基于ADASYN和图像分析的非侵入式负荷辨识方法研究

Research on non-intrusive load identification method based on ADASYN and image analysis

中文关键词: 智能电能表  非侵入式负荷  自适应合成采样  马尔可夫变迁场  密集连接网络  负荷辨识

英文关键词: intelligent energy meter  non-intrusive load  adaptive synthetic  markov transition field  dense connectivity network  load identification

中文摘要:

      为推广应用智能电能表负荷辨识技术及解决传统非侵入式负荷辨识算法在不平衡采样数据上辨识精度较低的问题,提出了一种基于自适应合成采样（adaptive synthetic,ADASYN）和图像分析的非侵入式负荷辨识方法。通过马尔可夫变迁场（mar?kov transition field,MTF）编码将一维功率数据转换成二维MTF特征图像,作为图像识别网络的输入。基于密集连接网络（denseconnectivity network,DenseNet）的深层信息挖掘能力,将二维图像输入至DenseNet121网络中提取特征信息,实现负荷类型的辨识。基于ADASYN算法对不平衡数据集进行过采样处理,消除数据不平衡分布带来的模型学习偏见。算例结果表明,ADASYN算法能够很好地解决非侵入式负荷监测数据不平衡问题,相对处理前的辨识准确率和 F1 得分,分别提升了0.247和0.267；同时,MTF图像具有明晰易辨的特征信息,结合DenseNet121网络强大的深层特征捕捉能力,其辨识准确率与 F1 得分均能达到0.952,有效提升了在不平衡采样数据上非侵入式负荷类型的辨识精度。

英文摘要:

      In order to popularize the load identification technology of smart meters and solve the problem of low identification accuracy of traditional non-intrusive load identification algorithm on unbalanced sampled data, a non-intrusive load identification method based on adaptive synthetic（ADASYN）and image analysis is proposed. 1D power data is converted into 2D MTF feature images by markov transition field（MTF）coding, which is used as the input of image recognition network. Based on the deep information mining capability of dense connectivity network（DenseNet）, 2D images are input into DenseNet121 network to extract feature information and realize the identification of load types. Based on ADASYN algorithm, the unbalanced data set is oversampled to eliminate the model learning bias caused by the unbalanced data distribution. The results show that ADASYN algorithm can solve the non-intrusive load monitoring data imbalance problem well, and its identification accuracy and F1 score are increased by 0.247 and 0.267, respectively. At the same time, MTF images have clear and easily distinguishable feature information. Combined with the powerful deep feature capture capability of DenseNet121 network, the identification accuracy and F1 score can both reach 0.952, which effectively improves the identification accuracy of non-intrusive load types on unbalanced sampled data.