It is crucial for the real-time scheduling operation of the power system to improve the refined photovoltatic power forecasting technology. It depends not only on the superiority and inferiority of the predictive model, but also relies on the similarity between the training sample day and the forecast day. A novel method based on MIE-LSTM is proposed for short-term photovoltaic power forecasting. The correlation metrics is established based on Mutual Information Entropy (MIE), the MIE between photovoltaic power and meteorological factors is calculated to reduce the dimension of high-dimensional meteorological data. Then, the weighted mutual information entropy between the multi-dimensional meteorological factors of historical day and those of predicted day is used to screen out similar day samples. Finally, the mapping relationship between meteorological factors and photovoltaic output is established via training Long-Short Term Memory (LSTM) neural network prediction model. Through forecasting and analyzing the power generation of a photovoltaic power station under different weather types, it is verified that the new combination method can achieve ideal forecasting precision. This work is supported by National Natural Science Foundation of China (No. 51607009), Beijing Natural Science Foundation (No. 3172015), and Scientific Research Project of Beijing Education Council (No. KM201911232016).