Wind and solar plants cannot fully predict their output a day ahead, so their marketers must sell a forecast on the spot market and then correct the position in the continuous intraday market until shortly before delivery. Probabilistic forecasts, imbalance-price models and optimisation- or RL-based bidding agents automate this loop, lowering balancing costs and raising the market value achieved for each plant, while human traders set risk limits and supervise.
Problem & value
- Today: under the German EEG, plants of a certain size must be sold via Direktvermarktung (direct marketing). The marketer sells the production, receives the market premium via the grid operator and carries the balancing-group responsibility: every deviation between sold and delivered energy is settled at the TSO imbalance price. Traders previously corrected positions manually; today much of the intraday work is automated.
- AI value: better forecast updates (nowcasting, ensemble weather), faster re-trading when a forecast changes, smarter handling of negative-price hours and curtailment, lower imbalance cost, and one engine that scales to thousands of small plants.
- Evidence: established marketers (Next Kraftwerke, Statkraft, Entrix, enspired) publicly describe algorithmic intraday trading as core to their business [1][2][3]. Published quantified uplift is portfolio-specific; measure it against your own baseline instead of relying on vendor claims.
Variants
- Direct marketer for wind/PV portfolios (EEG market premium, PPAs, merchant plants).
- Virtual power plant (VPP) / aggregator pooling small plants, batteries, CHP and flexible loads, also selling balancing reserve (see OPT-011, OPT-022).
- Utility or large operator trading its own fleet; PPA-backed industrial buyers shaping and balancing delivered profiles.
- Cross-border: other EU markets (Nordics, Netherlands, France, Poland) with own products and gate-closure times; SDAC/SIDC market coupling.
- Not in scope: securities and financial trading (OPT-010); private household optimisation (OPT-032).
How it works
- weather NWP + plant telemetry + market data
- probabilistic forecast per plant/portfolio (PRED-007)
- price & imbalance-price forecast (PRED-008)
- day-ahead bidding (stochastic optimisation / scenario-based)
- continuous-intraday agent (rule-based, MPC or RL) re-balancing the position
- risk limits & kill-switch (position, volume, price, order-rate caps)
- order gateway to the exchange (EPEX SPOT via member API)
- dispatch to plants/assets, settlement, REMIT & audit reporting
- Forecasting: gradient-boosted trees and neural nets on NWP ensembles plus live SCADA; quantile or distributional outputs matter more than point forecasts because the bid depends on the cost of error in each direction.
- Bidding: day-ahead as scenario-based stochastic optimisation (MILP/LP solver); intraday as model-predictive control against a limit-order-book simulator, increasingly trained with reinforcement learning in simulation using historical order-book replays.
- Human-in-the-loop: traders define strategy mandates, limits and the kill-switch; the desk monitors positions around storms, grid incidents and market disruptions where models extrapolate poorly. Pre-trade controls are not optional (see regulation).
Data preparation
The data is largely structured and machine-readable but fragmented: SCADA and meter data from plants, numerical weather predictions, EPEX SPOT and intraday order-book data, TSO imbalance and system-state data, plant master data and contract terms (EEG market premium, PPAs) which often sit as PDFs, Excel sheets or in the ETRM and DMS.
Typical gaps: plant metadata without consistent coordinates, hub height or capacity changes; missing or curtailed output not flagged; telemetry gaps and clock or time-zone errors; weather forecast archives that were not stored as issued; contract terms only in scanned or free-text documents; and duplicated plant IDs across systems.
Preparation steps: (1) inventory sources and owners (PREP-001, PREP-005); (2) harmonise plant master data and IDs, and mark curtailment and outages in the time series; (3) archive weather forecasts and market data as issued, so backtests avoid look-ahead bias; (4) extract contract terms from PDFs into structured fields (DOC-014, DOC-018, PREP-006); (5) define a KPI glossary for imbalance cost and achieved market value (PREP-012); (6) build a backtest set with traders (PREP-018). Tools: Postgres/TimescaleDB, Python, dbt. Data work is typically 25 to 35 percent of effort, since feeds are already digital.
Implementation paths
Which path for whom
- Enterprise (direct marketer, utility, aggregator): B, because bidding logic and forecasts are the core of the margin and must be tuned to your portfolio; buy a marketing service (A) only if you own plants but do not want a trading desk.
- Developer / integrator: C, because open-source forecasting, solver and RL libraries plus public market data allow building and backtesting strategies without licence lock-in.
Path A — Buy (off-the-shelf)
- Marketing as a service: hand the portfolio to a direct marketer (e.g. Next Kraftwerke, EU; Statkraft, EU; Entrix, EU; enspired, EU; Enel X, Axpo, Uniper) and pay a fee per MWh. Right for owners of plants who do not want a trading desk.
- Trading and portfolio-management software (ETRM / bidding platforms): e.g. Energy Exemplar PLEXOS, Nominal, Optimeering and Montel analytics (data/forecasts), Entrix and enspired optimisation modules, Fluence Mosaic (storage). Check REMIT support, EPEX connectivity and EU hosting.
- Choose A when: < several hundred MW and no appetite to staff a 24/7 desk.
Path B — Build on cloud APIs
- Market data and access: EPEX SPOT member trading interfaces (public market-data via EPEX SPOT and ENTSO-E Transparency Platform API for load, generation and cross-border flows), TSO imbalance data (e.g. regelleistung.net, netztransparenz.de).
- Weather and forecast data: ECMWF open data / IFS ensemble, DWD ICON open data, Open-Meteo, and weather-AI models such as ECMWF AIFS and Google's WeatherNext.
- Compute: managed time-series/ML services in EU regions (e.g. Azure ML, AWS SageMaker in Frankfurt, Vertex AI europe-west3); stream processing on Kafka. LLMs are not the trading brain here; they help with market-notice summarisation, analyst copilots and report drafting.
- Choose B when: you have a quant/data-science team but not the exchange-grade low-latency stack.
Path C — Self-host (open source)
- Forecasting: LightGBM/XGBoost, PyTorch Forecasting, Darts, plus open time-series foundation models (Chronos, TimesFM, Moirai) for baselines; DWD ICON and ECMWF open data.
- Optimisation: HiGHS, OR-Tools, Pyomo/linopy; open energy-system tools such as PyPSA and open-source RL (Stable-Baselines3, Ray RLlib) in a gym-style order-book simulator.
- Infra: Kafka/Redpanda, TimescaleDB or ClickHouse, Kubernetes. A CPU or modest GPU server (HW1) suffices; the demanding part is latency, resilience and 24/7 operations.
Path D — Custom training
Training own forecast and RL models on years of plant telemetry, order-book data and imbalance prices is the usual differentiator for larger desks. Needs multi-year, high-resolution histories, a faithful market simulator and strict backtesting against look-ahead bias.
State of the art (2026)
- Probabilistic and ensemble-based forecasting with ML post-processing of NWP is standard; AI weather models (AIFS, GraphCast-style) are entering operational ensembles and improve short-range wind/PV forecasts [4].
- Intraday market is now 15-minute-product dominated with highly automated participants, so speed, order-book awareness and execution quality matter; EPEX SPOT introduced 15-minute products in day-ahead coupling after SDAC moved to 15-minute MTU [5].
- RL and MPC hybrids: academic work shows RL agents competitive in simulated intraday markets, but production desks mostly use RL for execution or as a sub-module under hard risk limits rather than fully end-to-end.
- Negative prices and curtailment: growing PV share makes negative-price hours and market-premium rules (no premium during negative prices for newer plants) a core optimisation term.
- Limits: regime shifts (crises, rule changes) break backtests; simulators ignore market impact of own orders; forecast error is correlated across a portfolio; overfitting to historical order books is common.
Shared resources & combinations
- Forecast inputs: PRED-007 (load & renewable forecasting) and PRED-008 (price forecasting).
- Flexibility assets: OPT-011 (battery dispatch), OPT-021 (EV fleet charging), OPT-023 (district heating/CHP), OPT-032 (prosumers) can be pooled into the same VPP.
- Grid side: OPT-022 (congestion management) determines curtailment and redispatch effects.
- Monitoring: ANOM-001-type anomaly detection on telemetry and trades; LLM agents (AGENT-001) for market-notice triage and reporting.
- Same time-series/solver stack serves OPT-004 (pricing) and OPT-010 (financial trading).
Effort, cost & KPIs
- Path A: weeks to onboard plants; fee per MWh or revenue share.
- Path B: 6-12 months for a production-grade desk with a small quant team.
- Path C/D: 12+ months, plus 24/7 on-call operations and market-access fees.
- Cost drivers: data licences, exchange and clearing costs, development of simulators, 24/7 coverage, regulatory reporting.
- KPIs: forecast error (nRMSE/pinball loss), imbalance cost per MWh, achieved market value vs. reference market value (Marktwert), intraday trading margin, share of automated orders, risk-limit breaches, system uptime, REMIT reporting timeliness.
People & change
Start with management and the head of trading to set risk limits, mandate and budget. Traders and dispatchers are the key end users; involve them from backtesting on, since they must trust and be able to override bids. IT and IT security cover market-interface access, credentials and NIS2; legal and compliance cover REMIT, MiFID-style and market-abuse rules and contract points with data vendors; data owners release SCADA and contract data.
Works council: depends. If bids and orders are logged per trader and used to assess individual performance, §87(1) no. 6 BetrVG applies; with a pure portfolio-level system it may not, but §90(1) no. 3 still requires timely information when work processes are planned with AI, and §80(3) allows an expert. Public utilities involve the staff council. Personal data is minimal; the DPO confirms this and checks vendor AVV (GUIDE-005).
AI literacy: support traders and risk managers with training on forecast uncertainty, backtest limits and failure modes (Art. 4 AI Act, GUIDE-006). Communicate that the algorithm works inside hard position and loss limits set by humans, with a kill switch and daily review. Use one technology or one market (for example intraday for PV) as pilot group and run in shadow mode before live bidding. Typical concerns are fear of replacement and runaway algorithms; address them with transparent limits and audit logs. See GUIDE-009, GUIDE-011 and GUIDE-014.
Risks, regulation & compliance
- EU AI Act: typically minimal or not high-risk, since trading software is not an Annex III use case; but AI components that act as safety components in the management and operation of critical infrastructure (Annex III, 2: electricity supply) may be high-risk if they control grid operation. A marketer's bidding agent is normally not that, whereas TSO/DSO control software can be. Classify per deployment (not legal advice).
- REMIT (Regulation (EU) 1227/2011, strengthened by (EU) 2024/1106 "REMIT II"): bans insider trading and market manipulation in wholesale energy markets; algorithmic trading is explicitly addressed. Operators of algorithms must prevent manipulative behaviour (spoofing, layering, quote stuffing), keep audit trails and report orders and trades to ACER [6][7].
- MiFID II / EMIR: apply where derivatives/futures are traded (see OPT-010); exchange rules (EPEX SPOT) require pre-trade risk controls, order-rate limits and kill-switch functions.
- EEG / EnWG: direct-marketing obligations, market-premium rules, balancing-group contracts with TSOs, redispatch 2.0 data exchange.
- GDPR: mostly not applicable (no personal data), except for small private plant owners' contract and meter data.
- Cyber: NIS2 and sectoral rules for energy operators; protect order gateways, API keys and plant control channels.
- Mitigations: hard position and volume caps, independent risk control, full order logging, model validation, simulated stress tests, human override.
Getting started
- Baseline your portfolio: imbalance cost per MWh, achieved vs. reference market value, forecast error.
- Check build vs. buy: compare 2-3 direct-marketing offers or bidding platforms on your own data.
- Set up a forecast benchmark (ICON/ECMWF ensemble + SCADA) and measure pinball loss before trading logic.
- Build an order-book replay simulator; backtest a simple rule-based intraday strategy first.
- Go live in paper-trading, then with small capped volumes, with REMIT reporting and kill-switch in place.
Sources
- Next Kraftwerke — algorithmic trading and Direktvermarktung — https://www.next-kraftwerke.com/
- Statkraft — Energy trading and market access — https://www.statkraft.com/about-statkraft/
- EPEX SPOT — Intraday continuous market — https://www.epexspot.com/en/market-modules/intraday
- ECMWF — AIFS machine-learning forecasting system — https://www.ecmwf.int/en/about/media-centre/aifs-blog
- EPEX SPOT — Markets and products overview — https://www.epexspot.com/en
- ACER — REMIT portal — https://www.acer.europa.eu/remit-documents
- Regulation (EU) 2024/1106 (REMIT II) — https://eur-lex.europa.eu/eli/reg/2024/1106/oj
- ENTSO-E Transparency Platform — https://transparency.entsoe.eu/
- Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj