摘要
锂离子电池作为储能系统的关键,准确的电池模型对提高状态估计精度至关重要。该文提出一种基于动量反向传播神经网络的锂离子电池分数阶模型建模与荷电状态估计方法。在分数阶等效电路模型的基础上,构建具有半物理意义的神经网络模型,并使用动量反向传播算法,实现了对包含模型分数阶阶数的模型参数优化以及状态估计。为了验证所提算法的可行性,分别通过理论推导与实验结果证明。通过理论推导,证明所提方法误差会收敛。实验结果显示,相比于扩展卡尔曼滤波与无迹卡尔曼滤波算法,所提方法具有更高的精度,能够准确地预测电池的电压响应。综上所述,基于动量反向传播神经网络在锂离子电池建模和荷电状态估计中具有良好的应用前景,可为电池的性能评估和管理提供有效支持。
Boosting the adoption of new energy electric vehicles and energy storage technologies is pivotal in expediting the standardization of the"dual carbon"objectives.Lithium-ion batteries,serving as critical components in electric vehicles and energy storage systems,are indispensable for energy storage and buffering purposes.The precision of a battery model is vital for enhancing state estimation accuracy.Based on momentum back propagation neural networks,this paper introduces a fractional-order modeling and state-of-charge estimation technique for lithium-ion batteries,aiming to improve the interpretability of data-driven modeling.Firstly,a BPM neural network(BPMNN)-based battery model is constructed,which incorporates the physical information of a first-order fractional-order model.The weights of the BPMNN-based model represent the model parameters,including resistance,capacitance,and fractional order.With the hierarchical feature extraction capability of neural networks,the novel battery model enables the online optimization of fractional order.Secondly,the OCv-soC relationship is linearized using a Taylor expansion,and the linearized OCV-SoC relationship is employed to reconstruct the BPMNN-based battery model.Thirdly,a momentum gradient descent algorithm is utilized.Historical data is introduced using momentum,which improves the stability of the estimation.Moreover,the convergence of the proposed method is analyzed.The loss function of the BPMNN-based battery model converges to a positive value with a suitable learning rate.Finally,the proposed method is verified with a 2.55 A·h LiFePO4battery.The influence of temperature is compared with a UDDS profile at-5℃,25℃,and 45℃.The result shows that the external temperature influences the ohmic resistance,model accuracy,and fractional order.However,it seems to have no relationship with the SOC estimation.The fractional order descends faster as the temperature decreases.Compared to the EKF-based and UKF-based methods,the proposed method improves the model accuracy by up to 64%,and the SOC estimation accuracy is improved by up to 36.5%.Compared to the GRU-RNN-based SOC estimation method,the BPMNN-based method yields a better result in SOC estimation because the combination of physical information enables online optimization of the BPMNN-based battery model.As the memory space increases,the computational burden also grows,and the battery model becomes more accurate.The model achieves a balance between accuracy and computational burden when the memory space includes 50~70 steps.The following conclusions can be drawn.(1)The BPMNN-based online estimation method converges with a suitable learning rate,significantly improving the accuracy of the battery model and SOC estimation.The BPMNN-based method estimates SOC more accurately than the GRU-RNN method.(2)The external temperature influences the variation of ohmic resistance and model accuracy,but the SOC estimation is not affected.The fractional order descends more rapidly as the temperature decreases.(3)The memory space influences the computational burden and model accuracy.
作者
田佳强
张泽培
潘天红
黄大荣
刘兴华
Tian Jiaqiang;Zhang Zepei;Pan Tianhong;Huang Darong;Liu Xinghua(School of Electrical Engineering and Automation Anhui University,Hefei 230039 China;School of Artificial Intelligence Anhui University,Hefei 230039 China;School of Electrical Engineering Xi’an University of Technology,Xi’an 710048 China)
出处
《电工技术学报》
北大核心
2025年第24期8156-8170,共15页
Transactions of China Electrotechnical Society
基金
国家自然科学基金(62203352)
国家资助博士后研究人员计划(GZB20230001)
安徽省重点研发计划(202304a05020046)资助项目。
关键词
锂离子电池
分数阶模型
参数辨识
荷电状态估计
动量反向传播神经网络
Lithium-ion battery
fractional order model
parameter identification
SOC estimation
back propagation neural network with momentum
