ExDiffusion: Classifier-Guidance Diffusion Model for Extreme Load Scenario Generation With Extreme Value Theory

Extreme load scenario generation is essential for power system planning and security assessment. However, the challenges of limited data and complex external features in extreme events make it difficult for existing methods to generate realistic and diverse extreme load scenarios. To address the challenges of data scarcity and feature complexity, this paper proposes a novel scenario generation method, ExDiffusion, to generate diverse and high-quality load scenarios for both extreme and normal conditions, under limited data. This method enables effective modeling of the medium and tail behavior in time series load data by integrating the classifier-guidance diffusion model with the autoregressive model, leveraging Extreme Value Theory (EVT). Firstly, the method introduces an extremeness measure to synthesize load-level labels and augment the data via the diffusion model. A classifier-guidance diffusion model is then developed to reduce the dimensionality of the input features and guide the diffusion model to generate tailored scenarios, thereby reducing feature complexity and increasing the diversity of generated scenarios. Furthermore, extreme scenarios are modeled using the generalized Pareto distribution (GPD), refining the loss function of the diffusion model to facilitate the generation of realistic extreme load scenarios. The experiments on real-world data demonstrate that the proposed method has superior performance compared with existing methods in terms of accuracy, diversity, and effective generation of normal and extreme load scenarios, which demonstrates the potential of ExDiffusion as an efficient approach for extreme scenario generation.