Regional characteristics’ impact on the performances of the gated recurrent unit on streamflow forecasting

The gated recurrent unit (GRU) has obtained attention as a potential model for streamflow forecasting in recent years. Common patterns and specialties when employing it in different regions, as well as a comparison between different models still need investigation. Therefore, we examined the performances of GRU for one, two, and three-day-ahead streamflow forecasting in seven basins in various geographic regions in China from the aspect of robustness, overall accuracy, and accuracy of streamflow peaks’ forecasting. The robustness and accuracy of it are closely related to correlations between the input and forecasting target series. Also, it outperforms the benchmark machine learning models in more cases, especially for one-day-ahead forecasting (NSE of 0.88–0.96 except for the unsatisfactory result in the Luanhe River basin). The deterioration of its accuracy along the increasing lead time depends on the dominant time lags between the rainfall and streamflow peaks. Recommendations were proposed for further applications.