peak), x_s has no hedging role at all — the + minimum-variance term vanishes and x* is purely the tilt term. Default + posture: x_s ≈ 0 unless the basis view is strong; accepting such segments + to "complete the curve" in a negotiation should be priced explicitly as a + short-index position with the full W_s tail risk (evening spike exposure). + +**Multi-month, multi-window extension.** Stack months into a book; at each +window w, re-solve with (a) updated E[W_s], distributions from Models A/B, +(b) the existing book as the starting position, (c) transaction costs and +liquidity limits per window. The rolling problem is then a standard +stochastic-programming / model-predictive-control loop: solve, transact the +first stage, roll forward. Window timing (which window to transact in) is +handled by comparing the current window's achievable price against the +model's forecast of later windows' prices minus a liquidity/execution +discount — the calendar-flow patterns of DD report §1.5 (annual-window +pressure, month-end effects) enter as features of that forecast. + +### 2.4 Stage structure + +- **D1 — Segment basis forecaster.** Target: E[W_s] and quantiles per + province, delivery month, segment, at each window horizon (12m to D-2). + Features: Model A regime forecasts; provincial supply stack evolution + (capacity additions/retirements pipeline); coal price and band policy; + fleet renewable growth vs. demand growth; disclosed MLT window prices and + cleared volumes (the "forward curve" such as it is); calendar/window + dummies; policy-event flags. Baseline to beat: "W_s = current MLT price + level" (market-is-right) and "W_s = last year's realized" (naive). +- **D2 — Shape-risk quantifier.** Joint distribution of (G_s, W_s) per + segment: correlation estimation within regimes, scenario generation from + Model B ensembles pushed through Model A price scenarios. Output: the + covariance inputs and revenue distribution for any candidate book. +- **D3 — Portfolio optimizer.** The §2.3 program, with province rulebook + encoded (mandate bands, segment definitions, adjustment rights, position + limits). Output: target book per month/segment, trade list per window, + limit prices. +- **D4 — Window timing & negotiation support.** Fair-value marks per segment + (from D1), walk-away prices, counterparty-offer evaluation ("this curve at + this price = implied basis of X vs. our forecast of Y"), and + calendar-pressure timing signals. +- **D5 — Mechanism auction bidder.** The annual 机制电价 auction: bid low + enough to clear, high enough to beat the market alternative. Decision rule: + bid ≈ certainty-equivalent of the market revenue distribution (from + D1–D3's unhedged + optimally-hedged forecast) plus a margin for the + option value of mechanism coverage; clearing-price forecasting from prior + rounds, provincial caps, and competitor cost curves. + +### 2.5 Evaluation protocol + +- **D1:** pinball loss per segment/horizon vs. the two baselines above; + calibration checks concentrated on the downside tail (the basis moves that + hurt sellers). +- **D2/D3 economic backtest:** simulate the full rolling loop against + realized spot, comparing (i) mandate-minimum flat-curve naive strategy, + (ii) 100%-of-expected-production proportional hedge, (iii) the optimizer. + Metric: risk-adjusted ¥/MWh uplift and drawdown in the worst realized + month — clients in this market are loss-averse; the sales metric is as much + "avoided 2021-style blowups" as average uplift. +- **Regime-sliced:** separately score buyer's-market vs. tight-market + periods; a model tuned on 2024–26 buyer's-market data will misprice the + next tight cycle (encode the 2021–23 episode explicitly, however painful + the data work). +- **Point-in-time:** window-by-window information sets; disclosed MLT data + as-of each window, not as-revised. + +### 2.6 Interactions and portfolio risk + +- **B → D spec change:** Model B must ship monthly distributions with + intra-month correlation (scenario/ensemble form), not just daily quantiles. +- **D → DA optimizer:** the book {Q_c(t), P_c(t)} is an input to the daily + declaration problem; conversely, realized DA/RT settlement feeds D's + next-window re-solve. +- **Cross-client correlation (existential, as in Model C):** clients' + optimal tilts point the same direction (all renewable sellers, same + provincial basis view). Aggregate tilt exposure per province — total MWh + positioned away from minimum-variance across the client book — is a + first-class risk metric with stress tests against a policy-reset scenario + (band adjustment repricing the curve overnight). +- **Compliance:** negotiation-support outputs (fair values, walk-away + prices) are advisory and safe; anything resembling coordinated positioning + across clients in one province needs counsel review (the algorithmic- + collusion concern flagged in the DD report §4.5 applies to contract + markets too). + +--- + +## 3. Data requirements (Model D specific) + +Per priority province: MLT window results (listed/cleared prices and volumes +by segment — exchange disclosures, member portals), segment definitions and +their revision history, coverage-mandate notices, coal benchmark and band +notices, mechanism auction rules and past results, capacity/demand pipeline +(provincial energy bureau plans, project commissioning data), and — hardest — +a point-in-time archive of all of the above. The DD report's follow-up #4 +(data-supply mapping) is a prerequisite for D1. + +--- + +## 4. Build order + +| Phase | Deliverable | Depends on | +|---|---|---| +| D-1 | Province rulebook encoding (windows, segments, mandates, limits) for 2–3 priority provinces | DD follow-up #7 | +| D-2 | Segment basis atlas: historical W_s vs. contemporaneous MLT price levels, by province/month/segment — the empirical proof that the basis is persistent and directional | Data mapping | +| D-3 | D1 forecaster v1 (regression/GBM on the atlas features) + evaluation harness | D-2 | +| D-4 | D2 shape-risk quantifier using Model B monthly scenarios | Model B upgrade | +| D-5 | D3 optimizer + D4 negotiation-support pack (the sellable wedge: annual-window advisory) | D-3, D-4 | +| D-6 | D5 mechanism auction bidder (seasonal product, aligned to provincial auction calendars) | D-3 | +| D-7 | Rolling MPC loop + portfolio-tilt risk dashboard | D-5 | + +Note the commercial sequencing: D-2 (the atlas) is sellable as analytics +almost immediately — the same "empirical atlas first" pattern as the PJM +spread atlas — and D-5's annual-window advisory addresses the board-level +pain identified in the DD report before any real-time infrastructure exists. + +--- + +## 5. Open questions + +1. Segment granularity arbitrage: how finely do priority provinces define + segments, and where does hourly spot shape systematically escape the + segment definitions? (Determines D1's resolution.) +2. Adjustment-right valuation: monthly/rolling windows give the book option + value — should D3 price contracts with embedded flexibility as swaptions + (option-adjusted basis) rather than forwards? Likely yes at maturity; + defer to v2. +3. Green-power channel: when does bundling GECs (绿电交易) dominate plain MLT + for a given client, and does D3 need a green/plain allocation dimension? +4. Retail-side mirror: the same machinery values contracts for 售电公司 + buyers — is the buy-side a second market for Model D (and a hedge against + the seller-side cyclical position)? +5. Inter-provincial layer: as province-to-province spot and MLT expand, does + D need a cross-provincial arbitrage/allocation module for clients with + multi-province portfolios? +6. The 2021–23 dataset problem: how to encode the tight-cycle episode + (different band, different rules, renegotiations) without contaminating + the buyer's-market model — regime dummies vs. separate models. diff --git a/china/product_design_zh.md b/china/product_design_zh.md new file mode 100644 index 0000000..00e4b58 --- /dev/null +++ b/china/product_design_zh.md @@ -0,0 +1,161 @@ +# 中国产品设计：新能源电力交易的预测与优化模型 + +**工作设计文档 —— v0.1（2026年7月）** +**配套文档：** `../pjm/design_models_zh.md`（PJM 设计，模型 A–C）与 `market_vendor_dd_zh.md`（市场与供应商背景，尤其 1.4–1.6 节）。 + +范围：中国产品线。模型 A–C 以引用方式导入，其中国化调整在第一节汇总；模型 D——中长期头寸优化器，在 PJM 无对应物——在本文档中获得完整的一级规格说明，包括合同头寸问题的数学表述。 + +--- + +## 〇、总体架构：四个模型、六个决策时间尺度 + +``` + ┌──────────────────────────┐ + │ 模型 B：出力预测 │── 日度分位数 ────────────────┐ + │（为 D 新增月度尺度 │── 月度分布 ───────────────┐ │ + │ 出力分布） │ │ │ + └──────────────────────────┘ ▼ ▼ +┌──────────────────────┐ 市场状态 ┌──────────────┐ ┌──────────────┐ +│ 模型 A：现货价格 │───────────────────────────▶│ 模型 D： │ │ 日前申报优化器 │ +│ 水平/状态（分省， │── 月度分时段预测 ─────────▶│ 中长期头寸 │ │（报童模型 + │ +│ 日前与实时） │ │ 优化器 │ │ 协方差修正 + │ +└──────────────────────┘ │（年度/月度/ │ │ 两个细则项） │ + ┌──────────────────────────┐ │ 滚动） │ └──────────────┘ + │ 模型 C：日前–实时偏差价差 │─▶└──────────────┘ ▲ + │（分省） │ │ 剩余现货敞口 Q_open │ + └──────────────────────────┘ └────────────────────┘ +``` + +决策时间尺度：（1）机制电价年度竞价 →（2）年度中长期窗口 →（3）月度调整 →（4）月内/多日滚动 →（5）日前申报（报量报价）→（6）实时。模型 D 负责（1）–（4）；日前优化器负责（5）–（6），其中模型 B/C 的用法与 PJM 设计完全一致。 + +--- + +## 一、模型 A–C：中国化调整（引用导入） + +| 模型 | PJM 规格 | 中国增量 | +|---|---|---| +| A——价格水平/状态 | `../pjm/design_models_zh.md` 第 1 节 | 分省现货预测（日前与实时）；负电价/触底与触顶分类器；供需格局。 **新任务：按合同分段的月度尺度加权现货预测（D 的关键输入）。** 轻量结构化特征相应调整：省级供给堆栈、煤价、省间来电、机制/中长期披露数据 | +| B——出力分位数 | 第 2 节 | 两阶段管线与校准纪律不变。 **新要求：月度出力分布** （逐小时加总且保留月内相关性——优先采用集合/情景路线而非纯分位数回归）。双重变现：规避"两个细则"考核 + 竞价 | +| C——日前–实时价差 | 第 3 节 | 在日前对实时的偏差结算上沿用"状态分类 × 条件分布 × 协方差"架构。阻塞层大体被省级供需与新能源过剩状态取代（分区/统一定价）；"两个细则"考核项与价差一同进入优化器（即"BORD 类"机制的对应物——见 PJM 设计文档 4.1 节的对照） | + +时点纪律（PJM 文档 5.1 节）原样适用，仅以各省申报截止时刻替代美东 10:30 的日前闭市；且比 PJM 更难的是，"以当时为准"的存档必须涵盖 **中长期窗口披露数据** （挂牌价格、成交电量）——各省交易中心发布不规律且事后修订。 + +--- + +## 二、模型 D——中长期头寸优化器（完整规格） + +### 2.1 决策问题 + +对省份 p 中的电站（或组合），在每个签约窗口 w（年度、月度、滚动）决策： + +- **x_s** ——各交割月 m、各时段 s（峰/平/谷或更细，依省定义）持有的合同电量； +- **限价** ——各窗口双边/挂牌/竞价参与的价格上下限； +- **时机** ——在强制性约束范围内选择在哪些窗口成交； + +已知：模型 B 的月度出力分布、模型 A 的分时段现货预测、当前合同账簿、机制电量，以及该省规则手册（签约比例要求、曲线颗粒度、调整权、头寸限制）。 + +### 2.2 目标函数：结算恒等式 + +对每个交割小时 t（省略电站下标），合同账簿给出逐小时合同电量 Q_c(t) = Σ_s x_s·1[t ∈ s]、电量加权合同价 P_c(t)，机制电量 Q_m(t)、机制价 P_m： + +``` +R(t) = Q_da(t)·P_da(t) + [Q_rt(t) − Q_da(t)]·P_rt(t) （现货两层） + + [P_c(t) − P_ref(t)]·Q_c(t) （中长期差价；P_ref 通常=日前价） + + [P_m − P_ref_m(t)]·Q_m(t) （机制差价） + − A(t) （"两个细则"考核） +``` + +模型 D 针对现货相关量的分布优化 **合同条款** {x_s, P_c}；日前优化器随后在给定账簿下优化 Q_da。注意可分性：在 Q_c 固定时，日前/实时问题正是 PJM 的报童模型（外加考核项）。模型 D 的问题是外层的组合问题。 + +月度聚合：设月份 m 的时段小时集为 H_s(m)。按时段定义： + +- **W_s** = Σ_{t∈H_s} P_ref(t) / |H_s| ——该时段的已实现加权现货（"时段指数"），在签约时点是随机变量； +- **G_s** = Σ_{t∈H_s} Q_rt(t) ——电站在该时段的出力，随机，且与 W_s 相关（光伏在午间时段为负相关：全网出力高压低指数，恰逢自身出力也高）。 + +于是中长期层的月度盈亏为 Σ_s x_s·(P_c,s − W_s)·|H_s|（按度电口径），电站未套保的市场收入为 Σ_s G_s·（时段平均获得电价）。合同头寸的经济学归结为一个 **时段级基差头寸的组合** ：在每个时段以规模 x_s 做多基差 (P_c,s − W_s)。 + +### 2.3 核心优化 + +风险调整表述（按月，可跨月扩展）： + +``` +max_{x} Σ_s x_s·( P_c,s − E[W_s] )·|H_s| （基差期望盈亏） + − λ · Risk( Σ_s [ G_s·Ŵ_s + x_s·(P_c,s − W_s)·|H_s| ] ) +s.t. L_p ≤ Σ_s x_s·|H_s| / E[Σ_s G_s] ≤ U_p （签约比例区间） + x_s ≥ 0（或依省允许的分时段形式） （头寸/做空限制） + x_s ≤ κ_s·（实物基数） （与实物规模挂钩的规则） +``` + +其中 Risk(·) 为总收入的方差或 CVaR，λ 为客户风险厌恶系数，Ŵ_s 为电站在时段 s 的获得电价简记。在单时段的均值–方差情形下，得到两个经典结论： + +**最优套保比例。** 设出力 G（均值 μ_G）、时段指数 W（方差 σ_W²）、出力–指数协方差 σ_GW、合同价 P_c： + +``` +x* = [ β·μ_G ] + [ (P_c − E[W]) / (2λ·σ_W²·|H|) ] + 最小方差 投机倾斜 + 套保项 （基差观点） +其中 β = −Cov(G·W 收入, W)/Var(W)·(1/|H|) ≈ μ_G + Cov(G,W)/σ_W² +``` + +解读——以及使其超越教科书套保的三个中国特有修正： + +1. **最小方差套保并非预期出力的 100%。** 修正项 Cov(G,W)/σ_W² 在光伏的午间时段为 *负* （自身出力高 ⇔ 指数低），将纯套保头寸压到预期出力 *之下* ——由于电量集中在指数偏低的实现路径上，电站经由收入敞口"做空指数"的程度低于朴素假设。这是模型 C 协方差修正在合同市场的镜像；忽略它会系统性地对光伏过度套保。 +2. **投机倾斜项正是模型 A 基差预测的入口。** 在 PJM 我们将该项置零（远期 ≈ 有效）。在中国，依尽调报告 1.5 节，(P_c − E[W]) 持续非零且可预测——该倾斜是合法、有界的头寸，实践中由比例区间 [L_p, U_p] 与头寸规则而非优化器本身封顶。 +3. **电站无法出力的时段是纯粹的基差赌注。** 对 G_s ≈ 0 的时段（光伏之于晚高峰），x_s 完全没有套保功能——最小方差项消失，x* 只剩倾斜项。默认姿态：除非基差观点强烈，x_s ≈ 0；在谈判中为"补全曲线"而接受此类时段，应显式按"做空指数头寸"定价，并计入 W_s 的全部尾部风险（晚高峰飙升敞口）。 + +**多月、多窗口扩展。** 将各月叠成账簿；在每个窗口 w，用（a）模型 A/B 更新后的 E[W_s] 与分布、（b）以现有账簿为初始头寸、（c）各窗口的交易成本与流动性限制重新求解。滚动问题即标准的随机规划 / 模型预测控制（MPC）循环：求解、执行第一阶段、向前滚动。窗口择时（在哪个窗口成交）通过比较当前窗口可成交价格与模型对后续窗口价格的预测（扣除流动性/执行折价）完成——尽调报告 1.5 节的日历资金流规律（年度窗口压力、月末效应）作为该预测的特征进入。 + +### 2.4 阶段结构 + +- **D1——时段基差预测器。** 目标：分省、交割月、时段、各窗口时点（12 个月前至 D-2）的 E[W_s] 与分位数。特征：模型 A 的状态预测；省级供给堆栈演化（装机投退产管线）；煤价与区间政策；新能源装机增速对需求增速；已披露的中长期窗口价格与成交量（现有意义上的"远期曲线"）；日历/窗口虚拟变量；政策事件标志。须战胜的基线："W_s = 当前中长期价格水平"（市场正确论）与"W_s = 去年实现值"（朴素外推）。 +- **D2——形状风险量化器。** 各时段 (G_s, W_s) 的联合分布：分状态估计相关性，用模型 B 的集合情景穿过模型 A 的价格情景生成场景。输出：任意候选账簿的协方差输入与收入分布。 +- **D3——组合优化器。** 2.3 节的规划问题，内置省级规则手册（比例区间、时段定义、调整权、头寸限制）。输出：各月/时段目标账簿、各窗口交易清单、限价。 +- **D4——窗口择时与谈判支持。** 分时段公允价值标尺（来自 D1）、离场价格、对手方报价评估（"这条曲线在这个价格 = 隐含基差 X，对照我们的预测 Y"）、日历压力择时信号。 +- **D5——机制电价竞价器。** 年度机制电价竞价：报价须低到能入选、高到胜过市场化替代。决策规则：报价 ≈ 市场收入分布（来自 D1–D3 的未套保 + 最优套保预测）的确定性等值，再加机制覆盖之期权价值的边际；出清价预测基于往届结果、省级上限与竞争对手成本曲线。 + +### 2.5 评估协议 + +- **D1：** 按时段/时点的 pinball loss 对照上述两条基线；校准检验集中于下行尾部（伤害卖方的基差变动）。 +- **D2/D3 经济回测：** 对照已实现现货模拟完整滚动循环，比较（i）满足最低比例的平直曲线朴素策略、（ii）按预期出力 100% 比例套保、（iii）优化器。指标：风险调整后的 元/MWh 提升与最差实现月份的回撤——该市场的客户高度厌恶损失；销售指标与其说是平均提升，不如说是"避免 2021 式爆仓"。 +- **分状态切片：** 分别对买方市场与紧张市场时期打分；在 2024–26 买方市场数据上调优的模型会在下一个紧张周期中错误定价（无论数据工作多么痛苦，都要显式编码 2021–23 那段历史）。 +- **时点纪律：** 逐窗口的信息集；中长期披露数据以各窗口当时版本为准，而非事后修订版。 + +### 2.6 模型间交互与组合风险 + +- **B → D 的规格变更：** 模型 B 必须交付保留月内相关性的月度分布（情景/集合形式），而不只是日度分位数。 +- **D → 日前优化器：** 账簿 {Q_c(t), P_c(t)} 是每日申报问题的输入；反向地，已实现的日前/实时结算回馈 D 的下一窗口重求解。 +- **跨客户相关性（与模型 C 同样是生存级问题）：** 各客户的最优倾斜方向一致（同为新能源卖方、同一省级基差观点）。分省的合计倾斜敞口——客户账簿中偏离最小方差头寸的总 MWh——是一级风险指标，并须针对"政策重置情景"（区间调整一夜之间重定价整条曲线）做压力测试。 +- **合规：** 谈判支持类输出（公允价值、离场价格）属咨询性质、风险较低；任何形似在同一省份跨客户协同建仓的行为须经法务审查（尽调报告 4.5 节的算法合谋关切同样适用于合同市场）。 + +--- + +## 三、数据需求（模型 D 专属） + +按优先省份：中长期窗口结果（分时段的挂牌/成交价格与电量——交易中心披露、会员端口）、时段定义及其修订史、签约比例要求文件、燃煤基准价与区间文件、机制电价竞价规则与历届结果、装机/需求管线（省能源局规划、项目投产数据），以及——最难的——上述全部数据的"以当时为准"存档。尽调报告后续行动第 4 项（数据供给测绘）是 D1 的前置条件。 + +--- + +## 四、建设顺序 + +| 阶段 | 交付物 | 依赖 | +|---|---|---| +| D-1 | 2–3 个优先省份的规则手册编码（窗口、时段、比例要求、限制） | 尽调后续行动 #7 | +| D-2 | 时段基差图谱：历史 W_s 对照同期中长期价格水平，分省/月/时段——"基差持续且有方向"的经验证明 | 数据测绘 | +| D-3 | D1 预测器 v1（基于图谱特征的回归/GBM）+ 评估框架 | D-2 | +| D-4 | 使用模型 B 月度情景的 D2 形状风险量化器 | 模型 B 升级 | +| D-5 | D3 优化器 + D4 谈判支持包（可售卖的楔子：年度窗口咨询） | D-3、D-4 | +| D-6 | D5 机制电价竞价器（季节性产品，对齐各省竞价日历） | D-3 | +| D-7 | 滚动 MPC 循环 + 组合倾斜风险看板 | D-5 | + +注意商业化排序：D-2（图谱）几乎立即可作为分析产品售卖——与 PJM 价差图谱"先出经验图谱"的模式一致——而 D-5 的年度窗口咨询在任何实时基础设施建成之前，就直接命中尽调报告识别出的董事会级痛点。 + +--- + +## 五、开放问题 + +1. 时段颗粒度套利：优先省份的时段定义有多细？逐小时现货形状在哪些地方系统性地逃出时段定义？（决定 D1 的分辨率。） +2. 调整权估值：月度/滚动窗口赋予账簿期权价值——D3 是否应将含灵活性条款的合同按互换期权（期权调整基差）而非远期定价？成熟期大概率应当；暂缓至 v2。 +3. 绿电通道：对特定客户，何时捆绑绿证（绿电交易）优于普通中长期？D3 是否需要"绿电/普通"的配置维度？ +4. 零售侧镜像：同一套机器可为售电公司买方估值合同——买方侧是否构成模型 D 的第二个市场（并对冲卖方侧的周期性头寸）？ +5. 省间层：随着省间现货与中长期扩容，D 是否需要为多省组合客户提供跨省套利/配置模块？ +6. 2021–23 数据集问题：如何编码紧张周期（不同区间、不同规则、合同重谈）而不污染买方市场模型——状态虚拟变量还是分立模型？ diff --git a/pjm/design_models_en.md b/pjm/design_models_en.md new file mode 100644 index 0000000..f454a93 --- /dev/null +++ b/pjm/design_models_en.md @@ -0,0 +1,467 @@ +# Forecasting Models for Renewable Day-Ahead Bidding in PJM + +**Working design document — v0.1 (July 2026)** + +Context: We are building a bid-optimization business (Model 2: SaaS, with a likely +migration path toward Model 3: asset management / market participant services) that +helps renewable generators submit day-ahead (DA) offers into PJM. This document +captures the background and technical design of the three core forecasting models +that feed the bid optimizer, for further discussion and iteration. + +--- + +## 0. How the three models fit together + +The product's core loop is a **newsvendor-style quantity decision**: the optimal DA +commitment for a renewable plant is a quantile of its production forecast +distribution, where the quantile is chosen based on the expected DA–RT spread and +the asymmetry of settlement outcomes, corrected for the covariance between the +plant's production error and the spread. + +``` + ┌─────────────────────┐ + │ Model B: Production │──── production quantiles (per plant, per hour) + └─────────────────────┘ │ + ▼ +┌─────────────────────┐ regime features ┌──────────────────┐ +│ Model A: DA Price │──────────────────────────▶│ Bid Optimizer │──▶ DA offer curve +│ (level, coarse) │ │ (newsvendor + │ (quantity + price legs) +└─────────────────────┘ │ covariance) │ + ┌─────────────────────┐ └──────────────────┘ + │ Model C: DA–RT │──── spread quantiles (per node, per hour) + │ Spread │ + └─────────────────────┘ +``` + +Division of precision requirements: + +| Model | Precision needed | Role in the bid | +|---|---|---| +| A. DA price level | **Low** (regime/coarse) | Regime conditioning for Model C; negative-price / offer-floor tail probabilities; revenue expectations | +| B. Production | **High** (calibrated quantiles) | The quantity axis of the bid | +| C. DA–RT spread | **High** (conditional distribution, esp. tails) | Selects the production quantile to commit | + +Key insight driving this architecture: for a quantity-bidding renewable, the DA–RT +spread and own-production distribution jointly determine the optimal bid; the +absolute price level matters only (a) as a conditioning regime for the spread, +(b) in the tail near the plant's offer floor (negative prices / economic +curtailment), and (c) for reporting/hedging. + +--- + +## 1. Model A — DA Price Level (coarse / regime model) + +### 1.1 Purpose + +Not a desk-grade price forecast. Three narrowly-scoped jobs: + +1. **Regime classification** of tomorrow: e.g. {mild / normal / tight / scarce} and + {gas-marginal / coal-marginal / renewable-surplus}, used as conditioning + features by Model C. +2. **Offer-floor tail probabilities**: P(nodal LMP < floor) per hour, for both DA + and RT — the negative-price classifier. Determines the dispatch-probability + weighting of production (truncation of Model B's distribution) and the price + legs of the offer curve. +3. **Expected price levels** for client revenue reporting and hedge support. + +### 1.2 Why we do NOT replicate SCUC + +The DA price is the output of PJM's security-constrained unit commitment / +economic dispatch, so full structural replication (Dayzer / PROMOD-class) is +conceptually "correct" but practically dominated for next-day horizons: + +- **Offers are confidential** at bid time (published ~4 months later, masked). + Marginal-unit cost reconstruction errors land exactly where prices are set. +- **DA is a financial equilibrium, not a pure physical SCUC**: virtual + transactions (INCs/DECs/UTCs) push DA toward the market's RT expectation. A + physical replica misses this layer entirely. +- **Network model is CEII-restricted**; monitored constraints and operating limits + vary daily with operator judgment. Congestion — the nodal differentiator — is + the hardest component to replicate. +- **SCUC is combinatorial**: small input errors flip discrete commitment + decisions; even PJM's own solution is tolerance-gapped, not unique. +- **Out-of-market operator actions** are unobservable in advance. + +Empirical consensus: statistical/ML models beat structural replicas for +tomorrow's prices; structural models win for counterfactuals, congestion anatomy, +long horizons, and new regimes. + +### 1.3 Our approach: "mimic SCUC's logic, not its solution" (hybrid-lite) + +Build a **reduced supply-stack view** to generate features, then let a statistical +model learn the residual mapping: + +- **Stack tightness**: PJM load forecast minus available thermal capacity + (outage-feed adjusted) minus fleet renewable forecast. Strongly nonlinear — + matters above ~93% utilization. +- **Implied marginal fuel / regime flag** from the reconstructed stack and gas + signals. +- **Constraint watch-list** per client region: reconstructed from PJM's historical + binding-constraints feed + planned transmission outages (OASIS), not from a + power-flow model. + +Statistical layer: gradient-boosted models for expected level; a dedicated binary +classifier for negative-price hours (features: low load forecast, high fleet +solar/wind forecast, weekend/holiday, spring/fall, local congestion conditions). + +### 1.4 Data + +| Data | Source | Cost | +|---|---|---| +| DA/RT LMPs (energy/congestion/loss components) | PJM Data Miner (`da_hrl_lmps`, `rt_hrl_lmps`, `rt_fivemin_hrl_lmps`) | Free | +| Load forecast, metered load | PJM Data Miner | Free | +| Fleet wind/solar forecasts + actuals | PJM Data Miner | Free | +| Generation outages (aggregate MW) | PJM Data Miner | Free | +| Binding constraints | PJM Data Miner | Free | +| Planned transmission outages | PJM OASIS | Free | +| Nuclear unit status | NRC daily power reactor status | Free | +| Gas: Henry Hub spot, NYMEX settlements | EIA, CME | Free-ish | +| Gas: eastern basis (TETCO M3, Transco Z6, Dominion S, TCO) | Platts / NGI / ICE | Paid — highest-value paid dataset | +| Weather forecasts (temp, wind, irradiance) + ensemble spread | NOAA (GFS/HRRR/NDFD), ECMWF open data | Free (self-archive!) | +| Neighboring ISO prices/loads (MISO, NYISO) | Their public portals | Free | + +### 1.5 Known structural breaks to encode + +- Fleet turnover: retirements, new entry (a 2019-trained model misprices 2026). +- Reserve/ORDC market rule changes (flag dates; distrust pre-change history). +- PTC/ITC fleet composition drift (see §4): as neighboring PTC plants age out of + their 10-year windows, historical negative-price frequency becomes a biased + predictor. + +--- + +## 2. Model B — Probabilistic Production Forecast (per plant) + +### 2.1 Purpose and target + +For every hour of tomorrow, by ~10:00 AM ET (ahead of the 10:30 DA close): a +**calibrated quantile set** of the plant's *available* production, e.g. P10 / P25 / +P50 / P75 / P90. The bid optimizer commits a quantile of this distribution — a +point forecast is structurally insufficient. + +Horizon: ~14–38 hours at bid time ⇒ NWP (numerical weather prediction) carries +essentially all the signal; persistence is worthless at this range. + +### 2.2 Architecture: two-stage pipeline + +**Stage 1 — Weather-to-power model (per site).** +Learn the mapping NWP forecast → SCADA output with gradient-boosted trees (or +similar). This absorbs the plant's true power curve, wake losses, inverter +clipping, terrain effects, soiling, and the NWP model's local biases. + +> **Critical rule: train on forecast weather, not measured weather.** The model +> must learn NWP's error characteristics end-to-end; training on met-mast actuals +> then predicting from NWP degrades live performance. + +**Stage 2 — Distributional layer.** Three interchangeable routes: + +1. **Quantile regression** (pinball loss, one model per quantile) — default + starting point; native in modern GBM libraries. +2. **NWP ensembles** (e.g. ECMWF 51-member): push each member through Stage 1 → + 51 scenarios → empirical quantiles. Preserves temporal correlation across + hours (valuable for multi-hour bid coupling). +3. **Analog / error-dressing**: dress the point forecast with the historical error + distribution from similar conditions. + +### 2.3 Calibration — the core quality bar + +A forecast is calibrated if actuals fall below the stated P10 ~10% of the time, +etc. (verify with reliability diagrams). Miscalibration translates one-for-one +into settlement losses, because the entire bid strategy is "commit quantile q." + +- **Vendor recalibration layer**: even with purchased forecasts, recalibrate + in-house via quantile mapping against our own SCADA archive. Small project, + high ROI. Archive every vendor forecast ever received. +- **Curtailment contamination**: SCADA records *produced*, not *producible*. + Reconstruct available power (turbine-anemometry / unconstrained inverter + capability signals); flag curtailed intervals; never calibrate on contaminated + ground truth. +- **Availability separation**: outage-driven shortfalls are not weather error; + scale ground truth to available capacity or feed availability as a feature. + Note the asymmetry — unplanned outages only subtract (skews the lower tail). + +### 2.4 Build vs. buy + +- **Buy** (vendors: Solargis, Meteomatics, UL, DNV, many others; ~$1–3k/site/month) + when starting; always keep the in-house recalibration + evaluation layer. Score + vendors quarterly with pinball loss on our SCADA; running two vendors in + parallel and blending often beats either. +- **Build** at portfolio scale: pipeline cost amortizes; enables custom targets + (available power under our curtailment logic), guaranteed point-in-time + archive, direct optimizer integration. +- **Business-model advantage (cross-client learning)**: with many plants' SCADA + under management, weather-to-power models transfer across sites, NWP bias is + calibrated regionally, and fleet-wide forecast-error days (the spread driver) + become observable. Contracts must permit pooled/anonymized model training from + day one. + +### 2.5 Accuracy expectations (day-ahead horizon) + +- Solar point forecasts: ~5–10% of capacity RMSE (climate-dependent). +- Wind: ~8–15% of capacity (cube-law amplification of speed errors). +- But judge the **distribution**: pinball loss across quantiles + tail-focused + reliability, benchmarked vs. climatology-dressed persistence. + +--- + +## 3. Model C — DA–RT Spread Model (per node) + +### 3.1 Purpose + +The money model. Produces per-hour, per-node **conditional quantiles of +S = DA − RT**, which select the production quantile to commit (newsvendor), with +a covariance correction for the client's own production error. + +### 3.2 Statistical character of the target + +- **Near-zero unconditional mean by construction**: virtuals arbitrage away any + persistent gap; what remains is a small conditional risk premium (DA tends to + run rich into expected scarcity) plus transient inefficiencies. Low + signal-to-noise; expect modest R². Value lives in conditioning and tails. +- **Violently asymmetric tails**: DA is a smoothed expectation; RT is the spiky + realization (scarcity adders → $850+ prints; renewable surplus → negative RT). + Fat left tail (RT spike above DA), moderate right tail (RT crash). Gaussian + assumptions are disqualifying. +- **Weak day-over-day autocorrelation, strong conditional structure** (hour, + season, tightness, weather uncertainty). + +### 3.3 Decomposition (model the two components separately) + +``` +S_node = S_system (energy/hub component) + S_congestion (ΔDA−RT congestion at node) +``` + +- **S_system**: total supply-demand — load forecast error, fleet renewable + forecast error, post-DA forced outages, reserve scarcity. Built once, shared + across all clients. +- **S_congestion**: constraints binding in RT but not priced DA (or vice versa) — + transmission forced outages, unexpected flows. Often dominant for renewable + pockets and the *more predictable* component (RT congestion persistence during + outages). Per-node work; **our defensible IP layer**. + +Directly computable from Data Miner's LMP component breakdown for both markets. + +### 3.4 Architecture: regime classifier × conditional distribution × covariance + +**Stage A — Spike/regime classifiers (discrete events that dominate P&L):** +- P(RT spike above DA): tightness, reserve margin, extreme-temperature forecasts + *and their ensemble uncertainty*, high fleet renewable forecast (underdelivery + risk), recent forced outages, day type. +- P(RT crash / negative RT): fleet renewable forecast, low load, inflexible + baseload share, local constraint state. **Shares infrastructure with Model A's + negative-price classifier.** + +**Stage B — Conditional spread quantiles**, given regime probabilities: quantile +GBM on continuous features, or analog draws from regime-matched historical spread +distributions. + +**Stage C — Production–spread covariance correction (business-specific edge):** +fleet-overproduction days crash RT exactly when our clients overproduce — the +production error and S are negatively... [correlated such that overscheduling is +punished]. Implement by conditioning the spread model on the client's own +production forecast error, or estimate within-regime correlation and adjust the +newsvendor quantile analytically. Generic price shops skip this; we must not. + +### 3.5 Feature set, ranked by expected alpha + +1. **Tightness** (load forecast − available capacity), nonlinear above ~93%. +2. **Renewable forecast level and *revision velocity*** across recent NWP cycles + (late revisions ⇒ DA cleared on stale info ⇒ spread opportunity). +3. **Weather forecast uncertainty** (ensemble spread) — RT volatility fuel. +4. **Trailing RT-vs-DA congestion at the node** + constraint binding frequency, + cross-referenced with planned transmission outages. +5. **Aggregate cleared virtual volumes** (published with lag) — arbitrage + efficiency regime. +6. **Calendar interactions** (hour × season) + structural-break flags (rule + changes). + +### 3.6 Evaluation protocol + +- **Pinball loss vs. the brutal baseline S ≡ 0** ("market is efficient"). Beating + it consistently out-of-sample under point-in-time discipline is hard; a huge + backtest win ⇒ hunt for leakage first. +- **Economic backtest**: full newsvendor loop vs. naive P50 bidding; target + metric is $/MWh uplift (good implementations: ~$0.5–2/MWh). +- **Tail calibration specifically** (P5/P95), and **regime-sliced** evaluation — + average-fine models are often terrible in the ~30 days/year that drive annual + P&L. + +### 3.7 Portfolio risk (existential, not statistical) + +Clients are collectively long renewables ⇒ model errors are **correlated across +the book**. A wrong regime call on a fleet-overproduction day is wrong for every +wind client simultaneously. First-class risk metric from day one: aggregate MWh +leaning long DA across clients, stress-tested against spike scenarios. + +--- + +## 4. Cross-cutting: offer floors, tax credits, and the price legs + +The quantity bid (from Models B + C) pairs with **price legs** set by each asset's +true marginal cost, which is determined by its tax-credit election: + +| Asset type | Marginal cost | Rational offer floor | Runs at negative prices? | +|---|---|---|---| +| PTC (in 10-yr window) | ≈ −(PTC × tax gross-up) | ~−$25 to −$35/MWh | Yes, to the floor | +| ITC | ≈ $0 | ~$0/MWh | No | +| Post-PTC-window (yr 11+) | ≈ $0 | ~$0/MWh | No | + +- PTC (≈$27.50–30/MWh, inflation-adjusted, wage/apprenticeship-compliant, paid on + generation for 10 years) shifts the floor negative; ITC (30%+ of capex, paid on + investment) does not affect marginal cost. +- IRA made credits tech-neutral (45Y/48E) from 2025 — new high-CF solar + increasingly elects PTC ⇒ solar fleet also develops negative floors (structural + shift). OBBBA (July 2025) accelerated wind/solar phase-out (placed-in-service + generally by end-2027, with begin-construction safe harbor) — existing plants + keep locked-in credits. **Verify current guidance per project.** +- **Interaction with Model A's tail job**: expected production for bidding is + *dispatch-probability-weighted*: E[dispatched output] ≈ physical forecast × + P(LMP ≥ floor), computed per hour, separately for DA and RT (the + DA-negative/RT-positive and DA-positive/RT-negative cases settle very + differently). Naive use of the physical forecast overcommits most on exactly + the high-output days when curtailment risk peaks. +- Deviation settlement nuance: underdelivery against a DA position during + negative RT prices can be *profitable* (sell DA positive, buy back negative) — + which is precisely why deviation charges and must-offer rules exist. Any + autobidder logic near this line needs compliance review (Market Monitor + attention risk). Verify current PJM settlement/deviation rules; they change. + +### 4.1 BORD-style deviation charges (first-class optimizer input) + +"BORD" = PJM's **Balancing Operating Reserve Deviation** charges — a cost- +*allocation* mechanism that functions as a de facto penalty on deviating from +the day-ahead position. + +- **Mechanism.** PJM incurs uplift (make-whole payments to units committed or + dispatched whose LMP revenues don't cover offered costs). The *balancing* + portion — costs arising after DA close, i.e., from RT diverging from the DA + plan — is allocated substantially by cost causation to **deviations**: + generators off their DA schedule (either direction), load off its bid, + virtuals (which never deliver by construction). Each deviation-MWh attracts + a $/MWh charge rate that varies daily with uplift incurred and total + deviation-MWh. +- **Effect on the bid.** Adds a second term to the deviation leg of the + newsvendor objective: every MWh of |Q_rt − Q_da| picks up the charge. It is + a friction/transaction cost on leaning away from expected production — the + optimal quantile shifts back toward P50 as the expected charge rate rises. + The optimizer therefore needs the deviation charge rate (or a forecast of + it) alongside the spread forecast. +- **Loophole patch.** Deviation charges (plus Market Monitor scrutiny) are + what make systematically engineering profitable underdelivery at negative RT + prices costly — see the nuance bullet above. +- **Caveat.** PJM's uplift allocation has been litigated and reformed + repeatedly (cost buckets, deviation definitions, intermittent-resource + exemptions, netting). The settlement module must pull current parameters + from the live tariff, never from memory. "BORD-style" is shorthand for the + category: **ex-post, cost-causation-based charges on DA-vs-RT deviations.** +- **Contrast (for the China work).** China's 两个细则 achieves deviation + discipline through *administrative performance assessment* — penalties + against regulated forecast-accuracy/schedule-compliance thresholds, largely + independent of actual system balancing cost. PJM prices the externality + (optimize deviations against a forecastable stochastic *price*); China + grades the homework (engineer forecasts against a fixed *rulebook*). This is + also why China's forecasting market became a compliance-procurement market + with no PJM equivalent. + +--- + +## 5. Data and operational infrastructure (shared foundations) + +### 5.1 Point-in-time discipline (non-negotiable, applies to all three models) + +Every feature must reflect the information set available **before the 10:30 AM ET +DA close** (in practice, snapshot at ~10:00): + +- Weather: the NWP cycle actually available then (e.g. 00Z run), not later cycles, + never realized weather. +- PJM feeds: as-of snapshots (feeds get revised). +- Vendor forecasts: archive as-received (vendors overwrite). +- Plant availability: as known at bid time. + +Any post-close information in training data ⇒ backtests flatter, live +underperforms "mysteriously." The **as-of archive is a compounding business +asset** that cannot be reconstructed retroactively — competitors can't buy it. + +### 5.2 PJM access and licensing (business-model obligations) + +- Data Miner data is **internal-use only; redistribution of data or derivatives + requires a PJM redistribution license** (Associate Membership at minimum). + Whether our forecasts count as "derived data" must be resolved with PJM + + counsel **before product launch**. +- Rate limits: 6 connections/min (non-member) vs. 600 (member) — membership is + operationally necessary at portfolio scale anyway. +- Bid submission (Model 2 "software agent" variant / Model 3): Markets Gateway + API (separate, authenticated), credentialing, account-security rules for + third-party agents — confirm current requirements with PJM member services. +- Commercial data (gas indices, weather vendors): internal-use vs. redistribution + tiers priced very differently; contract accordingly. + +### 5.3 Settlement feedback loop + +Clients' actual PJM settlement statements are ground truth for validating Model C +and the economic backtest. Contract for access to settlement data (validation + +model training rights) in every client agreement. + +--- + +## 6. Build sequence (recommended) + +| Phase | Deliverable | Models touched | +|---|---|---| +| 1 | **Empirical spread atlas**: per node-cluster/hour/season spread distributions, negative-RT and spike frequencies. Zero ML; sellable as analytics; seeds Stage-B analog library. | C | +| 2 | **Point-in-time archive pipeline** (PJM feeds, NOAA/ECMWF, self-archived from day one). | A, B, C | +| 3 | **System regime classifiers** (spike/crash) on free PJM + NOAA data; negative-price classifier (shared A/C). | A, C | +| 4 | **Nodal congestion layer**: trailing RT-vs-DA congestion features + transmission-outage cross-reference for actual client nodes. | C | +| 5 | **Production pipeline**: vendor-based with in-house recalibration; migrate to in-house Stage-1/Stage-2 at portfolio scale. | B | +| 6 | **Joint production–spread covariance correction** per client (needs their SCADA). | B + C | +| 7 | **Newsvendor bid optimizer** + economic backtesting harness vs. P50-naive baseline. | All | +| 8 | Portfolio-correlation risk dashboard (aggregate DA lean, spike stress tests). | C / risk | +| 9 | Second-order upgrades: distributional deep learning, joint scenario generation, structural stack refinement. | All | + +--- + +## 7. Open questions for next discussion + +1. Newsvendor math in full: closed-form optimal quantile with the covariance + correction and deviation-charge asymmetry — worth deriving and unit-testing. +2. Target client nodes/regions: western wind (ComEd/AEP) vs. Mid-Atlantic solar + have very different congestion stories; prioritizes the Phase-4 constraint + library. +3. Vendor bake-off design for production forecasts (which two vendors, scoring + protocol, blend rule). +4. Hybrid/storage clients: brings the price-level *shape* forecast back as + first-order (arbitrage is a level-shape problem) — roadmap trigger point. +5. Intraday/rebidding scope: PJM rebid windows and RT strategy are out of scope + for v1 but affect architecture (temporal correlation needs from Model B). +6. Compliance review checklist for autobidder logic near deviation-settlement + edges; agent-framework clarification with PJM member services. +7. Must-offer / capacity-resource obligations per client asset — shifts optimal + quantiles and constrains offer flexibility; needs current-rules verification. + +--- + +## Appendix: Glossary + +| Term | Definition | +|---|---| +| LMP | Locational Marginal Price — nodal energy price, decomposed into energy + congestion + loss components | +| DA / RT | Day-ahead market (cleared ~10:30 AM ET for next day, hourly) / real-time market (5-minute dispatch and settlement) | +| DA–RT spread (S) | S = DA − RT price; the object of Model C; near-zero mean by arbitrage, fat asymmetric tails | +| Newsvendor bid | Optimal DA quantity = a quantile of the production distribution, where the quantile is set by the expected spread and settlement asymmetry | +| Quantile / pinball loss | Value below which the outcome falls X% of the time / the loss function that trains and scores quantile forecasts | +| Calibration | Property that stated probabilities are honest (P10 exceeded ~90% of the time); checked with reliability diagrams | +| SCUC | Security-Constrained Unit Commitment — the mixed-integer optimization the ISO runs to clear the DA market | +| Virtuals (INC/DEC/UTC) | Purely financial DA positions that never deliver physically; arbitrage DA toward expected RT | +| Uplift | Make-whole payments to units whose market revenues don't cover offered costs (startup, no-load, min-run) | +| BORD | Balancing Operating Reserve Deviation charges — PJM's allocation of real-time balancing uplift to DA-vs-RT deviations (§4.1) | +| ORDC / scarcity adders | Operating Reserve Demand Curve — administrative price adders when reserves run short; drives RT spikes | +| FTR | Financial Transmission Right — hedge/speculative instrument on DA congestion between two nodes | +| PTC / ITC | Production Tax Credit ($/MWh generated, 10 years; sets negative offer floors) / Investment Tax Credit (% of capex; floor ≈ $0) — see §4 | +| Offer floor | The price leg below which the plant prefers curtailment: ≈ −(PTC × tax gross-up) for PTC assets, ≈ $0 for ITC assets | +| Must-offer / RPM | Obligation of capacity resources to offer into the DA market / PJM's capacity market (Reliability Pricing Model) | +| BTM | Behind-the-meter (e.g., rooftop solar netted out of observed load) | +| NWP | Numerical Weather Prediction (GFS, HRRR, ECMWF); ensembles = many perturbed runs, source of forecast-uncertainty features | +| SCADA | Plant supervisory control and data acquisition — ground-truth production, availability, curtailment records | +| Point-in-time discipline | Every training/backtest feature must reflect only information available before the 10:30 AM DA close (§5.1) | +| Data Miner 2 / Markets Gateway | PJM's public data API / PJM's authenticated bid-submission API | +| QSE / scheduling agent | Entity with market-participant infrastructure that submits offers and handles settlement on an asset's behalf | diff --git a/pjm/design_models_zh.md b/pjm/design_models_zh.md new file mode 100644 index 0000000..4f3fabb --- /dev/null +++ b/pjm/design_models_zh.md @@ -0,0 +1,300 @@ +# PJM 新能源日前竞价的预测模型设计 + +**工作设计文档 —— v0.1（2026年7月）** + +背景：我们正在建设一项竞价优化业务（模式二：SaaS，并预留向模式三——资产管理/市场主体服务——迁移的路径），帮助可再生能源发电商向 PJM 提交日前（DA）报价。本文档记录支撑竞价优化器的三个核心预测模型的背景与技术设计，供后续讨论与迭代。英文原版：`design_models_en.md`。 + +--- + +## 〇、三个模型如何协同 + +产品的核心循环是一个 **报童式电量决策** ：可再生电站的最优日前承诺量是其出力预测分布的某个分位数，分位数的选取取决于预期的日前–实时（DA–RT）价差与结算结果的不对称性，并按电站出力误差与价差之间的协方差进行修正。 + +``` + ┌─────────────────────┐ + │ 模型 B：出力预测 │──── 出力分位数（分电站、分小时） + └─────────────────────┘ │ + ▼ +┌─────────────────────┐ 状态特征 ┌──────────────────┐ +│ 模型 A：日前价格 │──────────────────────────▶│ 竞价优化器 │──▶ 日前报价曲线 +│ （水平，粗颗粒） │ │ （报童模型 + │ （电量腿 + 价格腿） +└─────────────────────┘ │ 协方差修正） │ + ┌─────────────────────┐ └──────────────────┘ + │ 模型 C：DA–RT 价差 │──── 价差分位数（分节点、分小时） + └─────────────────────┘ +``` + +精度要求的分工： + +| 模型 | 所需精度 | 在竞价中的角色 | +|---|---|---| +| A. 日前价格水平 | **低** （状态/粗颗粒） | 为模型 C 提供状态条件；报价下限附近的负电价尾部概率；收入预期 | +| B. 出力 | **高** （经校准的分位数） | 竞价的电量轴 | +| C. DA–RT 价差 | **高** （条件分布，尤其尾部） | 选取承诺的出力分位数 | + +驱动该架构的核心洞察：对按电量竞价的可再生电站，DA–RT 价差与自身出力分布共同决定最优报价；绝对价格水平只在三处相关：（a）作为价差的条件状态，（b）报价下限附近的尾部（负电价/经济性弃电），（c）报告与套保。 + +--- + +## 一、模型 A——日前价格水平（粗颗粒/状态模型） + +### 1.1 目的 + +并非交易台级的价格预测。只承担三项窄范围任务： + +1. 对明日的 **状态分类** ：如 {宽松 / 正常 / 紧张 / 稀缺} 与 {气电边际 / 煤电边际 / 新能源过剩}，作为模型 C 的条件特征。 +2. **报价下限尾部概率** ：逐小时的 P(节点 LMP < 下限)，对日前与实时分别计算——即负电价分类器。它决定出力的"中标概率加权"（对模型 B 分布的截断）与报价曲线的价格腿。 +3. 用于客户收入报告与套保支持的 **预期价格水平** 。 + +### 1.2 为什么我们 **不** 复刻 SCUC + +日前价格是 PJM 安全约束机组组合/经济调度（SCUC/SCED）的输出，因此完整的结构化复刻（Dayzer/PROMOD 级）在概念上"正确"，但在次日尺度上实践中处于劣势： + +- **报价在竞价时保密** （约 4 个月后脱敏发布）。边际机组的成本重构误差恰好落在定价之处。 +- **日前市场不是纯物理 SCUC** ：虚拟交易（INC/DEC/UTC）把日前价推向市场对实时的预期。物理复刻完全错过这一金融层。 +- **网络模型属 CEII 受限信息** ；每日监控的约束与运行限额随调度员判断变化。阻塞——节点间的差异来源——是最难复刻的成分。 +- **SCUC 是组合优化** ：微小的输入误差会离散地翻转机组组合决策；即便 PJM 自己的解也带容差、并不唯一。 +- **市场外的调度员行为** 事前不可观测。 + +经验共识：对明日价格，统计/机器学习模型胜过结构化复刻；结构化模型的优势在反事实分析、阻塞解剖、长期尺度与新格局。 + +### 1.3 我们的路线："复刻 SCUC 的逻辑，而非它的解"（轻量混合） + +构建一个 **简化的供给堆栈视图** 生成特征，再由统计模型学习残差映射： + +- **堆栈紧张度** ：PJM 负荷预测减去可用火电容量（按停机数据调整）再减去全网新能源预测。强非线性——在利用率约 93% 以上才显著。 +- 由重构堆栈与气价信号得出的 **隐含边际燃料 / 状态标志** 。 +- 按客户区域维护的 **约束观察清单** ：由 PJM 历史绑定约束数据与计划检修（OASIS）重构，而非依赖潮流模型。 + +统计层：梯度提升模型预测期望水平；单独的二分类器识别负电价小时（特征：低负荷预测、高全网风光预测、周末/节假日、春秋季、局部阻塞状态）。 + +### 1.4 数据 + +| 数据 | 来源 | 成本 | +|---|---|---| +| 日前/实时 LMP（电能/阻塞/损耗分量） | PJM Data Miner（`da_hrl_lmps`、`rt_hrl_lmps`、`rt_fivemin_hrl_lmps`） | 免费 | +| 负荷预测、计量负荷 | PJM Data Miner | 免费 | +| 全网风/光预测与实际 | PJM Data Miner | 免费 | +| 发电停机（合计 MW） | PJM Data Miner | 免费 | +| 绑定约束 | PJM Data Miner | 免费 | +| 输电计划检修 | PJM OASIS | 免费 | +| 核电机组状态 | NRC 每日反应堆状态报告 | 免费 | +| 天然气：Henry Hub 现货、NYMEX 结算价 | EIA、CME | 基本免费 | +| 天然气：东部基差（TETCO M3、Transco Z6、Dominion S、TCO） | Platts / NGI / ICE | 付费——价值最高的付费数据 | +| 气象预报（温度、风、辐照）+ 集合离散度 | NOAA（GFS/HRRR/NDFD）、ECMWF 开放数据 | 免费（须自行存档！） | +| 邻区 ISO 价格/负荷（MISO、NYISO） | 各自公开门户 | 免费 | + +### 1.5 须编码的结构性断点 + +- 机组更替：退役、新投产（用 2019 年训练的模型会给 2026 年错误定价）。 +- 备用/ORDC 市场规则变更（标注日期；不信任变更前的历史）。 +- PTC/ITC 机组构成漂移（见第四节）：随着周边 PTC 电站陆续走完 10 年窗口，历史负电价频率对未来是有偏的预测量。 + +--- + +## 二、模型 B——概率化出力预测（分电站） + +### 2.1 目的与目标 + +在美东时间每日约 10:00（早于 10:30 日前闭市）前，给出明日每小时电站 *可用* 出力的 **经校准的分位数集** ，如 P10 / P25 / P50 / P75 / P90。竞价优化器承诺该分布的某个分位数——点预测在结构上就不够用。 + +预测时长：竞价时点约 14–38 小时 ⇒ 数值天气预报（NWP）承载几乎全部信号；该尺度下持续性外推毫无价值。 + +### 2.2 架构：两阶段管线 + +**第一阶段——气象到功率模型（分场站）。** +用梯度提升树（或类似方法）学习"NWP 预报 → SCADA 出力"的映射。它自动吸收电站的真实功率曲线、尾流损失、逆变器削顶、地形效应、积灰，以及天气模型在该位置的系统偏差。 + +> **关键规则：用预报气象训练，而非实测气象。** 模型必须端到端地学习 NWP 的误差特性；用测风塔实测训练、再用 NWP 输入预测，实盘表现会退化。 + +**第二阶段——分布层。** 三条可互换的路线： + +1. **分位数回归** （pinball loss，每个分位数一个模型）——默认起点；现代 GBM 库原生支持。 +2. **NWP 集合** （如 ECMWF 51 成员）：把每个成员穿过第一阶段模型 → 51 条出力情景 → 读出经验分位数。附加好处：集合保留了 *跨小时的时间相关性* （低风情景整日皆低），对关心多小时形态的竞价优化很有价值。 +3. **类比/误差修饰法** ：以点预测为中心，套上相似条件（相似预报水平、季节、天气格局）下的历史误差分布。 + +### 2.3 校准——原文所说的"金子" + +若声称的概率是诚实的（长期看，实际出力低于"P10"的频率约为 10%），预测即为校准的。用可靠性图检验：分箱历史、统计超越频率、对照名义值。 + +为何这是最值得死磕的事：整套竞价策略就是"承诺分位数 q"。若分位数失准——比如你的"P50"实际只有 40% 的时间被超越（常见的乐观偏差）——那么每一笔报价都系统性偏斜，你在结构性地超卖日前、在实时回购。校准误差一比一转化为结算损失。而实践中的馈赠是：即便预测购自供应商， **你也可以自行重新校准** ——只需你的 SCADA 存档与供应商预测存档，做分位数映射（对照其声称分位数与你的实现超越频率并调整）。小项目、高回报——这就是"历史 SCADA 对预测误差是金子"的含义。保存收到过的每一份预测；这份存档会持续升值。 + +两个悄悄毒化校准工作的因素： + +- **弃电污染。** SCADA 记录的是 *实发* ，不是 *可发* 。若被 PJM 经济性调减（或局部约束限电）、或被你自己的下限逻辑压住，记录出力低估了能力。用被污染的数据校准，模型会"学会"大风/大晴天发得更少。必须重构"可用功率"——现代 SCADA 多可由机舱测风或逆变器无约束能力计算潜在功率信号；以其为真值，并标注弃电时段。 +- **可用率与气象误差分离。** 某串机组检修停运不是气象预报误差。要么把真值折算到可用容量，要么把计划可用率作为特征输入。把停机误差混入气象误差会以错误的理由撑肥分位数——且注意其不对称性：非计划停机只做减法，专门加厚分布的下尾。 + +### 2.4 自建还是外购 + +- **外购** （供应商：Solargis、Meteomatics、UL、DNV 等众多；约 1–3 千美元/场站/月）适合起步；但务必保留内部的 *重校准与评估* 层。每季度用 pinball loss 在自己的 SCADA 上给供应商打分；并行采购两家、一年后留优（或加权融合——两家独立供应商的简单平均常常双双胜出）。 +- **自建** 适合组合规模化之后：管线成本被摊薄；可定制目标（自有弃电逻辑下的可用功率）；保证时点存档；与优化器直接集成。 +- **商业模式优势（跨客户学习）** ：管理多个电站的 SCADA 后（合同条款须允许），气象到功率模型可跨场站迁移、NWP 偏差可按区域而非按站校准、全网预测误差日（价差的驱动因素）变得可观测。客户合同从第一天起就要明确允许汇聚/脱敏的模型训练；事后补条款很痛苦。 + +### 2.5 精度预期（日前尺度） + +- 光伏点预测：约为容量的 5–10%（RMSE 口径，随气候而异）。 +- 风电：约 8–15%（功率约与风速三次方成正比，速度误差被放大）。 +- 但评判要看 **分布** ：跨分位数的 pinball loss + 聚焦尾部的可靠性，并对照气候学修饰的持续性外推这一朴素基线。 + +--- + +## 三、模型 C——DA–RT 价差模型（分节点） + +### 3.1 目的 + +赚钱的模型。输出分小时、分节点的 **S = DA − RT 条件分位数** ，用于选取承诺的出力分位数（报童模型），并按客户自身出力误差与价差的协方差进行修正。 + +### 3.2 目标变量的统计性格 + +- **无条件均值近零，由机制决定** ：虚拟交易会套走任何持续的可预测缺口；残留的是小幅条件风险溢价（日前在预期稀缺时段偏贵）与瞬时无效率。低信噪比；预期 R² 平平。价值在条件结构与尾部。 +- **剧烈不对称的尾部** ：日前是平滑的预期；实时是带刺的实现（稀缺加价可打出 850 美元以上；新能源过剩带来负实时价）。左尾肥（实时飙过日前）、右尾中等（实时崩塌）。对 S 的高斯假设是失格的。 +- **日间自相关弱、条件结构强** （小时、季节、紧张度、气象不确定性）。 + +### 3.3 先分解再建模 + +``` +S_node = S_system（枢纽/电能分量） + S_congestion（节点的日前–实时阻塞差） +``` + +- **S_system** ：总量供需——负荷预测误差、全网新能源预测误差、日前闭市后的强迫停机、备用稀缺。全 PJM 共用，一次建成服务所有客户。 +- **S_congestion** ：实时绑定而日前未定价的约束（或反之）——输电强迫停运、意外潮流。对新能源聚集区常为主导且 *更可预测* （停运期间实时阻塞有持续性）。逐节点工作； **我们可防御的 IP 层** 。 + +两个市场的 LMP 分量在 Data Miner 中均有披露，分解可直接由历史计算。 + +### 3.4 有效的架构：状态分类 × 条件分布 × 协方差 + +**阶段 A——尖峰/状态分类器（主导盈亏的离散事件）：** +- P(实时飙过日前)：紧张度、备用裕度、极端温度预报 *及其集合不确定性* 、高全网新能源预报（欠发风险）、近期强迫停机、日型。 +- P(实时崩塌/负实时价)：全网新能源预报、低负荷、非灵活基荷占比、局部约束状态。 **与模型 A 的负电价分类器共享基础设施。** + +**阶段 B——条件价差分位数** ：给定状态概率，用连续特征做分位数 GBM，或从状态匹配的历史价差分布做类比抽样。 + +**阶段 C——出力–价差协方差修正（本业务的独有优势）** ：全网超发的日子恰好压垮实时价、也恰好是客户超发的日子——出力误差与 S 负相关，使超量申报受罚。实现方式：把客户自身出力预测误差作为价差模型的条件变量，或在状态内估计相关性并对报童分位数做解析修正。通用价格商会跳过这一步；我们不能。 + +### 3.5 特征集（按经验预期的阿尔法排序） + +1. **紧张度** （负荷预测 − 可用容量），约 93% 利用率以上呈强非线性。 +2. **新能源预报水平及其 *修订速度*** ：近几轮 NWP 周期间的变化（临近的重大修订 ⇒ 日前基于陈旧信息出清 ⇒ 价差机会）。 +3. **气象预报不确定性** （集合离散度）——实时波动的燃料。 +4. **节点近期的实时对日前阻塞差** + 约束绑定频率，交叉对照输电计划检修。 +5. **虚拟交易成交总量** （滞后披露）——套利效率的状态变量。 +6. **日历交互项** （小时 × 季节）+ 结构断点标志（规则变更）。 + +### 3.6 评估协议 + +- **对残酷基线 S ≡ 0（"市场有效"）的 pinball loss** 。在时点纪律之下持续地在样本外战胜它是真难事；回测大胜 ⇒ 先查泄漏再庆祝。 +- **经济回测** ：完整报童循环对照朴素 P50 竞价；关键指标是 美元/MWh 提升（好的实现约 0.5–2 美元/MWh）。 +- **专门的尾部校准** （P5/P95），以及 **分状态切片** 评估——平均表现尚可的模型往往在每年驱动盈亏的那约 30 天里一塌糊涂。 + +### 3.7 组合风险（生存级，而非统计级） + +客户整体做多新能源 ⇒ 模型误差在 **全账簿相关** 。对全网超发导致实时崩塌的那一天判断失误，就是对每个风电客户同时失误。第一天起的一级风险指标：跨客户合计的日前多头倾斜 MWh，并对尖峰情景做压力测试。 + +--- + +## 四、横切议题：报价下限、税收抵免与价格腿 + +电量腿（来自模型 B + C）须与 **价格腿** 配对；后者由各资产的真实边际成本决定，而边际成本取决于其税收抵免选择： + +| 资产类型 | 边际成本 | 理性报价下限 | 负电价时是否发电？ | +|---|---|---|---| +| PTC（10 年窗口内） | ≈ −（PTC × 税收放大） | 约 −25 至 −35 美元/MWh | 是，直至下限 | +| ITC | ≈ 0 | 约 0 美元/MWh | 否 | +| PTC 窗口期满（第 11 年起） | ≈ 0 | 约 0 美元/MWh | 否 | + +- PTC（约 27.5–30 美元/MWh，随通胀调整、满足工资/学徒要求，按发电量支付 10 年）把下限推为负值；ITC（资本开支的 30% 以上，按投资支付）不影响边际成本。 +- IRA 自 2025 年起使抵免技术中立（45Y/48E）——高利用率光伏日益选择 PTC ⇒ 光伏机组也开始出现负下限（结构性变化）。OBBBA（2025 年 7 月）加速风光退坡（一般须 2027 年底前投运，含开工安全港）——存量电站保留已锁定的抵免。 **逐项目核实最新指引。** +- **与模型 A 尾部任务的交互** ：竞价用的预期出力是 *中标概率加权的* ：E[中标出力] ≈ 物理预测 × P(LMP ≥ 下限)，逐小时、对日前与实时分别计算（"日前负/实时正"与"日前正/实时负"两种情形结算迥异）。朴素地使用物理预测，恰在弃电风险最高的大发日超量申报最多。 +- 偏差结算的微妙之处：在负实时价期间对日前头寸欠发可能 *有利可图* （以正价卖日前、以负价买回）——这正是偏差费用与强制申报规则存在的原因。任何逼近该边界的自动竞价逻辑都需要合规审查（市场监测机构关注风险）。核实 PJM 现行结算/偏差规则；它们会变。 + +### 4.1 BORD 类偏差费用（优化器的一级输入） + +"BORD" = PJM 的 **平衡运行备用偏差** 费用——一种成本 *分摊* 机制，实际效果是对偏离日前头寸的准罚则。 + +- **机制。** PJM 会产生补偿成本（对被组合或调度、但 LMP 收入不足以覆盖其报价成本的机组做整体补偿）。其中 *平衡* 部分——日前闭市后产生、即实时偏离日前计划的成本——按成本因果大量分摊给 **偏差方** ：偏离日前计划的发电（无论方向）、偏离申报的负荷、以及从不实际交割的虚拟交易。每偏差 MWh 承担一个按日浮动的 美元/MWh 费率，与当日补偿成本及总偏差 MWh 挂钩。 +- **对竞价的影响。** 给报童目标函数的偏差腿增加第二项：每一 MWh 的 |Q_rt − Q_da| 都要分摊费用。它是对"偏离预期出力下注"的摩擦/交易成本——预期费率越高，最优分位数越向 P50 回拢。优化器因此需要偏差费率（或其预测）作为价差预测之外的并列输入。 +- **漏洞补丁。** 偏差费用（加上市场监测机构的审视）使系统性制造"负实时价下的有利欠发"变得昂贵——见上文微妙之处一条。 +- **注意。** PJM 的补偿成本分摊规则经历过反复诉讼与改革（成本科目、偏差定义、间歇性资源豁免、轧差规则）。结算模块必须从现行费率表拉取参数，绝不能凭记忆。"BORD 类"是对这一 *类别* 的简称： **事后、基于成本因果、针对日前对实时偏差的费用。** +- **对照（服务于中国业务）。** 中国的"两个细则"通过 *行政绩效考核* 实现偏差约束——按监管设定的预测精度/计划执行阈值罚分，与当日系统实际平衡成本基本无关。PJM 给外部性定价（对可预测的随机 *价格* 优化偏差）；中国给作业打分（对固定的 *规则手册* 优化预测）。这也是中国功率预测市场成为合规采购市场、而 PJM 从未产生同等强制需求渠道的原因。 + +--- + +## 五、数据与运营基础设施（共享地基） + +### 5.1 时点纪律（不可妥协，适用于全部三个模型） + +每个特征都必须反映 **美东时间 10:30 日前闭市前** （实践中约 10:00 快照）可得的信息集： + +- 气象：当时实际可得的 NWP 周期（如 00Z 场次），不用更晚周期，绝不用实测。 +- PJM 数据：以当时快照为准（数据会被修订）。 +- 供应商预测：按到达原样存档（供应商会覆盖）。 +- 电站可用率：以竞价时点已知为准。 + +训练数据混入闭市后信息 ⇒ 回测虚高、实盘"莫名"跑输。 **"以当时为准"的存档是复利型商业资产** ，无法事后重建——竞争对手花钱也买不到。 + +### 5.2 PJM 准入与许可（商业模式的义务项） + +- Data Miner 数据 **仅限内部使用；数据或衍生品的再分发须取得 PJM 再分发许可** （至少 Associate Membership）。我们的预测是否算"衍生数据"，必须在产品上线 **之前** 与 PJM 及律师厘清。 +- 速率限制：非会员每分钟 6 次连接对会员 600 次——组合规模化后出于运营需要本就该入会，同时也顺带打开再分发路径。 +- 报价提交（模式二"软件代理"变体 / 模式三）：Markets Gateway API（独立且需认证）、凭证管理、第三方代理的账户安全规则——与 PJM 会员服务确认现行要求。 +- 商业数据（气价指数、气象供应商）：内部使用与再分发是两档定价迥异的授权；照此签约。 + +### 5.3 结算反馈闭环 + +客户的 PJM 结算单是验证模型 C 与经济回测的真值。每份客户协议都要约定结算数据的获取权（验证 + 模型训练用途）。 + +--- + +## 六、建设顺序（建议） + +| 阶段 | 交付物 | 涉及模型 | +|---|---|---| +| 1 | **经验价差图谱** ：按节点簇/小时/季节的价差分布、负实时价与尖峰频率。零机器学习；可直接作为分析产品售卖；为阶段 B 的类比库播种。 | C | +| 2 | **时点存档管线** （PJM 数据、NOAA/ECMWF，从第一天起自行存档）。 | A、B、C | +| 3 | 基于免费 PJM + NOAA 数据的 **系统级状态分类器** （尖峰/崩塌）；负电价分类器（A/C 共享）。 | A、C | +| 4 | **节点阻塞层** ：近期实时对日前阻塞特征 + 客户实际节点的输电检修交叉对照。 | C | +| 5 | **出力预测管线** ：供应商方案 + 内部重校准；组合规模化后迁移为自建的两阶段管线。 | B | +| 6 | 分客户的 **出力–价差联合协方差修正** （需其 SCADA）。 | B + C | +| 7 | **报童竞价优化器** + 对照 P50 朴素基线的经济回测框架。 | 全部 | +| 8 | 组合相关性风险看板（合计日前倾斜、尖峰压力测试）。 | C / 风险 | +| 9 | 二阶升级：分布式深度学习、联合情景生成、结构化堆栈精化。 | 全部 | + +--- + +## 七、下一步讨论的开放问题 + +1. 完整的报童数学：含协方差修正与偏差费用不对称性的闭式最优分位数——值得推导并做单元测试。 +2. 目标客户节点/区域：西部风电（ComEd/AEP）与中大西洋光伏的阻塞故事迥异；决定阶段 4 约束库的优先级。 +3. 出力预测供应商比选设计（选哪两家、打分协议、融合规则）。 +4. 混合/储能客户：价格水平的 *形态* 预测重回一级重要性（套利是水平–形态问题）——路线图的触发点。 +5. 日内/改报范围：PJM 改报窗口与实时策略不在 v1 范围内，但影响架构（模型 B 的时间相关性需求）。 +6. 逼近偏差结算边界的自动竞价逻辑的合规审查清单；与 PJM 会员服务厘清代理框架。 +7. 各客户资产的强制申报/容量资源义务——影响最优分位数并约束报价灵活性；须按现行规则核实。 + +--- + +## 附录：术语表 + +| 术语 | 释义 | +|---|---| +| LMP | 节点边际电价——分解为电能 + 阻塞 + 损耗三个分量 | +| DA / RT | 日前市场（美东约 10:30 闭市出清次日逐小时）/ 实时市场（5 分钟调度与结算） | +| DA–RT 价差（S） | S = 日前价 − 实时价；模型 C 的对象；套利使均值近零、尾部肥且不对称 | +| 报童竞价 | 最优日前电量 = 出力分布的某个分位数，分位数由预期价差与结算不对称性决定 | +| 分位数 / pinball loss | 结果有 X% 概率落于其下的数值 / 训练并评估分位数预测的损失函数 | +| 校准 | 声称概率诚实（P10 约有 90% 时间被超越）的性质；用可靠性图检验 | +| SCUC | 安全约束机组组合——ISO 出清日前市场所解的混合整数优化 | +| 虚拟交易（INC/DEC/UTC） | 纯金融性的日前头寸，从不物理交割；将日前价套向实时预期 | +| 补偿成本（Uplift） | 对市场收入不足以覆盖报价成本（启动、空载、最小运行）的机组的整体补偿 | +| BORD | 平衡运行备用偏差费用——PJM 将实时平衡补偿成本分摊给日前–实时偏差方（见 4.1 节） | +| ORDC / 稀缺加价 | 运行备用需求曲线——备用短缺时的行政性价格加成；实时尖峰的推手 | +| FTR | 金融输电权——对两节点间日前阻塞的套保/投机工具 | +| PTC / ITC | 生产税抵免（按发电量美元/MWh、10 年；造就负报价下限）/ 投资税抵免（按资本开支比例；下限约 0）——见第四节 | +| 报价下限 | 低于该价格电站宁可弃电的价格腿：PTC 资产约为 −（PTC × 税收放大），ITC 资产约为 0 | +| 强制申报 / RPM | 容量资源向日前市场申报的义务 / PJM 容量市场（可靠性定价模型） | +| BTM | 表后（如从观测负荷中净掉的屋顶光伏） | +| NWP | 数值天气预报（GFS、HRRR、ECMWF）；集合 = 多组扰动初值的场次，预报不确定性特征的来源 | +| SCADA | 电站监控与数据采集——出力、可用率、弃电记录的真值 | +| 时点纪律 | 所有训练/回测特征只能反映美东 10:30 日前闭市前可得的信息（见 5.1 节） | +| Data Miner 2 / Markets Gateway | PJM 的公开数据 API / PJM 的认证报价提交 API | +| QSE / 调度代理 | 持有市场主体基础设施、代资产提交报价并处理结算的实体 | diff --git a/pjm/market_primer_en.md b/pjm/market_primer_en.md new file mode 100644 index 0000000..d9f45e1 --- /dev/null +++ b/pjm/market_primer_en.md @@ -0,0 +1,197 @@ +# PJM Market Primer: Written for Readers Familiar with China's Power Market + +**Working document — v0.1 (July 2026)** +**Positioning:** the mirror of Section 1 of `../china/market_vendor_dd_en.md` +(the China market primer) — it explains PJM for readers who know China's +market, with Chinese concepts as the constant point of comparison. Chinese +version: `market_primer_zh.md`. + +--- + +## 1. Institutional frame: who governs, who plays + +PJM Interconnection is the largest US Regional Transmission Organization +(RTO), spanning 13 eastern states plus Washington DC, serving ~65 million +people, with peak load on the order of 150 GW and installed capacity above +180 GW. The frame differs fundamentally from China's "provincial markets + +national push": + +- **The regulator is FERC (federal)**, not provinces/states — PJM's market + rules (tariff/operating agreement) are FERC-approved, and rule changes go + through a federal legal process (stakeholder proceedings + FERC rulings): + public, slow, litigable. Contrast: Chinese rules issue as NDRC/NEA and + provincial documents — fast and flexible. +- **PJM itself is a member-governed nonprofit**: generators, transmission + owners, retailers, state consumer advocates — and, crucially, **financial + institutions are a formal market-participant category**. That single fact + is the key to most PJM–China differences. +- **One market, one rulebook**: 13 states share the same clearing engine and + price system. There is no "one province, one policy." Inter-RTO trade + (PJM–MISO, PJM–NYISO) resembles China's inter-provincial layer, but with + greater volume and mechanism maturity. + +## 2. Market architecture: spot-native + +China is "MLT as ballast, spot as corrector"; PJM is the inverse — **the +spot market IS the market**, and all forward arrangements are off-exchange +financial contracts settling against it. + +- **Day-ahead (DA)**: closes 10:30 AM ET daily, clears all 24 hours of + tomorrow via SCUC + economic dispatch, hourly prices. Analog: the DA + declaration/clearing in Chinese spot provinces — but with no "physical" + MLT volume occupying space; the DA market reprices everything. +- **Real-time (RT)**: 5-minute dispatch and settlement. DA + RT form the + **two-settlement system**: DA positions settle at DA prices; deviations of + actuals from DA settle at RT — the same logic as Chinese spot provinces' + DA settlement + RT deviation settlement. +- **Nodal prices (LMP)**: ~13,000 pricing nodes, each decomposed into + energy + congestion + loss. Contrast: most Chinese provinces clear at a + unified provincial price or coarse zones (generation-side nodal in + Guangdong/Shanxi excepted) — PJM's spatial granularity is two orders of + magnitude finer, which is why "congestion" is a standalone business (see + FTR). +- **Caps and scarcity pricing**: offer cap $1,000/MWh (to $2,000 with cost + justification), plus the Operating Reserve Demand Curve (ORDC) — an + administrative adder that drives RT prices up when reserves run short. + Negative prices are fully legal and common. Contrast: China's declaration + caps anchor to industrial peak retail rates (~¥1–1.5/kWh) with + administratively compressed tails; PJM scarcity spikes reach an order of + magnitude higher — which is precisely why US volatility can fund + optimization fees. + +## 3. The financial layer: species absent from China + +PJM's DA market is open to **purely financial participation** — the deepest +structural difference from China: + +- **Virtual transactions (INC/DEC/UTC)**: entities with no physical assets + submit virtual supply (INCs) or virtual demand (DECs) into DA, clearing DA + and mandatorily unwinding at RT — speculation/arbitrage on the DA–RT + spread, at meaningful volume shares. Effect: any persistent, predictable + DA–RT gap gets arbitraged away; the spread's mean is pinned near zero. + Contrast: China has no such mechanism — the core reason the MLT–spot basis + can persistently diverge (see China report §1.5). +- **FTRs (Financial Transmission Rights)**: financial rights on DA + congestion differences between two nodes, auctioned by PJM — the hedging + and speculation instrument for congestion risk. Contrast: China's + unified provincial pricing makes the instrument largely unnecessary; + inter-provincial congestion rents flow to grid/government mechanisms. +- **OTC forwards/futures**: Nodal Exchange and ICE list PJM hub power + futures (peak/off-peak, monthly, years out), liquidity concentrated at + Western Hub; plus bilateral PPAs/virtual PPAs (CfDs against hub or node + prices) and bank hedge structures (proxy revenue swaps, etc.). Contrast: + functionally equivalent to China's MLT contracts, but (a) fully voluntary, + no coverage-ratio mandates; (b) priced by financial traders with + arbitraged basis convergence; (c) settled **outside** the market + operator — PJM doesn't know your hedge exists. China's MLT lives *inside* + the exchange: quasi-mandatory, priced by physical counterparties, + settlement-integrated, and — a silver lining — visible in disclosure data. + +## 4. Capacity market and ancillary services + +- **RPM capacity market**: rolling three-year-forward auctions procure + capacity obligations; cleared resources carry a **must-offer obligation** + (must bid into DA) and performance assessment in scarcity hours (Capacity + Performance — underperformance penalties are substantial). Contrast: + China's coal capacity payment (容量电价, from 2024) is administratively + priced availability compensation — no auction, no symmetric performance + penalty; similar function (a revenue pillar outside the energy market), + very different mechanism. +- **Ancillary services**: regulation and synchronized/primary reserves + **co-optimized** with energy in DA/RT with endogenous prices. Contrast: + Document 394 pushes China the same direction, but product scope and + co-clearing depth are still evolving. + +## 5. Settlement and deviations: BORD vs. the Two Detailed Rules + +PJM disciplines deviations through **Balancing Operating Reserve Deviation +(BORD) charges**: real-time balancing uplift is allocated by cost causation +to parties who deviated from DA positions (generation deviations, load +deviations, virtuals), at a daily rate scaled to actual costs incurred. +China disciplines forecast accuracy and schedule compliance through the +administrative 两个细则 assessments, decoupled from actual daily system +cost. One-line contrast: **PJM prices the externality; China grades the +homework.** Product implication: in PJM, deviations are optimized against a +forecastable stochastic *price*; in China, forecasts are engineered against +a fixed *rulebook* — which is why China developed a compliance-procurement +forecasting market and the US never did. (Details: `design_models_en.md` +§4.1.) + +## 6. Renewables: tax credits, not mechanism prices + +The US never had "guaranteed volume, guaranteed price": renewables were +always market-absorbed (PPAs are voluntary commercial contracts, not policy +procurement). Policy support flows through **tax credits**: + +- **PTC**: ~$27.5–30/MWh of generation for 10 years ⇒ a rational *negative* + offer floor (≈ −$25 to −$35/MWh) — the plant pays to generate because the + credit outweighs the negative price. +- **ITC**: a one-time 30%+ credit on capex ⇒ no effect on marginal cost; + offer floor ≈ $0. +- Contrast with China's **mechanism price**: both are revenue stabilizers, + but the mechanism price is a *price-type* instrument (a CfD stabilizing + the realized price) while tax credits are *tax-type* (leaving price + formation untouched but distorting offer floors). A neat mirror: China's + mechanism auctions (lowest bids win) and America's PTC-shaped negative + floors each create a distinctive supply-curve signature at the low end of + their respective markets. +- Curtailment in PJM is **economic** (you're dispatched down when price + falls below your offer floor — a market-clearing outcome), not an + administrative quota; "curtailment rate" statistics are therefore not + directly comparable across the two systems. + +## 7. Data and access: the transparency gap + +- **Data Miner 2**: PJM's public data platform — full nodal DA/RT LMPs with + component decomposition, load and forecasts, fleet wind/solar forecasts + and actuals, outages, binding constraints — web access without login, API + with free registration. Contrast: China has no equivalent; provincial + disclosures vary and sit behind member portals. In China, *data + acquisition* is a moat; in PJM data is free and **insight** is the moat. +- Bid submission uses the separately authenticated **Markets Gateway**; + participant status involves registration and credit/collateral (logic + parallel to Chinese exchange membership + performance bonds). +- Note the asymmetry, though: PJM unit offers publish only at a 4-month lag, + masked, and the network model is CEII-restricted — "transparent" is not + "fully transparent." + +## 8. Quick-reference comparison table + +| Dimension | PJM | China (spot provinces) | +|---|---|---| +| Regulator | FERC (federal); procedural, slow rule change | NDRC/NEA + provincial; document-driven, fast | +| Market scope | One market, one rulebook across 13 states | One province one policy; 31 provincial markets | +| Dominant settlement layer | Spot (DA+RT two-settlement) | MLT majority; spot settles deviations | +| Forwards/MLT | Voluntary OTC financial contracts (PPA/futures), outside the operator | Quasi-mandatory exchange contracts, curved, settlement-integrated | +| Forward price formation | Financial traders, arbitraged convergence | Physical counterparties inside coal-benchmark ±20%; persistent basis | +| Financial participants | Formal category (virtuals, FTR, futures) | Essentially absent | +| Spatial pricing | ~13,000 nodal LMPs | Unified provincial / zonal mostly | +| Price tails | $1,000–2,000/MWh caps + ORDC adders; negative common | ~¥1–1.5/kWh caps; floors at 0 or slightly negative | +| Capacity compensation | RPM auctions + must-offer + performance penalties | Capacity payment (administered, coal-centric) | +| Deviation discipline | BORD cost-causation allocation (priced) | Two Detailed Rules administrative assessment (ruled) | +| Renewable support | PTC/ITC tax credits (shaping negative floors) | Mechanism-price CfD (auction-set) | +| Public data | Data Miner, free and comprehensive | Provincial, member-gated, inconsistent | +| Service-provider moat | Nodal congestion insight + point-in-time archive | Provincial rules library + data acquisition + relationships | + +## 9. Three counter-intuitive notes for readers coming from China + +1. **"No MLT market" does not mean "no hedging"** — hedging is everywhere, + just grown outside the market operator and carried by the financial + system. Analyzing a PJM plant's revenue requires asking about its + PPA/hedge structure, which public data won't show (contrast China: the + contracts live inside the exchange, so aggregate data is visible). +2. **Price volatility is a feature, not a bug** — ORDC scarcity adders are + deliberate design, paying for reliability through a few extreme hours. + The Chinese instinct of "caps for stability" misreads where the + commercial opportunity comes from in PJM: optimization fees are + ultimately funded by volatility. +3. **Slow rules reshape the moat** — PJM rules evolve on a multi-year cycle + with precedent to consult, so model risk is dominated by *market* risk; + Chinese rules evolve quarterly, making *rule* risk itself a modeling + object. The same company needs different organizational muscles in the + two markets. + +--- + +*Glossary: appendix of `design_models_en.md`; China-side concepts: appendix +of `../china/market_vendor_dd_en.md`.* diff --git a/pjm/market_primer_zh.md b/pjm/market_primer_zh.md new file mode 100644 index 0000000..2e1162e --- /dev/null +++ b/pjm/market_primer_zh.md @@ -0,0 +1,83 @@ +# PJM 市场入门：写给中国读者 + +**工作文档 —— v0.1（2026年7月）** +**定位：** 本文是 `../china/market_vendor_dd_zh.md` 第一节（中国市场入门）的镜像——为熟悉中国电力市场的读者解释 PJM，处处以中国概念作对照。英文版：`market_primer_en.md`。 + +--- + +## 一、机构框架：谁在管、谁在玩 + +PJM Interconnection 是美国最大的区域输电组织（RTO），覆盖美国东部 13 个州加华盛顿特区，服务约 6,500 万人口，尖峰负荷约 15 万 MW 量级、装机约 18 万 MW 以上。与中国"省级市场 + 国家层面统一推进"的格局根本不同： + +- **监管者是联邦能源监管委员会（FERC）** ，而非省/州政府——PJM 的市场规则（费率表/运行协议）由 FERC 批准，规则变更走联邦法律程序（利益相关方流程 + FERC 裁决），公开、缓慢、可诉讼。对照中国：规则由国家发改委/能源局与省级主管部门以文件形式发布，节奏快、弹性大。 +- **PJM 本身是会员制非营利机构** ：发电商、输电企业、售电/零售商、金融交易商、州消费者代表都是会员并参与治理投票。 **金融机构是正式的市场主体类别** ——这一点是理解 PJM 与中国差异的钥匙。 +- **一个市场、一套规则** ：13 个州共用同一出清引擎、同一价格体系。不存在"一省一策"。跨 RTO（PJM 对 MISO、NYISO）的交易类似中国的省间交易，但体量与机制成熟度更高。 + +## 二、市场架构：现货为纲 + +中国是"中长期为主、现货为补充"；PJM 恰好倒过来—— **现货市场就是市场本身** ，所有远期安排都是围绕现货结算的场外金融合同。 + +- **日前市场（DA）** ：每日美东 10:30 闭市，出清次日 24 个小时。安全约束机组组合（SCUC）+ 经济调度，逐小时价格。对照：相当于中国现货省份的日前申报出清，但没有中长期电量的"物理"占位——日前市场对所有电量重新定价。 +- **实时市场（RT）** ：5 分钟调度、5 分钟结算。日前与实时构成 **两结算体系** ：日前头寸按日前价结算，实际与日前的偏差按实时价结算——与中国现货省份"日前结算 + 实时偏差结算"逻辑一致。 +- **节点电价（LMP）** ：约 1.3 万个定价节点，每个节点价格分解为电能 + 阻塞 + 损耗。对照：中国多数省份按全省统一价或粗分区出清（广东、山西发电侧节点价除外）——PJM 的空间颗粒度细两个数量级，"阻塞"因此是一门独立的生意（见 FTR）。 +- **限价与稀缺定价** ：报价上限 1,000 美元/MWh（成本证明下可至 2,000），配合运行备用需求曲线（ORDC）——备用紧张时行政性加价推动实时价格飙升。负电价完全合法且常见。对照：中国的申报上限锚定工商业尖峰电价（约 1–1.5 元/kWh），尾部被行政压缩；PJM 的稀缺尖峰可达中国上限的十倍以上（按购买力口径亦然），这正是美国市场"波动率养活优化服务"的原因。 + +## 三、金融层：中国市场没有的物种 + +PJM 日前市场对 **纯金融参与** 开放，这是与中国最深刻的结构差异： + +- **虚拟交易（INC/DEC/UTC）** ：无实物资产的主体可在日前申报虚拟供给（INC）或虚拟需求（DEC），日前成交、实时强制反向平仓——本质是对日前–实时价差的投机/套利。成交量占比可观。效果：任何持续、可预测的日前–实时价差都会被套走，价差均值被钉在零附近。对照：中国无此机制，这正是中长期–现货基差可以持续偏离的核心原因（见中国报告 1.5 节）。 +- **金融输电权（FTR）** ：对两节点间日前阻塞价差的金融权利，通过 PJM 组织的拍卖取得——阻塞风险的套保与投机工具。对照：中国的省内统一定价使该工具基本无必要，省间阻塞收益归于电网/政府机制。 +- **场外远期/期货** ：Nodal Exchange 与 ICE 挂牌 PJM 枢纽电力期货（峰/谷、月度、数年期限），流动性集中在 Western Hub。加上双边 PPA/虚拟 PPA（对枢纽或节点价结算的差价合约）、银行对冲结构（代理收入互换等）。对照：功能上等价于中国的中长期合同，但（a）完全自愿、无签约比例要求；（b）由金融交易商定价、基差被套利收敛；（c）在市场运营机构 *之外* 结算——PJM 不知道你的套保存在。中国的中长期在交易中心 *之内* ：强制、实物主体定价、结算一体、数据可见。 + +## 四、容量市场与辅助服务 + +- **RPM 容量市场** ：三年期滚动拍卖采购容量义务，中标资源承担 **强制申报义务** （必须向日前市场报价）与稀缺时段的性能考核（性能资本化，欠绩罚款可观）。对照：中国 2024 年起的煤电容量电价是行政定价的可用性补偿，无拍卖、无对称的性能罚则；功能相似（给容量以能量市场外的收入支柱）、机制迥异。 +- **辅助服务** ：调频、同步/一次备用与电能量在日前/实时 **联合出清** ，价格内生。对照：中国按 394号文方向也在推动辅助服务与现货联合出清，但品种与联合深度仍在演进。 + +## 五、结算与偏差：BORD 对"两个细则" + +PJM 通过 **平衡运行备用偏差（BORD）费用** 约束偏差：实时平衡产生的补偿成本（uplift）按成本因果分摊给偏离日前头寸的主体（发电偏差、负荷偏差、虚拟交易），费率按日随实际成本浮动。中国通过"两个细则"行政考核约束预测精度与计划执行，力度与当日系统实际成本脱钩。一句话对照： **PJM 给外部性定价；中国给作业打分。** 产品含义：在 PJM，偏差是对一个可预测随机价格的优化对象；在中国，预测是对一套固定规则手册的工程对象——这也是中国出现"合规采购型"功率预测市场、而美国没有的原因。（详见 `design_models_zh.md` 4.1 节。） + +## 六、新能源的处境：税收抵免而非机制电价 + +美国从未有过"保量保价"：新能源自始按市场消纳（PPA 是自愿商业合同，不是政策收购）。政策支持走 **税收抵免** ： + +- **PTC（生产税抵免）** ：按发电量约 27.5–30 美元/MWh、支付 10 年 ⇒ 理性报价下限为负（约 −25 至 −35 美元/MWh）——愿意付钱发电，因为抵免比负电价损失更值钱。 +- **ITC（投资税抵免）** ：按资本开支 30% 以上一次性抵免 ⇒ 不改变边际成本，报价下限约 0。 +- 对照中国的 **机制电价** ：同为收入稳定器，但机制电价是"价格型"工具（差价合约稳定电价），税收抵免是"税收型"工具（不触碰市场价格形成，却扭曲报价下限）。有趣的镜像：中国的机制电价竞价（谁的报价低谁入选）与美国 PTC 塑造的负报价下限，都在各自市场的低价端制造了独特的供给曲线形态。 +- 弃电在 PJM 是 **经济性** 的（价格低于你的报价下限即被调减，属市场出清结果），而非行政配额；"弃电率"概念因此与中国口径不可直接比较。 + +## 七、数据与准入：透明度的落差 + +- **Data Miner 2** ：PJM 的公开数据平台——日前/实时全节点 LMP（含分量分解）、负荷与预测、全网风光预测与实际、停机、绑定约束等，网页免登录、API 免费注册。对照：中国无对应物，各省交易中心披露口径不一且多在会员端口内——在中国"数据获取本身是护城河"，在 PJM 数据免费而 **洞察是护城河** 。 +- 报价提交走独立认证的 **Markets Gateway** ；市场主体资格涉及注册、信用保证金（对照中国交易中心会员 + 履约保函，逻辑相通）。 +- 但注意不对称：PJM 的机组报价延迟 4 个月脱敏公开、输电网络模型属 CEII 受限——"透明"不等于"全透明"。 + +## 八、中美速查对照表 + +| 维度 | PJM | 中国（现货省份） | +|---|---|---| +| 监管 | FERC（联邦），规则变更程序化、缓慢 | 发改委/能源局 + 省级，文件驱动、节奏快 | +| 市场范围 | 13 州一个市场、一套规则 | 一省一策，31 个省级市场 | +| 主导结算层 | 现货（日前+实时两结算） | 中长期为主、现货结算偏差 | +| 远期/中长期 | 场外自愿金融合同（PPA/期货），运营机构之外 | 交易中心内的准强制合同，带曲线、结算一体 | +| 远期价格形成 | 金融交易商套利定价，基差收敛 | 实物主体在煤电基准 ±20% 区间内协商，基差可持续 | +| 金融参与者 | 正式类别（虚拟交易、FTR、期货） | 基本缺位 | +| 空间定价 | 约 1.3 万节点 LMP | 全省统一/分区为主 | +| 价格尾部 | 上限 1,000–2,000 美元/MWh + ORDC 稀缺加价；负价常见 | 上限约 1–1.5 元/kWh；下限 0 或略负 | +| 容量补偿 | RPM 拍卖 + 强制申报 + 性能罚则 | 容量电价（行政定价，煤电为主） | +| 偏差约束 | BORD 成本因果分摊（价格化） | 两个细则行政考核（规则化） | +| 新能源支持 | PTC/ITC 税收抵免（塑造负报价下限） | 机制电价差价合约（竞价形成） | +| 公开数据 | Data Miner 免费全量 | 分省、会员制、口径不一 | +| 服务商护城河 | 节点阻塞洞察 + 时点存档 | 省级规则库 + 数据获取 + 关系与资质 | + +## 九、给中国读者的三个"反直觉"提示 + +1. **"没有中长期市场"不等于"没有套保"** ——套保无处不在，只是长在市场运营机构外面、由金融体系承接。分析 PJM 电站的收入，必须问它的 PPA/对冲结构，而这在公开数据里看不到（对照中国：合同在交易中心内，总量数据反而可见）。 +2. **价格波动是特性不是缺陷** ——ORDC 稀缺加价是刻意设计，用少数极端小时的高价支付可靠性成本。中国式"限价维稳"的直觉在 PJM 语境下会误判商业机会的来源：优化服务的费根本上由波动率支付。 +3. **规则慢即是护城河的形状不同** ——PJM 规则以年为单位演进且有判例可循，模型的"规则风险"低、"市场风险"高；中国规则以季度演进，"规则风险"本身就是要建模的对象。同一家公司在两个市场需要的组织能力并不相同。 + +--- + +*术语表见 `design_models_zh.md` 附录；中国侧概念见 `../china/market_vendor_dd_zh.md` 附录。* diff --git a/pjm/vendor_dd_en.md b/pjm/vendor_dd_en.md new file mode 100644 index 0000000..866161c --- /dev/null +++ b/pjm/vendor_dd_en.md @@ -0,0 +1,418 @@ +# Competitive Due Diligence Report: Autobidder & Bid-Optimization Vendors + +**Prepared for:** [Renewable DA bidding venture — working name TBD] +**Date:** July 10, 2026 +**Scope:** Six vendors identified as the closest direct or adjacent competitors to a +Model-2 (bid-optimization SaaS) business serving renewable generators in PJM: +Fluence Mosaic, Ascend Analytics (SmartBidder), Tyba Energy, Gridmatic, Dexter +Energy, and enspired. +**Status:** Working draft based on public sources (company sites, press releases, +SEC filings, funding databases, trade press) retrieved July 2026. Private-company +figures (funding, headcount, revenue) are as reported by third-party databases and +should be re-verified before being relied on for any commercial decision. Pricing +is generally not public for any vendor; pricing intelligence requires primary +research (reference calls, RFP participation). + +--- + +## 1. Executive summary + +The six vendors split into three strategic archetypes: + +| Archetype | Vendors | Balance-sheet posture | Primary asset class | +|---|---|---|---| +| **Software arm of a hardware major** | Fluence Mosaic | Public parent (NASDAQ: FLNC) | Storage first, renewables second | +| **Analytics incumbent extending into bidding** | Ascend Analytics; Tyba (younger variant) | PE-backed / VC-backed | Storage first | +| **AI trader/optimizer (takes or manages positions)** | Gridmatic; enspired; Dexter (signals-only variant) | Bootstrapped-profitable / VC-backed | Storage & renewables | + +Key findings for our positioning: + +1. **Every vendor is storage-led or storage-heavy.** The economic center of gravity + of this market is battery arbitrage/ancillary co-optimization in ERCOT and + CAISO. Standalone wind/solar DA bidding — our newsvendor problem — is a + secondary feature for all six. None markets a PJM-first, nodal-congestion-aware + renewable DA offering. +2. **PJM presence is thin and recent.** Gridmatic entered PJM via retail (Ohio, + then Pennsylvania, Feb 2026); Ascend reaches PJM via the Tenaska partnership; + Fluence Mosaic's announced market coverage emphasizes CAISO/ERCOT/NEM. The + Europeans (Dexter, enspired) are US-entrant, not US-established. +3. **The winning configuration is emerging: algorithms + 24/7 desk + market + participant infrastructure.** Ascend×Tenaska demonstrates it explicitly; + Gridmatic and enspired embody it natively. A pure-software Model 2 without an + execution path will face this configuration in every competitive deal. +4. **Two funding realities.** Capital is available for this category (Tyba $18M, + Dexter ~€40M lifetime, enspired >€75M lifetime, Ascend PE recap), but + Gridmatic's bootstrapped-profitable trading model shows an alternative: + monetize the models directly in the market rather than through SaaS fees. + +--- + +## 2. Vendor profiles + +### 2.1 Fluence Mosaic (Fluence Energy, Inc. — NASDAQ: FLNC) + +**What it is.** AI-powered automated bidding software for grid-scale storage and +renewable (wind/solar) assets, sold as part of the Fluence IQ digital platform, +plus a newer "Mosaic Trading Solutions" services arm offering bid-to-bill support. + +**Origin & history.** Fluence (founded 2018 as a Siemens–AES JV; IPO 2021) +acquired Advanced Microgrid Solutions (AMS) in late 2020. AMS built the +algorithmic-bidding engine — at acquisition it was bidding ~1.7 GW (largely in +Australia's NEM and CAISO) with reported renewable revenue uplift of >10% from +optimized wholesale bidding. Early flagship: PG&E's 182.5 MW / 730 MWh Moss +Landing BESS in CAISO. + +**Scale & markets.** Marketing materials cite ~16 GW of assets deployed or awarded +on the platform (up from ~22.3 GW "contracted or under management" claimed in a +Jan 2024 release — figures are not directly comparable across releases and mix +deployed/awarded definitions; treat with caution). Announced market coverage: +CAISO (standalone, co-located, and since 2024 hybrid resources), ERCOT, and +Australia's NEM. PJM is not prominent in public Mosaic materials — a notable gap. + +**Product notes.** Probabilistic ML price forecasts; DA/RT co-optimization across +energy and ancillary products; risk-tolerance parameterization; warranty/ +degradation constraints for batteries; manual override; technology-agnostic +(non-Fluence hardware supported). + +**Financial position (parent).** FY2025 revenue $2.3B (down from $2.7B FY2024), +FY2025 net loss $68M; Q1 FY2026 revenue $475M (+154% YoY) at ~5% GAAP gross +margin; Q2 FY2026 revenue $465M, net loss $29M, record $5.6B backlog, and reported +momentum with data-center/hyperscaler MSAs. The digital segment is reported to be +tracking toward ~$180M ARR by FY2026 year-end (analyst figure, not audited +segment disclosure). Stock rated ~Hold by consensus. Net: hardware business is +volatile and thin-margin; the software segment is strategically important to the +equity story, meaning continued investment in Mosaic is likely. + +**Strengths.** Brand and global installed base; deep battery operations data; +public-company resources; integrated hardware+software+services sell; anchor +customers (AES, PG&E). + +**Weaknesses / openings for us.** Storage-first DNA — standalone renewable DA +bidding is an adjacent use case, not the core; limited public evidence of PJM +depth or nodal congestion specialization; big-company sales motion poorly suited +to small IPP long tail; potential channel conflict (Fluence competes with some +prospective customers' hardware vendors). + +**Sources.** fluenceenergy.com/mosaic-intelligent-bidding-software; Fluence blog +(AMS acquisition, 2020–21); ir.fluenceenergy.com (Jan 2024 hybrid-CAISO release; +FY2025 results Nov 2025; Q1/Q2 FY2026 8-Ks); stockanalysis.com/flnc; Trefis. + +--- + +### 2.2 Ascend Analytics — SmartBidder + +**What it is.** Bid optimization and scheduling platform for standalone storage, +co-located, and hybrid renewable+storage projects, with probabilistic and +AI-assisted models incorporating nodal price dynamics, asset constraints, and +risk-based optimization; delivery ranges from self-service software to fully +managed service. Sits inside a broader analytics suite: BatterySIMM (storage +valuation — widely treated as a de facto standard), PowerSIMM (portfolio/resource +planning), PowerVAL, AscendMI market intelligence, plus consulting. + +**Origin & ownership.** Founded 2002 (Boulder, CO) by Dr. Gary Dorris; grew as a +consulting/analytics firm. Took a strategic growth investment led by PE firm +Rubicon Technology Partners with Galvanize Climate Solutions and Silversmith +(March 2024) — a scale-up recapitalization. Headcount grew from ~57 (2020) to +~162 (2024) per third-party data. + +**Scale & markets.** Customer base described as including some of the largest +utilities and renewables developers in North America and Europe. Public +SmartBidder wins skew ERCOT/CAISO storage (e.g., Tokyo Gas America's 174 MW +Longbow BESS in ERCOT, June 2025). Claims include revenue outcomes reaching +90–95% of perfect-foresight benchmarks and 50%+ uplift vs. comparable BESS +projects in ERCOT/CAISO (vendor claims; methodology not independently verified). + +**The Tenaska partnership (critical for us).** Ascend teamed with Tenaska Power +Services (Aug 2025) to combine SmartBidder's analytics/trading algorithms with +Tenaska's 24/7 real-time desk and QSE/market-participant infrastructure — Tenaska +manages 100+ GW of generation, storage, and retail load across all US ISOs/RTOs. +This is the exact Model-2→Model-3 stitch we identified: software vendor rents the +execution rail rather than building it. It also means Ascend can serve clients +who lack their own market participation — directly encroaching on the long-tail +segment we identified as whitespace. + +**Strengths.** Two-decade brand and analyst credibility; the valuation-to- +operations funnel (developers who used BatterySIMM to finance a project are +natural SmartBidder customers); PE capital for M&A and growth; Tenaska execution +path; genuine nodal/market-intelligence depth. + +**Weaknesses / openings.** Storage-centric product framing (SmartBidder is +consistently described as a storage bidding platform, with renewables as hybrid +add-ons); enterprise price points and consulting-heavy culture may leave sub-scale +renewable owners underserved; PE ownership pressures ARR growth, which typically +biases toward larger accounts. + +**Sources.** ascendanalytics.com/solutions/smartbidder; BusinessWire (Mar 2024 +investment); ZoomInfo/CB Insights/Dialectica profiles; PRNewswire (Tokyo Gas, +Jun 2025; Tenaska collaboration, Aug 2025). + +--- + +### 2.3 Tyba Energy + +**What it is.** AI-based forecasting and optimization platform for energy storage +(and increasingly hybrid/renewable) projects, with two products: Project +Simulation (development-stage revenue modeling; used on 100+ GW of projects and +credited in $1B+ of project financings) and Asset Operations (DA/RT price +forecasts with confidence intervals, automated bidding strategies and dispatch, +"autopilot or control center" positioning). + +**Origin & funding.** Founded ~2021 (CEO Michael Baker). $13.9M Series A (Feb +2025) led by Energize Capital, joined by Pear VC, Mobilize, Borusan and existing +investors; ~$18.2M total raised. Series A earmarked for expansion into new +markets and asset classes. + +**Scale & markets.** Supports >1 GWh of operating storage in ERCOT and CAISO; +customer base tripled in the year to Feb 2025; marquee names include +TotalEnergies (US short-term power), Intersect Power, White Pine Renewables, +Linea Energy. Performance claims: default low-risk strategy ~48% above median +ERCOT asset revenue; top-5% asset outcomes; strong spike-capture case studies +(Nov 10, 2024 ERCOT event). Notably for us, Tyba has published a co-located +wind+storage optimization case (shared-POI constraints) — evidence of movement +toward renewable-aware bidding — and rolled out Dynamic Price-Quantity bidding +(price-contingent offer curves rather than fixed quantities), conceptually +adjacent to our offer-curve framework. + +**Strengths.** Modern product velocity; strong early performance marketing; +credible investors; the simulation→operations funnel mirrors Ascend's at startup +speed; developer-friendly positioning could resonate with mid-size IPPs. + +**Weaknesses / openings.** Small (Series A) with ERCOT/CAISO focus — PJM depth +unproven; battery-first ("autopilot for batteries" is the company's own framing); +capacity-market/must-offer complexity of PJM is a heavier lift than its current +markets; limited services/execution arm — clients still need market access. + +**Sources.** tyba.ai (home, asset-operations, PQ-bidding guide, Series A +release); PRNewswire (Feb 2025); Energize Capital investment memo; Heatmap +interview via tyba.ai. + +--- + +### 2.4 Gridmatic + +**What it is.** An AI-first power company — not primarily a SaaS vendor. It +trades in all 7 US organized markets with fully automated bidding; signs tolling/ +offtake agreements with battery owners; operates a $50M energy storage fund +(capitalized to manage up to 500 MW in ERCOT/CAISO); runs a licensed retail +business (Gridmatic Retail) in ERCOT and PJM (Ohio, then Pennsylvania, Feb 2026); +and offers scheduling/bidding, DME (commercial decision-making), and non-binding +auto-bidder services across ERCOT, CAISO, PJM, SPP, and MISO (per Modo Energy's +optimizer directory). + +**Origin & economics.** Silicon Valley company; reportedly ran profitably for six +years before raising outside capital, and claims to be the most profitable +participant in ERCOT's wholesale market (the one US market where trading results +are publicly attributable) and operator of the top-performing CAISO battery. +Claims of up to 46% storage revenue uplift from its AI (2022 ERCOT backtest). +Builds proprietary weather models feeding deep-learning price forecasts across +thousands of nodes. + +**Relevance to us.** Gridmatic is the purest "monetize the model directly" +competitor: rather than selling software, it takes positions and offers offtake/ +tolling — for a renewable owner, a Gridmatic-style offtake can substitute +entirely for buying bidding software (the optimization is embedded in the +contract price). Its PJM expansion (retail + market operations) means it is +building exactly the PJM forecasting infrastructure we contemplate, with a +balance sheet and settlement data flywheel behind it. + +**Strengths.** Proven trading P&L (public in ERCOT); vertically integrated data +flywheel (trading + retail + storage operations feed the models); no dependence +on SaaS sales cycles; deep-learning nodal price forecasting at scale. + +**Weaknesses / openings.** Not a neutral software vendor — asset owners wanting +control and transparency (rather than an offtake counterparty) are not its +natural customers; its PJM entry is load/retail-led, and public evidence of +standalone wind/solar DA bidding services in PJM is limited; small team by its own +description — bandwidth constraints on bespoke service. + +**Sources.** gridmatic.com (home, about, news); Crunchbase; BusinessWire (storage +fund, Nov 2023; PJM/Ohio launch; Pennsylvania expansion, Feb 2026); Modo Energy +optimizer directory (2025). + +--- + +### 2.5 Dexter Energy + +**What it is.** Amsterdam-based (founded 2017) provider of AI forecasting and +short-term trade optimization for renewable and battery portfolios: power +forecasting (wind/solar/prosumption, asset-level; self-described top-3 accuracy), +real-time imbalance/price forecasting, and DA bidding strategies ("trading +signals") built as an ensemble of ML forecasts plus fundamental market signals +with value-at-risk overlays, drawing on 40+ external data sources and ~1,000 GB +of weather data daily. Positioning has evolved toward "Trading-as-a-Service." + +**Funding & scale.** €23M Series C (July 2025) led by Alantra's Klima fund with +Mirova; ~$40M lifetime funding per PitchBook; ~100+ employees; 80+ energy-company +clients including Centrica Energy, Luminus, Scholt Energy, Pure Energie. +European-market focused (day-ahead, intraday, imbalance markets across multiple +countries). + +**Relevance to us.** Dexter is the closest European analog to our Model 2: it +sells the forecast+signal layer to asset-backed traders rather than becoming the +market party. Its DA bidding-strategy product is philosophically identical to our +Model C (turn DA–imbalance spread risk into optimized offers). Its trajectory — +forecasting vendor → trade optimization → TaaS — is the migration path we +predicted. **US presence: not yet evident in public materials.** Its Series C is +explicitly earmarked for European expansion (new EU markets, battery products). +That makes Dexter a medium-term threat (or acquisition-track peer/partner) rather +than a present PJM competitor — but the European imbalance-market discipline it +embodies is exactly the skill set that transfers to DA–RT spread bidding. + +**Strengths.** Genuine probabilistic forecasting depth; large European client +proof base; ensemble ML + fundamentals architecture (mirrors our hybrid-lite +design); Series C capital. + +**Weaknesses / openings.** No US market machinery today (no ISO integrations, +nodal LMP/congestion modeling, or FERC/ISO compliance footprint in evidence); +European market structure differs materially (zonal prices, no nodal congestion, +different settlement) so its edge does not port one-for-one to PJM. + +**Sources.** dexterenergy.ai (home, solutions, Series C release, 2026 trading +outlook); PitchBook profile; Mirova/Alantra releases; Crunchbase. + +--- + +### 2.6 enspired + +**What it is.** Vienna-based (founded 2020) trading-as-a-service provider: a +fully automated, AI-driven (reinforcement-learning-emphasized) trading platform +that commercially optimizes flexible assets — BESS foremost, plus renewables and +consumption assets — simultaneously across wholesale (DA/intraday), reserve, and +ancillary markets. Clients hand over trading execution; enspired trades on their +behalf (the European route-to-market model, i.e., our Model 3 analog). + +**Funding & scale.** Series A €8.9M (2021); Series B €25.5M (May 2024, led by +Zouk Capital) extended to >€40M (Oct 2025, adding Future Energy Ventures). +Crossed 1 GW of BESS under management; entered six new markets in the year to +Oct 2025; publishes certified revenue benchmarks (unusual transparency in this +space). Clients reported include Gore Street Capital, TotalEnergies, Uniper, +Priogen. ~110+ employees. + +**US relevance.** The Series B extension is explicitly earmarked for expansion +beyond Europe, with the US and Asia named; Japan (via Banpu NEXT) is the first +non-European deployment. As of this writing there is no public evidence of live +US/PJM operations — so, like Dexter, a announced-intent entrant rather than an +incumbent. When it lands, expect a storage-led, full-service (Model 3) offering +requiring US market-participant infrastructure it must build or partner for. + +**Strengths.** Purpose-built TaaS operating model with 24/7 automated execution; +cross-market simultaneous optimization; certified-results transparency as a sales +weapon; strong European reference base. + +**Weaknesses / openings.** BESS-centric; zero US regulatory/ISO footprint today; +US entry will require QSE/participant buildout or partnership — 12–24 months of +runway during which a PJM-native offering can entrench; renewable DA bidding is +peripheral to its battery core. + +**Sources.** enspired-trading.com; ess-news.com (Oct 2025 Series B extension); +Vestbee, TFN, Dealroom, Tracxn profiles; Emerald Ventures release (May 2024). + +--- + +## 3. Comparative matrix + +| | Fluence Mosaic | Ascend SmartBidder | Tyba | Gridmatic | Dexter | enspired | +|---|---|---|---|---|---|---| +| **HQ / founded** | Arlington VA / 2018 (AMS 2013) | Boulder CO / 2002 | US / ~2021 | Silicon Valley / ~2017 | Amsterdam / 2017 | Vienna / 2020 | +| **Ownership** | Public (FLNC) | PE (Rubicon-led) | VC (Series A, $18M) | Founder/bootstrapped + fund vehicles | VC (Series C, ~$40M) | VC (Series B, >€75M lifetime) | +| **Archetype** | OEM software arm | Analytics → bidding | Startup SaaS | AI trader/marketer | Forecast+signals SaaS→TaaS | TaaS (trades for you) | +| **Asset focus** | Storage ≫ renewables | Storage ≫ hybrid | Storage ≫ hybrid | Storage + retail + renewables offtake | Renewables + BESS | BESS ≫ renewables | +| **US ISO coverage (public)** | CAISO, ERCOT (+NEM) | ERCOT/CAISO native; all ISOs via Tenaska | ERCOT, CAISO | All 7 markets (trading); PJM retail live | None yet | None yet (announced intent) | +| **PJM depth today** | Thin | Via Tenaska rail | Unproven | Building (retail-led) | None | None | +| **Execution path (Model 3)** | Mosaic Trading Solutions | Tenaska partnership | None public | Native (is the market participant) | No (signals only) | Native (EU); US TBD | +| **Standalone renewable DA bidding** | Legacy AMS capability (~10% uplift claim) | Hybrid-framed | Emerging (wind+storage POI case) | Via offtake structures | Core product (EU) | Peripheral | +| **Key claim (unverified vendor figures)** | 16 GW deployed/awarded | 90–95% of perfect foresight; 50%+ vs ERCOT/CAISO benchmarks | Top-5% outcomes; +48% vs ERCOT median | Most profitable ERCOT participant; +46% storage uplift | Top-3 forecast accuracy; up to 35% balancing-cost cuts | 1 GW+ BESS; certified revenues | + +--- + +## 4. Synthesis: what this means for our venture + +**4.1 The whitespace is confirmed, with a clock on it.** No vendor today offers a +PJM-native, renewable-first DA bidding product with nodal congestion intelligence +and capacity/must-offer awareness. But three separate forces are converging on it: +Ascend×Tenaska (top-down, via managed services), Gridmatic (via balance-sheet +offtake and PJM retail), and the Europeans (Dexter/enspired, via announced US +entry). Realistic exclusivity window: roughly 12–24 months of PJM-specific +buildout time that incumbents must also spend. + +**4.2 Competitive positioning implications.** +- *Against Fluence/Ascend/Tyba (software):* differentiate on PJM nodal depth + (spread atlas, congestion layer, negative-price classification) and on serving + energy-only renewables economically — their storage-optimized cost structures + and price points are misaligned with thin renewable fees. +- *Against Gridmatic/enspired (they trade for you):* differentiate on + transparency, client control, and alignment — we optimize the client's position + rather than becoming their counterparty. Expect procurement processes to + compare "SaaS fee" vs. "embedded offtake discount"; we need a clean + $/MWh-uplift evidence pack (our economic backtest protocol) to win that + comparison. +- *Against the Ascend×Tenaska configuration:* we likely need our own execution + rail early — a partnership with a CES-, Tenaska-, or similar QSE-class firm — + or we lose every deal where the client lacks market access. + +**4.3 Partnership / channel candidates surfaced by this research.** Tenaska is +taken (Ascend). Customized Energy Solutions and comparable QSE/asset-management +firms are the natural counterweight partners. Dexter is a conceivable technology +partner or future consolidator given zero US footprint and philosophical +alignment; enspired likewise on the Model-3 side. + +**4.4 Pricing intelligence gap.** None of the six publishes pricing. Anecdotally +the market spans roughly $1–3k/site/month for forecast feeds up to %-of-revenue +or $/kW-month structures for managed optimization (Tyba cites $16/kW-month *net +revenue* in one wind+storage case — an outcome figure, not a fee). Primary +research task: gather 5–10 datapoints via prospective-client interviews before +setting our own pricing. + +**4.5 Fundraising benchmarks.** Category rounds cluster at $10–25M (Series A/B) +for software plays; enspired's >€40M extension and Ascend's PE recap mark the +scale-up tier. Gridmatic demonstrates the self-funded trading alternative. For +our Model 2 with an eye to Model 3, the enspired capitalization path (A ~€9M → +B ~€25M+ext) is the closest comparable. + +**4.6 Claims hygiene.** Every uplift figure in Section 2–3 is a vendor marketing +claim on non-comparable baselines (median asset, manual bidding, perfect +foresight, backtests). Treat as positioning, not benchmarks. Our own evaluation +standard (pinball loss vs. S≡0; economic backtest vs. P50-naive; regime-sliced) +should be a sales asset precisely because the market's claims are unauditable — +enspired's "certified revenues" move shows transparency wins deals. + +--- + +## 5. Recommended follow-ups + +1. Primary pricing research: 8–10 structured interviews with renewable IPPs in + PJM on current bidding arrangements, fees, and pain points. +2. Deep-dive on Ascend×Tenaska deal mechanics (rev share? exclusivity? which + asset classes?) — determines whether QSE partners remain available to us. +3. Monitor triggers: Dexter/enspired US-entity registrations or ISO memberships; + Gridmatic generation-side (non-retail) PJM announcements; Fluence Mosaic PJM + market-coverage additions; Tyba post-Series-A market expansion announcements. +4. Secondary tier not covered here (recommend a v2 of this report): Tesla + Autobidder, Habitat Energy (Hartree), Stem, PCI/Hitachi/Yes Energy execution + platforms, Amperon and other forecast-layer vendors, and QSE incumbents (CES, + ACES, The Energy Authority, EDF Trading NA). +5. Re-verify all funding/headcount/scale figures against primary filings before + external use (investor decks, board materials). + +--- + +## Appendix: Glossary + +| Term | Definition | +|---|---| +| Autobidder | Software that computes and (often) submits market bids for an asset automatically | +| Bid-to-bill | Software category covering the full market-operations chain: bidding, scheduling, settlement, billing | +| Model 2 / Model 3 | Our shorthand: bid-optimization SaaS (client remains market participant) / asset-management services (vendor is participant of record) | +| QSE / scheduling entity | Qualified Scheduling Entity — holds market-participant infrastructure (credentials, credit, 24/7 desk) and transacts on assets' behalf | +| Route-to-market | European model: a service firm becomes the market party for a renewable asset, handling forecasting, bidding, and balancing risk | +| TaaS | Trading-as-a-Service — client hands over trading execution to the vendor's automated platform | +| Offtake / tolling | Contract where a counterparty buys the plant's output (offtake) or rents its capacity and takes market risk (tolling) — substitutes for buying bidding software | +| VPPA | Virtual PPA — financial CfD between generator and (usually corporate) buyer, settled against a hub/node price | +| DA / RT | Day-ahead / real-time markets; the DA–RT spread drives renewable bid optimization | +| Co-optimization | Jointly optimizing energy + ancillary-service offers across DA and RT (the core battery use case) | +| Perfect-foresight benchmark | Revenue achievable with perfect knowledge of prices; vendor claims like "90–95% of perfect foresight" are measured against it | +| Uplift claims | Vendor-reported revenue improvement vs. a baseline (median asset, manual bidding, backtest); baselines are non-comparable across vendors | +| ISO/RTO | Independent System Operator / Regional Transmission Organization (PJM, ERCOT, CAISO, MISO, SPP, NYISO, ISO-NE) | +| BESS | Battery Energy Storage System | +| IPP | Independent Power Producer | +| ARR | Annual Recurring Revenue (SaaS metric) | diff --git a/pjm/vendor_dd_zh.md b/pjm/vendor_dd_zh.md new file mode 100644 index 0000000..9dd03ee --- /dev/null +++ b/pjm/vendor_dd_zh.md @@ -0,0 +1,197 @@ +# 竞争尽职调查报告：自动报价与竞价优化软件供应商 + +**编制对象：** [可再生能源日前竞价创业项目 —— 名称待定] +**日期：** 2026年7月10日 +**范围：** 六家与"模式二"（面向 PJM 可再生能源发电商的竞价优化 SaaS）业务最直接或最相邻的竞争对手：Fluence Mosaic、Ascend Analytics（SmartBidder）、Tyba Energy、Gridmatic、Dexter Energy 及 enspired。 +**状态：** 工作草稿，基于 2026 年 7 月检索的公开资料（公司官网、新闻稿、SEC 文件、融资数据库、行业媒体）。私营公司数据（融资、员工数、收入）均转引自第三方数据库，在用于任何商业决策前须重新核实。所有供应商均未公开定价；定价情报需通过一手调研（客户访谈、参与 RFP）获取。 + +--- + +## 一、执行摘要 + +六家供应商可归为三种战略类型： + +| 类型 | 供应商 | 资产负债表定位 | 主攻资产类别 | +|---|---|---|---| +| **硬件巨头的软件部门** | Fluence Mosaic | 上市母公司（纳斯达克：FLNC） | 储能优先，可再生其次 | +| **分析软件老牌厂商向竞价延伸** | Ascend Analytics；Tyba（更年轻的同类） | PE 控股 / VC 投资 | 储能优先 | +| **AI 交易商/优化商（自持或代管头寸）** | Gridmatic；enspired；Dexter（仅提供信号的变体） | 自力盈利 / VC 投资 | 储能与可再生 | + +对我方定位的关键结论： + +1. **所有供应商均以储能为主导或高度侧重储能。** 该市场的经济重心在 ERCOT 和 CAISO 的电池套利与辅助服务联合优化。独立风电/光伏的日前竞价——即我们的"报童模型"问题——对六家而言都只是次要功能。目前没有任何一家提供"PJM 优先、具备节点阻塞感知能力"的可再生日前竞价产品。 +2. **各家在 PJM 的布局薄弱且起步不久。** Gridmatic 通过售电业务进入 PJM（先俄亥俄州，2026 年 2 月扩至宾夕法尼亚州）；Ascend 借助 Tenaska 合作触达 PJM；Fluence Mosaic 公开宣传的市场覆盖以 CAISO/ERCOT/澳洲 NEM 为主。两家欧洲公司（Dexter、enspired）尚处"拟进入美国"阶段，并未在美落地。 +3. **胜出的业务形态正在成形：算法 + 7×24 交易台 + 市场主体基础设施。** Ascend×Tenaska 的合作是明证；Gridmatic 与 enspired 则天然具备这一形态。缺乏执行通道的纯软件"模式二"，在每一笔竞标中都将面对这种组合。 +4. **两种融资现实并存。** 该赛道资本充裕（Tyba 1800 万美元、Dexter 累计约 4000 万欧元、enspired 累计逾 7500 万欧元、Ascend 获 PE 注资），但 Gridmatic 的"自力盈利的交易商"模式展示了另一条路：直接在市场中用模型变现，而非收取 SaaS 费用。 + +--- + +## 二、供应商档案 + +### 2.1 Fluence Mosaic（Fluence Energy, Inc. —— 纳斯达克：FLNC） + +**产品定位。** 面向电网级储能与可再生（风、光）资产的 AI 自动竞价软件，属 Fluence IQ 数字平台的一部分；另设较新的"Mosaic Trading Solutions"服务部门，提供从报价到结算（bid-to-bill）的支持。 + +**渊源与历史。** Fluence（2018 年由西门子与 AES 合资成立，2021 年 IPO）于 2020 年底收购 Advanced Microgrid Solutions（AMS）。AMS 打造了其算法竞价引擎——收购时约管理 1.7 GW 竞价资产（主要在澳洲 NEM 和 CAISO），并宣称优化后的批发市场竞价可为可再生资产带来 10% 以上的收入提升。早期旗舰案例：CAISO 市场中 PG&E 的 Moss Landing 储能电站（182.5 MW / 730 MWh）。 + +**规模与市场。** 营销材料称平台已部署或中标约 16 GW 资产（2024 年 1 月新闻稿曾称"已签约或管理" 22.3 GW——各期口径混用"部署/中标/签约"，不可直接比较，须谨慎对待）。公开的市场覆盖：CAISO（独立储能、共址，2024 年起支持混合资源）、ERCOT 及澳洲 NEM。公开材料中 PJM 并不突出——这是一个值得注意的空档。 + +**产品要点。** 概率化机器学习价格预测；日前/实时市场的电能与辅助服务联合优化；风险偏好参数化；电池质保与衰减约束；人工干预开关；技术中立（支持非 Fluence 硬件）。 + +**财务状况（母公司）。** 2025 财年收入 23 亿美元（低于 2024 财年的 27 亿美元），净亏损 6800 万美元；2026 财年 Q1 收入 4.752 亿美元（同比 +154%），GAAP 毛利率约 5%；Q2 收入 4.649 亿美元，净亏损 2920 万美元，在手订单创纪录达 56 亿美元，并披露与数据中心/超大规模客户签订主供协议的进展。据分析师口径，数字业务板块有望在 2026 财年末达到约 1.8 亿美元 ARR（非经审计的分部披露）。股票综合评级约为"持有"。小结：硬件业务波动大、毛利薄；软件板块对其股票故事具有战略意义，Mosaic 大概率将持续获得投入。 + +**优势。** 品牌与全球装机基础；深厚的电池运行数据；上市公司资源；硬件+软件+服务一体化销售；标杆客户（AES、PG&E）。 + +**弱点 / 我方机会。** 储能基因——独立可再生日前竞价只是相邻用例而非核心；PJM 深度与节点阻塞专长缺乏公开证据；大公司销售动作难以覆盖中小 IPP 长尾；潜在渠道冲突（Fluence 硬件业务与部分潜在客户的设备供应商存在竞争）。 + +**资料来源。** fluenceenergy.com/mosaic-intelligent-bidding-software；Fluence 博客（AMS 收购，2020–21）；ir.fluenceenergy.com（2024 年 1 月 CAISO 混合资源公告；2025 年 11 月年报；2026 财年 Q1/Q2 8-K）；stockanalysis.com/flnc；Trefis。 + +--- + +### 2.2 Ascend Analytics —— SmartBidder + +**产品定位。** 面向独立储能、共址及"可再生+储能"混合项目的竞价优化与调度平台，采用概率化与 AI 辅助模型，纳入节点价格动态、资产约束及基于风险的优化；交付形态从自助软件到全托管服务均可。其外围是更大的分析产品体系：BatterySIMM（储能估值，被业内广泛视为事实标准）、PowerSIMM（组合/资源规划）、PowerVAL、AscendMI 市场情报，以及咨询服务。 + +**渊源与股权。** 2002 年由 Gary Dorris 博士创立于科罗拉多州博尔德，以咨询/分析业务起家。2024 年 3 月获 PE 机构 Rubicon Technology Partners 领投、Galvanize Climate Solutions 与 Silversmith 参与的战略增长投资（属规模化阶段的资本重组）。第三方数据显示员工数从 2020 年约 57 人增至 2024 年约 162 人。 + +**规模与市场。** 客户群据称包括北美与欧洲最大的公用事业公司和可再生开发商。SmartBidder 公开案例偏重 ERCOT/CAISO 储能（如 2025 年 6 月东京燃气美国公司在 ERCOT 的 174 MW Longbow 储能项目）。宣传口径包括：收入达到"完美预见"基准的 90–95%，较 ERCOT/CAISO 同类储能项目提升 50% 以上（供应商自述，方法论未经独立验证）。 + +**Tenaska 合作（对我方至关重要）。** Ascend 与 Tenaska Power Services 于 2025 年 8 月合作，将 SmartBidder 的分析与交易算法同 Tenaska 的 7×24 实时交易台及 QSE/市场主体基础设施结合——Tenaska 在美国所有 ISO/RTO 管理逾 100 GW 的发电、储能与售电负荷。这正是我们提出的"模式二→模式三缝合"：软件商租用执行轨道而非自建。这也意味着 Ascend 已能服务缺乏市场准入的客户——直接侵入我们认定为空白的长尾市场。 + +**优势。** 二十余年品牌与分析公信力；"估值→运营"客户漏斗（用 BatterySIMM 完成融资的开发商是 SmartBidder 的天然客户）；PE 资本支持并购与扩张；Tenaska 执行通道；真正的节点级市场情报深度。 + +**弱点 / 我方机会。** 产品叙事以储能为中心（SmartBidder 一贯被描述为储能竞价平台，可再生只是混合场景的附加）；企业级定价与咨询导向的文化可能使中小可再生业主得不到匹配服务；PE 股东对 ARR 增长的诉求通常导致偏向大客户。 + +**资料来源。** ascendanalytics.com/solutions/smartbidder；BusinessWire（2024 年 3 月投资公告）；ZoomInfo / CB Insights / Dialectica 档案；PRNewswire（2025 年 6 月东京燃气；2025 年 8 月 Tenaska 合作）。 + +--- + +### 2.3 Tyba Energy + +**产品定位。** 面向储能（并日益覆盖混合/可再生）项目的 AI 预测与优化平台，两条产品线：Project Simulation（开发期收入建模；已用于逾 100 GW 项目并助力逾 10 亿美元项目融资）与 Asset Operations（带置信区间的日前/实时价格预测、自动竞价策略与调度，定位为"自动驾驶或控制中心"）。 + +**渊源与融资。** 约 2021 年成立（CEO：Michael Baker）。2025 年 2 月完成 1390 万美元 A 轮，由 Energize Capital 领投，Pear VC、Mobilize、Borusan 及老股东跟投；累计融资约 1820 万美元。A 轮资金明确用于拓展新市场与新资产类别。 + +**规模与市场。** 支撑 ERCOT 与 CAISO 逾 1 GWh 在运储能；截至 2025 年 2 月的一年内客户数增长两倍；标杆客户含 TotalEnergies（美国短期电力）、Intersect Power、White Pine Renewables、Linea Energy。业绩宣称：默认低风险策略较 ERCOT 中位资产收入高约 48%；资产收益位居前 5%；对价格尖峰的捕捉案例突出（2024 年 11 月 10 日 ERCOT 事件）。与我方尤为相关的是：Tyba 已发布共址"风电+储能"优化案例（共享并网点约束）——显示其正向"可再生感知型竞价"演进——并推出 Dynamic Price-Quantity 竞价（价格条件化的报价曲线而非固定电量），概念上与我们的报价曲线框架相邻。 + +**优势。** 现代化的产品迭代速度；强势的早期业绩营销；投资人可信；"仿真→运营"漏斗以创业公司速度复刻 Ascend 打法；开发者友好的定位对中型 IPP 具吸引力。 + +**弱点 / 我方机会。** 体量小（A 轮），聚焦 ERCOT/CAISO——PJM 深度未经验证；电池优先（"电池的自动驾驶"是其自我定位）；PJM 的容量市场/强制申报复杂度远高于其现有市场；缺乏服务/执行部门——客户仍需自行解决市场准入。 + +**资料来源。** tyba.ai（官网、asset-operations、PQ 竞价指南、A 轮公告）；PRNewswire（2025 年 2 月）；Energize Capital 投资备忘；tyba.ai 转载的 Heatmap 访谈。 + +--- + +### 2.4 Gridmatic + +**业务定位。** 一家"AI 优先"的电力公司——并非以 SaaS 为主业。其在美国全部 7 个电力市场以全自动竞价开展交易；与电池业主签订租赁/包销（tolling/offtake）协议；运营一支 5000 万美元储能基金（额度可管理 ERCOT/CAISO 至多 500 MW）；持牌开展售电业务（Gridmatic Retail，覆盖 ERCOT 与 PJM——先俄亥俄州，2026 年 2 月扩至宾州）；并按 Modo Energy 优化商名录，在 ERCOT、CAISO、PJM、SPP、MISO 提供调度/竞价、商业决策（DME）与非绑定自动报价服务。 + +**渊源与经济性。** 硅谷公司；据报道在引入外部资本前已自力盈利六年，并宣称是 ERCOT 批发市场中最赚钱的参与者（ERCOT 是美国唯一交易结果可公开归属的市场），且运营 CAISO 表现最佳的电池。宣称其 AI 可为储能带来至多 46% 的收入提升（2022 年 ERCOT 回测）。自建气象模型，驱动覆盖数千节点的深度学习价格预测。 + +**与我方的关联。** Gridmatic 是"直接用模型变现"最纯粹的竞争者：不卖软件，而是自持头寸并提供包销/租赁——对可再生业主而言，Gridmatic 式包销可完全替代购买竞价软件（优化能力已内嵌于合同价格）。其 PJM 扩张（售电+市场运营）意味着它正在搭建我们设想中的那套 PJM 预测基础设施，且背后有资产负债表与结算数据飞轮。 + +**优势。** 经市场验证的交易盈利（ERCOT 公开可查）；纵向一体化的数据飞轮（交易+售电+储能运营反哺模型）；不依赖 SaaS 销售周期；规模化的深度学习节点价格预测。 + +**弱点 / 我方机会。** 并非中立软件商——希望保留控制权与透明度（而非接受一个包销对手方）的资产业主并非其天然客户；其 PJM 切入以负荷/售电为先，独立风光在 PJM 的日前竞价服务缺乏公开证据；按其自述团队精干——定制化服务的带宽受限。 + +**资料来源。** gridmatic.com（官网、关于我们、新闻）；Crunchbase；BusinessWire（2023 年 11 月储能基金；PJM/俄亥俄上线；2026 年 2 月宾州扩张）；Modo Energy 优化商名录（2025）。 + +--- + +### 2.5 Dexter Energy + +**产品定位。** 总部阿姆斯特丹（2017 年成立），为可再生与电池组合提供 AI 预测与短期交易优化：功率预测（风/光/产消者，资产级；自述精度居市场前三）、实时不平衡/价格预测，以及日前竞价策略（"交易信号"）——采用机器学习预测与基本面市场信号的集成（ensemble）并叠加在险价值（VaR）流程，依托 40 余个外部数据源、每日处理约 1000 GB 气象数据。定位正演进为"交易即服务"（TaaS）。 + +**融资与规模。** 2025 年 7 月完成 2300 万欧元 C 轮，由 Alantra 旗下 Klima 基金领投、Mirova 参与；PitchBook 口径累计融资约 4000 万美元；员工 100 余人；服务 80 余家能源企业，含 Centrica Energy、Luminus、Scholt Energy、Pure Energie。聚焦欧洲市场（多国日前、日内与不平衡市场）。 + +**与我方的关联。** Dexter 是我们"模式二"最接近的欧洲镜像：向资产型交易商出售"预测+信号"层，而不自己成为市场主体。其日前竞价策略产品与我们的模型 C 在哲学上同构（将"日前–不平衡"价差风险转化为优化报价）。其发展轨迹——预测商 → 交易优化 → TaaS——正是我们预判的迁移路径。 **美国布局：公开材料中尚无迹象。** 其 C 轮资金明确用于欧洲扩张（新的欧盟市场、电池产品）。因此 Dexter 是中期威胁（或潜在并购轨道上的同行/伙伴），而非当下的 PJM 竞争者——但它所体现的欧洲不平衡市场功力，恰是可迁移到日前–实时价差竞价的技能组合。 + +**优势。** 扎实的概率预测能力；庞大的欧洲客户验证；"集成机器学习+基本面"架构（与我们的轻量混合设计互为镜像）；C 轮资本。 + +**弱点 / 我方机会。** 当前不具备任何美国市场机制能力（无 ISO 对接、无节点 LMP/阻塞建模、无 FERC/ISO 合规布局的公开证据）；欧洲市场结构差异显著（分区电价、无节点阻塞、结算机制不同），其优势无法一比一移植到 PJM。 + +**资料来源。** dexterenergy.ai（官网、solutions、C 轮公告、2026 交易展望）；PitchBook 档案；Mirova/Alantra 公告；Crunchbase。 + +--- + +### 2.6 enspired + +**业务定位。** 总部维也纳（2020 年成立）的"交易即服务"提供商：一套全自动、以强化学习为卖点的 AI 交易平台，对灵活性资产——以电池储能为首，兼及可再生与用电侧资产——在批发（日前/日内）、备用与辅助服务市场进行跨市场同步商业优化。客户将交易执行整体托付，由 enspired 代其交易（即欧洲的 route-to-market 模式，对应我们的"模式三"）。 + +**融资与规模。** A 轮 890 万欧元（2021）；B 轮 2550 万欧元（2024 年 5 月，Zouk Capital 领投），2025 年 10 月扩至逾 4000 万欧元（新增 Future Energy Ventures）。管理电池储能突破 1 GW；截至 2025 年 10 月的一年内进入六个新市场；公开发布经认证的收入基准（业内罕见的透明度）。据报道客户包括 Gore Street Capital、TotalEnergies、Uniper、Priogen。员工约 110+。 + +**美国相关性。** B 轮扩募资金明确用于欧洲以外扩张，点名美国与亚洲；首个非欧部署是日本（与 Banpu NEXT 合作）。截至本文撰写，尚无美国/PJM 实际运营的公开证据——与 Dexter 一样属于"官宣意向型"进入者而非在位者。一旦落地，预计将以储能为先、以全托管（模式三）形态出现，且必须自建或合作获得美国市场主体基础设施。 + +**优势。** 专为 TaaS 打造的运营模式，7×24 全自动执行；跨市场同步优化；以"认证业绩"作为销售武器；欧洲客户背书扎实。 + +**弱点 / 我方机会。** 以电池储能为核心；当前美国监管/ISO 布局为零；进入美国需自建或合作 QSE/市场主体能力——这 12–24 个月的窗口期正是 PJM 原生产品建立壁垒的时机；可再生日前竞价处于其电池主业的边缘。 + +**资料来源。** enspired-trading.com；ess-news.com（2025 年 10 月 B 轮扩募）；Vestbee、TFN、Dealroom、Tracxn 档案；Emerald Ventures 公告（2024 年 5 月）。 + +--- + +## 三、对比矩阵 + +| | Fluence Mosaic | Ascend SmartBidder | Tyba | Gridmatic | Dexter | enspired | +|---|---|---|---|---|---|---| +| **总部 / 成立** | 弗吉尼亚州阿灵顿 / 2018（AMS 2013） | 科罗拉多州博尔德 / 2002 | 美国 / 约2021 | 硅谷 / 约2017 | 阿姆斯特丹 / 2017 | 维也纳 / 2020 | +| **股权结构** | 上市（FLNC） | PE（Rubicon 领投） | VC（A 轮，1800 万美元） | 创始人/自力盈利 + 基金载体 | VC（C 轮，累计约 4000 万美元） | VC（B 轮，累计逾 7500 万欧元） | +| **类型** | 设备商软件部门 | 分析商→竞价 | 创业 SaaS | AI 交易商/电力营销商 | 预测+信号 SaaS→TaaS | TaaS（代客交易） | +| **资产侧重** | 储能 ≫ 可再生 | 储能 ≫ 混合 | 储能 ≫ 混合 | 储能 + 售电 + 可再生包销 | 可再生 + 储能 | 储能 ≫ 可再生 | +| **美国 ISO 覆盖（公开）** | CAISO、ERCOT（+NEM） | ERCOT/CAISO 原生；经 Tenaska 覆盖全部 ISO | ERCOT、CAISO | 全部 7 个市场（交易）；PJM 售电已运营 | 暂无 | 暂无（官宣意向） | +| **当前 PJM 深度** | 薄弱 | 借 Tenaska 轨道 | 未验证 | 建设中（售电先行） | 无 | 无 | +| **执行通道（模式三）** | Mosaic Trading Solutions | Tenaska 合作 | 无公开信息 | 原生（自身即市场主体） | 无（仅信号） | 原生（欧洲）；美国待定 | +| **独立可再生日前竞价** | AMS 传承能力（宣称约 10% 提升） | 以混合项目为框架 | 萌芽（风+储 POI 案例） | 通过包销结构实现 | 核心产品（欧洲） | 边缘 | +| **核心宣称（未经验证的厂商口径）** | 16 GW 部署/中标 | 达完美预见的 90–95%；较 ERCOT/CAISO 基准 +50% | 前 5% 收益；较 ERCOT 中位数 +48% | ERCOT 最赚钱参与者；储能 +46% | 预测精度前三；平衡成本最高降 35% | 逾 1 GW 电池储能；认证收入 | + +--- + +## 四、综合研判：对本项目的意义 + +**4.1 空白得到确认，但有时限。** 目前没有任何供应商提供"PJM 原生、可再生优先、具备节点阻塞情报与容量/强制申报意识"的日前竞价产品。但三股力量正向该空白汇聚：Ascend×Tenaska（自上而下，经托管服务）、Gridmatic（凭资产负债表包销与 PJM 售电）、欧洲双雄（Dexter/enspired，官宣入美）。现实的独占窗口：约 12–24 个月——即在位者同样必须投入的 PJM 专项建设周期。 + +**4.2 竞争定位启示。** +- *对 Fluence/Ascend/Tyba（软件类）：* 以 PJM 节点深度（价差图谱、阻塞层、负电价分类器）以及"经济地服务纯电能可再生资产"实现差异化——其面向储能优化的成本结构与价位，与可再生的微薄费率错配。 +- *对 Gridmatic/enspired（代客交易类）：* 以透明度、客户控制权与利益一致性差异化——我们优化的是客户自己的头寸，而非成为其对手方。预计采购流程会比较"SaaS 费用"与"内嵌于包销折价"两种方案；我们需要一套干净的 $/MWh 提升证据包（即我们的经济回测协议）来赢下这类比较。 +- *对 Ascend×Tenaska 组合：* 我们大概率需尽早搭建自己的执行轨道——与 CES、Tenaska 同级的 QSE 类公司合作——否则凡客户缺乏市场准入的交易我们都将出局。 + +**4.3 本次调研浮现的合作/渠道候选。** Tenaska 已被 Ascend 锁定。Customized Energy Solutions 及同级 QSE/资产管理公司是天然的制衡型合作方。鉴于 Dexter 尚无美国布局且理念契合，可作为潜在技术伙伴或未来整合方；enspired 在"模式三"一侧同理。 + +**4.4 定价情报缺口。** 六家均不公开定价。经验区间大致为：预测数据订阅约 1–3 千美元/场站/月，至托管优化的收入分成或 $/kW·月结构（Tyba 在一个风+储案例中提到 16 美元/kW·月的*净收入*——这是业绩数字而非收费标准）。一手调研任务：在制定我方定价前，通过潜在客户访谈采集 5–10 个数据点。 + +**4.5 融资基准。** 该品类软件公司的融资集中于 1000–2500 万美元（A/B 轮）；enspired 逾 4000 万欧元的扩募与 Ascend 的 PE 资本重组代表规模化梯队。Gridmatic 展示了自我造血的交易路线。对我们"模式二并展望模式三"的路径而言，enspired 的资本化节奏（A 轮约 900 万欧元 → B 轮约 2500 万欧元+扩募）是最接近的可比对象。 + +**4.6 数据口径卫生。** 第二、三节中的每一个"提升"数字都是基于不可比基准（中位资产、人工竞价、完美预见、回测）的厂商营销口径，应视为市场定位话术而非基准。我们自己的评估标准（对 S≡0 的 pinball loss、对 P50 朴素策略的经济回测、分市场状态切片）恰恰应成为销售资产——正因市场上的宣称无法审计，enspired 的"认证收入"打法证明了透明度能赢单。 + +--- + +## 五、后续行动建议 + +1. 一手定价调研：与 PJM 可再生 IPP 进行 8–10 次结构化访谈，了解现行竞价安排、费用与痛点。 +2. 深挖 Ascend×Tenaska 合作机制（收入分成？排他性？覆盖哪些资产类别？）——决定 QSE 伙伴对我们是否仍然可得。 +3. 监测触发信号：Dexter/enspired 的美国实体注册或 ISO 会员资格；Gridmatic 在 PJM 发电侧（非售电）的公告；Fluence Mosaic 新增 PJM 市场覆盖；Tyba A 轮后的市场扩张公告。 +4. 本报告未覆盖的第二梯队（建议出 v2 版本）：Tesla Autobidder、Habitat Energy（Hartree 旗下）、Stem、PCI/日立/Yes Energy 等执行平台、Amperon 等预测层供应商，以及 QSE 在位者（CES、ACES、The Energy Authority、EDF Trading NA）。 +5. 在对外使用（投资人材料、董事会文件）前，依据一手文件重新核实所有融资/员工数/规模数据。 + +--- + +## 附录：术语表 + +| 术语 | 释义 | +|---|---| +| 自动报价（Autobidder） | 自动计算并（通常）向市场提交报价的软件 | +| Bid-to-bill | 覆盖市场运营全链条的软件品类：报价、调度、结算、开票 | +| 模式二 / 模式三 | 本项目简称：竞价优化 SaaS（客户仍为市场主体）/ 资产管理服务（服务商成为注册市场主体） | +| QSE / 调度实体 | 合格调度实体——持有市场主体基础设施（资质、信用保证金、7×24 交易台），代资产开展交易 | +| Route-to-market | 欧洲模式：服务商成为新能源资产的市场主体，承担预测、竞价与平衡风险 | +| TaaS（交易即服务） | 客户将交易执行整体托付给服务商的自动化平台 | +| 包销 / 租赁（Offtake / Tolling） | 对手方买断电站出力（包销）或租用其容量并承担市场风险（租赁）——可替代购买竞价软件 | +| VPPA（虚拟购电协议） | 发电商与（通常为企业）买方之间的金融差价合约，对枢纽/节点价格结算 | +| DA / RT | 日前市场 / 实时市场；日前–实时价差是新能源竞价优化的核心驱动 | +| 联合优化（Co-optimization） | 跨日前与实时对电能量与辅助服务报价进行联合优化（电池储能的核心场景） | +| 完美预见基准 | 在完全预知价格情况下可实现的收入；"达到完美预见的 90–95%"一类宣称以此为分母 | +| 收益提升宣称 | 服务商自报的相对某基准（中位资产、人工竞价、回测）的收入改善；各家基准不可比 | +| ISO/RTO | 独立系统运营商 / 区域输电组织（PJM、ERCOT、CAISO、MISO、SPP、NYISO、ISO-NE） | +| BESS | 电池储能系统 | +| IPP | 独立发电商 | +| ARR | 年度经常性收入（SaaS 指标） |