Fleet-level network lifetime prediction for electric vehicle batteries: physics-guided spatio-temporal graph neural networks with interaction effects

Abstract Battery lifetime prediction has traditionally focused on individual cells or packs, using model-based estimators or data-driven methods. These approaches, however, do not generalize to Electric Vehicle (EV) fleets, where degradation is influenced by shared charging infrastructure, overlapping routes, and correlated thermal and geographic exposures. This work introduces the fleet network lifetime prediction problem, which aims to forecast when a specified fraction of a fleet will cross a critical State-of-Health (SoH) threshold. To address these challenges, we model the fleet as a time-varying operational graph and propose a Physics-Guided Fleet Spatio-Temporal Graph Neural Network (PG-FSTGNN) that jointly learns individual SoH trajectories and produces calibrated fleet-level Time-To-Failure (TTF) metrics. To compensate for the scarcity of long-term fleet data, we develop a modular simulator that integrates calendar and cycle aging, charging-station queuing, thermal exposure, and operational policies such as load balancing, C-rate throttling, and battery swapping. Across 12 baselines, PG-FSTGNN consistently achieves superior prediction accuracy for $$\:TT{F}_{20}$$ , $$\:TT{F}_{40}$$ , and $$\:TT{F}_{60}$$ , maintains robust calibration even under severe sensor missingness, and provides interpretable insights into fleet-level degradation dynamics, evaluated using the NASA Prognostics Center of Excellence (PCoE) battery dataset and the CALCE (Center for Advanced Life Cycle Engineering) Li-ion battery aging dataset. By combining physics-guided degradation modeling, graph-based learning, and realistic fleet simulation, this framework provides a scalable and reliable solution for EV fleet battery management and supports more effective planning, maintenance scheduling, and infrastructure optimization.