Leveraging historic streamflow and weather data with deep learning for enhanced streamflow predictions

Abstract

Streamflow information is crucial for effectively managing water resources. The declining number of active gauging stations in many rivers is a global concern, necessitating the need for reliable streamflow estimates. Deep learning techniques offer potential solutions, but their application in southern Africa remains largely underexplored. To fill this gap, this study evaluated the predictive performance of gated recurrent unit (GRU) and long short-term memory (LSTM) networks using two headwater catchments of the Steelpoort River, South Africa, as case studies. The model inputs included rainfall, maximum, and minimum temperature, as well as past streamflow, which was utilized in an autoregressive sense. The inclusion of streamflow in this way allowed for the incorporation of simulated streamflow values into the look-back window for predicting the streamflow of the testing set. Two modifications were required to the GRU and LSTM architectures to ensure physically consistent predictions, including a change in the activation function of the GRU/LSTM cells in the final hidden layer, and a non-negative constraint that was used in the dense layer. Models trained using commercial weather station data produced reliable streamflow estimates, while moderately accurate predictions were obtained using freely available gridded weather data.