Multivalued Daily Runoff Prediction in Small Watershed Based on Improved CREST Model

This study aims to address a limitation of the CREST distributed hydrological model, which has shown insufficient responsiveness to sudden changes in runoff peaks during long-term daily runoff series in small watersheds, resulting in decreased prediction accuracy. We propose an enhancement by coupling the CREST model with a Gated Recurrent Unit (GRU) neural network, further enriched with the Sparse Multi-headed Self Attention mechanism. The enhanced model is designated as CSSA-GRU (CREST-GRU-Sparse Multi-headed Self Attention). The CSSA-GRU model utilizes runoff data derived from the CREST model and observed daily runoff as inputs, leveraging both GRU and Sparse Multi-headed Self Attention mechanisms for enhanced accuracy in coupled daily runoff predictions. As a case study, we apply the CSSA-GRU model to the Mi River basin in Shandong Province. With a forecasting horizon of 1–3 days, CSSA-GRU demonstrates significant superiority over both the original CREST model and other GRU variants. This is evidenced in its performance on both single-value daily runoff and multi-value coupled daily runoff predictions, where the Nash-Sutcliffe Efficiency Coefficients (NSE) are markedly improved, registering at 0.936, 0.846, and 0.767 respectively.