Temporal convolutional optimization network for SOC estimation of power batteries
Electric vehicles,as the main development direction of new energy vehicles,are entering a new stage of accelerated development.However,in the actual operating conditions of electric vehicles,factors such as driving mode,road conditions,aging of battery lines,and environmental temperature lead to significant errors in power batteries state of charge (SOC ) estimation.This can cause drivers to experience range anxiety,ultimately affecting the sales of electric vehicles.Therefore,it is crucial to embed an accurate algorithm for estimating SOC in the battery management system.At present,most of the research on SOC estimation are based on individual battery cells,while data of power batteries is multi-scale data composed of multiple battery cells.The accuracy of power batteries SOC estimation is not only affected by the internal characteristics of the battery,but also by the connection method,driving conditions,and usage mode.Then,battery data has temporal characteristics,and battery capacity will degrade over time.This poses a challenge for traditional machine learning algorithms or ordinary neural networks in capturing temporal features.Although existing studies employ RNNs,such as LSTM,for processing time series data,the serial nature of RNNs restricts their ability to process only one time step at a time.The computing resources and time will significantly increase with the increase of time steps.In addition,some studies use hybrid neural networks,such as CNN-LSTM.Because relying solely on CNN for network modeling is not ideal,it is often set at the forefront of the neural network for feature extraction.However,this method disrupts the temporal structure of the data,causing LSTM to lose the significance of extracting temporal features.To address these challenges,this paper proposes a temporal convolutional optimization network (TCON)for real-time SOC of power batteries.Firstly,a non-normalized temporal convolutional network (TCN)model is established.It can perform parallel computation to extract temporal information,with the advantages of fewer parameters and high accuracy.Secondly,a time optimization module (TOM) is designed to address the issue of significant fluctuations in TCN output.It optimizes the TCN output by generating timing optimization weights,effectively suppressing data noise and further improving estimation accuracy.Subsequently,this paper adopts a joint analysis method that simultaneously considers vehicle state,driver behavior,and battery system parameters to estimate SOC.Finally,the Spearman correlation coefficient method is utilized for screening parameters,while the Hyperband optimization algorithm is employed to determine the model's hyperparameters.The model is validated using real-time operational data from electric vehicles,and experimental results show that compared to TCN,the proposed method reduces the error by 18.3％ with only a 5.8％ increase in parameters.The model achieves a mean absolute error of less than 1％ and a root mean square error of less than 2％ in SOC estimation,thereby alleviating driver range anxiety to a certain extent.
power batteriesSOC estimationrange anxietytemporal convolutional optimization network
国家科技期刊平台
NSTL
万方数据
