Machine learning-based methods are widely adopted in the state-of-charge estimation of lithium-ion batteries due to easy application. However, they will sometimes cause a phenomenon of abrupt errors since they perform… Click to show full abstract
Machine learning-based methods are widely adopted in the state-of-charge estimation of lithium-ion batteries due to easy application. However, they will sometimes cause a phenomenon of abrupt errors since they perform the data mapping without considering the physical mechanism. Here, a physics-constrained neural network (NN) is proposed, which simultaneously minimizes the data mapping loss and also the physical constraints loss. Experimental results show that the problem of abrupt errors is significantly reduced in the proposed scheme compared with NN. Further analysis shows that the reduced error is attributed to the physical constraints between two consecutive time steps to force the estimation to follow the model equation. This article presents a combined data-model method, which also shows generalization capability to apply to other machine learning-based methods.
               
Click one of the above tabs to view related content.