Forecast-Led Scheduling Against Curtailment: It Would Help Only Where the Grid Allows

The problem

Starting point. The earlier AI in Practice paper found that machine learning forecasts of wind and solar output run in Great Britain's control room and in Australian wind farms, but that most reported gains are back-tests by the operators or vendors using the tools⁹. One example is a national solar forecast whose average miss fell from 650 MW to 233 MW in a back-test (operator-reported, June 2025)¹⁰. A forecast describes the surplus; it does not by itself stop any power being switched off. What this paper adds. It asks what could be done with such forecasts: schedule storage, electrolysers (equipment that splits water into hydrogen using electricity) and other flexible demand in advance, so that surplus wind and solar has somewhere to go.

Great Britain. In 2024/25 the system operator spent £1.7bn on thermal constraints, which arise when power lines cannot carry all the power that generators offer. The cost rose 64% and the volume of actions rose 81% to 13.5 TWh (operator-reported, April 2024 to March 2025)¹. Wind curtailment volumes rose to 13% of hypothetical wind outturn¹. The operator attributes this to the large share of wind capacity connected in Scotland, a constrained part of the network, and says the trend continues through the 2020s as new connections outpace network build¹.

California. The grid operator CAISO curtailed 3.4 million MWh of utility-scale wind and solar in 2024, 29% more than in 2023, and solar made up 93% of it (grid-operator data reported by the U.S. Energy Information Administration)². Most was in spring, when solar output was high and demand low².

United States Midwest. In 2023 the Southwest Power Pool and the Midcontinent operator curtailed an hourly average of 800 MW of wind, against under 200 MW in 2019. The Energy Information Administration states that curtailment there is more often caused by congestion³.

Germany. Congestion management measures totalled 30,320 GWh in 2024, 12% below 2023⁴. That total covers redispatch and countertrading, not renewable curtailment alone⁴. The regulator's report gives the cost as about €2.9bn in one section and €2.7bn in its foreword, and the two are not reconciled⁴.

Two different problems sit under the single word curtailment. The explanation given for California centres on spring oversupply, with more solar than demand can use at midday. In Great Britain and the Midwest the explanation centres on power lines that cannot carry output from where it is made to where it is used¹²³.

What already works

The building blocks exist separately, in different places and for different purposes.

Demand flexibility so far. Great Britain's Demand Flexibility Service paid households and businesses to cut demand in 22 events in winter 2022/23. About 1.6 million took part, providing roughly 350 MW of flexibility during events, at a total cost of about £10.5m⁷. The article gives the demand reduction as 2.92 GWh in one place and 3,300 MWh in another and does not reconcile them⁷. About one in two participants reported demand destruction, meaning demand that was cancelled rather than moved to another time, in at least one event⁷. These figures are from 2022/23 and were about cutting demand at peak times, the opposite direction from absorbing a surplus. The service has since changed. It became year-round on 27 November 2024 and added bi-directional flexibility, the ability to ask for more demand as well as less, and zonal procurement on 9 April 2026 (operator-reported). The operator's page says the capability to procure constraint management actions is due from 7 October 2026; whether that launch has taken place could not be confirmed⁸.

Modelled benefit of turn-up demand. The operator's Demand for Constraints assessment, prepared with the consultancy Arup and published in March 2025, modelled flexible demand switching on during constraint periods. Across the cases modelled, constraint cost savings ranged from £0.5bn to £2.5bn over 2025 to 2035 in 2023 prices. The largest case saved 8% (£2.4bn) over 2028 to 2035 against a counterfactual⁵. Constraints affected 17% of periods in a year on average, and Scottish boundaries were constrained for 29% of the year⁵. These are model outputs (projected), not measured results. In plain terms, even in the most favourable case modelled, flexible demand removes a small share of constraint cost.

Electrolysers. A peer-reviewed review published in April 2026 cites a prediction that UK wind curtailment would reach 7.72 TWh a year by 2030 (projected). It notes that most Great Britain wind capacity, and so most curtailment, is in Scotland, far from the main demand centres⁶. The same review reports that electrolysers currently take 24 to 36 months to produce and install and that there are difficulties with where electrolysis can be placed⁶.

What does not yet exist. No programme was found that uses a forecast of renewable output, or of a grid constraint, to schedule a flexible load in advance with the stated aim of reducing curtailment. Each ingredient runs, but their combination was not found in the sources read.

The possibility

One way this could work: a day ahead, a grid operator or an aggregator (a company that pools many small customers) publishes a forecast of where and when surplus renewable output is likely. Flexible loads such as electrolysers, large batteries, heat stores, vehicle fleets and batch computing would bid to turn up in those hours. The operator would accept bids from loads that sit where they can relieve the problem, at a price below what it would otherwise pay to have generators switch off. With reliable forecasts, loads could prepare in advance, for example by leaving battery space or hydrogen storage empty.

The two causes of curtailment would produce two versions. Where the cause is oversupply, as in California in spring, almost any flexible load on the system could help, and batteries already do part of this². Where the cause is a bottleneck, as in Scotland, only load placed behind the bottleneck could help.

The role of AI. AI would supply one input and possibly one decision step. The input is a machine learning forecast of output and of the hours in which a boundary would be constrained. The decision step is an optimiser that picks which loads to call. Ordinary systems would carry everything else: metering, contracts, settlement, the electrolyser or battery itself, hydrogen buyers and, above all, the wires. If the surplus is caused by a bottleneck, a flexible load placed on the wrong side of it removes none of the waste. On this reasoning AI would be a small part of the possibility, and the decisive parts are location, contract design and network build.

What would have to be true

1. Flexible load behind the bottleneck. Most Great Britain wind, and so most curtailment, is in Scotland¹⁶. A turn-up service would need electrolysers, batteries or other load in or near the constrained zones, not in southern England.
2. Forecasts accurate at the lead time the load needs. Electrolysers and heat stores need notice. The earlier paper found that gains concentrate in the first hours: the operator's extension targets a 20% cut in error up to four hours ahead but only 10% day-ahead¹⁰. An independent analysis found that combined day-ahead wind and solar error in 16 European countries roughly doubled between 2017 and 2021 as capacity grew (no control group; older than three years; no update found)¹¹. A scheduler that relies on day-ahead notice would depend on the weaker part of the evidence.
3. A price that pays the load to turn up. The Demand Flexibility Service paid an average of about £3,330 per MWh across 22 sessions in 2022/23⁷, a scarcity price for cutting demand. The operator's assessment envisaged long-term contracts in which it can ask a customer to increase consumption when the network is constrained⁵. The price would have to sit below the cost of curtailment, which the sources reviewed do not give per MWh.
4. Equipment in place in time. Electrolysers take 24 to 36 months to produce and install⁶, so the load would have to be ordered years before the surplus it is meant to absorb.
5. A use for the energy beyond one day. In California, in spring, more solar is produced than can be used within a day, and without more transmission or long-term storage high curtailment can still occur².
6. Customers who shift rather than cancel demand. About one in two Demand Flexibility Service participants reported demand destruction in at least one event⁷.

What could stop it

What the evidence shows

• Network, not forecast. The Great Britain operator calls wind curtailment a major driver of balancing costs because much wind is connected in a constrained Scotland¹. The Energy Information Administration finds congestion the more common cause in the Midwest³ and says high curtailment can persist in California without more transmission or long-term storage².
• Scale of the lever. The modelled saving from flexible demand in the largest case is 8% of constraint cost over 2028 to 2035⁵. The operator projects balancing costs peaking at about £8bn in 2030, with up to £4bn avoidable if critical network projects are brought forward¹.
• Participation. The Demand Flexibility Service required a smart meter or half-hourly metering and under-represented younger people, lower-income households, renters and city residents; the 2023 article described it as still in a trial stage⁷.
• Lead times and siting. Electrolysers take 24 to 36 months to install and have siting difficulties⁶.
• Thin evidence on the forecasts themselves. The earlier paper found few independent checks of machine learning forecast gains⁹.
• No failed attempts found. Searches for failed or cancelled attempts to schedule flexible load against curtailment found none specific to AI, and found no attempt of this kind to evaluate.

Reasoning

The challenge question is whether the forecast is the missing piece at all. If curtailment in Scotland is mostly a question of wire capacity, the cheaper lever might be transmission, and a rule-based schedule such as running a load whenever wind is high might do most of what a forecast-driven one would. A forecast would add value only where a load needs notice, which is plausible for electrolysers and heat stores but less so for batteries that can respond in seconds.

Who bears the risk also matters. If a forecast misses, the customer who turned up bears the cost of power that was not surplus, and the operator bears the cost of curtailment that was not avoided. If the customer is paid to turn up, consumers would be paying for flexibility on top of the constraint payments they already fund. Incentives could work against the idea if the party that sets the forecast also profits from the schedule.

Finally, the simpler solutions already exist. Trading avoided about 8% of California's curtailment in 2024², and batteries are growing quickly¹². The possibility would be most useful where these do not reach, which the sources suggest is the long, regional or seasonal surplus.

Signs to watch

• Whether the Demand Flexibility Service procures constraint management actions as the operator's page says is due from 7 October 2026, and whether it publishes volumes, prices and locations⁸.
• Whether the operator publishes outcomes from its Demand for Constraints work against the modelled £0.5bn to £2.5bn savings⁵.
• The next annual balancing costs report, showing wind curtailment against the 13% of hypothetical wind outturn in 2024/25¹.
• California's 2025 curtailment figures against the 3.4 million MWh of 2024².
• The next German monitoring report, showing congestion management volume against 30,320 GWh in 2024⁴.
• Results from the operator's weather-ensemble forecasting work against its targets of a 20% error cut up to four hours ahead and 10% day-ahead¹⁰.