Prediction of battery cooling processes in electric vehicles by Physics-Informed Neural Networks

点击翻译按钮获取中文摘要

This project exploits the latest developments in the field of machine learning and computational modelling and uses the so-called 'Physics-Informed Neural Network' (PNN) to model the time-dependent cooling processes in real battery packs. This leads to substantial reduction in the computational cost and therefore allows for utilisation of the method as a predictive design tool in industry.Battery thermal management (BTM) is an ongoing major challenge hindering the wide application of electric vehicles (EV). It is well established that batteries can operate optimally only in a small temperature range, while there are a number of time-dependent factors that can heavily influence battery temperature. For example, the drive cycles and thus battery discharge scenarios are highly unsteady, turning BTM to a dynamic problem. Nonetheless, a vast majority of the existing models of battery cooling are limited to steady cases. More importantly, they are predominantly focused on a single or a small number of battery cells. Yet, in practice, an EV often uses thousands of cells. These shortcomings have contributed to the formation of a major gap between capabilities of the existing models and the practical needs of industry.Performance prediction of BTM systems requires careful consideration of electrochemistry, fluid dynamics and heat transfer. Addition of unsteadiness and complex configuration of a battery packs renders the task quite complicated. This makes the conventional modelling approaches expensive and therefore impractical. To resolve this issue, this project exploits the latest developments in the field of machine learning and computational modelling and uses the so-called 'Physics-Informed Neural Network' (PNN) to model the time-dependent cooling processes in real battery packs. This leads to substantial reduction in the computational cost and therefore allows for utilisation of the method as a predictive design tool in industry.The project includes generation of limited training and testing data by using CFD integrated with the existing electrochemical models. In parallel, PNN models are developed through combining deep learning techniques with the governing physical equations of the system. This builds upon a recently developed PNN-based flow simulator in our group and advances that to include electrochemistry. The PNN will be validated and refined using the data generated by conventional modelling.