Data Generation Scheme for Photovoltaic Power Forecasting Using Wasserstein Gan with Gradient Penalty Combined with Autoencoder and Regression Models

Machine learning and deep learning (DL)-based forecasting models have shown excellent predictive performances, but they require a large amount of data for model construction. Insufficient data can be augmented using generative adversarial networks (GANs), but these are not effective for generating tabular data. In this paper, we propose a novel data generation scheme that can generate tabular data for photovoltaic power forecasting (PVPF). The proposed scheme consists of the Wasserstein GAN with gradient penalty (WGAN-GP), autoencoder (AE), and regression model. AE guides the WGAN-GP to generate input variables similar to the real data, and the regression model guides the WGAN-GP to generate output variables that well reflect the relationship with the input variables. We conducted extensive comparative experiments with various GAN-based models on different datasets to verify the effectiveness of the proposed scheme. Experimental results show that the proposed scheme generates data similar to real data compared to other models and, as a result, improves the performance of PVPF models. Especially the deep neural network showed 62% and 70% improvements in mean absolute error and root mean squared error, respectively, when using the data generated through the proposed scheme, indicating the effectiveness of the proposed scheme in DL-based forecasting models