The mid- and long-term runoff prediction with satisfactory accuracy plays an important role as basic information in water resources planning & management and optimal operation of water conservancy projects. Combination and bias reduction are two common post-processing approaches in runoff forecast. Applying them in turn, considering the complicated non-stationary and nonlinear characteristics of runoff, a new mid- and long-term runoff prediction method by connecting time-varying weight combination and Bayesian correction is proposed. This method was used to study the annual and monthly inflow prediction of the Longjiang Reservoir in Yunnan Province. The results show that time-varying weight combination balances the performance difference of the established random forest (RF) and support vector machine (SVM) models in the modeling period and the test period. As a consequence of Bayesian correction, the prediction metrics are close to or better than the best of the predictions of the two individual stages. The proportion of correctly classified hydrological year type reaches 77.2% by employing the forecasted annual runoff, and the Nash-Sutcliffe efficiency coefficient of predicted monthly runoff series is over 0.90. Overall, the method put forward in this study has achieved positive effects in improving the forecast performance.