Power System Security Research based on a Novel Reward-Guided Diffusion Model for Extreme Weather Forecasting
Yi‐Cheng Zhang, Dexi Xia, Fei Yuan · Journal of Physics Conference Series · 2025
Abstract Recently, the frequency of extreme weather events impacting power grids has increased, presenting significant challenges to the security, quality, and stability of electrical systems. Effectively predicting the impact of extreme weather events on distribution networks has become crucial for maintaining the safe operation of power systems. Current ensemble forecasting methods utilize numerical weather prediction to forecast weather by simulating atmospheric laws with parameterized models. These methodologies require the simultaneous operation of numerous model instances, leading to substantial computational resources and thereby increasing operational costs for power systems. Simultaneously, despite ongoing enhancements in meteorological forecasting technology, ensemble forecasting tends to produce blurry predictions, often underestimating the severity of extreme weather events. Moreover, there is a lack of extreme weather data. To address these issues, a novel Reward-Guided Diffusion (RGD) model is proposed to augment data of the extreme weather for further forecasting. By integrating the diffusion model with the reward regression model which can capture finer-grained atmospheric features of the extreme weather, the diffusion model can be finetuned by the feedback from the reward regression model in an online manner. The extreme weather data can be augmented by the proposed RGD model and further enhance the prediction for occurrence of extreme weather events more accurately at the later stage. As a result, by providing more reliable early warning information to power systems, this method strengthens the grid’s resilience to risks and ensures the secure and stable operation of electrical systems.