A K-means clustering algorithm was proposed and a conditional Wasserstein generative adversarial network with gradient penalty（CWGAN-GP）to address the problem of imbalanced photovoltaic generation data caused by the low occurrence probability of extreme weather. A prediction approach combining bidirectional long short-term memory（BFLSTM）with convolutional neural network was introduced and incorporating channel attention mechanism to enhance the PV power prediction performance by integrating spatio-temporal features and dynamically adjusting the importance of feature channels. Firstly，correlation analysis and K-means algorithm were utilized to select and label various environmental factors. Then，extreme weather labels with fewer samples after clustering were selected，and CWGAN-GP was used for data augmentation.Finally，the augmented dataset was used to train the CNN-SE-BiLSTM prediction model for PV power prediction under extreme weather conditions.Simulation modeling was conducted using data from a certain PV power station，and the results demonstrate that augmenting the original extreme weather training set with CGAN-GP helps improve the prediction accuracy of the model. Moreover，CNN-SE-BiLSTM shows higher prediction accuracy among five weather categories compared to other traditional models，indicating that the proposed method is suitable for ultra-short-term photovoltaic power prediction.