Short-term prediction of photovoltaic power generation for small sample based on generative adversarial networks

Accurate prediction of photovoltaic (PV) power generation is the basis for rational dispatch of PV energy, but insufficient sample data of power generation from distributed PV power plants is an important factor that reduces the prediction accuracy. Expanding PV power generation samples using data enhancement algorithms is an important method to solve the problem of small-sample PV power generation prediction, so a data enhancement method based on Wasserstein’s adversarial generative network (WGAN) algorithm and a fusion algorithm for small-sample PV power generation prediction are proposed. The WGAN algorithm is trained using the source domain dataset, and a deep migration learning algorithm is introduced to migrate and optimize its training parameters in order to complete the effective expansion of the small sample dataset. Using the fusion model, local features are extracted by bi-directional temporal convolutional network (BiTCN), combined with bi-directional gated recurrent unit (BiGRU) to model the long-range dependence of the PV power generation to realize multi-scale spatio-temporal feature mining. The experimental results using distributed PV dataset in a certain area of Anhui show that the method improves the accuracy of PV power generation prediction for small samples, which is of practical significance for power generation prediction of new PV power plants.