Order to improve agricultural water precision and improve the basic particle swarm optimization (PSO) convergence performance, the paper proposed PSO-SVR multivariate agricultural water prediction model based on the adaptive mutation (Adaptive Variation, AV).It took national agricultural water consumption forecast as a case study from 2000 to 2011. First of all, it selected 3 typical function test the performance of AVPSO algorithm, and compared with the basic PSO algorithm; secondly choose 4 influence factors such as grain crops sown area, the flood area as that of agricultural water forecast. AVPSO algorithm is used to optimize the SVR penalty factor and kernel parameter, construction of AVPSO-SVR agricultural water consumption forecast model, and construct the basic PSO-SVR, GA-SVR, GA-BP and traditional BP model as a model for comparison; finally,it used the case before 8 years and after 4 years of data for the training and prediction of each model. The results showed that ①the global searching ability of AVPSO algorithm has been significantly improved,which can avoid the premature convergence problem. The AVPSO-SVR model of the relative mean absolute error of prediction of water with 4 years of agricultural instance and the maximum relative error absolute values were 0.48%, 0.78%.The prediction accuracy and generalization ability are better than those of PSO-SVR, GA-SVR, GA-BP and traditional BP model. AVPSO algorithm can effectively carry on the optimization of SVR penalty factor and kernel parameter.