Item

Abstract

Long short-term memory (LSTM) networks have been proved to work favorably in precipitation-runoff modeling due to their outstanding performance in processing long-length temporal dynamics. However, the other hydrological processes and the spatial information are rarely incorporated in current LSTM hydrological models, which hinders the models from making full use of hydrological data. This study proposes a spatiotemporal deep-learning (DL)-based hydrological model, by coupling the 2-Dimension convolutional neural network (CNN) and LSTM and introducing actual evaporation (E) as an additional training target. We use three large mountainous basins on the Tibetan Plateau to test the proposed CNN-LSTM model and compare the results to the LSTM-only model. Further analyses are conducted to explore meteorological and hydrological information hidden in the CNN and LSTM based on the linear regression model. Results indicate that both LSTM and CNN-LSTM hydrological models perform well on runoff (Q) and E simulation, with Nash-Sutcliffe efficiency coefficients (NSE) higher than 0.82 and 0.95, respectively. Involving CNN in the LSTM-only Q model to capture spatial information enhances the overall and peak Q modeling performance. Multi-task simulation with LSTM-only models show better accuracy in the estimation of Q volume and performance. On the other hand, we conclude that CNN models can capture the basin spatially-averaged values from 2-D spatial meteorological data. The internal cells of LSTM models for Q (E) contain the E (Q) process. This study provides an essential tool for accurate flood forecasting and management in large mountainous basins.