STL‐Decomposition Functional Deep Learning Ensemble Models for Hydro Power Forecasting

ABSTRACT Reliable short‐term forecasting of renewable energy generation is a cornerstone for optimizing virtual power plant (VPP) operation and ensuring grid stability. Hydroelectricity, characterized by dispatchability, low operating cost, and minimal greenhouse gas emissions, presents unique challenges due to the nonlinear and stochastic dynamics of water inflows. This study introduces a hybrid ensemble framework that integrates seasonal–trend decomposition using loess (STL), functional data analysis (FDA), and deep learning (DL) architectures. Hourly hydroelectric production data are additively decomposed via STL into interpretable components, which are partitioned into subsets and individually modeled using DL. The component‐specific forecasts are subsequently fused through FDA to construct ensemble predictors. Seven ensemble strategies are bench marked on real data from the Edea hydro power plant. The seasonal–random ensemble achieved superior accuracy (mean absolute error (MAE) = 2.376 MWh, root mean squared error (RMSE) = 3.827 MWh, nash‐sutcliffe efficiency (NSE) = 0.892), demonstrating that decomposition–based ensemble learning substantially enhances predictive skill in hydroelectric forecasting.