基于随机森林的高压电缆局部放电特征寻优

高压电缆局部放电(简称局放)新特征的构建与优化选择是提升识别精度、优化识别效率、增强监测参数可视化效果的重要手段。提出了一种基于随机森林的局放特征优选新方法。在实验室构建了5种类型的电缆人工缺陷,通过加压测试获取局放原始数据,并提取了3500个局放脉冲和3500个典型干扰信号脉冲,构建了1235个局放特征。基于上述样本,开展了基于随机森林的特征寻优,分别获得了局放和干扰信号特征排序结果和不同类型局放信号的特征排序结果,并通过反向传播神经网络(back propagation neural network,BPNN)和支持向量机(support vector machine,SVM)对优选排序结果进行了验证。结果表明,局放和干扰识别的有效特征参数主要是表征信号快慢的特征和小波组合特征;不同类型局放识别的有效特征参数主要是小波组合特征。结果证明,随机森林算法是一种有效的电缆局放特征优选方法,并有望推广到其他电力设备局放的特征寻优。

Random Forest Based Feature Selection for Partial Discharge Recognition of HV Cables

New feature construction and optimal feature selection of partial discharge(PD) pattern recognition for HV cables contribute not only to improvement of pattern recognition accuracy and efficiency, but also to PD parameter visualization of HV cable condition monitoring and diagnostics. In the paper, a novel random forest(RF) algorithm based optimal feature selection for PD pattern recognition of HV cables is proposed. Based on five types of artificial defects of 11 kV ethylene-propylene(EPR) cables, 3500 sets of transient PD pulses and 3500 sets of typical interference pulses are extracted. 1235 features in total are extracted. Then the RF based optimal feature selection for PD pattern recognition of HV cables is carried out. The feature ranking results of PD signals and interference signals as well as the ranking results of different PD signals are obtained and evaluated with pattern recognition methods based on back propagation neural network(BPNN) and support vector machine(SVM). Results show that,the wavelet combination features and the features describing"fast pulses" and "slow pulses" are effective features for identification of PD and interference signals. The wavelet combination features are attested to be effective for recognition of different types of PD signals. The RF method is proven effective for PD features selection of HV cables and is prospective to be applied to PD feature selection of other HV power apparatus.