A physics-guided neural network with battery constraint mechanism for state-of-energy estimation of lithium-ion batteries

Numerous studies have investigated the application of neural network models for state-of-energy (SOE) estimation. However, purely data-driven approaches fail to capture the underlying battery operation mechanism during modeling; hence the model parameters can only be optimized through a trial-and-error process, which limits the model's performance. This paper investigates a physics-guided neural network with a battery constraint mechanism for SOE estimation of lithium-ion batteries (LIBs). A physics-guided neural network incorporating the power integration method (PIM) is designed for SOE estimation, and an incremental PIM-based constraint is constructed using the reference SOE trajectory available during training. By analyzing the principles of model parameter optimization, an adaptive weight adjustment method is proposed for the PIM-based incremental constraint. During the model inference stage, a PIM-based monitoring mechanism is employed in the sliding-window, and a consistency score is defined to mitigate prediction drift during SOE estimation, thereby improving the reliability of online predictions from the pre-trained model. The proposed method is validated under various cycle profiles, including tests with 18650-type LIBs, electric vehicle power batteries, and real-world BEV data. The ablation studies further demonstrate its effectiveness.