A process-driven deep learning hydrological model for daily rainfall-runoff simulation

Abstract

Although deep learning (DL) models, especially long-short-term memory (LSTM), demonstrate greater accuracy than process-based models in rainfall-runoff simulation, the predictions from process-based models are more physical than DL models. The main reason is that DL models have almost no process understanding capabilities like process-based models beyond their fitting capability. In this study, we developed a process-driven DL model under a unified DL architecture to improve the process awareness of DL models. To implement the model, a conceptual hydrological model (EXP-HYDRO) is implanted into a recurrent neural network (RNN) cell as a process driver for providing multi-sub-process variables related to the runoff process, and an Entity-Aware LSTM (EA-LSTM) cell is incorporated as a post-processor layer, resulting in the Process-driven RNN-EA-LSTM (PRNN-EA-LSTM). Under the assistance of the process driver, the model performance of PRNN-EA-LSTM on the 531 catchments from the Catchment Attributes and Meteorology for Large-sample Studies dataset is more robust than the pure DL model, and better than using only EXP-HYDRO as the input of EA-LSTM (i.e., EXP-HYDRO-EA-LSTM). Specifically, the median Nash-Sutcliffe efficiency (NSE) of PRNN-EA-LSTM in local and regional simulation is 0.03 and 0.02 higher than LSTM and 0.01 higher than EXP-HYDRO-EA-LSTM. Additionally, PRNN-EA-LSTM significantly enhances the low flow simulations and reduces the catchments number with negative NSE. This study demonstrates that process-based models can help DL models better represent the rainfall-runoff relationship under a unified architecture. Consequently, integrating the adaptability of process-based models into the DL architecture is anticipated to bolster the process understanding capabilities of DL models.