Hybrid CNN-LSTM models for river flow prediction. Issue 5 (13th April 2022)

• Record Type: Journal Article
• Authors: Li, Xia; Xu, Wei; Ren, Minglei; Jiang, Yanan; Fu, Guangtao
• Is Part Of: Water Supply. Volume 22:Issue 5(2022)
• Page Start: 4902
• Page End: 4919
• Publication Date: 2022-04-13
• Subjects: CNN -- flow prediction -- hydrological modelling -- LSTM -- shortcut layer
• DOI: 10.2166/ws.2022.170
• Languages: English
• ISSNs: 1606-9749

Abstract

Abstract: River flow prediction is a challenging problem due to highly nonlinear hydrological processes and high spatio-temporal variability. Here we present a hybrid network of convolutional neural network (CNN) and long short-term memory (LSTM) network for river flow prediction. The hybridization enables accurate identification of the spatial and temporal features in precipitation. A shortcut layer is used as an additional channel of passing input features through the deep network to increase feature diversity. The flows in Hun River Basin, China are predicted using the trained hybrid network and are compared with the results from the Soil and Water Assessment Tool (SWAT) model. The results demonstrate the learning efficiency of the hybrid network is greatly affected by its structure and parameters, including the number of convolutional layers and LSTM cell layers, the step size of pooling and training data size. Further, the shortcut layer can effectively solve the diversity reduction problem in a deep network. The hybrid network is shown to have a similar predictive performance to SWAT but is superior in wet seasons due to its nonlinear learning ability. This study shows that the hybrid network has great promise in learning nonlinear and high spatio-temporal variability in river flow forecasting. HIGHLIGHTS: Developed a hybrid convolutional neural network and long short-term memory network (CNN-LSTM) for hydrological process prediction. The performances of the network structures and the effects of shortcut layers are evaluated separately. CNN-LSTM has good predictive accuracy compared to Soil and Water Assessment Tool (SWAT) model.