Generalized streamflow forecast model using deep learning

Abstract:

Recent studies have shown deep learning models in hydrology can perform better than machine learning models and physically-based models on multiple stations. However, most studies are based on training custom model for each station, and the generalization ability of deep learning in hydrology is not studied comprehensively. We developed a generalized model with distributed structure for the streamflow forecast for the next 120 hours. Our generalized model has been successfully applied to 125 USGS gauges in the State of Iowa at a higher median accuracy than 125 single models on each basin with considering watershed-scale features including area, time of concentration, slope, and soil types. Furthermore, theoretically, the proposed model can be applied to any new gauges in the State of Iowa without training in an acceptable accuracy, and more data from additional gauges may increase the accuracy in future studies. Our study shows different regional models could be better than one model when some watershed features (e.g., area) are highly heterogeneous in the dataset. The performance of our model also suggests that the deep learning models with pretreatments such as the construction of a distributed dataset can achieve better results. Our study suggests that deep learning studies in hydrology could include domain knowledge and physical features for better performance in the future.