A novel physics-informed transformer-based hybrid model for satellite on-board battery health prognostics

The energy balance of a satellite is centered around its battery subsystem. The importance to closely monitor the health conditions of the batteries on-board a satellite is evident. To provide such monitoring, traditional methods demand significant amounts of on-board sensor data to be transmitted back to ground, then be processed with sufficient ground-based processing power. In recent years, the rapid proliferation of Low-Earth Orbit (LEO) constellations with tens of thousands of LEO satellites planned to be launched into orbit is out-pacing the development of ground-based operation management capacity, while simultaneously saturating the satellite-ground telemetry link. This, in turn, introduces unprecedented challenges for constellation-level health management. The industry has shown an urgent shift to on-board autonomies, leveraging the abundance of in-situ data while off-loading the workload on the ground system, in order to enable satellites to partially govern their own operational status. In this paper, we introduce the Physics-Informed Transformer for Battery Health (PITBH), a novel deep learning architecture designed with on-board integration constrains in mind. The model’s innovations are threefold: A hybrid Linformer-GRU architecture that effectively utilizes the intrinsic advantages of both models while maintaining a small cost footprint; A delta-prediction paradigm that targets cycle-to-cycle capacity changes rather than absolute SOH, fitting the practical requirements of on-board systems; The integration of Differential Capacity Analysis (DCA) into the loss function, providing physical constraints that the model enforces. The resulting framework is computationally efficient, making it a promising candidate for engineering implementation. Evaluation demonstrates that despite its fundamentally different cumulative integral approach, PITBH achieves highly competitive accuracy compared to established models, validating it as a powerful and practical solution for intelligent battery management in future on-board missions.