Transformer Based State Of Charge Estimation For Lithium-ion Batteries

As one of the most commonly used batteries in electric vehicles,lithium-ion batteries are an important component of electric vehicles,and their state and performance are inseparable from real-time monitoring and control by the battery management system.State of Charge(SOC)estimation of lithiumion batteries is an important part of the battery management system.Accurate SOC estimation enables the battery management system to fully perform its functions,accurately predict the remaining range of electric vehicles,prevent overcharging and over-discharging of batteries,effectively prevent battery aging,and play a key role in the safe operation of electric vehicles.Currently,one of the mainstream methods for SOC estimation is based on data-driven deep learning methods,which often use recurrent neural networks.However,its main shortcomings include inapplicability to complex operating conditions,insufficient accuracy in low-temperature state estimation,and poor applicability to lithium-ion batteries with different chemical properties.The main reason for this is the insufficient attention of the recurrent neural network structure to the long-distance correlation of time series data.The emergence of Transformer models provides an opportunity to solve these problems.This paper proposes a SOC estimation algorithm based on an improved Transformer model,which attempts to improve the estimation accuracy under complex operating conditions and low-temperature states,and enhance the applicability to lithium-ion batteries with different chemical properties.The main contribution of this paper are as follows(1)Obtain dynamic charge-discharge data of two mainstream batteries,including Lithium Iron Phosphate battery(experiment)and Lithium-ion battery with ternary cathode(open source),under various temperature conditions.Analyze the changes in current,voltage,temperature,and SOC curves under complex operating conditions,and provide insights for further research.(2)Based on the Transformer architecture,improvements have been made to the SOC estimation,and comparisons have been made with other algorithms under different temperature,dynamic driving conditions,and different material batteries.The results validate that the proposed method has better accuracy,robustness,and generalization performance for SOC estimation.In particular,significant improvements in accuracy have been achieved in low-temperature conditions.(3)To address the problem of poor adaptability of the model to lithium-ion batteries with different chemical properties,a pre-training model was constructed,and transfer learning was performed on batteries with different chemical properties.By comparison,it was demonstrated that the learned weights during pre-training could be transferred to new lithium-ion batteries with different chemical properties,resulting in almost identical performance.This indicates that the model has good generalization ability and strong adaptability.This paper proposes a Transformer-based SOC estimation method for lithium-ion batteries,which overcomes some of the limitations of traditional recurrent neural network structures.The maximum estimation error at low temperature(0℃)is reduced from 12.6%(LSTM)to 3.6%(Transformer),and the average absolute error is reduced from 1.8%to 0.98%.The maximum estimation error under complex operating conditions at multiple temperatures is reduced by 3.33%,and the average absolute error is reduced by 0.82%.The estimation accuracy is significantly improved.The lithium-ion battery is used as the research object in this study.