A hydrological process-based neural network model for hourly runoff forecasting

Neural network models have been widely used in runoff forecasting, but are often criticized for their lack of physical interpretability. In this study, we present a simple but useful approach to developing hydrological models by designing neural networks based on the principles of runoff generation and concentration, which we refer to as a Hydrological Process-based Neural Network (HPNN) model. The Convolutional neural network (CNN) and softmax function are used because of their similar formula to the conventional runoff generation and unit hydrograph approach used in hydrology. We apply the HPNN model and four other benchmark models to forecast runoff in two catchments (Yutan and Chenda) in China. Results show that the HPNN model has higher computational efficiency, its parameters are interpretable and closely linked to the processes of runoff generation and concentration, and the HPNN model outperforms conventional GRU-based models.