Abstract

The increasing deployment of smart meters and IoT sensing infrastructures has produced large volumes of high-resolution building-energy data, offering new opportunities for understanding consumption behavior, improving operational efficiency, and supporting energy planning. Extracting meaningful structure from this data remains challenging due to its high dimensionality, nonlinear temporal patterns, and the absence of labels. Existing clustering approaches often struggle with instability, sensitivity to preprocessing choices and distance metrics, and inconsistencies across internal validation indices, limiting their ability to reliably characterize underlying consumption behavior patterns. Moreover, most clustering studies treat the process as a black box, providing little insight into why specific profiles are grouped together, which restricts their usefulness for practitioners and decision-makers.

To address these limitations, this research develops a comprehensive analytical framework that combines traditional clustering methods, deep time-series representation learning, and explainable artificial intelligence to analyze building-energy load profiles. Cluster quality across all analyses is assessed using five internal validation indices. First, we evaluate K-Means, K-Medoids, Fuzzy C-Means, and Gaussian Mixture Models, both with and without dimensionality reduction, examining their robustness under variations in intra- and inter-cluster characteristics, including outliers, overlapping profiles, density shifts, skewness, kurtosis, and sub-clustering. Multiple techniques are used to estimate the optimal number of clusters for each method. Decision-tree–based explanation models, specifically axis-aligned and sparse oblique, are applied to produce human-explainable rules linking profile features to cluster assignments. Second, we develop a deep time-series clustering pipeline across seven encoder architectures, five representation-learning losses, and seven clustering losses on both univariate and multivariate building-energy datasets, tested with two different cluster configurations. To determine whether the additional complexity of deep clustering is justified, we compare its performance against four traditional clustering algorithms. To overcome the sensitivity of deep clustering to hyperparameters, we integrate Population-Based Training as an evolutionary optimization strategy. Explainability is incorporated through prototype–criticism analysis, providing representative and atypical profiles that summarize each cluster’s internal structure.

The results show that dimensionality reduction has minimal impact on overall clustering quality but can enhance separability in overlapping or variable-density settings. The explainability analysis further revealed a trade-off between completeness and simplicity: axis-aligned trees achieved full cluster coverage at the cost of greater rule complexity, whereas sparse oblique trees produced simpler rules but occasionally failed to cover specific clusters. Overall, deep clustering methods consistently capture more coherent and better-separated patterns than traditional algorithms, while the proposed interpretability modules offer clear and actionable explanations of consumption behavior. Together, these contributions provide a scalable, unsupervised, and transparent approach for transforming raw building-energy time series into meaningful behavioral archetypes, enabling improved energy management, personalized feedback, and data-driven planning.