ENSO-Informed Hydroelectric Discharge Forecasting with Gradient Boosting and Temporal Fusion Transformer

Abstract

Accurate weekly discharge forecasting is essential for hydroelectric reservoir operation, energy planning, and sustainable water resource management. However, long-term hydrological records are inherently non-stationary due to large-scale climate variability, making accurate forecasting challenging for conventional data-driven models. This study proposes a hydro-climatic regime-aware forecasting framework that combines an ENSO phase indicator with hydrological lag variables, storage-derived descriptors, and inflow variability features. The framework is evaluated on an 85-year (1938–2023) weekly Hoover Dam dataset using XGBoost, LightGBM, and the Temporal Fusion Transformer (TFT).An ablation study was conducted using XGBoost to distinguish the predictive contribution of the ENSO indicator from that of the complete Regime-Aware feature set. The results show that adding the ENSO indicator alone provides only marginal improvement (R2 = 0.683 versus 0.686 for the Baseline model), whereas the complete Regime-Aware feature set substantially improves forecasting performance, achieving R2 = 0.823 while reducing MAE from 313.5 m3/s to 232.1 m3/s. Among all evaluated models, TFT Regime-Aware achieves the highest overall accuracy (R2 = 0.919, MAE = 214.8 m3/s). A policy change-point experiment further indicates that administrative policy information provides negligible additional predictive benefit once the proposed hydro-climatic feature representation is available. These findings suggest that integrating climate context with hydrological state variables improves discharge forecasting under non-stationary hydro-climatic conditions.