A novel parameter adaptive method for state of charge estimation of aged lithium batteries

Lithium-ion batteries have been identified as the most promising and suitable choice for commercial application in electric vehicles (EVs). Lithium-ion batteries have the benefits of low weight, excellent cycle performance, low self-discharge rate and high energy density. Generally, electric vehicles are integrated with battery management systems (BMSs), whose main function is to estimate the state of charge (SOC) of the batteries. This is important in ensuring not only the reliability of the battery operations but also the performance and safety of the car. Thus, an accurate and effective SOC estimation is imperative.

Numerous estimation methods have been developed. They are mainly based on models, ampere-hour counting, intelligent learning or open circuit voltage (OCV) characteristics based on their operating principles. Among them, the model-based methods, especially those using extended Kalman filter (EKF) and unscented Kalman filter (UKF), have drawn significant research attention owing to their robustness in measuring noises and estimating SOC. Unfortunately, their practical application is often restricted by their intrinsic Gaussian noise distribution problem.

Generally, model-based methods used for lithium-ion batteries can be grouped into two, namely, lumped equivalent circuit models (ECMs) and detailed electrochemical models. The former has been extensively investigated due to its low computation costs and better real-time performance. However, ECM methods face two main challenges when used to estimate SOC. First, it uses recursive least square (RLS) to identify different circuit component parameters. RLS has limited computational ability and does not adapt to mutations leading to errors. Second, they use experimentally obtained empirical functions of SOC to estimate SOC, which may lead to unreliable and inaccurate results under different degradation states since the evolution of OCV with respect to SOC changes as the batteries age.

To address these problems, Professor Depeng Kong, Mr. Shuhui Wang and Professor Ping Ping from China University of Petroleum (East China) developed a novel parameter adaptive method for SOC estimation in aged lithium batteries. This approach was based on an improved RLS method and neural-network-based OCV method that were used to identify the model parameters. The accuracy of SOC estimation with self-update ability was improved by designing an RLS with an activation zone (RLS-AZ). Additionally, the nonlinear relationship between SOC, OCV and battery capacity was developed using a back-propagation neural network (BPNN) to enhance the accuracy of OCV and OSC estimation. Lastly, the SOC estimation strategy was implemented by UKF via online identification of the process parameters. The work is currently published in the Journal of Energy Storage.

Results showed that when the error between the real and simulated voltages fell within the activation zone – defined as the interval described by the error between the measured and predicted terminal voltages – the covariance matrix of the RLS was reset. This allowed for effective definition of the SOC/OCV/battery nonlinear relationship, which further allowed the effective use of UKF to estimate the SOC under degradation accurately. Based on these improvements, the proposed methods produced better SOC estimation results than traditional methods, with a mean accuracy error of about 0.8%. For aging batteries, the SOC estimation error of the new method was about 0.3%, which was three times lower than that of traditional methods. Furthermore, it exhibited excellent characteristics suitable for voltage prediction.

In summary, applying a parameter adaptive method for accurate SOC estimation of an aging battery was demonstrated using an improved UKF-based SOC estimator coupled with improved RLS-AZ and BPNN techniques. The robustness of the proposed method was verified. With high accuracy and good robustness, this SOC estimation scheme is generally appropriate for electric vehicle applications. In a statement to Advances in Engineering, the authors noted that the study provided useful insights that would contribute to developing more advanced SOC estimation methods, especially those based on deep learning techniques.