Deep learning architectures optimization for precise streamflow prediction: assessing activation–optimizer synergies

Abstract

Precise streamflow estimation plays a pivotal role in water resources management. Predicted streamflow has certain uncertainty owing to the inaccurate selection of activation functions and optimizers. This study aims to assess various activation functions (ReLU, Tanh, Sigmoid, ELU, SoftMax, and Leaky_ReLU) and optimizers (Adam, Nadam, Adagrad, Adadelta, FTRL, RMSProp, and SGD) in simulating streamflow using long short-term memory (LSTM), stacked-LSTM, Bidirectional LSTM (Bi-LSTM), and gated recurrent unit (GRU) deep learning (DL) models to improve prediction accuracy. About 42 distinct architectures for LSTM, Stacked-LSTM, Bi-LSTM, and GRU were developed for the Swat River basin, employing various activation functions and optimizers. The DL models were trained and tested using 70 and 30% of observed data, respectively. The min–max scaling technique was employed for data scaling, utilizing various optimizers and activation functions across different batch sizes, epochs, layers, and patience values. The model's performance was assessed using statistical performance indicators, including the coefficient of determination (R2), mean squared error (MSE), Nash–Sutcliffe efficiency (NSE), and root mean squared error (RMSE). The Bi-LSTM model was identified as superior model under Adam–ReLU optimizer–activation combination via compromise programming (R2: 0.83 during training, 0.78 during testing). The best-performing Bi-LSTM model was used to predict future streamflow under two shared socioeconomic pathways (SSPs) namely SSP245 and SSP585 indicating alterations in projected streamflow. The Bi-LSTM model demonstrated its reliability for predicting streamflow and its potential for integration with climate projections, making it a valuable tool for water resource management in the study area.