Extreme events pose a significant threat to the operational security of distribution networks, making accurate and efficient probabilistic risk assessment crucial. To address the low efficiency of traditional Monte Carlo simulation (MCS) and the insufficiency accuracy of existing surrogate models in low-probability regions, this paper proposes a risk assessment method that balances computational efficiency and accuracy. First, a non-parametric probabilistic model of distributed generation output is built using kernel density estimation (KDE), while dynamic probability modeling of equipment states is performed using exponential distributions and weather factors. Next, the input samples are divided into branch fault and non-fault categories. For non-fault samples, a polynomial chaos expansion (PCE) surrogate model is used for rapid probabilistic power flow calculation. For fault samples, a topology reconfiguration model is employed to calculate optimal power flow and capture risks under topology changes. Based on this, three risk indicators of node voltage over-limit, branch power flow overload, and load loss are constructed. The weights of these indicators are determined by the subjective and objective combination weighting method of game theory, and the comprehensive risk evaluation value is obtained. Simulation results on the IEEE33 and IEEE118 node distribution systems show that the proposed method effectively handles topology uncertainties caused by extreme events, improving both the efficiency and accuracy in risk assessment.