Renewable Output Forecasting: Large Reported Gains, Few Independent Checks

The use case

In a back-test reported by Great Britain's electricity system operator, the average miss of its national solar forecast fell from 650 MW to 233 MW for forecasts up to two hours ahead¹. That is the use case in one number: predicting how much power wind and solar farms will deliver, from minutes to days ahead. Operators use the forecasts to decide how much spare capacity to hold, and inaccurate forecasts raise balancing needs and market costs².

What counts as AI here. Machine learning models trained on past generation and weather data, such as random forests and neural networks. What does not. Physics-based weather simulation, which feeds almost every system described below, and older statistical conversions. Historically, Great Britain's operator forecast solar output from just two inputs, installed solar capacity and solar irradiance (the sunlight reaching the ground)³. European operators do not publish their forecasting methods, so their results cannot be credited to AI or to older techniques². Demand forecasting, price forecasting and automated bidding are out of scope and appear only where they bear on the evidence.

Where it is running

Six deployments or datasets carry most of the published evidence. Only the first has results from 2025, and none was found for other regions.

What the evidence shows

Solar nowcasting in Great Britain (operator-reported). A nowcast is a forecast for the next few hours. The first test, on historical data, compared 650 MW mean absolute error (the average size of the miss) for the existing national forecast with 233 MW for the best project forecast, 2.8 times better¹. Later figures reported forecasts about 30% more accurate than the project's earlier model over four to eight hours, about 40% more accurate than one existing NESO forecast and over 40% more accurate than a second over zero to eight hours, and a NESO target of a 20% cut in mean absolute error as met¹. The report does not say how the later figures were measured and describes no live control group¹.

Claimed savings (vendor-reported, November 2025). Open Climate Fix says its tool halved forecast errors after adoption, worth at least £30 million a year in imbalance cost and potentially £150 million a year by 2035 if solar targets are met (projected)⁴. The baseline is not stated, and "halved" is not reconciled with NESO's "about 40%" in either source¹⁴.

Earlier NESO result (operator-reported, July 2019). NESO reported a 33% improvement in solar forecast accuracy from an ensemble of machine learning models, without stating the metric, baseline or test period³. This is older than three years; the 2025 figures above are the newer evidence.

Wind at Google (operator-reported, February 2019). The company reported that machine learning raised the value of its wind energy by roughly 20% on 700 MW, compared with a baseline of no time-based commitments to the grid⁷. That baseline is not a conventional forecast, so the figure blends the value of forecasting with the value of committing to deliveries in advance. The figure is older than three years; no independent follow-up evaluation and no newer figure were found.

Correcting a commercial forecast (measured on historical data, one island). On a small Mediterranean island, machine learning cut the normalised mean absolute error (the average miss as a share of capacity) of a commercial solar forecast from 8.6% to 3.34% with the best model, but of a wind forecast only from 10.25% to 9.84%. The models removed bias but not the swings in wind output⁸.

Independent view of the whole system (measured without a control group, 2017 to 2021). Day-ahead forecasts from 16 European operators were often three to four times more accurate than a naive "same as yesterday" baseline for wind, and about twice as accurate on average for solar². Those forecasts are not identified as AI-based. The data are older than three years and no newer comparison was found.

Why accuracy turns into money (peer-reviewed, German data, 2019). Higher wind and solar forecast errors raised imbalance volumes and could pass into higher spot prices, with wind errors affecting prices more than solar errors⁹.

Why solar and wind differ. The older British solar forecast had two inputs³, and traditional weather forecasts update roughly every six hours, against 15 minutes for the machine learning model that reads satellite images of cloud⁴. The solar gain mixes new data, faster updates and a new learning method, and the reports do not separate the three. Wind day-ahead accuracy depends on upstream weather forecasts, and storms produce the largest errors²; a correction can remove bias but not weather the forecast missed⁸.

Australia. In the government-supported trial, self-forecasts accredited by the market operator were on average more accurate than its own system, but the summary gives no figures⁵. A commentator who monitors generators reported improved wind-farm dispatch forecasts most of the time¹⁰.

What it takes: cost, integration, governance

• Money. The NESO project was extended three times, adding £355,000, £80,000 and £103,500; the report gives no total budget¹.
• Data and computing. Ensemble weather data (many runs of a weather model from slightly different starting points) from the European weather centre exceeded 100 GB a day, and downloading a month's archive took 60 to 120 hours. Six of twelve weather variables carried most of the information, halving download time and storage. Fifty forecast versions for one grid supply point ran in about five seconds, and a national run was estimated at about 25 minutes¹.
• Deployment. The project found large models hard to put into production and data processing slower than planned; high turnover in NESO's forecasting team affected communication; service availability exceeded 99.5%¹.
• Upkeep. Trial participants reported that models benefit from 12 months of training data to capture seasons, that site maintenance and communications outages degrade forecasts, and that self-forecasting is not "set and forget"⁵. One operator built a switch to withhold its own forecast and fall back on the market operator's; it was used under 0.1% of the time at one wind farm and never at the other⁶.
• Market rules. Australian generators face "causer pays" charges that penalise output departing from dispatch targets in the wrong direction for grid frequency. A commentator put these at 2–10% of energy revenue for some solar farms and 1–3% for many wind farms in 2020, and suspected that some self-forecasts add deliberate offsets to lower them¹⁰. The regulator, the AER, updated its compliance bulletin in February 2025 with expectations on how self-forecasts are used¹¹.
• EU rules. Whether the EU AI Act's high-risk category for electricity infrastructure covers forecasting systems could not be confirmed, because the official legal text could not be read. No view is given.

Where it falls short

The headline gains are measured against older or weaker baselines, mostly on historical data, and mostly by the parties running or selling the tools.

Baselines. The British gain is measured against a two-input solar forecast and the Google gain against no commitments at all³⁷. Neither source reports a comparison with a best-practice older method.

Test design. Most figures are back-tests, replaying past data through a model, reported by NESO, a vendor or the generator that uses the tool¹⁴⁷. The independent evidence is thinner: a European analysis that does not isolate AI² and a single-island study⁸.

Horizon. NESO's own extension targets a 20% cut in error up to four hours ahead but only 10% for day-ahead forecasts¹, and the independent day-ahead wind gain was small⁸. The gains appear to concentrate in the first hours.

Error growth. Between 2017 and 2021 the combined day-ahead error of wind and solar forecasts across 16 European countries roughly doubled as capacity grew, and the 95th percentile of wind error rose from about 6 GW to almost 10 GW². Wind forecast error rose by about 100 MWh per hour for each extra GWh per hour generated². This is not evidence that AI failed, since methods are unpublished, but it shows that more accurate forecasts do not shrink the absolute problem as capacity grows. The data are older than three years and no update was found.

Incentives. A commentator reported that some self-forecasts appear to add deliberate offsets to reduce market charges, and that under certain constraints these under-forecasts can reduce a farm's output; some operators reportedly revert to the market operator's forecast in those intervals, which removes the benefit self-forecasting was meant to deliver¹⁰. These are a commentator's observations, not regulator findings.

People and process. Reported gains depend on data pipelines, retraining and staff continuity, not on the model alone¹⁵.

Bad news searched for. Searches for withdrawals, lawsuits, regulator bans and critical evaluations of machine learning forecasts of renewable output found none specific to AI. The nearest finding is regulatory attention to how self-forecasts are used¹¹. No independent evaluation of the £30 million claim or of Google's 20% figure was found, and no source reviewed follows an AI-specific gain over several years of live operation.

Phase: In wide use — why

The hypothesis is kept. Machine learning forecasts run in the control room of Great Britain's operator, and Australian wind farms have submitted machine learning forecasts into dispatch since September 2019¹⁴⁶. The label is stronger than the evidence behind it. The share of forecasts built on machine learning rather than older methods could not be measured, because European operators do not publish methods² and the Australian trial summary names ten providers, not a count of farms⁵. The evidence supports routine use of forecasting automation and documented use of machine learning at a handful of named sites, with little proof of benefit from anyone outside the parties involved.

What to watch next

• Whether NESO publishes results from its weather-ensemble work, against its stated targets of a 20% cut in error up to four hours ahead and 10% for day-ahead¹.
• Whether NESO or its regulator publishes an independent measurement of the £30 million a year saving claimed by Open Climate Fix⁴.
• Whether the AER publishes enforcement action on how self-forecasts are used, following its February 2025 bulletin¹¹.
• Whether European operators' forecast errors are re-analysed beyond 2021, showing if absolute errors kept growing².

Questions to ask before adopting

• Against which baseline was the accuracy gain measured, and was it a live trial or a back-test?
• Which part of the gain comes from machine learning and which from new data such as satellite images?
• Does the gain hold at the horizons and in the weather, such as storms and fast cloud movement, that cost most?
• How will forecasts be kept honest if market charges reward tuning them for something other than accuracy?
• Who retrains and monitors the model, and what happens when site outages or staff turnover interrupt it?