River streamflow has gradually developed into a non-stationary and non-linear complex process under the influences of climate change and human interferences. A major technical issue associated with this environmental changing is how to predict accurately the future change in river runoff. At present, a new prediction system, namely decomposition-prediction-reconstruction, has been widely used in the mid- and long-term prediction of runoff series. Its prediction efficiency, however, is unsatisfactory due to large errors in its prediction of high-frequency components that are decomposed using the empirical mode decomposition (EMD). To forecast the high-frequency components in the runoff at the Dingjiagou gauge station on the Wuding River, this study has adopted three approaches: the radial basis function (RBF)neural network, autoregressive (AR)model, and mean generating function (MGF)model. Based on these models, a comprehensive prediction was also made using the Bayesian model averaging (BMA)method. In this paper, we confirm the accuracy of BMA and demonstrate its effective control on the prediction error of high-frequency components through a comparison of its errors with those of the three single models. Thus, this study comes to a conclusion that the BMA method is an effective approach to improve the prediction accuracy of runoff series and would provide valuable references for similar issues in forecasting non-stationary time series.