Explainable deep representation learning for clustering building-energy time series

Large volumes of smart-meter data can be transformed into actionable knowledge about building energy consumption. In the absence of labeled data, clustering offers a way to partition daily load shapes into coherent and decision-relevant groups. However, traditional clustering methods struggle with the high dimensionality and temporal variability of energy data. Motivated by recent advances in deep learning for time-series clustering, this study investigates deep clustering pipelines that learn latent features directly from signals before grouping them in the learned representation space. In this context, seven architectures ranging from convolutional to recurrent networks, five pretext losses, and seven clustering losses were evaluated on two real-world datasets with different optimal cluster counts: a univariate dataset of 470,120 daily electricity profiles and a multivariate dataset of 4261 daily profiles combining electricity, chilled water, steam, and hot-water consumption. The proposed deep clustering pipelines were compared against four traditional clustering algorithms and model performance was assessed using five internal evaluation metrics. To address hyperparameter sensitivity, we optimize the top three models per dataset using an evolutionary strategy. To enhance explainability in energy analytics, where actionable insights are as important as clustering quality, we addressed the black-box nature of deep learning models using explainable artificial intelligence (XAI). Prototype–criticism analysis was applied to characterize each cluster through typical and atypical daily load profiles, revealing how buildings differ in their operational dynamics and demand patterns. The proposed deep clustering pipelines achieved higher internal evaluation scores than traditional methods and produced clearer separations of energy-use regimes, providing explainable and decision-relevant insights for benchmarking, demand management, and retrofit planning.