Optimal bus load grouping method based on source–load-constrained multi-view clustering and TS2Vec-TabM

Conventional bus classification relies heavily on expert judgment and a single view of load-curve shape, making it difficult to account for source-load attributes and operating mechanisms. This paper proposes an optimal bus load grouping method that combines source-load-constrained multi-view clustering with TS2Vec-TabM. Annual 15-min bus load curves are first normalized by capacity and reconstructed at multiple temporal scales to derive typical-day and seasonal curves, statistics, and customer-composition features. TS2Vec then extracts self-supervised temporal representations from the seasonal curves. These representations are combined with engineered statistics and source-load attributes to form a multi-view feature space. Physical anchors are constructed from renewable-energy penetration, the capacity share of impulsive loads, and the number of impulsive events. Must-link and cannot-link constraints are incorporated into the multi-view distance to cluster buses into conventional, renewable-integrated, and impulsive-load groups. Finally, the parameter-efficient ensemble model TabM performs three-class classification and outputs class probabilities and confidence scores. Experiments on real bus data show that the proposed clustering method achieves an adjusted Rand index of 0.965, normalized mutual information of 0.936, and clustering accuracy of 0.974. In the classification stage, the proposed TS2Vec-TabM framework achieves an accuracy of 0.985, a Macro-F1 of 0.979, and a log loss of 0.057, demonstrating strong overall grouping and classification performance.