Event-preserving autoencoder for compression of transient-rich load profiles in smart metering systems

Abstract

Given the vast amounts of data generated by smart meters in smart grids, developing effective compression schemes is essential to address storage and communication challenges. However, the suitability of deep learning models for compressing smart meter data with events characterized by sharp transitions, jumps, and surge peaks from appliance activities remains underexplored. In this paper, we underscore that generalization performance deteriorates substantially when autoencoder approaches fail to preserve the shift equivariance property, indicating that the model overlooks essential features and instead relies on non-generalizable patterns. Furthermore, we observe that autoencoders incur significant reconstruction errors when handling transient periods due to Gibbs-like and staircase artifacts. To address these challenges, we propose an event-preserving autoencoder framework that maintains shift equivariance and is robust to inherent transient load profiles in smart meter data. To this end, it combines a convolution-based architecture with an event-preserving regularization term. Evaluations on public datasets demonstrate that our method outperforms existing autoencoders and a state-of-the-art transformer for time-series data, establishing it as an efficient and reliable solution for compressing smart meter data with events. Furthermore, we analyze how quantization and entropy-constrained loss influence reconstruction quality and compression ratio performance, providing insights into the trade-offs in practical deployments.