Artificial intelligence is used in photovoltaic (PV) systems mainly to forecast how much power panels will produce, and the evidence on what a better forecast is worth comes from grid studies, not from AI marketing. A 2026 study by Lawrence Berkeley National Laboratory and the National Renewable Energy Laboratory estimated that a more accurate day-ahead solar forecast would cut production costs by USD 87 million per year and reserves held for forecast error by 30 percent in a modeled Southeastern U.S. grid. That study improved the forecast with a weather model, not with machine learning, so its numbers show the size of the prize, not the winner.
Key takeaways
- Solar forecasts are judged against a simple baseline, persistence, which assumes tomorrow's output at each hour equals today's.
- In the Berkeley Lab study, forecast improvements saved USD 87 million per year in the base case, and the saving depended on how flexible the rest of the grid was.
- A peer-unreviewed 2026 preprint tested seven nowcasting models on 6,434 Swiss PV stations, and satellite-based models beat a physics-based weather ensemble at short lead times.
- A 2022 meta-analysis found that the best input data changes with the forecast horizon, so no one method wins everywhere.
What is being predicted
A PV array converts sunlight to electricity, so its output rises and falls with cloud cover, haze, and the angle of the sun. Grid operators schedule other power plants a day ahead and adjust within the day. If solar output comes in below the forecast, another source must make up the difference, and operators hold spare capacity called reserves for that purpose. Forecasts therefore matter at three time scales: minutes ahead (nowcasting, mostly from cloud movement), hours ahead (intraday), and a day or more ahead.
The standard way to score a forecast is skill, which compares its error with a reference forecast. The reference is usually persistence. A method that cannot beat persistence adds nothing, however advanced it sounds. Machine learning enters at several points: reading satellite or sky images to move clouds forward in time, correcting a weather model's known biases, or converting predicted sunlight into power for a specific plant.
What a better forecast is worth: a grid study
The clearest quantitative evidence is a Berkeley Lab study by Kahrl and coauthors, published in Renewable Energy in March 2026 and supported by the U.S. Department of Energy's Solar Energy Technologies Office. The authors modeled the solar-rich Southeast, covering Alabama, Georgia, Kentucky, North Carolina, South Carolina, and Tennessee. In their scenario solar made up 45 percent of installed capacity and 34 to 36 percent of generation, alongside 43 GW of storage equal to 33 percent of peak demand.
They compared two forecasts: a persistence forecast and one based on the North American Mesoscale model, a numerical weather prediction system run by the U.S. National Centers for Environmental Prediction. The better forecast reduced production costs by USD 87 million per year in the base scenario (USD 0.13 per megawatt-hour of load, 2023 dollars) and cut the reserves needed to cover forecast error by about 30 percent. Neither forecast was a machine learning model.
Savings depend on the rest of the grid
The same study ran other scenarios, and the results show why a forecast has no fixed value. The figures below are from the paper.
| Scenario | What changes | Effect of the better forecast on yearly production cost |
|---|---|---|
| Base | Limited trading between regions | Saves USD 87 million (0.9 percent) |
| High coordination | No trading charge between regions, shared reserves | Saves USD 230 million (2.5 percent) |
| Flexible gas | Gas plants can start and stop in real time | Costs rise by USD 41 million (0.5 percent) |
In the last case the paper reports that the better forecast raised costs slightly. The authors conclude that regional coordination complements better forecasts while gas flexibility substitutes for them. A claim that AI forecasting "saves" a given amount should therefore name the grid it was measured on.
Machine learning against physics and persistence
A January 2026 arXiv preprint by Lanzilao and Meyer, not yet marked as peer reviewed on its page, compared seven intraday nowcasting models using production data from 6,434 PV stations across Switzerland. Satellite-based approaches outperformed the IFS-ENS numerical weather prediction ensemble, especially at short lead times. The deterministic network IrradianceNet had the lowest root mean square error, while the probabilistic models SolarSTEPS and SHADECast gave better-calibrated uncertainty ranges. For daily national generation, the satellite models kept relative error below 10 percent on 82 percent of days in 2019 and 2020. Forecast skill generally fell with elevation, a reminder that mountainous terrain is harder.
A 2022 meta-analysis by Nguyen and Müsgens, also a preprint, screened 1,447 papers from scholarly search, reviewed 320 in full, and pooled 4,687 data points. It found that forecast horizon has the largest effect on skill, that satellite and sky images matter most for intraday forecasts, and that weather-model data and local measurements matter most for day-ahead forecasts. It also reported that adding training data helps until about 2,000 days, after which overfitting appears.
Reading skill from reported errors
The Berkeley Lab paper's appendix reports a normalized mean absolute error of 2.8 percent for the weather-model forecast and 5.8 percent for persistence across the whole region. Skill relative to persistence is one minus the ratio of the two errors, so 1 - 2.8/5.8 gives about 52 percent. This is a calculation from the reported figures, not a number the paper states, and the paper itself notes that accuracy comparisons across studies are difficult because data, sites, and metrics differ.
Questions to ask about an AI solar forecast claim
- What is the baseline: persistence, a weather model, or the operator's current forecast?
- Which horizon was tested: minutes, hours, or a day ahead?
- Was the model tested on sites and years it was not trained on?
- Which error metric was used, and is it normalized to plant capacity?
- Is the savings figure from a real market or a simulation, and what flexibility did the grid have?
Other uses of AI in PV plants, such as detecting faulty modules from drone images or tuning inverters, were not assessed here because no primary source with a measured result was reviewed. This article is general information and not engineering or investment advice.
FAQ
Does AI make solar panels more efficient?
No source reviewed here shows that. Forecasting predicts output but does not change how much sunlight a panel converts. Its benefit is better scheduling of the rest of the grid.
What is a persistence forecast?
It assumes the future repeats the present, for example that tomorrow's output at noon equals today's noon output. Because it is easy to compute, it is the usual baseline a new method must beat.
Is the USD 87 million saving a measured result?
No. It is a simulation result for a modeled grid, using 2023 dollars, and it differs sharply across the scenarios studied.
Are there Canadian figures?
The studies reviewed here concern the U.S. Southeast and Switzerland. No Canadian measured result was verified for this article.
Related reading
- Optimization problems in AI
- AI in agriculture
- Traditional AI compared with machine learning
- AI hardware and what it measures
More in AI and Society.
Last verified: October 2026
Sources
- Getting brighter: Impacts of improved day-ahead solar forecasts in high-solar, high-storage electricity systems, Kahrl et al., Renewable Energy 259, Berkeley Lab, March 2026
- Intraday spatiotemporal PV power prediction at national scale using satellite-based solar forecast models, Lanzilao and Meyer, arXiv:2601.04751, January 2026
- A Meta-Analysis of Solar Forecasting Based on Skill Score, Nguyen and Müsgens, arXiv:2208.10536, 2022