Hybrid hydrological modeling: Integration of machine learning and conventional hydrology

This study introduces a novel hybrid modeling framework that combines the process-based HEC-HMS hydrological model with artificial neural networks (ANN) and long short-term memory (LSTM) algorithms to improve daily streamflow simulations in the Kedah River Basin, Malaysia. Unlike conventional standalone models, the hybrid approach integrates physical process understanding with advanced machine learning to address the limitations of each method in data-limited, tropical environments. Model performance was evaluated using widely recognized statistical metrics at two streamflow stations representing both regulated and natural flow regimes. Results show that the hybrid HEC-HMS-LSTM model achieved significantly higher predictive accuracy than either process-based or data-driven models alone, particularly in capturing peak flows and overall streamflow dynamics. For example, the hybrid model improved the Nash-Sutcliffe Efficiency from 0.34 to 0.66 at the regulated station and from 0.41 to 0.68 at the natural station during validation. These findings highlight the robustness and transferability of hybrid approaches for hydrological forecasting in tropical catchments. The novelty of this work lies in its systematic integration and evaluation of HEC-HMS with deep learning methods for tropical river basins, offering practical value for flood prediction and water resources management in regions where high-quality data are limited.