Study on Real-Time Battery Temperature Prediction Based on Coupling of Multiphysics Fields and Temporal Networks

Real-time temperature prediction is essential for ensuring the thermal safety of Lithium-ion batteries (LIBs), yet its industrial application faces challenges due to fluctuations in operating conditions such as temperature, voltage range, capacity degradation, and current rates (C-rates). To address this, we introduce a novel framework, Transformer-GPR, which merges the physical battery model with a Transformer-based network. This integration facilitates the offline training of hyperparameters, enhancing real-time temperature prediction accuracy. Additionally, we employ two residual models using Gaussian Process Regression (GPR) to correct for local temperature deviations. The Transformer-GPR framework is designed to predict temperature accurately across the entire lifecycle of LIBs with limited data and under varied operational conditions. It has been benchmarked against several existing methods, showing superior interpretability, accuracy, and transferability. Validation with operational data from a pure electric vehicle confirmed the model's efficacy; it precisely predicted temperature change sequences, with an RMSE of 0.048, an MAE of 0.036, and a maximum error of 0.28, using training inputs from similar vehicles.