Integrating transformers into physics-informed neural networks: An approach to lithium-ion battery state-of-health prognostics

Résumé fourni par la source

Accurate prediction of the State of Health for lithium-ion batteries is critical to ensure the operational safety and prolong the service life of energy storage systems. However, conventional purely data-driven models and physics-informed neural network often suffer from limited accuracy and poor generalization when dealing with battery samples exhibiting diverse electrochemical characteristics. To address these challenges, this paper proposes a Physics-Informed Neural Network integrated with Transformer for state of health prediction ( PI-TNet ). The framework introduces a Convolutional Data Processor to extract multi-dimensional features of electrochemical processes, while incorporating the Verhulst model as learnable parameters within a Vision Transformer architecture. This hybrid design enables simultaneous capture of long-term temporal dependencies in sequential cycling data and precise modeling of battery degradation mechanisms. Extensive cross-dataset validation using both NASA and CALCE battery datasets demonstrates that PI-TNet significantly outperforms existing methods in handling long-sequence prediction tasks while achieving better accuracy and generalization performance. • PI-TNet: A Transformer-integrated Physics-Informed Neural Network for SOH prediction. • The Verhulst model boosts interpretability with physical constraints. • PI-TNet outperforms models like MVIP-Trans and MyA+ by over 50% in key metrics.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.