To enhance the accuracy of PV(photovoltaic) power output prediction by effectively capturing local temporal variations, we propose a PV generation forecasting model that integrates SSIM(structural similarity) algorithm with the Informer model for efficient long-term sequence prediction. As a case study, daily meteorological data from the Yanpeng PV Station in Yunnan Province were converted into vectors. These vectors were then transformed into Gram matrices and subsequently into daily meteorological pixel maps. SSIM algorithm was applied to identify and filter days exhibiting similarity to the target prediction day. On this basis, feature selection of meteorological elements influencing PV output was carried out. The Informer model was then employed to construct the PV output prediction model, enabling forecasts across multiple time intervals. The results show that the graph similarity day method effectively identifies similar days for the target prediction day, and the constructed Informer PV output prediction model demonstrates robust predictive performance under different weather conditions. Compared to traditional forecasting methods, the root mean square error on sunny days is reduced to 0.66, and the daily accuracy improves by 1.63% to 3.92%.