KAN-GAT-PFL: A KAN-Enhanced Graph Attention Network with Personalized Federated Learning for Lithium-Ion Battery Early Cycle-Life Prediction

Abstract

Accurate early-stage cycle-life prediction of lithium-ion batteries is critical yet challenged by faint early degradation signatures and data privacy constraints across heterogeneous clients. This paper proposes KAN-GAT-PFL, a novel privacy-preserving framework tailored for distributed battery health management. We intrinsically integrate Kolmogorov-Arnold Networks (KAN) into a Graph Attention Network (GAT) coupled with an LSTM, substituting linear weights with learnable B-spline basis functions to dynamically capture highly non-linear, localized electrochemical variations with high parameter efficiency. Concurrently, a multi-stage personalized federated learning strategy, incorporating performance-weighted parameter aggregation and local fine-grained domain adversarial training, is developed to alleviate client drift while preserving strict data privacy. Extensive evaluations on diverse multi-client datasets demonstrate that the proposed KAN-GAT-PFL consistently outperforms state-of-the-art benchmarks, achieving up to a 40% accuracy improvement.