Time-series clustering: A benchmark study on energy data with insights into demand response

This study presents a comprehensive benchmarking framework for time-series clustering, addressing the lack of standardized guidelines for selecting appropriate approach for clustering tasks. The framework derives and evaluates 15 clustering pipelines comprising multiple well-known clustering techniques and diverse distance metrics, with the constraint of using the same number of clusters for all pipelines to ensure comparability and consistency in the evaluation process. The procedure for standardizing clustering labels, generating ensemble outcomes, and incorporating a stability score is introduced to provide a comprehensive and rigorous evaluation of clustering pipelines. The challenge posed by arbitrary label assignments across different pipelines is resolved through a standardization process that aligns the clusters for meaningful comparison. The ensemble approach mitigates inconsistencies in clustering results and addresses the limitations of traditional clustering validity metrics, leading to more stable and reliable groupings. Additionally, the inclusion of a novel stability score adds a critical layer of evaluation, enabling the identification of the most consistent and accurate clustering outcomes. The results highlight limitations of traditional quality metrics, while showcasing the strong performance for various pipelines with more than 90% of similarity with ensemble results. This would aid in informed selection of the best pipelines for specific applications. The framework is further discussed in the context of optimizing clustering for demand response strategies in smart grids, highlighting its real-world relevance. To support transparency and reproducibility, the study provides open-source code for validating and applying it to various time-series datasets, offering a robust tool for benchmarking clustering across domains. • Comprehensive benchmarking framework for time-series clustering introduced. • Evaluates 15 clustering pipelines, with a fixed cluster count for fair comparison. • Novel stability score identifies the most accurate and stable clustering outcomes. • Validity metrics have limitations in giving consistent results for similar datasets. • Made source codes & datasets open to promote reproducibility in clustering research.