Forecast-free real-time V2B control of battery swapping stations as distributed battery assets via imitation learning
Applied Energy, Vol. 427, 129239 · Jan 2027
Read at publisherDOI: 10.1016/j.apenergy.2026.129239Abstract
Battery swapping stations are emerging as distributed storage assets for coordinating transportation and building energy use. Their integration with renewable generation creates opportunities to improve flexibility and reduce electricity costs across increasingly electrified urban distribution networks. Conventional energy management systems combine forecasts of demand and renewable generation with repeated online optimization to schedule battery charging and vehicle-to-building (V2B) dispatch. However, forecast errors and online optimization costs limit their suitability for real-time operation under uncertainty and incomplete data. We propose an imitation-learning framework that distills perfect-foresight mixed-integer linear programming (MILP) decisions into a Static–Temporal Attention Fusion (STAF) network, supported by multivariate normal reconstruction of incomplete fleet data. The framework is evaluated in a 72-hour closed-loop simulation using two Gogoro stations and an academic building in Taipei, with synthesized photovoltaic generation. Across 20 training seeds, STAF achieves a median electricity-cost gap of 7.7% relative to the perfect-foresight benchmark without unmet swaps under a common swap-readiness safeguard. Median decision latency is 0.79 ms, approximately 1200 times faster than the benchmark. STAF also achieves lower median cost than forecast-augmented MILP at the tested nominal 20% and 30% error settings, supporting its potential for low-latency control under forecast uncertainty in this V2B system.
Keywords
Distribution networks · Distributed energy resources (DERs) · Battery swapping station (BSS) · Vehicle-to-building (V2B) · Imitation learning · Forecast-free control · Energy management system (EMS)