A method of PV array short-term power prediction is proposed based on improved similarity of fuzzy clustering algorithm. First, the weights of the meteorological factors are obtained through path analysis. An improved similarity is constructed integrating weighted similarity coefficients and weighted distance coefficient according to the weight of each factor. Second, the history day samples are divided into several categories by making fuzzy similarity matrix. The classification of the forecasting day is got by pattern recognition. Then the BP neural network prediction model using differential evolution algorithm to optimize BP neural network weights and threshold value is constructed based on the history data of the classification and the meteorological factors of the forecasting day. Experimental results demonstrate that the model has higher prediction accuracy compared with the traditional prediction model based on similar day selected. It is conducive to the operation of PV system operation and its security economic dispatch.