Abstract: Aiming at the nonlinear and non-stationary characteristics of runoff sequences, we develop a combined model of extreme-point symmetric mode decomposition (ESMD) and Elman neural network, and apply it to annual and monthly runoff forecasts at eight stations in the upper reaches of the Yangtze River. First, ESMD is used to decompose a runoff sequence into modal components and trend remainders; then, the Elman neural network model is used to predict each of the stationary sequences; lastly, final prediction results are obtained by adding and reconstruction. The results show this combined model has forecast accuracy higher than that of a single model. Compared with the ESMD-BP neural network combination model, for annual runoff forecasts, it has an average reduction of 3.6% in mean absolute percentage error (MAPE) and 7.8% in root mean square error (RMSE), and an average increase of 5.0% in determination coefficient for the eight stations; while for monthly runoff forecasts, the MAPE is decreased by an average of 3.0% and the RMSE decreased by an average of 2.8%. Our combined model, characterized by decomposition-prediction-reconstruction, improves prediction accuracy.

Key words: extreme-point symmetric mode decomposition, Elman neural network, time scale, runoff forecast, non-stationary series, upper reaches of Yangtze Rive