With the popularization of user side distributed photovoltaic power generation equipment, higher requirements are put forward for distributed photovoltaic output forecasting and regulation technology. To solve the problems of low generalization ability and high sample dependence of traditional photovoltaic output forecasting methods, a user distributed short-term photovoltaic forecasting model based on meteorological coupling characteristics and improved XGBoost algorithm is proposed. Mutual information and principal component analysis are used to select characteristics and reduce dimensions to obtain highly correlated and decoupled meteorological characteristic variables. The parallel integration of XGBoost forecasting model based on Bagging algorithm improves the model generalization ability. An evaluation index of combined forecasting accuracy based on mean absolute error (MAE） and mean arctangent absolute percentage error (MAAPE） is proposed. In the example analysis, the mean MAE of the model in this paper is 6.934kW, and the mean MAAPE is 16.73%. The relative error is less than 10% in more than half of cases. Compared with the traditional BP neural network or random forest forecasting model, forecasting accuracy is greatly improved and has good practical application ability.