To improve the accuracy of the mediumlong term runoff forecasting of the whole watershed, a combined method integrating information entropy and improved extreme learning machine (ELM) is proposed. Firstly, the comprehensive runoff index is constructed based on the controlling area of different hydrological stations to characterize the abundance and drought in the basin. Secondly, the partial mutual information (PMI) approach is applied to calculate the correlation between multiple factors and the comprehensive runoff index for inputs of the forecasting model. Finally, an improved ELM model by combining Kfold cross validation with improved particle swarm optimization (IPSO) is proposed to optimize parameters of ELM, together denoted as IPSOELM, for mediumlong term forecasting. In a case study of the Yalong River basin, the proposed model was compared with classical forecasting models, i.e., backpropagation neural networks (BPNN), support vector machines (SVM), ELM and PSOELM models. The results shows that the performance evaluation indexes of the proposed model perform much better than the above four datadriven models in terms of Emape, Ermse, Edc, Eqr, and Ere. The five forecasting models demonstrate better results for the D1 dataset compared to that of the D2 dataset.