Zhongrun Xiang, Ibrahim Demir, Ricardo Mantilla, Witold F. Krajewski 

Recent studies have shown that deep learning models in hydrological applications significantly improved streamflow predictions at multiple stations compared to traditional machine learning approaches. However, most studies lack generalization; i.e. researchers are training separate models for each location. The spatial and temporal generalization ability of deep learning models in hydrology that can be gained by training a single model for multiple stations is evaluated in this study. We developed a generalized model with a multi-site structure for hourly streamflow hindcasts on 125 USGS gauges. Considering watershed-scale features including drainage area, time of concentration, slope, and soil types, the proposed models have acceptable performance and slightly higher median NSE value than training individual models for each USGS station. Furthermore, we showed that the trained generalized model can be applied to any new gauge in the state of Iowa that was not used in the training set with acceptable accuracy. This study demonstrates the potential of deep learning studies in hydrology where more domain knowledge and physical features can support further generalization.