State of charge estimation for lithium-ion batteries based on a digital twin hybrid model
Chao Ji, Guang Jin, Ran Zhang
Abstract
The estimation of the state of lithium-ion batteries is a critical aspect of battery management systems. However, the accuracy of such estimates deteriorates over time due to complex chemical reactions occurring within the battery as a result of repeated charging and discharging. To address this issue, this paper introduces a digital twin hybrid model (DTHM), integrating both an equivalent circuit model (ECM) and a neural network model (NNM). The method employs a residual technique to merge and enhance these models throughout battery operations, utilizing periodic operational outcomes to formulate calibration rules and dynamically calibrate parameters. Specifically, the augmented adaptive unscented Kalman filter method is applied within the ECM to reflect the dynamic behaviors and internal state of the battery. For the NNM, an importance sampling mechanism augments the training effect of neural network. Moreover, a dynamic calibration strategy, informed by digital twin technology, mitigates parameter uncertainties due to the internal evolution of the battery. Experiments conducted under diverse working and temperature conditions reveal that the DTHM achieves a synchronization error of less than 0.2% when aligning the physical and digital models, which attests to its high fidelity. Furthermore, at the early cycle, the DTHM demonstrated a maximum mean absolute error and a maximum root-mean-square error in estimating battery SOC of 0.0046 and 0.0058, respectively. Towards the end cycle, these errors were 0.0057 and 0.0065, respectively. Compared to statistical results from other state-of-the-art models, the DTHM consistently exhibits superior stability and robustness, particularly in estimating the state of batteries at various aging stages. • Introducing digital twins into battery field, a state estimation model framework responding to battery aging is proposed. • Two decision rules are proposed to distinguish the performance fluctuations caused by battery aging and model distortion, which solves the problem of model imbalance. • In the neural network model, a resampling method is proposed to generate auxiliary samples to prevent overfitting and improve the prediction accuracy of the model.