Synchrophasor Data Anomaly Detection Using Unsupervised Transformer Autoencoder

Abstract

Detecting anomalies in synchrophasor measurement data is critical for improving situational awareness and supporting better decision-making in power grid operations and control, especially as cyber-physical threats continue to grow. This work proposes an unsupervised Transformer Autoencoder (TAE) for anomaly detection in Phasor Measurement Unit (PMU) data with spatio-temporal characteristics. The model leverages a self-attention mechanism to capture complex temporal dependencies and integrates an autoencoder structure for unsupervised learning. Unlike conventional autoencoder-based methods that require entirely clean training data, the proposed TAE can be trained on datasets containing mostly normal measurements with a small fraction of anomalies, making it more practical for real-world applications. In addition, a block-wise Z-score normalization scheme with moving window is introduced to improve robustness, enabling the model to better identify spatio-temporal variations in PMU data. The model is validated using a high-resolution PMU dataset generated from a cyber-power testbed that includes a wide range of realistic cyber and physical anomalies. Extensive experiments show that the proposed TAE delivers consistent accuracy and strong generalization across different thresholds and window sizes when applied to realistic datasets. Results indicate performance of the proposed TAE algorithm with precision, recall and F1-Score of 97.20%, 98.58%, and 97.89%, respectively. © 2025 IEEE.