The connection of a high penetration of distributed photovoltaic (DPV) increases power system risk. Given the characteristics of complicated forms of power distribution, first a self-adaptive Gaussian mixture model based on the improved affinity propagation clustering is proposed to optimize the probability fitting iteration process of DPV joint output. Then a Nataf transformation method is given based on an improved third-order polynomial normal estimation process, and combined with a cumulant method and a Cornish-Fisher series, the probabilistic power flow calculation in conditions of correlation variables of DPV is realized. Finally, indicators of voltage limit and line overload are adopted to calculate the operational risk of the power system. Scenarios of different levels of DPV penetration are simulated using a modified IEEE 14-node power system. Compared with the Monte Carlo simulation, the results show that the method proposed has a higher accuracy on the probability distribution calculation of the status variables of the power grid. It is shown that the evaluation results can effectively reflect the effects of different penetration levels of DPV on the level of power system risk.