Multi-Model Stacking Ensemble with Multi-Perspective Interpretability Analysis for Solar Power Forecasting

Abstract

Accurate photovoltaic (PV) power forecasting is important for grid dispatch, energy-storage management, and electricity-market trading. This paper presents a PV power forecasting framework that combines multi-model ensemble learning with multi-perspective interpretability analysis. A 22-dimensional feature set was constructed from irradiance, meteorological, temporal, and lagged-power variables. Five supervised regressors, a zero-shot Chronos-Bolt-Small time-series foundation model, and six additional forecasting baselines were evaluated at two PV plants. The three ensemble schemes used only the five supervised regressors and were compared on the final 20% of the 2019 development data; the complete 2020 period was used only for final evaluation. The selected methods achieved normalized RMSE values of 0.0350 and 0.0376 at Sites 1 and 2, respectively. SHAP, PDP/ICE, LIME, and permutation importance were applied to the ensemble model at each site. Recent power-history features dominated this one-step-ahead forecasting task with a 15 min horizon, while irradiance features provided additional site-dependent information.