Bi-LSTM and Style-Based Generative Adversarial Network for Stochastic Simulation of Photovoltaic Power Generation Based on Weather

Abstract

Renewable energy systems represented by photovoltaic power generation systems have strong weather sensitivity, and weather factors are important reasons for the high uncertainty. This paper proposes a power flow analysis method based on a weather scenario generation model, which generates a massive amount of hourly weather scenarios with real probability characteristics, temporal features, and diverse variations throughout the year to fully analyse power flow. The proposed model, named “BL-StyleGAN”, combines generative adversarial networks (GAN), bidirectional long short-term memory networks (Bi-LSTM), and style transfer strategies to accurately learn the probability, diversity, and temporality of real temperature, direct radiation, and diffuse radiation data. Compared with other deep generative models based on GANs, the proposed model has significant advantages in learning the temporal and diverse characteristics of weather data. We applied this weather scenario generation model to the power grid in a certain location in Guangdong Province, China for power flow analysis. Experimental results demonstrate that the proposed BL-StyleGAN model surpasses other deep generative models in the accuracy of power flow analysis.