A Multivariate and Multistage Streamflow Prediction Model Based on Signal Decomposition Techniques with Deep Learning

Accurate streamflow forecast plays an important role in the flood control operation of reservoir and water resource management, and how to build a prediction model with high accuracy is a research hotspot. In order to improve the prediction accuracy, this paper establishes a multivariate and multistage streamflow prediction model based on the combination of mode decomposition and deep learning models, including the variational mode decomposition (VMD), improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMD), convolutional neural network (CNN), and long short-term memory (LSTM). Taking the Wei River Basin (WRB) of China as an example, streamflow in six hydrometric stations from the main stream of the WRB is used to validate the model. The results show that the proposed model have good prediction skills, and the prediction results of multistage models are better than single-stage models; however, the most complex models do not have the best results. The VMD also preformed better prediction skills than ICEEMD, and the optimal model was VCL (VMD-CNN-LSTM). The root-mean-square error, peak percentage of threshold statistic, and Nash-Sutcliffe efficiency coefficient of VCL at Huaxian station are 43.82 m3/s, 10.02%, and 0.94, respectively. The models proposed in this paper are suitable for streamflow forecasting, which provide reference for streamflow forecasting and sustainable watershed management.