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

This study evaluates deep learning methods for daily streamflow prediction in data-scarce contexts, using two headwater catchments of the Steelpoort River (Olifants River basin, South Africa) as case studies. Gated Recurrent Unit (GRU) and Long Short-Term Memory (LSTM) networks are trained using daily rainfall and temperature data, together with historic streamflow in an autoregressive manner. The approach incorporates predicted streamflow values into the look-back window to generate continuous predictions across the testing period, enabling both gap-filling and short-term forecasting. To ensure physically consistent outputs, the authors modify standard GRU/LSTM architectures by (i) changing the activation function in the final hidden layer and (ii) applying a non-negative constraint in the dense layer to prevent negative streamflow simulations. Weather inputs are sourced from (1) an ARC station record and (2) gridded products (CHIRPS rainfall and NASA POWER/NASAP variables), combined with DWS streamflow data from gauging stations B4H007 and B4H001. Training spans 1 Oct 1979–30 Sep 1997 and testing spans 1 Oct 1997–22 Feb 2002. Results show that models using ARC station data achieve reliable predictions, while models using gridded datasets provide moderately accurate predictions, indicating potential for low-cost streamflow estimation where gauges are inactive or incomplete. The work contributes a practical method for southern Africa where monitoring networks have declined.