Monitoring lithium-ion battery (LIB) temperature in real time is crucial for ensuring system safety, especially across different chemistries, testing, and environmental conditions. In particular, low temperatures, accurate temperature prediction is crucial because it supports early detection of potentially unsafe operating conditions, such as those associated with lithium plating. From such prediction, the battery management system must intervene before any damage to the LIBs occurs, thus improving both safety and performance in cold environments. This work introduces an enhanced physics-informed neural network (PINN) model that leverages thermal laws and entropy effects to accurately predict the surface temperature of commercial LIBs cycled at 1C at different temperature conditions (5 °C, 25 °C, and 45 °C). As an understudied temperature range, focus is given to low temperature cycling. Unlike prior PINN approaches, the proposed model incorporates Joule heating, reversible entropic heat, and radiative losses directly into its loss function. The model also uses adaptive weighting to balance physical constraints with data-driven terms. A charge-consistency term further enhances prediction stability by linking state of charge behavior to thermal dynamics, while enforcing unit consistency in the final loss function. The experimental results from three commercial cell chemistries (LFP, LCO, and NMC) show that the proposed model achieves mean absolute errors as low as 0.07 °C (LFP at 5 °C). Such performance was obtained when the model is trained and evaluated within the same chemistry using only 30% of the dataset. These results fall within the typical uncertainty range of thermocouple measurements (±1 to ±2 °C), demonstrating that the proposed approach is robust and shows close agreement with the measured temperature signals, thereby supporting improved safety and performance in lithium-ion battery systems.
Pereira et al. (2026) studied this question.