Comparing Clustering Approaches for Smart Meter Time Series: Investigating the Influence of Dataset Properties on Performance

TL;DR

This study systematically compares various clustering methods for smart meter time series data, emphasizing dataset properties' effects on performance and identifying robust approaches like Dynamic Time Warping and $k$-sliding distance.

Contribution

It introduces a comprehensive framework for evaluating clustering methods on synthetic datasets tailored to smart meter data, considering dataset variability and robustness.

Findings

DTW and $k$-sliding outperform traditional methods

Robust clustering combinations identified with $k$-medoids and hierarchical clustering

Dataset properties significantly influence clustering performance

Abstract

The widespread adoption of smart meters for monitoring energy consumption has generated vast quantities of high-resolution time series data which remains underutilised. While clustering has emerged as a fundamental tool for mining smart meter time series (SMTS) data, selecting appropriate clustering methods remains challenging despite numerous comparative studies. These studies often rely on problematic methodologies and consider a limited scope of methods, frequently overlooking compelling methods from the broader time series clustering literature. Consequently, they struggle to provide dependable guidance for practitioners designing their own clustering approaches. This paper presents a comprehensive comparative framework for SMTS clustering methods using expert-informed synthetic datasets that emphasise peak consumption behaviours as fundamental cluster concepts. Using a phased methodology, we first evaluated 31 distance measures and 8 representation methods using leave-one-out classification, then examined the better-suited methods in combination with 11 clustering algorithms. We further assessed the robustness of these combinations to systematic changes in key dataset properties that affect clustering performance on real-world datasets, including cluster balance, noise, and the presence of outliers. Our results revealed that methods accommodating local temporal shifts while maintaining amplitude sensitivity, particularly Dynamic Time Warping and $k$-sliding distance, consistently outperformed traditional approaches. Among other key findings, we identified that when combined with $k$-medoids or hierarchical clustering using Ward's linkage, these methods exhibited consistent robustness across varying dataset characteristics without...