Invariant Feature Extraction for Power Quality Disturbances Using Deep Learning - Compact and Interpretable Features Enhance Performance on Downstream Tasks

Abstract

ABSTRACT Feature extraction is a crucial step for tasks such as classification, clustering and predictive risk estimation. In power quality analysis, feature selection is often performed by domain experts. However, crafting a small yet sufficiently informative feature set is often challenging, particularly when evaluation data is scarce. In this thesis, we present an autoencoder-based feature extraction method tailored for spectrograms of three-phase current and voltage waveforms. The extracted feature set is invariant with respect to time translation and sample length, and we demonstrate that sufficient information is retained to reconstruct the spectrogram with high accuracy. By applying Uniform Manifold Approximation and Projection (UMAP) with a specific pseudometric, invariance with respect to phase permutations can be obtained while further reducing the dimensionality of the extracted features. The interpretability of the extracted feature sets is evaluated by observing changes in reconstructed signals when the latent variables are perturbed. The usefulness of the features is measured through three tasks: clustering, fault prediction and root cause disturbance classification. For the fault prediction task, we demonstrate that training an LSTM model with UMAP features significantly increases the AUC value compared to training the same model with manually selected features by a domain expert. In the classification task, both autoencoder features and UMAP features result in a higher macro 𝐹1-score than manual features when training a neural network classifier, regardless of the training set size. The improvements are particularly notable for the smallest training sets. Additionally, using a model-free label propagation method on these features further enhances performance. Interestingly, we find that pretraining the autoencoder improves reconstruction fidelity, even when the pretraining dataset consists of audio recordings that are quite different from power quality measurements.