Battery intelligent temperature warning model with physically-informed attention residual networks

The rapid development of electric vehicles demands improved thermal safety management of lithium-ion batteries. Traditional physical models require extensive offline parameter identification, struggling to balance computational efficiency and model fidelity, while data-driven methods, though precise, lack interpretability and require large datasets for varied conditions. To address these challenges, we propose the Physics-Informed Attention Residual Network (PIARN), which integrates an improved nonlinear dual-capacitor model and a thermal lumped model within a physics-guided recurrent neural network, enhancing both interpretability and generalizability. The residual attention network, comprising channel attention and time-series blocks, analyzes online measurements and hidden physical states to infer complex nonlinear dynamic responses, significantly improving accuracy. While a simplified physical model captures primary dynamics, the residual attention block corrects for missing nonlinear relationships. An adaptive weighting method accelerates network convergence by addressing voltage and temperature loss function magnitude discrepancy. Validation on three dynamic datasets demonstrates PIARN's ability to accurately predict battery voltage and temperature using sparse discharge data, showcasing strong generalization across varied conditions. Additionally, a cost-effective online iterative training framework is designed, enabling precise battery modeling and lifecycle tracking of aging and thermal status, with temperature prediction root mean square error as low as 0.1 °C and nearly 100 % accuracy in thermal warnings after multiple iterations. Thus, the novel PIARN model significantly enhance the accuracy of online temperature predictions and thermal warnings, thereby improving battery thermal management.