A hybrid method for day-ahead photovoltaic power forecasting based on generative adversarial network combined with convolutional autoencoder

Abstract

Photovoltaic (PV) generation has high impact on the decarbonization pathways of power systems. Accuracy of day-ahead PV power forecasting has become crucial in the operation and control of power system with high PV penetration. This paper develops a hybrid approach based on generative adversarial network (GAN) combined with convolutional autoencoder (CAE) to improve PV power forecasting accuracy. Self-organizing map method is first utilized as data pre-processing to classify target days into different weather types based on solar irradiance. With the ability of GAN to reduce the burden of loss and the advantages of CAE to extract multi-scale effective features from the weather and PV power, PV power forecasting model consisting of GAN and CAE is proposed. The developed method has been tested on a real dataset in a Chinese PV station and compared with base reference PV forecasting methods. Numerical testing results demonstrate the effectiveness of our method with high accuracy.