Abstract: Aiming at the problem of low prediction accuracy caused by the lack of original data of new photovoltaic power station, we propose a short-term photovoltaic power prediction method based on Wasserstein generative adversarial network with gradient penalty (WGAN-GP) and improved long-term and short-term memory network. First, WGAN-GP is used to learn the distribution law of original photovoltaic data, then the generator in WGAN-GP generates new samples of high quality similar to the original data, so as to enhance the training set data. Secondly, the crisscross optimization algorithm (CSO) is used to optimize the fully connected layer parameters of the long short-term memory (LSTM) network, and the LSTM-CSO combination model is constructed to predict the photovoltaic power. The simulation model is established with the data of a photovoltaic power station in Australia. The experimental results show that the prediction accuracy of the model can be effectively improved by using the sample training prediction model after data enhancement. The greater the proportion of the original training set data expansion is, the higher the accuracy of the prediction model for photovoltaic power prediction will be. At the same time, LSTM-CSO has higher prediction accuracy than LSTM in different meteorological days of each season type. A spring test set is taken as an example, and results show that the root mean square error of LSTM-CSO model in sunny, cloudy and rainy days in spring is reduced by 5.62%, 3.44% and 10.44% respectively compared with LSTM model.