Silent Signals: Applying AI to Reveal Hidden Power Quality Trends

Abstract

Power quality (PQ) monitoring is essential for ensuring the efficient operation of modern electrical grids, where increasing electrification and the integration of nonlinear loads introduce complex disturbances. While traditional methods focus on short-term events, long-term PQ trends remain underexplored due to the high dimensionality and volume of the data. This thesis investigates the application of unsupervised machine learning techniques to discover long-term anomalous patterns in PQ data without relying on labelled datasets. Daily statistical features were extracted from root mean square (RMS) voltage, total harmonic distortion (THD) voltage, and voltage unbalance factor (VUF), collected from two locations in Sweden. Three models were implemented and evaluated: A Self-Organising Map (SOM), Autoencoder (AE), and Variational Autoencoder (VAE). Anomaly detection was carried out using model-specific error metrics: quantisation error (QE) for SOM and reconstruction error (RE) for AE-based models alongside a density-based outlier score (OS). The results demonstrate that all models are capable of identifying anomalous days in long-term PQ behaviour. Distance-based methods (QE and RE) showed higher consistency and better alignment with meaningful deviations compared to densitybased OS. The VAE approach exhibited the strongest performance in distinguishing anomalous patterns, likely due to its regularised latent space, which enhances sensitivity to deviations from normal behaviour. The models showed limited agreement on specific anomalies, this highlights that anomaly detection is strongly modeldependent, as each model captures different aspects of the data structure. Overall, the study confirms that unsupervised machine learning is a promising tool for long-term PQ analysis, enabling automated detection of subtle anomaly disturbances. The findings support the integration of such methods into PQ monitoring systems to improve data-driven decision-making and facilitate a transition toward more proactive grid management.