A robust method for non-stationary streamflow prediction based on improved EMD-SVM model

Monthly streamflow prediction can offer important information for optimal management of water resources, flood mitigation, and drought warning. The semi-humid and semi-arid Wei River Basin in China was selected as a case study. In this study, a modified empirical mode decomposition support vector machine (M-EMDSVM) model was proposed to improve monthly streamflow prediction accuracy. The accuracy was improved by introducing polynomial fitting to amend the error caused by the boundary effect existing in the counting process of empirical mode decomposition (EMD). Meanwhile, the computational process of the EMD was analyzed to confirm the decomposition method for the EMD. The root mean square errors, mean absolute error, mean absolute percentage error and Nash-Sutcliffe efficiency coefficient were adopted as the standards to evaluate the performance of the artificial neural network (ANN), SVM, WA-SVM, EMD-SVM, and M-EMDSVM models. Meanwhile, the performance of the M-EMDSVM model with different lengths of training dataset was compared and analyzed. Moreover, the monthly streamflow series with various non-stationary levels were simulated to investigate the prediction capacity of the M-EMDSVM model. Results indicated that: (1) the ANN model had the worst performance among the five models at all stations, whereas the EMD-SVM model performed better than the WA-SVM with better metric values; (2) for strong non-stationary series, the performance of the M-EMDSVM model was superior to the EMD-SVM; (3) for weak non-stationary series, the performance of the M-EMDSVM model was similar with the EMD-SVM. Generally, the findings of this study showed that more accurate prediction of strong non-stationary streamflow could be achieved using the proposed modified EMD-SVM model than single SVM model.