基于物理信息神经网络的牵引变流器直流支撑电容参数辨识方法  

Parameter Identification of DC-Link Capacitor in Traction Converter Based on Physical Information Neural Network

作　　者：向超群[1] 尹雪瑶 伍珣 曹忠林 刘元才 Xiang Chaoqun;Yin Xueyao;Wu Xun;Cao Zhonglin;Liu Yuancai(School of Transportation Engineering Central South University,Changsha 410075 China;Tianjin Rail Transit Operation Group Co.Ltd,Tianjin 300000 China)

机构地区：[1]中南大学交通运输工程学院,长沙410075 [2]天津轨道交通运营集团有限公司,天津300000

出　　处：《电工技术学报》2024年第15期4654-4667,共14页Transactions of China Electrotechnical Society

基　　金：国家自然科学基金资助项目(52072414)。

摘　　要：为了解决车载牵引变流系统直流支撑电容器故障预测问题,该文提出一种基于物理信息神经网络的直流支撑电容器参数辨识方法。该方法只需要利用直流环节预充电过程的直流支撑电容器两端电压及采样频率,无需拟合曲线,无需严格对齐时间轴就可以获得较为准确的电容参数辨识结果。与此同时,为了克服在采集数据时因条件所限造成的数据量稀疏与分布不均问题,该文利用循环一致性生成对抗网络算法增强数据,使该方法可以适用于同一拓扑下宽范围电容区间的电容容值预测,降低了模型训练要求。实验结果表明:在正常条件下,该方法的辨识相对误差约在1%以下,并且降低采样频率能够缓解信噪比对该方法的影响。该方法为解决直流支撑电容参数辨识问题提供了新思路。At present,the fault monitoring dilemma of DC-link capacitors in traction converters mainly focuses on the noise interference of sensor measurement,the aging speed of the capacitor is affected by various environmental factors,and the accuracy and stability requirements are high.Therefore,this paper proposes a DC-link capacitor parameter identification method for traction converter based on physical information neural network(PINN)and capacitor pre-charging model.The sampling frequency of the voltage sensor is very low,and there is no need for the capacitor pre-charging voltage curve to be strictly aligned with the time axis,effectively reducing the influence of the measurement signal-to-noise ratio on the prediction results.Moreover,the amount of capacitance data required is minimal using the cycle consensus generative adversarial network(CycleGAN)algorithm,which can be applied to a wide range of capacitance intervals under the same topology.This method can be applied to rail transit applications.The architecture design of the adaptive physical information neural network model in this method is described,including the construction of partial differential equations and boundary conditions,physical constraint loss function modeling,network structure,and training process.Regarding the adaptive weight PINN execution part,the AdamW algorithm is used to update the adaptive weight and linear network weight of each loss component.After the number of iterations is set,the L-BFGS algorithm performs a new round of iterative optimization on the linear network weight part of the current PINN network.However,the adaptive weight is no longer updated.During the training process using the PINN network,when the number of iterations is greater than the specified number limit and the mean square error of the test set is less than the set number,the CycleGAN model is used to generate data,and the current PINN network is used as the constraint condition label.After adding the generated training set,the training results are monitored b

关 键 词：直流支撑电容器 参数辨识 物理信息神经网络 循环一致性生成对抗网络 直流 环节预充电工况 

分 类 号：TM464[电气工程—电器]