Theoretical model-inspired ensemble learning for maximum available energy estimation based on re-constructed driving patterns for solid state lithium batteries

Solid-state batteries are considered to be one of promising future battery technology due to their high energy density. The maximum available energy determines the battery operating state and boundaries, where the precise estimation can effectively enhance the vehicle range estimation capability, alleviating users range anxiety. To address the issue of low estimation accuracy of solid-state battery maximum available energy due to the lack of consideration of driver behavior, in this article, we propose an integrated learning framework guided by theoretical models named Bidirectional Gated Recurrent Unit-Adaboost ensemble learning model. 11 types of driving pattern features are quantified based on historical driving data, and then use a self-attention transformer network to derive specific equivalent driving cycles for constructing user profile. Finally, the equivalent circuit model combined with the user driving cycles provides reference values for ensemble learning, which are used as one of the inputs to guide the training of the final Adaboost learning framework. This article highlighted an integrated learning framework based on theoretical models, which leverages existing empirical knowledge and theoretical models to drive the training of complex networks. Owing to the guidance from theoretical input during training, the network would not far imprecise when issuing the dataset without pre-training, thereby enhancing the interpretability of the neural network. The presented method contributes to improve the precision on battery available energy, and helps to enhance the precision on estimation of driving range. It hopes that this method may offer an alternative approach to coupling traditional models with innovative yet opaque neural networks and be applied in engineering practices aimed at addressing issues such as range anxiety.