A hybrid deep learning approach for streamflow prediction utilizing watershed memory and process-based modeling

Abstract

Accurate streamflow prediction is essential for optimal water management and disaster preparedness. While data-driven methods’ performance often surpasses process-based models, concerns regarding their ‘black-box’ nature persist. Hybrid models, integrating domain knowledge and process modeling into a data-driven framework, offer enhanced streamflow prediction capabilities. This study investigated watershed memory and process modeling-based hybridizing approaches across diverse hydrological regimes – Korean and Ethiopian watersheds. Following watershed memory analysis, the Soil and Water Assessment Tool (SWAT) was calibrated using the recession constant and other relevant parameters. Three hybrid models, incorporating watershed memory and residual error, were developed and evaluated against standalone long short-term memory (LSTM) models. Hybrids outperformed the standalone LSTM across all watersheds. The memory-based approach exhibited superior and consistent performance across training, evaluation periods, and regions, achieving 17–66% Nash–Sutcliffe efficiency coefficient improvement. The residual error-based technique showed varying performance across regions. While hybrids improved extreme event predictions, particularly peak flows, all models struggled at low flow. Korean watersheds’ significant prediction improvements highlight the hybrid models’ effectiveness in regions with pronounced temporal hydrological variability. This study underscores the importance of selecting a specific hybrid approach based on the desired objectives rather than solely relying on statistical metrics that often reflect average performance.