Horizon-Dependent Solar Irradiance Forecasting with Boosted Trees, and Seasonal Baselines Based on Measurements in Sudan

Abstract

Solar irradiance forecasting accuracy depends on the prediction horizon because the use of recent observations, meteorological variables, and seasonal patterns changes with lead time. However, weak persistence baselines, temporally unreliable validation, and inconsistent test samples can overstate the advantage of complex models. This paper presents a leakage-safe, horizon-specific solar irradiance forecasting framework combining boosted trees, validation-weighted convex forecast combinations, and seasonal reference models. Direct forecasts are evaluated at 10-min and 1-, 3-, 6-, 12-, and 24-h horizons. Models are tuned using expanding-window validation; preprocessing is fitted only to the training data; and every method is evaluated on a predefined canonical test support. XGBoost achieves the lowest root mean square error (RMSE), 10.78 W/m2, at 10 min, whereas CatBoost achieves the lowest RMSE, 16.08 W/m2, at 1 h. At 3, 12, and 24 h, the lowest RMSE is obtained by seasonally guided forecast combinations. At 6 h, the XGBoost and two-day seasonal-mean combination ranks first by RMSE, although its gain over the two-day seasonal mean results in a larger mean absolute error (MAE). Daily-block significance tests show an improvement over the reference models at 10 min, while improvements over seasonal references at longer horizons are not statistically significant after adjustment. The framework provides a reproducible benchmark for horizon-dependent solar irradiance forecasting.