19. IMPROVING SHORT-TERM RESERVOIR INFLOW FORECASTING USING A HYBRID HYPE-ANN FRAMEWORK: A CASE STUDY OF THE PLEIKRONG RESERVOIR

Abstract

Accurate reservoir inflow forecasting is vital for hydropower operation and water resources management in monsoon-driven basins. However, process-based hydrological models often exhibit systematic errors, particularly under high-flow conditions. To address this, we developed a hybrid modeling framework that integrates the HYPE (Hydrological Predictions for the Environment) model with an Artificial Neural Network (ANN) for daily inflow forecasting to the Pleikrong Reservoir in Vietnam. The HYPE model was first calibrated and validated using daily hydrometeorological data from 1994 to 2022. ANN models with different hidden-layer architectures (three, four, and five layers) were then employed to post-process HYPE outputs. Results show that the hybrid HYPE-ANN approach substantially improved forecast accuracy compared to HYPE alone. The four-layer ANN (512 - 256 - 128 - 64 neurons) achieved the best performance, with CC = 0.93, KGE = 0.92, and NSE = 0.87 for one-day-ahead forecasts, while maintaining stable results for two-and three-day lead times. These findings highlight the effectiveness of hybrid process-based and data-driven approaches for short-term inflow prediction using daily data. The proposed framework offers a reliable and computationally efficient tool to support hydropower operation and adaptive water resources management in data-scarce basins.