Study On Li-Ion Battery State Of Charge Estimation Based On Real-Word Evs Data

Lithium-ion battery has become the mainstream choice of electric vehicle power battery because of its high energy density and low self-discharge rate.State of Charge(SOC)is a key parameter in the battery management system.An accurate and reliable SOC value can not only relieve the driver’s mileage anxiety,but also prevent the battery from overcharging and overdischarging,thus prolonging battery life.Therefore,accurate prediction of battery SOC is the key to promote the further development of battery management system technology,which is of great significance to the management of automotive power batteries.The commonly used battery SOC estimation methods are based on laboratory data,which cannot correctly reflect the real operating conditions of the battery.Therefore,this thesis proposes an estimation method of lithium-ion battery SOC which is based on the real-word electric vehicles(EVs)data stored in the New Energy Vehicle Data Center.The main work is as follows.(1)The battery SOC values provided by the cloud platform are calculated by the on-board BMS system,and there is a certain error compared with the unknown real SOC values,and a method to correct the SOC of the dataset is proposed for this problem.First,the SOC value corresponding to the battery when it is fully charged is corrected to 100,and then the SOC value of the charging segment of the dataset is corrected.Then,temperature,charge/discharge rate and depth of discharge are jointly used as accelerated factors for battery pack capacity decay with cumulative cycle time.We propose an improved capacity decay model that is applicable to the real-world EVs data.The improved capacity decay model is then used to calculate the capacity decay coefficients of individual battery discharge segments.Finally,the SOC values of the discharged segments of the dataset are corrected based on the improved capacity decay model,and the corrected SOC is used as the label for neural network model training.(2)A method of data pre-processing is proposed.It includes processing with data missing values,outliers,noise points,and data magnitudes.Among them,the KNN algorithm is used to fill in the missing values in the dataset because of the time-series nature of the battery data.Box plots are used to visualize the distribution of different features,and the outliers are processed in terms of their physical significance.Experimentally,the dataset is denoised with sym8 wavelets.Finally,the data are normalized to remove the effect of magnitudes on the model during training.(3)A model of temporal convolutional network plus dual attention mechanism(DA-TCN)is proposed to achieve battery SOC prediction for the characteristics of large amount and high dimensionality of real-world EVs data in cloud platform.To make the model more sensitive to important information in the input feature dimensions and time series during the training process,an attention mechanism is applied to the data input dimensions before they are fed to the TCN layer.It can assign appropriate weights to the input feature dimensions in advance based on the correlation between the features and the prediction elements.In addition,by keeping the intermediate results of the training of the input sequences in the TCN layer,another attention mechanism is applied,which assigns attention weights to the time steps of the data,and then the weighted sequences are used as the input of the next layer.It is experimentally demonstrated that applying the dual attention mechanism on the temporal convolutional network can help to improve the accuracy and robustness of the model.(4)Based on two years of real operation data of a brand of electric vehicle.One year of the data is used as the training dataset,and the rest of the data is used as the test dataset.We compare and analyze the estimation results of the proposed model by the experiments under different combinations of hyperparameters,different time intervals,and different temperatures.The experimental results prove that the proposed model has high accuracy and excellent robustness.Finally,a comparison test with RNN,GRU,LSTM and BP neural network is conducted.These algorithms are commonly used to predict the battery SOC.The results show that the estimation results of the proposed model outperform most of other neural network models.Its optimal MAE,MSE,and RMSE can reach 0.56%,0.44%,and 0.67%,respectively.