Coupling deep learning and physically-based hydrological models for monthly streamflow predictions

Authors: XU, WENXIN - Wuhan University; CHEN, JIE - Wuhan University; XU, CHONG-YU - University Of Oslo; Zhang, Xunchang; XIONG, LIHUA - Wuhan University; LIU, DEDI - Wuhan University

Submitted to: Water Resources Research
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 1/19/2024
Publication Date: 2/23/2024
Citation: Xu, W., Chen, J., Xu, C., Zhang, X.J., Xiong, L., Liu, D. 2024. Coupling deep learning and physically-based hydrological models for monthly streamflow predictions. Water Resources Research. 60.Article e2023WR035618. https://doi.org/10.1029/2023WR035618.
DOI: https://doi.org/10.1029/2023WR035618

Interpretive Summary: Long term streamflow forecasts are crucial for reservoir operation and water resource allocation. This study proposes a new hybrid model for monthly streamflow predictions by coupling a physically-based distributed hydrological model with an artificial intelligence-based deep machine learning (DL) computer algorithm. Specifically, a simplified hydrological model is driven by bias corrected global climate model (GCM) projections to generate soil moistures for the forecasting months. Then, the model-simulated soil moisture along with other predictors from multiple sources are used as inputs into the trained DL model to predict future streamflow. The proposed hybrid model is applied to predict 1-, 3-, and 6-month ahead reservoir inflows to the Danjiangkou Reservoir in China. The results show the hybrid model consistently performs better than either the hydrological model or the DL models for lead times up to 6 months. Overall, the new hybrid model is efficient and has great potential for monthly streamflow predictions, especially when training data are limited. This work provides a new approach to hydrologists for forecasting streamflow.

Technical Abstract: This study proposes a new hybrid model for monthly streamflow predictions by coupling a physically-based distributed hydrological model with a deep learning (DL) model. Specifically, a simplified hydrological model is first developed by optimally selecting grid cells from a distributed hydrological model according to their soil moisture characteristics. It is then driven by bias corrected general circulation model (GCM) predictions to generate soil moistures for the forecasting months. Finally, model-simulated soil moisture along with other predictors from multiple sources are used as inputs of the DL model to predict future streamflows. The proposed hybrid model, using the simplified Variable Infiltration Capacity (VIC) as the hydrological model and the combination of Convolutional Neural Network and Gated Recurrent Unit (CNN-GRU) as the DL model, is applied to predict 1-, 3-, and 6-month ahead reservoir inflows (Danjiangkou Reservoir in China). The results show that the original VIC model performs remarkably better than the CNN-GRU model in the studied watershed, because DL has difficulty in predicting the peak flows when training samples are insufficient and weakly correlated. However, the hybrid model consistently performs better than VIC and CNN-GRU models with great improvement in Kling-Gupta efficiency (KGE) values for lead times up to 6 months. In addition, input predictors have great influences on the performance of the CNN-GRU model, but the hybrid model is capable of minimizing their influence by introducing model-simulated soil moisture. Overall, the new hybrid model is efficient and has great potential for monthly streamflow predictions, especially when training data are limited.