Abstract: An improved statistical scale-up modeling method for PV cluster output forecasting based on empirical orthogonal function（EOF）and density-base spatial clustering of applications with noise（DBSCAN）is proposed. To solve the problems of inconsistent output characteristics of PV power stations in the subgroup of traditional statistical upscaling method,Pearson correlation coefficient and EOF were used to optimize the feature space,and then the PV power stations in the region were divided into clusters according to DBSCAN model,so as to enhance the consistency of the output characteristics of PV power stations after clustering. Aiming at the problems of extracting and predicting the dynamic characteristics of time series of weight coefficients of days to be forecasted,a similar day selection algorithm based on dynamic time warping（DTW）was proposed. Finally,the gate recurrent unit（GRU）neural network model is built to predict the power output of PV power stations. Experimental results show that the mean absolute percentage error（MAPE）,root mean square error（RMSE）and mean square error（MSE）of the cluster forecasting method are about 6.33%,13.93 and194.25 kW. The effectiveness and accuracy of the proposed method are verified by the measured data of actual PV power stations.