复杂噪声环境下电能质量复合扰动特征选择

Feature selection of composite power quality disturbances under complex noise environment

为满足复杂噪声环境下海量电能质量信号高效分类需要,提出一种可应用于复杂噪声环境的电能质量信号特征选择新方法。首先,采用贝叶斯优化方法优化随机森林参数;之后,以具有随机噪声的原始特征向量训练随机森林分类器,训练过程中通过Gini重要度量化比较各特征分类效果;然后,采用序列前向搜索方法,以分类准确率为主要标准,兼顾特征维数,确定最优特征子集;最后,以最优特征子集建立随机森林分类器,识别15种电能质量信号。仿真对比实验证明,在信噪比30 dB以上噪声环境下,新方法分类准确率在99.33%以上,20 dB噪声环境下分类准确率为94.60%。此外,通过葡萄牙某配电网实测电能质量数据开展实验,证明了新方法在实际工业应用中的有效性。

In order to meet the needs of high efficiency classification of complex power quality signals,a new feature selection method of power quality signal under the complex noise environment is proposed. Firstly,the parameters of random forest（ RF） are optimized by Bayesian optimization method. The original feature vectors with random noise are used to train RF classifiers,in which the classification ability of features are compared by Gini importance. Then,the sequential forward search method is utilized to obtain the optimal feature subset based on the classification accuracy recognized by RF and feature dimension. Finally,RF classifier is set up by the optimal feature subset and applied to identify 15 kinds of power quality signals. The simulation results show that,the proposed method can achieve the classification accuracy above 99. 33% and 94. 60% under the condition of 30 dB and 20 dB,respectively. Besides,the experimental results with real power quality data from Portugal distribution network prove that the effectiveness of the propsoed method for practical industrial applications.