Knowledge-Integrated GAN Model for Stochastic Time-Series Simulation of Year-Round Weather for Photovoltaic Integration Analysis

The temporal stochastic simulation of weather has become crucial for stochastic production simulation in power systems with a high proportion of photovoltaic (PV) generation. Generative artificial intelligence (AI) has become the central technology for the stochastic simulation of weather sequences. This research presents a novel approach for the stochastic simulation of annual weather scenarios driven by data and knowledge fusion, in light of the uncontrollability in content production by generative AI technologies. The fusion work encompasses the establishment of the monthly weather generative adversarial network (MWGAN), a generative scenario quality enhancement approach based on statistical probability knowledge, and a suite of statistical machine learning methods for evaluating generated weather scenarios. Weather data from a location in Guangdong, China, over 48 years, was utilized to validate the proposed method for stochastic simulation of annual weather scenarios. By comparing the proposed model to five cutting-edge generative adversarial networks (GAN) and two shallow statistical machine learning models, the superior performance of the proposed model in simulating the temporality and probability of weather is confirmed. The Wasserstein distance is reduced by an average of 46% when compared to the best GAN used for comparison (the smaller the distance, the more accurate the simulation of uncertainty). The effectiveness of the proposed weather sequence stochastic simulation method for probabilistic load flow calculation is verified using an actual distribution network in Guangdong, China.