There is a problem of low accuracy of photovoltaic power prediction due to lack of historical data of non-sunny weather types. Thus a short-term photovoltaic power prediction model based on an Wasserstein generative adversarial network with gradient penalty (WGAN-GP) and CNN-LSTM-Attention is proposed. First, the K-means++ clustering algorithm is used to divide the historical photovoltaic data into several weather types, and WGAN-GP is used to generate new high-quality samples that conform to the distribution law of each weather type to enhance the training set data. Second, this paper combines the advantages of both a convolutional neural network (CNN) for feature extraction and of long short-term memory (LSTM) for time series prediction to improve the ability of prediction models to learn the long-term mapping relationship between photovoltaic power and weather data. In addition, the attention mechanism (Attention) is introduced to compensate for the deficiency of LSTM in retaining key information when the input sequence is long. Experimental results show that the prediction accuracy is effectively improved after each type of weather sample expansion based on WGAN-GP. Compared with the three classical prediction models, the proposed CNN-LSTM-Attention model has higher prediction accuracy.