Research On SOC/SOH Estimation Algorithm Of Lithium-ion Batteries Based On Recurrent Neural Network

At present,promoting the development of pure electric vehicles is an important way to change China’s energy structure and promote the transformation and upgrading of transportation.Because of its own advantages,lithium-ion battery has been the first choice of power battery for pure electric vehicles.As the power source of pure electric vehicles,effective battery management is the core technology.State of charge(SOC)and state of Health(SOH)are both key technologies in battery management system(BMS).SOC reflects the remaining charge of the current battery.SOH is a quantitative index of battery aging degree,which is the basis of battery management and of great significance to ensure the safe and reliable operation of pure electric vehicles.Accurate and reliable SOC estimation is not only helpful for drivers to fully grasp the current remaining mileage situation and arrange driving strategy reasonably,but also of great significance for the whole system to formulate charging and discharging strategy reasonably and prolong the service life of battery itself.Accurate SOH estimation helps you learn about the current battery health status in real time,effectively avoiding faults and improving the security and stability of the entire system.Therefore,based on the estimation of SOC and SOH of lithium-ion batteries,the following studies are carried out in this thesis:(1)Research on lithium battery SOC estimation method based on bidirectional Gated Cyclic neural network(BiGRU).Firstly,BiGRU network is proposed to capture the nonlinear mapping relationship between battery measurable data(including voltage,current and temperature)and SOC for the timing of measurable data of lithium ion power battery.Then,based on the public data set of lithium iron phosphate battery from the Center for Advanced Life Cycle Engineering(CACLE)at the University of Maryland,DST conditions were used as training sets,US06 and FUDS conditions were used as test sets to verify the generalization ability of the proposed method.In order to fully reflect the estimation performance of BiGRU network,one-way gated recurrent neural network was trained and compared under the same conditions.The final experimental results show that compared with the one-way gated cyclic neural network,BiGRU has better estimation performance under different temperature and two working conditions.The average absolute error of BiGRU’s SOC estimation is about 1.1%,and it also has good generalization ability.(2)Research on SOC estimation method based on BiGRU and adptive particle filter(APF).According to the previous experiments,BiGRU network has a good estimation performance in predicting battery SOC,but due to the "plateau" characteristics of lithium iron phosphate batteries,the estimation of the middle segment fluctuates greatly.To solve this problem,this paper proposes a BiGRU network with adaptive particle filter to improve the stability of global estimation.Firstly,BiGRU network is used to learn the nonlinear mapping relationship between SOC and battery measurable data,and then adaptive particle filter(APF)is used to smooth the output of BiGRU network to achieve stable and reliable SOC estimation.This method simplifies the tedious process of BiGRU network parameter adjustment and does not need to build a complex battery model.Finally,US06 and FUDS conditions at different temperatures were used to verify the performance of the proposed method.The experimental results show that compared with the single BiGRU network,the average absolute error of the improved SOC estimation model is reduced by more than 50%,the maximum error converges to less than 2%,and it can quickly cope with the situation of inaccurate initial SOC value and different initial noise.(3)SOH estimation based on convolutional autoencoder(CAE)and gated recurrent neural network(GRU).In this paper,GRU network is proposed to estimate the macro-scale characteristics of SOH estimation.Firstly,CAE was used to extract the health features from the original battery data,and then the GRU network model was trained with the extracted health features to determine the battery health state prediction model.Based on the above process,this paper firstly analyzed the phenomenon and mechanism of battery aging based on NASA battery capacity decay data,providing theoretical support for subsequent battery health estimation.At the same time,combining with the actual working state of power battery,selecting phase constant current charge of the battery voltage range between 3.5 v～4.2 v voltage and temperature data of data analysis along with the change of battery aging trend,and according to the characteristics of non-uniform sampling data by linear interpolation method is uniformly distributed data,avoid the negative effect on the subsequent sequence processing model.Then,the processed health features were input into THE GRU to establish the lithium battery health state estimation model.Finally,NASA’s battery aging data were used to verify the results.The overall average error and root mean square error of the four estimated results were less than 1.2%and 1.6%,respectively.The results show that the proposed method can effectively track the trend of battery aging.