A Diffusion-based Generative Machine Learning Paradigm for Contingency Screening

TL;DR

This work tackles the computational burden of contingency screening in power systems by reframing it as a generative task. It introduces DDPM-CS, a diffusion-based model that learns from historical contingencies to generate likely-worst system states, treating grid configurations as image-like data and using a forward diffusion with a U-Net-based reverse denoiser. The method integrates Continuation Power Flow to define voltage-collapse margins via a nose curve and the direct margin $\max_\lambda$, enabling physically meaningful evaluation of generated contingencies under the $N-1$ criterion. Empirical results on IEEE-6, IEEE-14, and IEEE-30 demonstrate that DDPM-CS can produce high-risk contingencies and corresponding critical load profiles with low mean absolute error, suggesting potential for real-time online security assessment and broader generative AI applications in power systems.

Abstract

Contingency screening is a crucial part of electric power systems all the time. Power systems frequently encounter multiple challenging operational dilemmas that could lead to the instability of power systems. Contingency analysis is effort-consuming by utilizing traditional numerical analysis methods. It is commonly addressed by generating a whopping number of possible contingencies or manipulating network parameters to determine the worst scenarios. This paper proposes a novel approach that diverts the nature of contingency analysis from pre-defined scenario screening to proactive-unsupervised screening. The potentially risky scenarios of power systems are generated from learning how the previous ones occurred. In other words, the internal perturbation that initiates contingencies is learned prior to being self-replicated for rendering the worst scenarios. By leveraging the perturbation diffusion technique, a proposed model is built to point out the worst scenarios instead of repeatedly simulating one-by-one scenarios to define the highest-risk ones. Empirical experiments are implemented on the IEEE systems to test and validate the proposed solution.

Figures (5)

• Figure 1: The illustration of the proposed model DDPM-CS: The forward process adds noise to the data until the fully noisy state is obtained. The reverse process utilizes a U-NET neural network to estimate the added noise. Finally, the sampling procedure generates new data based on the predicted noise.
• Figure 2: The flowchart of the DDPM-CS's implementation.
• Figure 3: Evaluation and frequency analysis for IEEE-6 case study across 100 samples.
• Figure 4: Evaluation and frequency analysis for IEEE-14 case study across 100 samples.
• Figure 5: Evaluation and frequency analysis for IEEE-30 case study across 100 samples.