Abstract: In order to make full use of historical information, maximize the effect of the model, and improve the accuracy of photovoltaic power prediction, we proposed a photovoltaic power prediction method based on graph similarity days and particle swarm optimization-extreme gradient boosting tree (PSO-XGBoost). The daily vector composed of weather features was converted into a Gram matrix to fully explore the relationship between the various vectors, and then the Ram matrix was converted into an image. Moreover, the structural similarity algorithm (SSIM) was adopted to find the best similar historical days, and the photovoltaic power of the historical day, the irradiance, temperature, and humidity of the day to be predicted were selected as the input variables of the extreme gradient boosting tree. In order to give full play to the predictive ability of the model, the particle swarm algorithm was adopted to optimize the extreme gradient boosting tree to determine optimal hyperparameters, and the predicted value of photovoltaic power was finally output in each period. The actual data of photovoltaic power plants were used for verification. The results show that, compared with the unimproved XGBoost model, the method proposed in this paper can be employed to reduce the root mean square error (RMSE) by 31.6% and the mean absolute error (MAE) by 31.6% in sunny days; the RMSE is reduced by 24.1% and the MAE by 40% in cloudy weather; RMSE is reduced by 25% and MAE is reduced by 38.5% under rainy weather, which effectively improves the prediction accuracy and generalization ability of the model.