Explainable real-time data driven method for battery electric model reconstruction via tensor train decomposition

The global energy transition is based on renewable energy sources and batteries to store electrical energy. Efficient use of batteries requires accurate state estimation algorithms and proper control. In this paper, a data-driven method for estimating a battery dynamic model using a Tensor Train (TT) is designed and tested. The method intrinsic efficiency in handling high-dimensional data allows one to develop an algorithm, that can use a batch of observable variables to reconstruct a system dynamics with negligible computational costs and reliable accuracy, as well as without any preliminary characterization test data. Here, the method is applied to reconstruct a dynamic battery model from operational data and tested upon a solid state lithium-ion battery cell. Furthermore, the explanatory power of the method is demonstrated by extracting the open circuit voltage and the impedance in the form of a relaxation times distribution, as well as the activation energy related to temperature dependence, and its accuracy is further validated against the results of standard battery characterization tests. Due to intrinsic scalability and low computational costs, the method can become a part of AI-driven battery management systems, thus improving battery durability and safety, as well as helping to optimize time-consuming battery characterization tests. • TT-based method allows to reconstruct a battery electric model with error within 5%. • Battery operation data only are used without time-consuming characterization tests. • Meaningful information (OCV, impedance, and activation energy) is extracted. • The method computational efficiency allows one to use it in real-time systems.