Abstract: The power prediction of photovoltaic power generation cluster is of great significance to the optimal dispatch of regional photovoltaic power generation. The research on photovoltaic cluster power prediction content is rarely available in the existing literature. In order to improve the power prediction accuracy of photovoltaic power station clusters, a short-term power prediction method for photovoltaic clusters based on K-means clustering is proposed. Based on the characteristics of photovoltaic power generation at the station, the clustering kind is divided, and the bias-compensation long-short-term memory (BC-LSTM) network is used for power prediction. The results show that the BC-LSTM can improve the prediction accuracy by about 0.6% compared with long short-term memory network (LSTM) network, and the cluster accumulation method can also be adopted to improve about 0.5% compared with the statistical upscaling method and the accumulation method.