Industries / Finance and business services / Finance and banking / Energy trading
Forecasting
Predicting demand, wind and solar output, and prices, from minutes to years ahead, so that traders and grid operators know what to buy and sell.
7
Milestones achieved, from 1995 to 2025.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1995
Neural networks forecast demand
The Electric Power Research Institute (EPRI) paid Southern Methodist University to build ANNSTLF, the Artificial Neural Network Short-Term Load Forecaster. It learns from past hourly demand and weather and forecasts each hour's demand from one hour to seven days ahead. EPRI's 1995 report put its errors at 2 to 4% and said it was in use at twenty utilities, and by the developers' 1998 count 35 utilities in the US and Canada used it.
Credit
A. Khotanzad of Southern Methodist University, with EPRI
1998
Wind forecasts enter daily dispatch
In the late 1990s western Denmark had about 1,000 MW of wind turbines on a system whose demand ran from 1,200 to 3,700 MW. The utilities Elsam and Eltra installed WPPT, the Wind Power Prediction Tool, in their control centres in October 1997 and used it in daily operation from January 1998, feeding its forecasts "directly in the economic load dispatch and the day-to-day electricity trade". The physical model from Risø National Laboratory had served forecasts online since January 1997.
Credit
Lars Landberg (Risø); H. Madsen and T. S. Nielsen (Technical University of Denmark); Elsam; Eltra
2012
Open contests set benchmarks
The IEEE Working Group on Energy Forecasting ran the Global Energy Forecasting Competition 2012 on Kaggle: 107 teams forecast hourly demand for 20 zones and their total, and 134 teams forecast wind power. The organizers then published the data and the answers, so anyone can test a method against the winners, who improved on the benchmark model by about 30%. GEFCom2014 moved to probabilistic forecasts, which give a range of outcomes with their likelihoods, in four tracks on load, price, wind and solar, and drew 581 participants from 61 countries.
Credit
Tao Hong, Pierre Pinson and Shu Fan; IEEE Power & Energy Society
2015
An eclipse tests solar forecasts
On 20 March 2015 a partial solar eclipse crossed Europe in mid-morning, and grid operators forecast both the drop in solar output and its rebound. German operators bought about twice their usual reserves and sold the expected solar output in advance on the quarter-hour market, so that in real time they had only the forecast error to cover. Across Europe's grid operators, estimated solar output fell from about 22 GW at the start of the eclipse to about 14 GW at its peak and reached 35 GW by noon. In Britain demand came in about 200 MW above forecast, and pumped-storage hydro covered it.
Credit
the transmission system operators of ENTSO-E; report by ENTSO-E and SolarPower Europe
2021
A shared test for prices
Four researchers found that new methods for forecasting electricity prices "are often compared using unique, not publicly available datasets", over short periods and against weak baselines. Their 2021 paper in Applied Energy published six years of day-ahead prices for each of five markets (Nord Pool, PJM, and EPEX Belgium, France and Germany), two strong reference models and an open-source Python library, epftoolbox, so that any new method can be scored the same way.
Credit
Jesus Lago, Grzegorz Marcjasz, Bart De Schutter and Rafał Weron
2025
AI weather model goes operational
On 25 February 2025 the European Centre for Medium-Range Weather Forecasts (ECMWF) put its machine-learning model, AIFS Single, into operation beside its physics-based model, on a 28 km grid with round-the-clock support. ECMWF calls AIFS "the first fully operational weather prediction open model using machine learning". Earlier, in a paper first posted in December 2022, Google DeepMind's GraphCast had shown that a model trained on past weather could beat the most accurate operational physics forecast on 90% of 1,380 test targets.
Credit
ECMWF; Remi Lam and colleagues at Google DeepMind (GraphCast)
2025
Pretrained models tested on prices
Amazon, Google and others have released pretrained time-series models, meant to forecast series they were never trained on. Independent tests on European day-ahead power prices found them roughly level with standard methods. A June 2025 study of 2024 prices in five countries found that none statistically beat a simple model of daily and weekly cycles, and an August 2026 study of Germany, Poland and Spain found that only TabPFN consistently beat models built for price forecasting, though it was not always the most profitable in a simulated battery test.
Lessona model that leads general benchmarks still has to be tested on the market where it will be used.
Credit
Timothée Hornek, Amir Sartipi, Igor Tchappi and Gilbert Fridgen; Arkadiusz Lipiecki and Rafał Weron
Watch for
AI solar forecast halves errors
Open Climate Fix says its Quartz Solar forecast has halved solar forecast errors and saves at least £30 million a year in balancing costs, and that Britain's National Energy System Operator (NESO) uses it in the control room to decide how much reserve power to buy. NESO's own project page gives its target as a 20% cut in mean absolute error, the average size of the miss.
Achieved whenNESO or an independent reviewer publishes measured control-room errors and cost savings that match the claim.
Claimed by
Open Climate Fix, November 2025
One model for any series
Amazon's Chronos-2 is a pretrained model that forecasts series it was not trained on and can take in related series and outside inputs such as weather. Its authors say it is the most accurate on three general benchmarks. An independent test on Belgian prices found mixed results: against the best ensemble of other machine-learning methods, its mean absolute error was 5% lower on day-ahead prices but 10% higher on imbalance prices.
Achieved whenindependent load and price benchmarks, such as epftoolbox, show it consistently beating models built for those markets.
Claimed by
the Chronos-2 authors at Amazon, October 2025
Open
01
AI weather scored in megawatts
A daily public scorecard for three to five European bidding zones that turns open weather forecasts, from ECMWF's AIFS and its physics model IFS to other open AI models, into day-ahead wind and solar output forecasts in MW. It scores each against actual output, side by side with the official forecast that each grid operator must publish by 18:00 Brussels time the day before.
The briefCredit · unclaimed02
A benchmark for gas demand
An open benchmark for forecasting daily gas demand a day ahead, on the model of epftoolbox for power prices: fixed data for several countries, fixed test years, reference models from a temperature regression to a pretrained model such as Chronos-2, and one scoring script for single-number and probabilistic forecasts.
The briefCredit · unclaimed
Open 01 · the brief
AI weather scored in megawatts
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A daily public scorecard for three to five European bidding zones that turns open weather forecasts, from ECMWF's AIFS and its physics model IFS to other open AI models, into day-ahead wind and solar output forecasts in MW. It scores each against actual output, side by side with the official forecast that each grid operator must publish by 18:00 Brussels time the day before.
Why it is still open
EU rules require grid operators to publish "a forecast of wind and solar power generation (MW) per bidding zone", so the official forecasts are public. Published work scores AI models on weather variables or for a single country: a 2024 TU Delft thesis compared GraphCast with physics models for the Netherlands, a 2026 code repository evaluates AIFS wind speed, and Open Climate Fix covers British solar. Searches of arXiv and GitHub found nothing that scores AI weather models in MW against the grid operators' forecasts.
Built with what exists
AIFS open data from ECMWF; the ENTSO-E Transparency Platform for the grid operators' forecasts and actual output; Python with gradient-boosted models to turn weather into MW; one rented GPU to run open AI weather models.
First step
Pick one zone, pull a year of the grid operator's forecasts and actual output, write the scoring script, and publish one week of AIFS-based forecasts.
Done looks like
90 days in a row of published forecasts for at least three zones, scored by mean absolute error as a share of installed capacity next to the grid operator's forecast, with any day re-runnable by an outsider from the archive to the same score.
Open 02 · the brief
A benchmark for gas demand
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An open benchmark for forecasting daily gas demand a day ahead, on the model of epftoolbox for power prices: fixed data for several countries, fixed test years, reference models from a temperature regression to a pretrained model such as Chronos-2, and one scoring script for single-number and probabilistic forecasts.
Why it is still open
The Global Energy Forecasting Competitions covered load, price, wind and solar, not gas. A 2021 doctoral thesis on forecasting German gas imports and consumption found that taking the methods further "would require standard data sets known from forecasting competitions (e.g. Kaggle, M4)". My searches found open gas demand data, such as Ireland's daily national demand since 2018, but no competition or shared test with fixed years and reference scores.
Built with what exists
Gas Networks Ireland's daily demand data, under the Irish Public Sector Information Licence; national demand series from other gas grid operators; ERA5 temperatures from the Copernicus Climate Data Store; Python, with epftoolbox as a template; pretrained time-series models such as Chronos-2 and TimesFM.
First step
Build the Irish series for 2018 to 2025 with daily temperatures, and score a heating-degree-day regression, which ties demand to how far each day's temperature falls below a set base, on 2024 and 2025.
Done looks like
at least five countries, each with two held-out test years, and scores for at least four reference models in mean absolute percentage error and pinball loss, a score for probabilistic forecasts. An outsider who reruns the code gets the same scores.