Explainable AI workflow targets trustworthy solar power forecasting
A new arXiv paper proposes a reproducible machine learning workflow for day-ahead solar plant forecasting, pairing automated model search with explainable AI and testing it on real data.
In plain words
Researchers describe a repeatable process for choosing and checking machine learning models that predict how much power a solar plant will produce a day ahead. On a test plant in Porto, the best models roughly halved the error of a persistence forecast, and a calibration step made the models' uncertainty ranges much more reliable.
What was announced
The paper argues that published accuracy figures for solar forecasting are often hard to compare because many studies validate with random data splits, omit a reference forecast, and present feature attributions without explaining their limits. As an alternative, the authors review major machine learning families, from tree ensembles to deep and foundation models, and set out a workflow combining automated model search with explainable AI. They then test it on an openly available dataset: a 16.32 kWp plant in Porto, Portugal, forecast one day ahead using a 4 km numerical weather prediction grid.
Over a full test year, the best models reached a normalised root mean square error of 11.6% of installed capacity, about half the error of smart persistence. Input feature choices mattered more than the choice among strong algorithms, and shuffled cross-validation proved mildly optimistic. After conformal calibration, nominal 80% prediction intervals covered 81.7% of observations, up from 64.5% before. SHAP rankings stayed stable across retraining periods but shifted when correlated inputs were removed, so the authors caution that attributions describe the model, not the plant itself.
What it does not show
- The source is only the paper's abstract, so the full methodology, comparisons and deployment discussion cannot be verified from the supplied passages.
- The results come from a single test plant and one test year, so performance on other sites or climates is not established.
- The paper lists edge deployment and climate change as open problems, meaning those areas remain unresolved.
For security and IT teams
- Status
- The workflow is a research proposal validated on an openly available dataset, not a reported production deployment.
Sources
Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.