Research on ultra-short-term power prediction for photovoltaic modules based on improved generative adversarial networks

Aiming at the problem of inaccurate power generation efficiency prediction due to the strong randomness and volatility of photovoltaic power output, this paper proposes an improved autoregressive feedforward neural model based on Generative Adversarial Networks (GAN) to predict the ultra-short-term power of photovoltaic modules. Firstly, identifying the internal parameters of photovoltaic modules by the improved Coati Optimization Algorithm(COA) with Token-Ring integrated with Sequential Least Squares Programming(SLSQP), and selecting the internal and external factors that have tremendous impact on the output power based on Grey Relation Analysis(GRA); then, considering the global time correlation and long-term dependence of the data, a Generator Network with Self-Attention mechanism module as encoder and Long Short Term Memory(LSTM) module as decoder and a Discriminator Network are constructed for recursive multi-step training. At the same time, building a pre-trained SVM to construct Hinge loss, Discriminator loss and Generator loss to generate accurate PV power sequences. The results show that the proposed method can improve the prediction accuracy of output power of photovoltaic module and fault identification accuracy effectively and has certain engineering practicality.