Online SOC Estimation Of Lithium-Ion Battery Based On EKF-NARX Neural Network

Compared with traditional batteries,lithium-ion batteries have the advantages of high energy density and good power performance,and are widely used in electric vehicles and various energy storage systems.In the battery management system,the most important task is to accurately estimate the amount of remaining battery power,so as to avoid overcharging or over-discharging of the battery,damaging the battery performance or affecting the electric equipment.Since lithium ion is a nonlinear system,it is always a challenge to estimate the battery charge state accurately under complex working conditions.Model-based lithium-ion battery state estimation method has strong robustness,but under different working conditions,model parameters will produce deviations,affecting the estimation accuracy.Although the estimation method based on neural network can obtain higher accuracy,the estimation accuracy still decreases under non-training conditions.To solve the above problems,this paper combines neural network with equivalent circuit model,and optimizes the output value of the model by using extended Kalman filter,which not only achieves accurate estimation of the polarization voltage parameters of lithium battery,but also maintains high accuracy in online estimation of battery charge state under non-training conditions.The main research work completed is as follows:Firstly,the internal structure and charge-discharge principle of lithium-ion battery were studied.The charging and discharging experiment platform of lithium-ion battery was built,and the internal characteristics of lithium-ion battery such as open circuit voltage,ohmic internal resistance and polarization voltage were studied based on the HPPC discharge condition.The U_OCV-SOC curve and R₀-SOC curve were obtained by analyzing the variation trend of battery parameters under different charge states.Secondly,the equivalent circuit model and neural network model based on lithium-ion battery are analyzed,and the hybrid equivalent circuit model combining first-order equivalent circuit model and neural network is designed.The neural network is used to estimate the battery polarization voltage,and the structure of the equivalent circuit model is simplified.In order to estimate the polarization voltage,a NARX（Nonlinear Autoregressive with External Input）neural network with memory capacity for historical information was used to estimate the polarization voltage.In order to simplify the complexity of training conditions,NARX neural network was designed for the training of mixed charging-discharging conditions and continuous charging-discharging conditions.The off-line estimation of battery charge state was carried out for different NARX neural networks under DST and UN/ECE operating conditions,and the best network structure was obtained.Thirdly,aiming at the problem that NARX neural network is sensitive to the accuracy of input open circuit voltage,Extended Kalman Filter（EKF）is used to optimize the estimated value of the model’s charged state,reduce its estimation error,and convert it into open circuit voltage input to the model to estimate the polarization voltage at the next moment.Achieve the purpose of online estimation.The effectiveness of the algorithm is verified in training condition,DST condition and UN/ECE condition.Finally,the hardware circuit composition and code realization flow of the lithium-ion battery experiment platform are described.The design of microcontroller,battery terminal voltage and current acquisition module and SPI communication protocol and the design idea of control program are introduced.