A novel hybrid machine learning coulomb counting technique for state of charge estimation of lithium-ion batteries

Accurate state of charge (SOC) estimation is crucial to ensure safe and reliable operation of battery systems. However, the capacity of an operating battery is difficult to measure, and some existing SOC estimation methods either rely too much on the established battery physical model, such as the Kalman filter and particle filter algorithm, or the number of training samples required is too large, and the online estimation performance is poor, such as neural network and Gaussian regression algorithm, cannot achieve good online estimation results. To solve these problems, this paper proposes a novel hybrid SOC estimation method, movIRVM-Coulomb, which integrates moving mean, incremental learning, relevance vector machine (movIRVM), and coulomb counting method. The hybrid algorithm uses the movIRVM algorithm and Coulomb counting method to estimate SOC alternately so that the estimated value is kept in the accuracy of the first estimation, which avoids the error accumulation problem when using a single algorithm to predict multiple times and improves the generalization performance of the algorithm. Firstly, the offline estimation model based on the RVM algorithm and the online estimation model based on the incremental learning RVM(IRVM) algorithm are established by using only 10 % to 30 % training data approximately. Secondly, aiming at the problem that the estimation result of the RVM algorithm fluctuates violently, this paper combines the moving average strategy to make the geometric average of SOC value by setting a reasonable moving window. This method further improves the estimation accuracy, fitting ability and generalization performance of the RVM algorithm. Finally, jointing the coulomb counting method to further improve the online estimation accuracy. The effectiveness of the proposed algorithm is verified by a single-cell public data set, battery pack experimental test data set and Advisor simulation data set. The experimental results highlight that the estimation accuracy of the movIRVM-Coulomb algorithm is >30 % improved compared with that of the movIRVM algorithm and its root means square error can be restricted within 2 % in a wide temperature range and variable dynamic conditions, including operating conditions of US06, UDDS, NYCC and 1015. Overall, it is manifesting that the movIRVM-Coulomb algorithm requires only a few training samples to have high estimation accuracy, strong generalization ability, and excellent robustness.