Advanced Search
  • Abstract

    Accurate battery state estimation and health prognostics are essential for the safety, reliability, and lifecycle management of electric vehicles and stationary energy-storage systems. Battery degradation is nonlinear, path-dependent, and strongly affected by operating conditions, which complicates state of charge (SOC) and state of health (SOH) estimation as well as remaining useful life (RUL) prediction. This review critically examines Transformer architectures for battery state estimation and health prognostics. It links architectural choices to their intended functions: global attention for long-horizon dependence, local-global hybrids for curve features, observer and filter feedback for online correction, sparse or compressed attention for efficiency, physics-informed losses for trajectory consistency, and generative pre-training for transfer. Reported results are interpreted against their datasets, input availability, computational evidence, and validation protocols rather than ranked directly across non-identical experiments. Particular attention is given to feature engineering, laboratory-to-field dataset shift, cross-condition generalization, uncertainty quantification, reproducibility, and deployment constraints. The evidence indicates that Transformers are most appropriate when long context or multi-source transfer is central; convolutional neural networks, recurrent networks, filters, or physics-based models can be preferable for short windows, small datasets, or tightly constrained embedded estimation. Progress therefore depends on calibrated uncertainty, leakage-resistant benchmarking, cross-chemistry and fleet validation, and transparent reporting of both accuracy and resource cost.
  • loading
  • Related Articles

Related Articles
    Show full outline

    Catalog