Hybrid Neural Hierarchical Interpolation Time Series with STL Optimized by Multi-Agent HPO for Flow Forecasting in Hydroelectric Power Plants

Accurate forecasting of reservoir levels in hydroelectric power plants is essential for efficient energy generation, operational safety, and sustainable water management. This study proposes a hybrid forecasting framework that integrates Seasonal-Trend Decomposition using Loess (STL) with the Neural Hierarchical Interpolation Time Series (NHITS) model optimized through Multi-Agent Hyperparameter Optimization (HPO). The STL filter is employed to remove high-frequency noise and preserve underlying signal trends, enhancing the quality of inputs for the predictive model. The NHITS architecture leverages hierarchical multi-scale processing and interpolation-based reconstruction to capture both short- and long-term temporal dependencies, while the Multi-Agent HPO framework ensures optimal hyperparameter configuration through cooperative agent-based exploration. The proposed method was evaluated using turbine flow data from the Santo Antônio Hydroelectric Power Plant in Brazil, achieving superior performance compared to state-of-the-art benchmarks, including transformer-based and recurrent models, across very short- and short-term forecasting horizons. The model demonstrated substantial reductions in RMSE, MAE, MAPE, and MSLE, while maintaining competitive inference times. The results confirm that the STL–NHITS architecture optimized via Multi-Agent HPO offers an efficient, robust, and interpretable solution for hydropower forecasting, providing a valuable tool for intelligent management of renewable energy systems.