Unsupervised learning of latent features for fault root cause analysis

An increasingly electrified society requires power system operators to improve network resiliency. Root cause analysis of power disturbances aims to identify minor disturbances to enable preventive maintenance, reduce outages, and improve efficiency. Machine learning (ML) offers a compelling solution, automating analysis while leveraging the growing availability of disturbance data from digitialised power systems. Standard ML models rely on informative features computed from raw data, but manually engineering such features requires extensive domain expertise. We propose an autoencoder-based feature extractor that uses unlabelled disturbance recordings for automatic feature extraction. The autoencoder learns compact feature representations by encoding input signals to a low-dimensional latent space before reconstructing them. The model learns to encode the data into a compact and informative feature space by optimising a reconstruction loss. The training process uses all available data without the need for manual labelling. We demonstrate the effectiveness of this approach by training a power disturbance classifier on the learned features. The results show improved performance compared to models trained on hand-crafted features, even with smaller latent dimensions. This highlights the potential of automated feature extraction for improving root cause analysis in power systems.