摘要
基于数据驱动的电容式电压互感器(Capacitor Voltage Transformer,CVT)误差状态预测技术面临着数据不平衡,小样本数据不足的困境。针对该问题,整合DBSCAN聚类算法和基于径向的过采样算法(Radial-Based Oversampling,RBO),并在此基础上提出了一种引入权重分布RBO的过采样算法(Weight Distribution and Radial-Based Oversampling,WD-RBO)。该算法首先使用DBSCAN对CVT超差样本聚类,过滤数据集中的噪声;其次,通过簇密度分布函数和质心势能,计算超差样本各个簇的采样权重,确定每个簇中合成新样本的数量;最后,使用改进的RBO过采样算法合成新超差样本。实验结果表明,与SMOTE和RBO等过采样算法相比,WD-RBO使得预测模型的准确率和AUC值分别提高到0.967和0.996,且运行时长满足应用要求。WD-RBO可以有效解决类间不平衡和类内不平衡,提高模型对CVT误差状态的预测性能。
Data-driven capacitor voltage transformer(CVT)error state prediction technology faces the dilemma of data imbalance and insufficient data in small samples.To solve this problem,we integrated the DBSCAN clustering algorithm and the radial-based Oversampling algorithm(RBO),and on this basis,we proposed a Weight Distribution and RadialBased Oversampling algorithm(WD-RBO).The algorithm first uses DBSCAN to cluster the CVT out-of-tolerance samples to filter the noise in the data set;secondly,it calculates the sampling weight of each cluster of the out-of-tolerance samples through the cluster density distribution function and the centroid potential energy,and determines the number of new samples synthesized in each cluster;Finally,it uses the improved RBO oversampling algorithm to synthesize new out-of-tolerance samples.The experimental results show that compared with oversampling algorithms such as SMOTE and RBO,WDRBO improves the accuracy and AUC value of the prediction model to 0.967 and 0.996,respectively,and the running time meets the application requirements.WD-RBO can effectively solve the imbalance between the classes and the imbalance within the classes,and improve the prediction performance of the model for the CVT error state.
作者
梁东
石英
谢长君
刘红丽
孙宇峰
LIANG Dong;SHI Ying;XIE Chang-jun;LIU Hong-li;SUN Yu-feng(School of Automation,Wuhan University of Technology,Wuhan 430070,China)
出处
《武汉理工大学学报》
CAS
2021年第5期92-98,共7页
Journal of Wuhan University of Technology
基金
国家重点研发项目(2020YFB1506802)
作者简介
梁东(1995-),男,硕士生.E-mail:944839147@qq.com
