Daily Runoff Forecasting Using Ensemble Empirical Mode Decomposition and Long Short-Term Memory

Hydrological series data are non-stationary and nonlinear. However, certain data-driven forecasting methods assume that streamflow series are stable, which contradicts reality and causes the simulated value to deviate from the observed one. Ensemble empirical mode decomposition (EEMD) was employed in this study to decompose runoff series into a trend with several stationary components. To verify the effect of EEMD on LSTM method accuracy, two models, LSTM and EEMD-LSTM, were established to simulate the daily inflow data of the Three Gorges Reservoir from 2005 to 2017. The GM (Group by Month)-EEMD-LSTM model, which divides the data from 2005 to 2017 by month, was used as a control to explore the simulation characteristics of the EEMD-LSTM model for different monthly series. The input of each model was the variable with the highest mutual information number with the target sequence. It was selected from the predictor set through the mutual information method. The results demonstrate that decomposing the streamflow series through EEMD can effectively improve the Nash coefficient and mean absolute percentage error values of the LSTM model. The water balance accuracy was slightly reduced at the annual scale, and it improved at the monthly scale after decomposition. The LSTM model did not exhibit any clear rules for the simulation results of each month. However, the EEMD-LSTM model exhibited superior simulation results for high-flow months, while the GM-EEMD-LSTM model performed more optimally in the low-flow months. The monthly results of the two models were significantly improved compared to those of the undecomposed LSTM models. In practical applications, the two models can be combined for a more optimal prediction.