Breaking ‘Chicken-Egg’: Cross-city Battery Swap Demand Prediction via Knowledge-guided Diffusion

The battery swap market has experienced rapid expansion in recent years. The deployment of battery swap stations (BSSs) is a fundamental issue that needs to be addressed during expansion in new cities. Intuitively, deploying BSSs according to user demand distribution is sensible. However, the absence of historical data makes it challenging to predict the demand in new cities before actual deployment, resulting in a ‘Chicken-Egg’ dilemma. To address this dilemma, we propose STKDec, a spatio-temporal knowledge-guided conditional diffusion model for cross-city battery swap demand prediction. STKDec leverages historical data from source cities with deployed BSSs to predict user demands for target cities planning BSS deployment. The core of STKDec is a novel station-level context representation-based condition module. The key idea of this module is to extract shared spatiotemporal context representations of all stations and use these representations to bridge the source and target cities for demand prediction. Specifically, we derive the spatial context representation of stations through an urban knowledge graph construction and a designed multi-relation-aware graph convolutional network while extracting the temporal context representation using a temporal information encoding strategy. Extensive experiments on real-world battery swap datasets demonstrate that STKDec averagely outperforms state-of-the-art methods by 30.7 %.