Abstract: There are a large number of redundant and irrelevant features in multi-dimensional numerical weather prediction (NWP) data when forecasting short-term photovoltaic power generation, which not only affect the accuracy of the forecast, but also increase the complexity of the model. Therefore, a short-term photovoltaic power prediction model based on multi-feature analysis and extraction was proposed. After using K-means++ clustering to select historical data with similar weather types to the predicted date as training sample, the unstable characteristic data were processed by using the feature of filtering with first difference, and new features were constructed at the same time. Factor analysis method was introduced to extract effective features considering the correlation between features and output power, and the common factors which are far less than the number of features were used as the input data of the prediction model. Finally, XGBoost was used to predict the photovoltaic power. The simulation results of a photovoltaic power station show that the root mean square errors (RMSE) of the proposed prediction model are 5.33%, 6.13% and 9.5% in clear day, clear to overcast day and rainy day, respectively. And the prediction accuracy of the proposed model can be improved by 3%~10% compared with the traditional method in non-sunny days. The research results can provide a reference for photovoltaic power prediction under complex weather.