Abstract: In order to improve the accuracy of short-term prediction results for photovoltaic power generation, a photovoltaic power generation short-term prediction method based on the long short term memory（LSTM） neural network optimized by improved grey wolf optimization（IGWO） algorithm is proposed. The similar days are searched for by using cosine similarity and the feature quantity and training set for photovoltaic power generation prediction are determined. The GWO algorithm is improved by using nonlinear convergence factors and differential evolution strategies. An IGWO algorithm with better convergence performance is obtained. The hyperparameters of LSTM are optimized by using the IGWO algorithm, and a short-term prediction model for photovoltaic power generation based on IGWO-LSTM is established. Simulation analysis is conducted by using the operating data of a small photovoltaic power plant. The results show that the root mean square error of the IGWO-LSTM model for predicting photovoltaic power in sunny, cloudy, and rainy weather is 2. 11 kW, 2. 48 kW and 2. 74 kW, respectively,and the The average relative error is 3. 43%, 4. 81% and 6. 33%, respectively. The prediction effect is superior to other methods, verifying the practicality and effectiveness of the proposed method.