Lithium-ion Battery Modelling And SOC Estimation Based On NARX Neural Networks

Due to the high energy density and long cycle times,lithium-ion batteries have become one of the most widely used energy storage devices in new energy industries such as electric vehicles and wind power factory.In order to ensure the safety and reliability of the battery pack as well as prolong the actual usable life of the battery pack,the battery management system is required to manage the battery state.In a battery management system,accurate and real-time estimation of the state of charge(SOC)of each cell in the battery pack is the core function for the battery management system to implement an efficient energy balance control strategy,which needs a a precise battery model as the prerequisite.The existing battery equivalent circuit models have the problem of parameter identification and the difficulty in keeping track of the dynamic characteristics of the battery under complicated conditions.In order to solve the problem of more accurately modeling the nonlinear polarization voltage inside battery,lots of experiments were carried out to study the characteristics of polarization effect in this paper.Based on which,a nonlinear autoregressive exogenous(NARX)neural network is used to replace the higher-order RC networks in the traditional equivalent circuit model.Using the proposed method,the parameter identification problem of the traditional equivalent circuit model is then transformed into the problem of training the NARX neural network with massive working data.The generalization accuracy of the trained model is verified by long-time random current excitation.The experimental results show that the equivalent model based on NARX neural network simulates the polarization voltage variation of lithium-ion battery better than the traditional equivalent circuit model.Based on the proposed battery equivalent model,the state equation and the observation equation are established,and then the extended Kalman filter is used to estimate the battery SOC.Under the cyclic dynamic stress test condition,the maximum estimation error of SOC is 3.5%,which indicates that the extended Kalman filtering algorithm based on the NARX neural network model has high SOC estimation accuracy.