HYBRID MODELING BASED ON DECOMPOSITION AND RESAMPLING FOR FORECASTING POWER GENERATION IN HYDROELECTRIC POWER PLANTS

This study proposes a hybrid model with decomposition to forecast hydroelectric generation in the months of January and February 2026, in multiple steps forward (1, 6 and 12), for a plant located on the border between Brazil and Paraguay. The model incorporates two decomposition methods , Singular Spectrum Analysis (SSA) and Wavelet, a resampling method, the Circular Block Bootstrap (CBB), and the Light Gradient Boosting Machine (LGBM), Extreme Gradient Boosting (XGBoost), Categorical Gradient Boosting (CatBoost), Random Forest (RF) and K-Nearest Neighbors (KNN) prediction models, applied independently. Single models, non-decomposing hybrids and decomposing hybrids are compared. The results reveal the superiority of hybrid models with decomposition in most horizons, with emphasis on the WAV-SSA-XGBoost model, which obtained Symmetric Mean Absolute Percentage Error (SMAPE) of 4.09% and 4.02% for the January and February forecast, respectively, in the horizon of 6 steps ahead.