A parallel weighted ADTC-Transformer framework with FUnet fusion and KAN for improved lithium-ion battery SOH prediction

Abstract

This paper delves into the extraction and integration of local and global features for lithium-ion battery State of Health (SOH) prediction, proposing an innovative parallel weighted architecture—ADTC-Transformer. This framework combines Adaptive Dilated Temporal Convolution (ADTC) with a Transformer encoder to effectively capture and balance local and global dependencies while dynamically optimizing feature contributions through a weighted fusion mechanism. Additionally, the traditional U-shaped network (Unet) is enhanced by incorporating a Feature Pyramid Network (FPN), forming the FUnet module, which significantly strengthens the fusion and utilization of multi-scale features. Building on this, the Kolmogorov–Arnold Network (KAN) is introduced as the final prediction module, leveraging Kolmogorov–Arnold representation theory to model complex high-dimensional features through local interpolation and global nonlinear transformations. This enables the KAN module to capture intricate temporal dependencies and interactions across a wide range of feature scales, thus improving the model’s ability to predict long-term SOH. Experimental results demonstrate that the proposed method markedly improves prediction accuracy across NASA, CALCE, and WRBD datasets, excelling particularly in long-term SOH prediction for lithium-ion batteries. This provides robust support for battery health management and performance optimization.