Unsupervised Detection of Spatiotemporal Anomalies in PMU Data Using Transformer-Based BiGAN

TL;DR

This work tackles unsupervised detection of spatiotemporal anomalies in PMU data by integrating a window-attention Transformer into a Bidirectional GAN (T-BiGAN) to learn normal grid behavior and flag deviations. It combines a Transformer-based encoder and generator with a joint discriminator to achieve cycle-consistency in the latent space, and uses a composite anomaly score with adaptive thresholding for real-time operation. On a realistic hardware-in-the-loop PMU benchmark, the approach attains ROC-AUC around 0.95 and average precision around 0.996, outperforming a wide range of supervised and unsupervised baselines and demonstrating strong detection of subtle frequency and voltage deviations. The method offers a scalable, label-free solution for wide-area monitoring, with practical implications for live grid resilience and edge deployment.

Abstract

Ensuring power grid resilience requires the timely and unsupervised detection of anomalies in synchrophasor data streams. We introduce T-BiGAN, a novel framework that integrates window-attention Transformers within a bidirectional Generative Adversarial Network (BiGAN) to address this challenge. Its self-attention encoder-decoder architecture captures complex spatio-temporal dependencies across the grid, while a joint discriminator enforces cycle consistency to align the learned latent space with the true data distribution. Anomalies are flagged in real-time using an adaptive score that combines reconstruction error, latent space drift, and discriminator confidence. Evaluated on a realistic hardware-in-the-loop PMU benchmark, T-BiGAN achieves an ROC-AUC of 0.95 and an average precision of 0.996, significantly outperforming leading supervised and unsupervised methods. It shows particular strength in detecting subtle frequency and voltage deviations, demonstrating its practical value for live, wide-area monitoring without relying on manually labeled fault data.

Figures (7)

• Figure 1: Feature distribution before normalization.
• Figure 2: Distribution after log-compression and scaling.
• Figure 3: Schematic of the proposed Transformer-augmented BiGAN framework.
• Figure 4: Transformer block structure (LN: Layer Normalization; MLP: Multi-Layer Perceptron).
• Figure 5: ROC (up) and precision–recall (down) curves for the proposed T-BiGAN.