AI use case

AI forecasting of wind and solar generation for grid balancing and trading

AI that predicts how much power a wind or solar asset will generate over the next hours to a day ahead, combining weather forecasts with historical turbine or panel output, so a grid operator or generator can commit to a delivery schedule instead of treating renewable output as unplannable.

By Len Debets · Last verified 29 September 2026 · 2 public deployments

About 20%
Reported revenue uplift
Google, organization claim.
USD 1.5 million to USD 16 million
Indicative value per year
A 500 megawatt wind portfolio selling into a day ahead power market. Worked example, see how it is calculated.

What problem does it solve?

Wind and solar generation is hard to predict, and National Grid Electricity System Operator (the operator of the Great Britain grid) says this is because it is "weather dependent and connected at a local rather than national level". Google's own account of its wind fleet describes a related problem from the generator's side: wind's variable nature makes it an unpredictable source of electricity, less useful than one that can reliably deliver power at a set time, so a generator that cannot commit to a schedule in advance captures less value from the same megawatt hours.

How does it work?

  1. Train on weather and output history. A model learns the relationship between weather variables, such as wind speed and solar irradiance, and the asset's actual historical output.
  2. Forecast output ahead of delivery. The model runs against an updated weather forecast to predict output over the commit horizon; Google's wind model, for example, forecasts output 36 hours ahead of actual generation.
  3. Turn the forecast into a commitment or a balancing input. For a generator, the forecast feeds a recommendation for the delivery commitments to make to the grid; for a system operator, it feeds the wider forecast used to balance supply and demand in real time.
  4. Submit and monitor. A trader, scheduler or the system operator's forecasting team reviews and submits the resulting plan, then compares delivered output against both the forecast and the commitment through the day.
  5. Retrain on the error. Forecast error by weather regime and season feeds back into the model so it keeps improving as more storms, calm periods and seasons are observed.
Audience
Back office
Autonomy
Assist
Adoption
Early adopters
Channels
Internal tools, API and system to system

What is it worth?

Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.

Value benchmarks for AI forecasting of wind and solar generation for grid balancing and trading
KPIMedianReported rangeData pointsClaimed by
Revenue upliftToo few to pool
about 20%
11 organization

Value drivers: Revenue growth, Risk and loss reduction.

Indicative value

A 500 megawatt wind portfolio selling into a day ahead power market

USD 1.5 million to USD 16 million

Additional value captured from forecasting driven delivery commitments per year

How this is calculated

Formula: annualRevenue * valueUplift. The low scenario uses every low input, the high scenario every high input.

InputLowHighBasis
Annual wholesale revenue from the portfolio without time based commitments annualRevenue, USD per year30,000,00080,000,000Editorial assumption based on typical wind capture prices and load factors; replace with your own portfolio revenue.
Share of revenue gained from forecasting and time based commitments valueUplift, fraction of revenue0.050.2The high end matches Google's own reported figure of roughly 20% more value from its wind fleet; the low end is an editorial assumption for a portfolio with a less mature forecasting and trading setup.

What it leaves out: Assumes the portfolio can act on the forecast with time based commitments, which needs a market that rewards scheduled delivery; it leaves out the cost of the weather data, the forecasting model and the trading desk time, and neither deployment on this page reports the cost side of that trade.

Who already uses it?

2 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.

Google

United States · Technology and software · 2019

ProductionGrade B

Google and DeepMind applied a neural network, trained on historical turbine data and widely available weather forecasts, to 700 megawatts of wind power capacity across a group of wind farms in the central United States. The model predicts wind power output 36 hours ahead of actual generation, and recommends the optimal hourly delivery commitments to make to the power grid a full day in advance, so the wind fleet can be scheduled like a conventional generator instead of treated as unplannable.

  • Revenue uplift: about 20%, to date, as reported February 2019
    "To date, machine learning has boosted the value of our wind energy by roughly 20 percent, compared to the baseline scenario of no time-based commitments to the grid."
    Claimed by: organization

National Grid Electricity System Operator

United Kingdom · Energy and utilities · 2019

PilotGrade B

National Grid Electricity System Operator, the system operator for the Great Britain electricity grid, worked with The Alan Turing Institute to replace a simple two variable solar forecast with a random forest model trained on around 80 weather and irradiance variables, combined with other machine learning methods into a multi model ensemble. The project, funded through Ofgem's Network Innovation Allowance, produced a solar forecasting system the operator's own page describes as 33% more accurate; the page does not state which forecast horizon this covers.

No outcome disclosed.

How do you implement it?

A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.

Data you need

  • Historical output per asset at a fine enough time resolution to match the commit horizon
  • Weather forecasts covering wind speed, solar irradiance and related variables at the asset's location
  • Asset metadata, such as turbine or panel specifications and site layout
  • Market rules for how a commitment or bid is scored and penalised

Systems to integrate

  • Weather data provider feed
  • Energy trading and risk management system, or the equivalent balancing system for a system operator
  • Asset monitoring or SCADA system for actual output
  • Market bidding or nomination platform

Complexity: Medium

The forecasting model itself is well understood machine learning; the harder part is wiring it into the market and grid systems that turn a forecast into an actual commitment or balancing action, and getting enough clean historical output data per asset to train on.

  1. 1

    Start on one asset class and one horizon

    Google applied its model to 700 megawatts of wind capacity in the central United States; pick one asset class and one forecast horizon rather than every asset and every horizon at once.

  2. 2

    Build the forecast, not the trading decision, first

    Get the output forecast accurate and measured against actuals before automating what the forecast changes about a commitment or a bid.

  3. 3

    Connect the forecast to a real commitment or balancing action

    A forecast that never changes what gets submitted or scheduled produces no value regardless of its accuracy; wire its output into the actual commitment, bid or balancing decision, not just a dashboard that sits next to it.

  4. 4

    Compare against a clear baseline

    Measure value against a stated baseline, such as flat or average delivery with no time based commitments, the same baseline Google used to report its own result.

  5. 5

    Keep a person on the commitment for as long as the stakes justify it

    Early on, a trader or scheduler should review the recommended commitment before it is submitted, moving to closer to automatic submission only once forecast error is well understood by season and weather regime.

Guardrails

  • A person or a risk system checks a recommended commitment against exposure and penalty limits before it is submitted
  • Forecast confidence, not just a single point prediction, is shown so a scheduler can see how much to trust a specific window
  • Model changes are back tested against historical weather and output data before going live

KPIs to instrument

  • Forecast error against actual output, by asset, season and horizon
  • Value captured from time based commitments against a flat delivery baseline
  • Imbalance penalties or curtailment avoided per period

Human in the loop

Traders and schedulers review recommended delivery commitments before they are submitted, and a system operator's forecasting team checks the model's output against its own judgment during unusual weather before it feeds the live balancing process.

Common failure modes

A weather regime the model has not seen
A model trained mostly on ordinary conditions can misjudge an unusual storm, heat wave or calm spell; monitor forecast error by weather regime and retrain as new extremes occur.
Treating a point forecast as certain
A single predicted number hides real uncertainty; publish a confidence range and size commitments to that range rather than to the midpoint alone.

What are the risks and rules?

EU AI Act

Minimal risk

A model that forecasts generation output to inform a trading commitment or a system operator's balancing input is not intended as a safety component in the management and operation of electricity supply under Annex III point 2; it informs a commercial or planning decision that a trader, scheduler or system operator makes, rather than directly protecting the physical integrity of the grid. It carries no specific obligation beyond the Article 4 AI literacy duty, though a system operator's own internal risk policies may still require testing and review before a forecast changes a live balancing action.

Controls to put in place

  • Documented intended purpose limiting the model to forecasting, with the commitment or balancing decision kept with a person or a separately governed system
  • Regular comparison of forecast against actual output, reviewed by the trading or forecasting team, not only by the model's own developers

Frequently asked questions

Does the AI decide how much power to commit to the grid?
The sources on this page do not say who signs off on the resulting commitment. Google says its model recommends optimal hourly delivery commitments; it does not say who approves them. National Grid ESO's page does not describe an approval step at all. As a matter of playbook advice rather than a reported fact, keep a person reviewing the recommended commitment until forecast error is well understood by season and weather regime.
How far ahead do these forecasts look?
Google's wind forecasting system predicts output 36 hours ahead of actual generation to support delivery commitments made a full day in advance. National Grid ESO's own page about its solar forecasting improvement does not state what horizon the forecast covers.
Does better forecasting reduce the need for conventional backup generation?
Neither deployment on this page reports a reduction in reserve or backup generation as a measured figure, so treat that as a plausible but unproven benefit rather than a checked result.

How to cite this page

Blits.ai AI Use Case Library, "AI forecasting of wind and solar generation for grid balancing and trading", last verified 29 September 2026, https://www.blits.ai/ai-use-cases/renewable-generation-forecasting. Licensed under CC BY 4.0. Method: how we verify use cases.

Changelog
  • 29 September 2026: First published

Related use cases

Energy and utilities

AI analytics for smart meter and AMI data

AI that turns the flood of readings from smart electricity, gas and water meters into usable information: it monitors meter and network health at scale, estimates which appliances drive a household's usage from the meter signal alone, flags unusual consumption, and targets efficiency and electrification programmes at the customers who will benefit most, instead of a utility treating every meter and every customer the same way.

Deployments
2 public, best grade C
Autonomy
Assist
Healthcare

AI command center for hospital bed and staff capacity planning

An AI powered operations center that predicts patient admissions, discharges and transfers across a hospital or health system, and helps a team of coordinators sitting in one room sequence real time bed assignments, staffing levels and patient moves, so patients get into the right bed faster and existing capacity is used fully without adding beds.

Deployments
2 public, best grade B
Reported cycle time reduction
38%
Johns Hopkins Medicine, organization claim
Retail and ecommerce

AI demand forecasting and automated replenishment for retail

Machine learning that forecasts demand for every item in every store or fulfillment center, day by day, from sales history, promotions, prices, weather and local events, and turns the forecast into automatic store and warehouse orders within limits set by planners, who handle the exceptions.

Deployments
4 public, best grade B
Autonomy
Supervised agent
Pharma and life sciences

AI for clinical trial site selection and feasibility

AI that scores and ranks candidate investigator sites and countries for a planned clinical trial by predicted enrollment speed, access to the eligible patient population and historical performance, so clinical operations teams choose and activate a shortlist of sites with a higher chance of meeting enrollment targets on time, instead of relying on which sites a study team happens to know.

Deployments
2 public, best grade B
Autonomy
Assist