Analysis of Multivariate Indoor Building Data: A Comparative Study of Time-Series Clustering Methods

Clustering multivariate time-series data is crucial for uncovering complex temporal patterns in dynamic environments, such as building indoor conditions and behavior where variables like temperature, humidity, and CO<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sub> concentration evolve simultaneously. This study conducts a comparison of the performance of six clustering techniques: KMeans with Euclidean and Dynamic Time-Warping (DTW) distances, HDBSCAN, KShapes, Self-Organizing Maps (SOM), and Bi-Clustering combined with different normalization methods, applied to real-world indoor sensor datasets. Results show that applying a first-order derivative transformation notably improved clustering quality across all methods by highlighting temporal dynamics. Bi-Clustering and KMeans (Euclidean) consistently outperformed others, achieving higher Silhouette scores, Calinski-Harabasz indices, and lower Davies-Bouldin scores. These findings underscore the necessity of multivariate time-series clustering in capturing latent dependencies across variables and time-relevant data that are often missed in univariate analyses. Moreover, incorporating derivative-based preprocessing enhances the ability of clustering algorithms to distinguish subtle temporal dynamics, making this approach highly relevant for applications in smart building monitoring and environmental control.