The correlation among input variables in power system affects the accuracy of probabilistic load flow (PLF) calculation. A PLF algorithm is proposed based on third-order polynomial normal transformation (TPNT) combining cumulants and Gram-Charlier expansion. The method that transforms a multivariate non-normal dependent random variables group into a multivariate standard normal independent one is used to calculate input variables and power flow based on correlation coefficient matrix and draw cumulative distribution curves of node status variables and load flow. The simulation results of the IEEE14-bus system with several correlative wind farms show that the proposed method is effective and accurate.