Uncertainty-Aware Representation Learning for Power Quality Waveforms Using Convolutional Autoencoders

Abstract

Power quality (PQ) monitoring relies on high-resolution waveforms that are inherently affected by measurement uncertainty. Conventional PQ indices provide compact descriptors but may fail to capture waveform morphology and transient behavior. This paper investigates an uncertainty-aware representation learning approach based on a one-dimensional convolutional encoder–decoder network, which maps waveform windows into a low-dimensional latent space treated as a measurement representation rather than as a feature space for classification. A synthetic PQ disturbance dataset is used for controlled pretraining, while a real-life PQ transient dataset is adopted for validation on measured signals. Measurement uncertainty is modeled through repeated uniform and Gaussian perturbations with relative amplitudes of 1 % and 2 %. The learned representation is evaluated in terms of reconstruction error, latent-space stability, comparison with principal component analysis (PCA), and correlation with classical PQ descriptors. Results show that, although the model is not optimized for waveform reconstruction, it provides a compact and stable latent representation that is robust to uncertainty and preserves meaningful relationships with physically interpretable PQ features. These findings support the use of uncertainty-aware latent representations for compact and interpretable monitoring of PQ waveforms.