Research on Water Level Prediction on CEEMDAN-GRU Model under the IMFs Recombination

This research proposes a model bases on the median filter- complete ensemble empirical mode decomposition with adaptive noise and gated recurrent unit network model (CEEMDAN-GRU) under IMFs recombination to predict the water level. We firstly use the median filtering method for data preprocessing and apply CEEMDAN method to decompose the historical water level sequence. Then we obtain 6 IMFs (Intrinsic mode function), recombine IMF1-IMF4 into high-frequency IMFs(H-F) and take IMF5-IMF6 as low-frequency IMFs by t-test method. This research also proposes a method of recombining the IMFs in the high-frequency IMFs to optimize the H-F. On this basis, the optimized H-F, low-frequency IMFs and residual are respectively predicted by the GRU neural network model, and finally the prediction results of the three IMFs are superimposed in equal proportions. Taking the monitoring Hong Ze Lake as an example, the performance of the optimized CEEMDAN-GRU model in the test set is increased by 2.06% and the RMSE increased by 13.86% MAE increased by 9.11 %, and MAPE increased by 11.98% compared with the unoptimized full-IMFs prediction model. Meanwhile, we compared the optimized CEEMDAN-GRU with CEEMDAN-LSTM and LSTM for further investigation. The results show that: the optimized CEEMDAN-GRU has stronger predictive performance compared to the other two models.