PV penetration is steadily increasing with the large-scale utilization of new energy sources. Accurate PV power prediction can bring more benefits to grid enterprises. Based on this, a random forest prediction model with multi-feature analysis extraction is proposed for ultra-short- term PV power prediction.Firstly, the collected PV data is pre-processed to clean up the missing and duplicate values. Then, correlation analysis is performed on the influencing factors and factors with strong correlation are selected. Next, feature engineering is performed on the screened factors and the processed feature vector is used as input of the prediction model. Finally, the random forest prediction model is built and compared with BP, RBF and MLP models. Empirical results show that the model proposed has better fit and higher prediction accuracy, which is of certain guidance for PV prediction work.