Since quintessential probabilistic photovoltaic (PV) models like Weibull and Beta distribution are not able to reflect accurately the changes of PV station's output characteristics in different scenarios, a data-driven scenario generation model for distribution system is proposed. The model which is based on the algorithm of local density clustering (LDC), takes into account the correlation of multiple PV stations and the influence of external conditions, thus realizing the accurate modeling of multiple PV stations' output power. Firstly, LDC is used to classify historical data into several clusters. Secondly, kernel density function and copula function are utilized to build joint probabilistic density functions (PDF) in each cluster, and the data of joint PDF are processed by applying Latin Hypercube Sampling (LHS). Then, the samples are used for probabilistic power flow calculation to set up the model which estimates operational condition of distribution network. Finally, the stability of model is analyzed, and influence of sampling size on model is also discussed. It is shown that the proposed method in the paper improves the accuracy of modelling for PV station's output in distribution network, and it also reduces the error of probabilistic power flow calculation.