Investigating The State-of-Charge Estimation Model For Lithium-Ion Battery Based On Temporal Convolutional Networks

State of charge(SOC)plays a fundamental role in energy management and control in Battery Management System(BMS)for lithium-ion batteries.Accurate estimation of SOC directly affects the reliability of electronic equipment using lithium-ion batteries.However,there are shortcomings in the existing SOC estimation algorithms,especially in accuracy and applicability.In order to improve the accuracy and generalization of SOC estimation,a SOC estimation model for lithium-ion batteries based on temporal convolutional networks was designed in this thesis.The proposed model showed excellent estimation accuracy in different battery data sets,as well as superior prediction accuracy in public CALCE data sets compared to most existing neural network-based SOC estimation models.The related research work is as follows:Firstly,the importance of accurate estimation of SOC for electric vehicle development was stated.From the perspective of traditional estimation,filter series and neural network series models,the advantages and disadvantages of the existing SOC estimation algorithms were analyzed and then the research objective of designing a SOC estimation model for lithium-ion battery based on temporal convolutional networks with higher accuracy and greater applicability was illustrated.Then Secondly,tetailed description of the internal components and functions in respect of the proposed time convolution,including the dilated causality convolution with memory,residual skips that prevent network degradation,activation function that increases the non-linearity and model framework that expands receptive field has been carried out.This paper,simultaneously,has analyzed the MAE,MSE,Huber loss function and Adam optimization algorithm that were employed in training model.Then,the non-linear relationships between the discharging current,working voltage,operating temperature,cycle index,internal resistance and the SOC of lithium battery have been verified respectively through batteries charge-discharge experiments,among which,the discharging current,working voltage and operating temperature have been selected as the temporal convolutional network input in according to the practical application requirements.The experimental battery discharge data with regard to the three kinds of operating conditions of DST,US06 and FUDS and the public battery data set of NASA and CALCE have been collected in order to train and verify the proposed model.Besides,the scrubbing and pretreatment methods of the battery statistics have been proposed for improving the training efficiency.Finally,the validity of the proposed model for estimating the SOC of lithium-ion batteries was proved through the CALCE battery data set.The estimation results obtained by the proposed method,with the best RMSE,MAE and MAPE being 0.30%,0.19% and 1.47% respectively,were superior to those by most existing neural networks based models.The generalization ability of the proposed model was proved through the NASA data set,and its estimation accuracy was equivalent to that of the estimation accuracy on CALCE data set.The accuracy of the proposed model for estimating the SOC in the fullcycle discharge of the lithium battery was proved through the discharge experimental data.Finally,based on the optimal hyperparameter combination as a benchmark,the negative impact of changing a single hyperparameter in the optimal hyperparameter combination on the model performance of estimation was explored.