Hybrid CNN-LSTM models for river flow prediction

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