Annual runoff forecasting based on multi-model information fusion and residual error correction in the Ganjiang River Basin

Abstract

Accurate forecasting of annual runoff time series is of great significance for water resources planning and management. However, considering that the number of forecasting factors is numerous, a single forecasting model has certain limitations and a runoff time series consists of complex nonlinear and nonstationary characteristics, which make the runoff forecasting difficult. Aimed at improving the prediction accuracy of annual runoff time series, the principal components analysis (PCA) method is adopted to reduce the complexity of forecasting factors, and a modified coupling forecasting model based on multiple linear regression (MLR), back propagation neural network (BPNN), Elman neural network (ENN), and particle swarm optimization-support vector machine for regression (PSO-SVR) is proposed and applied in the Dongbei Hydrological Station in the Ganjiang River Basin. Firstly, from two conventional factors (i.e., rainfall, runoff) and 130 atmospheric circulation indexes (i.e., 88 atmospheric circulation indexes, 26 sea temperature indexes, 16 other indexes), principal components generated by linear mapping are screened as forecasting factors. Then, based on above forecasting factors, four forecasting models including MLR, BPNN, ENN, and PSO-SVR are developed to predict annual runoff time series. Subsequently, a coupling model composed of BPNN, ENN, and PSO-SVR is constructed by means of a multi-model information fusion taking three hydrological years (i.e., wet year, normal year, dry year) into consideration. Finally, according to residual error correction, a modified coupling forecasting model is introduced so as to further improve the accuracy of the predicted annual runoff time series in the verification period.