Electricity Demand Forecasting: Machine Learning Helps Short Horizons, Not the Big Uncertainties

The use case

The forecast that moved most in 2025 was not produced by a new model at all. In April 2025 the Electric Reliability Council of Texas (ERCOT) reported that transmission providers had submitted data-centre load for 2030 of 77,965 MW, against 29,614 MW in the 2024 forecast (operator-reported, before ERCOT's own adjustments)². ERCOT then discounted the submissions using factors drawn from its own records: new large loads were delayed by 180 days because those energised in 2022–2024 ran about 220 days late, and data-centre demand was cut to 49.8% of the amount requested because that was the share observed for data centres with 2022–2024 in-service dates (operator-reported)². The method is econometric forecasting plus judgement, which is older technology.

Demand forecasting is the prediction of how much electricity customers will use, by hour, day or year, so that operators can schedule generation and planners can size the grid. Machine learning (ML), where a model learns patterns from historical data, is applied mainly to the short horizons, from minutes to a week ahead.

Scope. Counted as AI here: ML and deep learning models used to forecast load. Not counted: standard load profiles, ARIMA time-series models, naive benchmarks such as repeating yesterday's load, econometric long-term forecasts and human adjustments. Solar generation forecasting is mentioned once as an adjacent use; price forecasting is out of scope.

Where it is running

The documents opened confirm ML in operator forecasting in a few places, all described by the operators themselves, and none of them is a forecast of total system demand at ERCOT.

The NESO demand document is older than three years and no newer NESO demand figure was found. Australia's market operator (AEMO) widened its data-centre demand forecast in 2025 using advice from a consultancy; the report on it does not describe ML⁸.

What the evidence shows

Portugal, national day-ahead load. A study of the Portuguese transmission system forecast national load for the next day in 15-minute steps. On the 2019 test year the best deep-learning model (N-BEATS) reached a mean absolute percentage error (MAPE) of 2.001%, against 3.333% and 3.581% for a clustering-based baseline, about 40% lower than the better baseline (measured out of sample, single country, one test year, preprint with venue and date not shown)⁵. The mechanism is that deep networks capture daily and seasonal shape more flexibly than the pattern-matching baseline. The baseline is a research method, not the model any operator runs, so the figure does not measure improvement over current practice.

Germany, simulated future grids. A peer-reviewed conference study used measured smart-meter data from 3,511 households in one town over 34 months and simulated 2037 grids with more rooftop solar and home batteries. In the current state, a CNN-LSTM model cut root mean square error (RMSE) by 68.5% against standard load profiles (RMSE 182.4 against 579.9; out-of-sample test over 365 days)⁶. The comparison that matters is the simpler one: the naive day-before benchmark scored a MAPE of 6.4% against 6.2% for the CNN-LSTM and 6.7% for the LSTM in the current state, and in the first simulated 2037 scenario the day-before benchmark (23.2%) beat the LSTM (26.2%) (out-of-sample backtest on simulated data)⁶. The authors found no superior approach between the two ML models⁶.

Great Britain, operator-reported gains. NESO reported about 12% improvement in a core forecasting KPI between April 2018 and February 2020, and about 20% improvement in short-term national demand accuracy (operator-reported, June 2020; the KPI, test set and horizon are not defined)³. These gains follow a redesign of the national model as well as new ML at supply-point level, so they cannot be credited to ML alone³. A separate solar nowcast reached a mean absolute error of 233 MW against 650 MW for the existing ESO forecast, up to two hours ahead (operator-reported, July 2023, test period not stated)⁴.

What it takes: cost, integration, governance

Training cost is small for the models tested. The German study trained its models in 47 seconds and 1 minute 13 seconds on a laptop-class processor (measured)⁶, and the Portuguese deep models needed between 1.49 and 5.85 GPU hours for the two models reported in the text (measured)⁵. No money figures for operational ML forecasting were found.

The cost sits in integration. NESO reported that large ML models 'proved to be difficult to productionise', that data processing took longer than expected, and that it built two on-premises computers with six GPUs in total (operator-reported, July 2023)⁴. Keeping thousands of models current also means continuous retraining³.

No law or guidance specific to demand forecasting was identified in the sources opened. The governance that appears in the documents concerns inputs: ERCOT noted that new rules allow transmission providers' officer-attested load to be used in planning and that a good-cause exception is likely needed for transmission planning (operator-reported, April 2025)².

Where it falls short

The challenge question is what would make the use case look less successful than it is presented. Five points emerge.

Baselines. The Portuguese gain is against a research baseline, and in the German study the simplest benchmark matched or beat ML in several cases⁵⁶.

Regime change. In the Portuguese data, errors rose in spring and summer 2020, and the lockdown from 17 March to 3 May 2020 produced larger errors, especially for the LSTM and N-BEATS models, while a third architecture (TCN) lost less accuracy (measured; the numbers appear only in a figure not read)⁵. Models trained on past patterns fail when behaviour shifts. In the German simulation all methods did worse in future grid states, because demand near zero makes percentage errors unstable⁶.

Where the big errors are. The largest forecast uncertainties in the documents are not short-horizon patterns. AEMO gave itself a 40% margin of error on data-centre demand, with hyperscalers' share of data-centre consumption rising from 14% in FY18 to 55% in FY25 (news report of AEMO and Oxford Economics, August 2025; original not opened)⁸. These are handled by scenarios and adjustment factors, not ML²⁸.

Extreme weather. In February 2021 ERCOT averaged 34,000 MW of generation outages against a record winter peak of 69,871 MW, and 23,418 MW of firm load was shed (regulator staff report, tracking document of December 2022)⁷. The regulators' recommendations on winter peak load forecasts and reserve margins were being addressed by ERCOT, MISO and SPP, and the document does not state a forecast error or any ML role⁷. The event shows what short-horizon accuracy does not cover.

Evidence gaps. No independent evaluation of an operational ML demand forecast against audited actuals was found. Searches for failures, withdrawals, lawsuits and regulator action specific to ML demand forecasting returned none. The sources are single-country tests, simulations and operator documents with reliance disclaimers⁴.

Phase: In early use — why

The hypothesis was 'In wide use'. The evidence supports a lower phase. Confirmed ML in an operator's demand forecast appears in one operator-reported document from 2020, where supply-point models were still in trial and validation³. ERCOT's confirmed ML covers battery and flexible-load forecasts, while its system forecasts are econometric¹². Academic tests are encouraging but limited⁵⁶. Wide use by other utilities or vendors could not be established from sources opened, so the lower plausible phase was chosen.

What to watch next

• ERCOT states that its large-load adjustment factors can be updated as contract and officer-letter load is energised²; the next published update will show whether the 49.8% and 55.4% factors hold.
• AEMO's data-centre forecasts are to be applied in its upcoming Electricity Statement of Opportunities and Integrated System Plan⁸.
• NESO's solar nowcasting extension targeted a further 20% reduction in mean absolute error up to four hours ahead and 10% at day ahead; published live results would test the claim⁴.
• ERCOT's list of potential ML uses, price-responsive demand and net load forecast risk, is described as preliminary; a published trial result would mark a move from proposal to use¹.

Questions to ask before adopting

• Against which baseline, including the naive previous-day forecast, was the claimed accuracy measured?
• How did the model perform in past regime changes such as the 2020 lockdowns, and how quickly is it retrained?
• Which part of the forecast error comes from weather, large new loads or behind-the-meter generation, and does ML address that part?
• What data, computing and staff are needed to run and maintain the models in the control room?
• Who audits forecast error against actual demand, and are the results published?