Research On SOC Estimation And Low Temperature Migration Of Lithium-ion Batteries Based On Optimized Neural Networ

Implementing carbon neutrality and carbon emissions peak policies requires a high-level electric vehicle field.Lithium-ion batteries have been considered an essential component of electric vehicle power batteries.State of charge(SOC)estimation affects the working strategy,safety performance and life of battery,and is a key problem to be solved at present.With feature extraction and fitting capabilities,neural networks can achieve accurate SOC estimation without considering the electrochemical state of the battery itself.Taking lithium-ion batteries as the research object,this paper discusses SOC estimation and low-temperature migration based on optimized neural networks in depth.Specific research contents are as follows:First of all,this study outlines the definition of battery SOC and its relationship with other battery states.And then primarily reviews the progress of neural networks in SOC estimation applications,including principles,advantages,disadvantages,current status,and estimation errors.Neural network methods are classified into three categories:feedforward neural network methods,deep learning methods,and hybrid methods.Secondly,the types and characteristics of lithium-ion battery models are described.The method flow of parameter identification of the Thevenin model and the detailed evaluation of all methods are introduced.The Kalman filter-first and neural network-first methods are proposed.The network-first method,namely Kalman filter optimization neural network,is determined as the core method of this paper.The related work of experiment and data is introduced,which lays the foundation for the following experimental verification.Afterwards,in order to reduce the difficulty of super parameter selection of neural network and improve the stability and robustness of estimation,a new hybrid estimation method for lithium-ion battery SOC is proposed.The method is composed of an improved bidirectional gated recurrent unit(IBGRU)network and an unscented Kalman filter(UKF).Using attention mechanism and Dropout to improve bidirectional GRU networks.The proposed method is experimentally validated using data from UDDS and US06 driving cycles.The verification results show that the method can adapt to various working conditions and obtain good estimation accuracy and robustness,with Mean Absolute Error(MAE)and Root Mean Square Error(RMSE)less than0.83%and 1.12%,respectively.In the next part,it is difficult for simple current,voltage and temperature to meet the requirement of estimation accuracy at low temperatures,but in order to avoid manual processing of original data.Convolutional neural network(CNN)is introduced to automatically extract the characteristic parameters of the original data to meet the needs of low temperature estimation research.CNN can automatically extract the feature parameters into the"multi-time input"structure.The accurate estimation of SOC at low temperatures can be achieved by adjusting the weights.It has an excellent performance in all four driving cycles.The performance of the CNN-IBGRU-UKF network is better than the traditional GRU network and CNN-GRU network.Both MAE and RMSE are less than 0.0127 and 0.0171.The R~2determination coefficient is more significant than 0.99.In practical applications,the initial SOC may not always be 100%.The network still has a high estimated performance at SOC=80%and SOC=50%.Finally,after transfer learning with fine-tuning strategy,the method can also be applied to new lithium-ion batteries and achieve good estimation performance at new temperature conditions.The maximum errors are 4.98%and 5.76%at 25°C and-10°C,respectively.The US06 data for Samsung 21700 at 25°C and-10°C are combined with transfer learning to achieve SOC estimation for this new lithium-ion battery in less time.The training time is about 0.45 hours,much less than the 2.5 hours of the complete training.The experimental results show that the optimized neural network method can accurately and stably estimate the SOC of lithium-ion batteries under wide temperature and dynamic operating conditions.And it can be applied to other types of lithium-ion batteries through transfer learning.