When the support vector machine is used for load forecasting, the input samples of support vector machine have important effect to forecasting results. Support vector machine can study any non-linear relation, but if a group of non-distinct variables are selected as input variable set, the training time of support vector machine is lengthened and the errors become bigger. The non-linear relation of the load can be effectively explained only when a group of appropriate input variables are found. In this paper, the correlation coefficient idea is used to input variables selection of support vector machine short-term load forecasting model. The load values, which have bigger correlation coefficient with expectation output values, are chosen from effect factor sets as input variables. By mean of this method, a preferable input variables set can be gained, the correlation between the input variables and the forecasting points are bigger, and the forecasting results is more exact. The simulation results show that the method is effective.