Spatiotemporal Multi-Horizon Solar Irradiance Forecasting in High-Potential Microclimates: A Comparative Benchmark from Linear Baselines to Deep Learning and Foundation Models

Abstract

The operational integration of solar energy into modern power grids, particularly for day-ahead markets and battery storage scheduling, requires highly accurate, multi-horizon irradiance forecasts to mitigate cloud-induced intermittency. This study evaluates a spatiotemporal, multi-horizon (24-h) Global Horizontal Irradiance (GHI) forecasting framework across a high-resolution 167-node microclimate in Tula, Tamaulipas, Mexico, using a robust 2020–2024 baseline. We benchmarked six predictive paradigms: diurnal persistence, linear regression, tree ensembles (Extra Trees and XGBoost), a zero-shot foundation model (TimesFM), and a deep CNN–LSTM network, using historical clear-sky irradiance sequences and strictly causal meteorological features. On the independent 2024 test set, XGBoost achieved the lowest macro-average RMSE (120.87Wm−2), but Diebold–Mariano testing confirmed that it was statistically equivalent to the CNN–LSTM architecture. XGBoost showed significant superiority at ultra-short horizons (t+1 and t+2) by efficiently capturing nonlinear radiative inertia, whereas CNN–LSTM minimized errors under sustained overcast regimes and produced the sharpest empirical 95% prediction intervals. At the 24-h mark, ordinary linear regression converged, whereas the complex models did not, highlighting a physical transition toward deterministic solar geometry. Remarkably, zero-shot TimesFM outperformed diurnal persistence without any domain-specific training. Conditional Tree SHAP interpretability demonstrated that the XGBoost model dynamically reallocated predictive weight, shifting from contemporaneous irradiance under clear skies to precipitable water during volatile, intermittent conditions.