Short-term photovoltaic power prediction provides a reliable basis for photovoltaic participation in intraday scheduling. However, the inevitable extreme weather poses a serious challenge to the stability of photovoltaic output, and extreme fluctuations lead to a decrease in the accuracy of short-term power prediction. In this regard, this article proposes a short-term prediction method of photovoltaic cluster power that considers photovoltaic output clustering and improves MSE loss. Firstly, explore the correlation between photovoltaic output and meteorological processes, using deep attentional embedded graph clustering (DAEGC), cluster and partition photovoltaic output types to obtain historical samples with different fluctuation evolution trends. Secondly, based on the diurnal NWP meteorological information, the evolution types of photovoltaic fluctuations and power clustering results are identified to divide photovoltaic clusters. Finally, a photovoltaic power prediction model that integrates an improved weighted MSE loss function is adopted for short-term prediction of photovoltaic cluster power. The effectiveness of the proposed method was validated using photovoltaic cluster data from Inner Mongolia, China. The results show that the prediction method using dynamic division of clusters combined with improved MSE loss function effectively improves the PV cluster power prediction accuracy, and reduces the NRMSE by 9.26% and the NMAE by 4.30% on average compared with the direct overall prediction method.