CAPN: A Cluster-Aware Prediction Network for Multi-step Occupancy Forecasting of Battery Swapping Stations

Accurate forecasting of battery swapping station (BSS) occupancy is essential for battery scheduling, enhancing user experience, and enabling coordinated cluster management. Most existing methods fail to effectively capture spatial heterogeneity, limiting the performance of multi-step prediction. To address this challenge, we propose a two-stage Cluster-Aware Prediction Network (CAPN) to predict the BSS occupancy. The core of our approach has two main components. First, we design a Particle Swarm Optimization (PSO)-enhanced K-Means clustering module that integrates time-domain features, frequency-domain characteristics, and geographical information to group BSSs, effectively capturing the heterogeneity in their operational patterns. Second, we incorporate the resulting cluster embeddings into a Temporal Convolutional Network (TCN) via Feature-wise Linear Modulation (FiLM), enabling the model to adapt its predictions to the heterogeneous characteristics of different station clusters. On a real-world dataset comprising records from over 40 battery swapping stations, our proposed CAPN consistently outperforms other state-of-the-art methods, achieving an average 14.37% reduction in MAE for multi-step occupancy forecasting.