Accurate prediction and simulation of hydrological procedures are necessary to manage water resources. Complex hydrological systems are dependent to numerous variables such as spatial and temporal variables, different precipitation patterns and model parameters. Lack of knowledge in perception of water balance lead to uncertainty which increase in hydrological procedures such as rainfall-runoff recently. The proper uncertainty analysis must be considered in which the whole sources of uncertainty have been covered. In this study, hydrological SWAT model is incorporated to simulate rainfall-runoff in Chelgerd sub-basin and regression Support Vector Machine (SVR) is used to simulate inflow to Zayandehroud dam. To develop appropriate model with high accuracy, uncertainty is analyzed, decomposed and reduced. Using Evidence theory, the runoff uncertainty of physical SWAT model is analyzed and measured with the value of 19 percent while it is increased to 33 percent for empirical SVR model. Variance decomposition method is used to determined shares of uncertainty sources for SWAT model. Precipitation data, temperature data and model parameters as uncertainty sources with values of 40, 32 and 28 percent have the most effect in producing uncertainty, respectively. Data assimilation EnKF technique is implemented to improve physical and empirical models performance and decrease uncertainty. The model output data are evaluated by assessment criteria including; Nash-Sutcliff, RMSE, R 2 and Strife index which indicates the models performance improvement.