Investigating the applicability of long short-term memory model for streamflow prediction

Streamflow prediction, especially extreme events, poses a significant challenge due to the intricate and unpredictable nature of the rainfall-runoff process. Recently, promising results have been observed in time series problems by applying deep learning methods, including Long Short Memory (LSTM) and sequential modelling. This study investigates the application of the LSTM network to predict daily streamflow in the Dez River basin, Iran, during 2012–2019. Accordingly, observed precipitation, temperature, empirical evapotranspiration, and runoff were utilized as predictor variables. The performance of the LSTM model was compared with an established process-based approach, the Hron rainfall-runoff model, which served as a benchmark to evaluate the effectiveness of this innovative model. The models were evaluated using Kling-Gupta efficiency (KGE), Nash-Sutcliff efficiency coefficient (NSE), normalized root mean square error (NRMSE), and mean absolute percentage error (MAPE). Through evaluation and analysis, the NSE and MAPE indices were, respectively, 0.95 and 15.6% for the LSTM model in the validation stage. The results demonstrated that the LSTM model performed better than the Hron model in predicting daily streamflow. The superior performance of the LSTM network represents its efficiency in capturing and utilizing inherent temporal dependencies in hydrological data. This finding highlights the potential of the proposed model for improving the accuracy and reliability of real-time hydrological forecasts.