Study On Battery State-of-charge Estimation Methods Based On Bayesian Optimazation And Transfer Learning

With the increasing number of fossil fuel vehicles in China,the environmental pollution caused by vehicle exhaust emissions has been becoming more and more serious.Replacing fossil fuel vehicles with electric vehicles turns out to be an important way to alleviate environmental pollution and foreign oil dependence.In Hainan province,the government has planned to stop selling fossil fuel vehicles by2030.To develop electric vehicle(EV)technology has been the main trend.Reliable EV battery management technology is one of the crucial technologies for reliable EV technology.As an important characteristic index for battery management system,state of charge(SOC)plays an important role in evaluating EV’s remaining driving range,estimating battery health,balancing battery charging state,and predicting battery life.With the wide application of machine learning in various fields,the data-driven method has become one of the important methods for SOC estimation.Compared with the traditional SOC estimation methods,the data-driven SOC method can well reflect the relationship between the battery data and the battery state as long as it has sufficient and representative data.This work focuses on the SOC estimation method of EV batteries based on recurrent neural network(RNN).Bayesian optimization and transfer learning are employed to ensure the accuracy of the RNN model and reduce the demand of the model on the amount of training data.In current research of SOC estimation methods based on RNN,model optimization is mainly embodied in hyper-parameter optimization.The influences of hyper-parameters variation on model performance are studied respectively.Then the appropriate hyper-parameters are selected as the optimal parameters.This method can easily make the model approach a local optimum compared with Bayesian optimization.Based on the battery data,Bayesian optimization and transfer learning in this paper are employed to optimize hyper-parameters and modeling process of the RNN model,respectively.To solve the problem of battery SOC estimation,a process of battery SOC estimation based on RNN is designed in this work.The battery SOC estimation is achieved by two stages: data preprocessing and battery SOC estimation.In data preprocessing stage,K-nearest neighbor method,linear interpolation method and normalization are used to preprocess the input data.This preprocess method can solve the problem of missing battery data and uneven distribution.In the estimation stage,SOC estimation models at different temperatures were trained.Firstly,Bayesian search was applied to optimize the hyper-parameters of the model to improve the accuracy of the model under the limited number of training data.Then,transfer learning is employed to optimize the modeling process and reduce the amount of training data required by the model.Multiple battery experiment results show that the Bayesian optimization and transfer learning methods can optimize the traditional SOC estimation methods based on RNN,which can not only achieve a more accurate estimation of SOC,but also reduce the dependence of the training data amount.When the amount of training data is limited,the modeling strategy presented in this work can obtain a better estimation result.In addition,the modeling method in this work can be extended to battery packs or other types of batteries and has practical value.