Energy-Consumption-Informed Driving Behavior Clustering for Electric Vehicles: A GRA-Weighted K-Means Approach

Conventional driving-behavior clustering primarily distinguishes behavioral differences without explicitly considering feature associations with electric vehicle (EV) energy consumption, leaving the energy relevance of the resulting patterns unclear. Motivated by this challenge, this study develops an energy-consumption-informed clustering and analysis framework integrating gray relational analysis (GRA) with weighted K-means. Sliding windows extract aligned behavioral features and specific energy consumption (SEC) from real-vehicle controller area network (CAN) data. Acceleration standard deviation, mean accelerator-pedal position, and braking frequency characterize combined driving behavior. Using SEC as the reference sequence, GRA determines feature-association weights. These weights are incorporated into the K-means distance metric, allowing energy information to guide clustering without directly including SEC as a clustering variable. A random forest (RF) provides a supplementary assessment of temporal assignment reproducibility and feature contributions to reproducing cluster labels. Analysis of 99 model-development samples identifies economical and aggressive patterns with an average silhouette coefficient of 0.94. The aggressive pattern exhibits 25.2% higher mean SEC than the economical pattern, while RF-clustering agreement reaches 90.5% on 42 temporally held-out samples. Comparisons with unweighted K-means, Gaussian mixture modeling, and Ward-linkage clustering show more favorable clustering-validity indices and greater inter-cluster SEC separation within this dataset. The findings provide a basis for interpreting combined driving-behavior patterns and their energy associations under the sampled urban conditions.