Abstract: Mid-long term inflow prediction is a critical prerequisite and complicated issue for reservoir operation. This paper presents a wavelet decomposition parameter optimized support vector machine (WD-SVM-PSO) model based on previous studies of the data driven prediction theory, including historical inflows frequency division pre-process, classification based training, parameter optimization and cross validation. Its performance is optimized in terms of dataset refining, model parameters calibration, and training mechanism. Application to the annual inflows of the Xianghongdian reservoir in the Huai River basin during 1959-2014 shows that 93% of its predictions are acceptable due to its better generalization performance and it can significantly reduce the overfitting. And the controlled trial simulation reveals the effect of three key elements, ranked from top to down: data set pre-process, prediction model, model parameters. This study helps analyze and improve data driven prediction models and their accuracy and reliability of inflow prediction.