Integrating transformers into physics-informed neural networks: An approach to lithium-ion battery state-of-health prognostics
Year: 2025
Type: article
Abstract: Accurate battery health prediction is critical for ensuring the reliability and safety of devices. However, existing methods face limitations when processing long-sequence inputs and complex data patterns, which restricts their performance in prognostic tasks. These methods are often sensitive to data noise, limiting their prediction accuracy. They also struggle to extract meaningful features from complex patterns, resulting in a poorer representation of the underlying processes. Furthermore, traditional approaches typically fail to adequately incorporate the physical laws governing battery behavior, which reduces their accuracy and reliability in real-world applications. To address these challenges, this paper introduces a novel prognostic framework based on Physics-Informed Neural Networks (PINN), named PI-TNet. The framework integrates a multimodal information perception mechanism with a Local Information Processor (LIP) that employs convolutional attention to reduce noise and improve feature matching accuracy. Furthermore, it incorporates a Vision Transformer (ViT) to enhance global feature extraction from complex inputs and defines a loss function within the PINN architecture that includes dynamic hyperparameter adaptation. Experimental results demonstrate the effectiveness of the PI-TNet system in handling long-sequence inputs for prognostic tasks, significantly outperforming existing methods. Notably, on specific battery packs, it shows improvements exceeding 90\% over other neural network approaches
Authors: Yichen Zhan, Kehuan Yan, Xianghan Zheng
Institution: Fuzhou University
Language: English
Subfield: Automotive Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $3,920
Awards: G2447469041, G6680830185
OpenAlex ID: W4414815845