AI for Science
CatBoost led five machine-learning models at forecasting solar energy output in a 21,045-sample study
In a 2025 paper in PLoS ONE, researchers compared five machine-learning regression algorithms for forecasting solar energy generated by a photovoltaic panel. CatBoost made the most accurate predictions overall. A SHAP analysis found ambient temperature and humidity had the greatest influence on the amount of solar energy generated.
How did the study assess the models?
The comparison included five regression algorithms, which predict a continuous number: Gradient Boosting Regressor, XGBoost, K-Neighbors Regressor, LightGBM, and CatBoost. The researchers trained the models on 21,045 samples using six inputs: humidity, ambient temperature, wind speed, visibility, cloud ceiling, and pressure. They measured accuracy with R2, root mean square error (RMSE), and mean absolute error (MAE).
Which model did best?
CatBoost led the comparison. On the testing split, CatBoost reached an R2 accuracy score of 0.46, an RMSE of 4.748 watts, and an MAE of 3.583 watts. On the training split, the scores were an R2 of 0.608, an RMSE of 4.478 watts, and an MAE of 3.367 watts.
What factors mattered most?
The SHAP analysis showed that ambient temperature and humidity had the greatest influence on the solar energy generated from the photovoltaic panel.
Questions readers ask
What do R2, RMSE, and MAE mean?
In general, R2, the coefficient of determination, measures how closely predictions match observed values. RMSE and MAE measure prediction error in watts. Lower error values are better. CatBoost's testing error was 4.748 watts by RMSE and 3.583 watts by MAE.
Why does solar energy forecasting matter?
Solar output changes with weather and time. Accurate forecasting is needed to balance supply and demand, optimize energy storage, and manage grid stability, the paper says.
Limits and what's next
The study compared five regression algorithms on one dataset of 21,045 samples. The abstract names six weather-related inputs: humidity, ambient temperature, wind speed, visibility, cloud ceiling, and pressure. It does not name where or when the samples came from. The design was a model comparison, not an external field test, and the abstract reports no independent replication of the CatBoost result. The testing R2 of 0.46 was lower than the training R2 of 0.608, and the abstract does not explain that gap.
Source: Huu Nam Nguyen et al., “Solar energy prediction through machine learning models: A comparative analysis of regressor algorithms,” PLoS ONE, 2025.