Temperature distribution learning of Li-ion batteries using knowledge distillation and self-adaptive models

Temperature monitoring and estimation are essential in battery thermal management systems to optimize performance and extend the lifespan of batteries in electric vehicles (EV) and stationary energy storage systems. The presence of various data-driven models, each reflecting a facet (or part) of the thermal distribution, calls for the need for a unified model giving the holistic distribution. Considering the finite computation power on board an EV, the unified model must not be too large. Given such constraints, this study presents a novel framework for learning the temperature distribution of Li-ion batteries, employing a knowledge distillation approach combined with self-adaptive control. The proposed framework addresses the limitations of traditional temperature calculation methods, i.e., the requirement of precise physical parameters and the lack of real-time adaptability. Our approach integrates multiple neural network models, including lumped physics-based and field image-based types, into a Principal model that merges physical processes with data-driven insights. This Principal model then distills its knowledge into a Student model optimized for deployment in resource-constrained environments, such as electric vehicles. Furthermore, an online self-adaptation mechanism enables the Student model to adjust to changing operational conditions without the need for retraining. The proposed framework significantly enhances the accuracy and efficiency of temperature distribution estimation in Li-ion batteries, improving the overall temperature monitoring system within battery management systems.