Abstract: [Objectives]　To make full use of historical information, maximize the optimization of model effect and improve the accuracy of PV power prediction. [Methods]　A short-term PV power prediction model considering weather-coupled similar days is proposed. Firstly, fuzzy C-means clustering (FCM) is used to divide the dataset into different weather types, and the correlation weight factors are calculated according to the affiliation and features selection of the day to be predicted for each weather type, and the similarity of the historical days is calculated by grey relation analysis (GRA), and similar days with weather coupling are selected. The similarity of historical days is calculated by combining with gray relation analysis (GRA) and selecting the similar day sets with weather coupling, and the similar days are decomposed into modal components with different frequencies by variational mode decomposition (VMD) to realize further denoising; Secondly, in order to fully utilize the nonlinear fitting ability of the model, the red-tailed hawk algorithm (RTHA) is used to optimize the hyperparameters of the bidirectional long short-term memory network (BiLSTM) and construct the RTHA-BiLSTM model to predict the modal components. Finally, the prediction results of each modal component are superimposed to obtain the final prediction results. [Results]　Simulation experiments are conducted with actual data from an power station in Jiangsu, China, and the proposed method can reduce the root mean square error by 11.1%, 6.5%, and 13.9% under sunny, cloudy, and rainy days, respectively, compared with the combined model without similar days. [Conclusions]　The effectiveness of model training through similar days selected by the method to improve the prediction accuracy is verified.

Key words: photovoltaic power generation, photovoltaic power prediction, fuzzy c-means, grey relation analysis, similar days, red-tailed hawk algorithm