Assessing Physics-Informed Neural Networks in Streamflow Predictions for Tropical Monsoonal Basins in Sri Lanka

Abstract

Globally, escalating floods and droughts demand accurate streamflow prediction for effective water management and early warning. Yet, this remains challenging in tropical monsoon regions. Traditional process-based models (e.g., HEC-HMS) underperform given extensive data needs, while purely data-driven models (e.g., LSTMs) risk physical inconsistency during extreme events. Consequently, this study introduces a novel Physics-Informed Neural Network (PINN) embedding a generalized catchment water balance ordinary differential equation (ODE) directly into its loss function. The model was evaluated across two hydro-climatically contrasting Sri Lankan catchments: the Kalu Ganga (wet zone) and the Maduru Oya (dry zone), benchmarked against HEC-HMS and ANN, RNN, LSTM, and Transformer baselines. In the Kalu Ganga basin, the PINN achieved exceptional accuracy (Validation NSE = 0.949, KGE = 0.951), significantly outperforming the best data-driven baseline and HEC-HMS. In the more complex, regulated Maduru Oya basin, the PINN attained the lowest volumetric bias (+6.0%) and best KGE (0.721), demonstrating superior mass conservation compared to the LSTM's higher NSE but larger bias (+20.1%) and HEC-HMS's bias (-36.5%). The findings demonstrate that embedding physical laws as soft constraints acts as a volumetric anchor, significantly enhancing predictive reliability and mass conservation over purely empirical approaches in tropical environments.