Bridging data-driven and process-based approaches for hydrological modeling in the tropics: insights from the Kelani River Basin, Sri Lanka

Accurate streamflow prediction is essential for effective water resource planning and management. Although physics-based hydrological models such as SWAT and WEAP are commonly used for streamflow simulation, they often encounter limitations due to structural complexity, rigid conceptual assumptions, and sensitivity to parameter calibration. In this study, LSTM models are utilized as a data-driven alternative for monthly streamflow prediction in the Kelani River Basin (KRB), Sri Lanka. Three variations of the LSTM architecture, Vanilla LSTM, Stacked LSTM, and Bidirectional LSTM (Bi-LSTM), are assessed and compared against conventional physics-based models, including SWAT and WEAP. Results illustrated that LSTM models consistently outperform SWAT and WEAP during both calibration and validation phases. During calibration, LSTM models achieved high accuracy with NSE values nearing 0.95, R² between 0.95 and 0.96, PBIAS ranging from 2.03 to 4.56, and RSR between 0.21 and 0.23. Physics-based models exhibited lower performance (NSE: 0.71–0.74; R²: 0.74–0.83; PBIAS:23.67 to 4.7; RSR: 0.51–0.54). Validation results confirmed this trend, with LSTM models maintaining strong performance (NSE: 0.82–0.84; R²: 0.84–0.88; PBIAS:11.5 to –15.60; RSR: 0.40–0.43), while physics-based models displayed weaker predictive capability (NSE: 0.50–0.61; R²: 0.66–0.81; PBIAS:16.33 to –42.14; RSR: 0.62–0.71). Among the LSTM variations, Bi-LSTM demonstrated the best performance during calibration, while Stacked LSTM proved to be more effective during validation. The study underscores the robustness and reliability of LSTM models for monthly streamflow prediction, presenting a valuable approach for long-term water resource management in the KRB.