Real-Time Small-Signal Security Assessment Using Graph Neural Networks

TL;DR

The paper tackles real-time small-signal security assessment (SSA) for power systems under N-1 contingencies by introducing a graph neural network (GNN) that leverages PMU-derived signals and grid topology. The approach reduces input and model size, accelerates offline training and online inference, and offers robustness to partial PMU observability through graph-based aggregation and centrality-driven PMU placement. Empirical results on the IEEE 68-bus and NPCC 140-bus systems show over 95% accuracy, significant speedups (online inference in tens of milliseconds), and resilience to up to 20% missing PMU data, outperforming CNN-based approaches especially in large networks. The work demonstrates practical impact by enabling real-time SSA, informing PMU placement, and highlighting the benefits of exploiting the grid’s graph structure for stability analysis.

Abstract

Security assessment is one of the most crucial functions of a power system operator. However, growing complexity and unpredictability make this an increasingly complex and computationally difficult task. In recent times, machine learning methods have gained attention for their ability to handle complex modeling applications. Some methods proposed include deep learning using convolutional neural networks, decision trees, etc. While these methods generate promising results, most methods still require long training times and computational resources. This paper proposes a graph neural network (GNN) approach to the small-signal security assessment problem using data from Phasor Measurement Units (PMUs). Leveraging the inherently graphical structure of the power grid using GNNs, training times can be reduced and efficiency improved for real-time application. Also, using graph properties, optimal PMU placement is determined and the proposed method is shown to perform efficiently under partial observability with limited PMU data. Case studies with simulated data from the IEEE 68-bus system and the NPCC 140-bus system are used to verify the effectiveness of the proposed method.

Figures (8)

• Figure 1: A flowchart showing the steps involved for SSA using N-1 contingency analysis. The major computational burden is in performing the N-1 contingency analysis. This is where the proposed method is used.
• Figure 2: Operation of the proposed method similar to the method in b44. The Real-time SSA is carried out in 3 steps: first data generation, then offline training using graph data and lastly online testing using the trained GNN.
• Figure 3: GNN formed from a stack of 3 graph perceptrons b20
• Figure 4: IEEE 68-Bus system graph with 16 generators and 86 transmission lines b18. Buses with highest centrality are highlighted in red.
• Figure 5: NPCC 140-Bus system graph with 48 generators and 233 transmission lines b28. Buses with highest centrality are highlighted in red.