A Deep Learning-Based Method for Power System Resilience Evaluation

TL;DR

The paper tackles the challenge of evaluating power system resilience to low-probability, high-impact weather events without requiring detailed physical network models. It introduces a GRU-based encoder–MLP decoder that maps severe-weather scenarios to system performance measured by the resilience trapezoid Rs_{i,k}, enabling computation of unweighted Ru_k and weighted Rw_k that account for social vulnerability. By training on real outage data (EAGLE-I) and a synthetic Case B dataset, the approach demonstrates accurate resilience estimation and meaningful guidance for resilience enhancement and DER planning, with case studies showing high correlation to ground truth and simulation results. This data-driven, topology-agnostic framework offers a scalable tool for cross-region resilience comparisons and policy-oriented planning in the context of climate-related outages.

Abstract

Power systems are critical infrastructure in modern society, and power outages can cause significant disruptions to communities and individuals' daily lives. The resilience of a power system measures its ability to maintain power supply during highly disruptive events such as hurricanes, earthquakes, and thunderstorms. Traditional methods for quantifying power system resilience include statistics-based and simulation-based approaches. Statistics-based methods offer a retrospective analysis of system performance without requiring a physical model, while simulation-based methods necessitate detailed physical system information and often simplify real-world scenarios. This paper introduces a deep learning-based method for evaluating power system resilience using historical power outage data. The method leverages the generalization capabilities of deep learning models and incorporates socio-economic and demographic factors as weighting terms to highlight the impacts on vulnerable demographic groups. The effectiveness of the proposed method is demonstrated through two case studies: one with real historical outage data and the other with simulated outage records. This approach provides valuable insights into measuring power system resilience against hazardous weather events without requiring a physical model of the target systems. The evaluation results can further guide the planning of distributed energy resources for resilience enhancement.

Figures (14)

• Figure 1: Demonstration of the system performance curve under a hazardous weather event. $f(t)$ is the normalized system performance curve. $T_1$ and $T_2$ are the start and end times of the outage event, respectively.
• Figure 2: Model architecture for predicting $Rs_{i,k}$, i.e., the resilience of the power system $k$ under weather event $i$, where $\hat{Rs}_{i,k}$ is the predicted value of $Rs_{i,k}$.
• Figure 3: Training and validation losses of case study A, averaged over the 5 folds.
• Figure 4: Unweighted electric power resilience of 71 counties in Michigan.
• Figure 5: Weighted electric power resilience of 71 counties in Michigan.