Abstract: Long-term runoff forecasting for the reservoir is of great significance for studying the hydrological regime and guiding the regulations. In this paper, the mean inflow of annual, flood and dry seasons of the Longjiang Reservoir are selected as forecast elements. Random Forest (RF) is utilized to filter key predictors from circulation indices, sea temperature, air pressure and previous monthly runoff. Afterwards, models based on RF and support vector machine (SVM), which are calibrated using particle swarm optimization algorithm combined with cross-validation, are established to predict the inflow of the Longjiang Reservoir. Results show that climate factors in the north-central and western Pacific have generally implemented a greater influence on prediction, while the effect of the pre-monthly runoff is relatively low, however it can be comparable to some climate factors when used to predict runoff in the dry season. The average accuracy of RF and SVM is generally satisfactory, with the qualification rate of simulation and forecast exceeding 85% and the average absolute percentage error less than 15%. SVM shows stronger generalization ability compared to RF in this study case, while the ability of both models in predicting partial extreme inflow remains to be improved.