TriMa: Tri-strategy Margin-filtered Pseudo-labeling for Battery Swapping Demand Forecasting

Accurate prediction of battery-swapping demand is crucial for minimizing operational costs and maintaining service quality in electric motorcycle Battery-as-a-Service (BaaS) deployment. However, early-stage BaaS providers often face severe data scarcity and bias, especially in newly expanded regions. In these areas, newly deployed battery swapping stations (BSS) encounter cold-start problems due to limited historical labels, and existing regional models degrade under domain shift caused by distributional changes in contextual covariates. Although conventional data augmentation techniques like SMOTE are often employed to mitigate these challenges, they perform poorly in multi-class settings by exacerbating inter-class overlap near complex decision boundaries. To address this limitation, we introduce TriMa (Tri-strategy Margin-filtered Pseudo-labeling), an advanced pseudo-labeling framework designed to enhance predictive performance under sparse and imbalanced multi-class data distributions. TriMa constructs augmented samples exclusively from high-confidence pseudo-labels, which are derived through three complementary labeling strategies combined with margin-based filtering. Experiments on UCI datasets under simulated sparsity and bias, as well as on real-world operational data from a BaaS provider in Seoul, demonstrate the efficacy of our proposed method. Results show that TriMa consistently improves the F1 score and reduces performance variability relative to existing techniques, while remaining robust to label noise.