基于颜色编码的非侵入式负荷细粒度识别方法

Non-intrusive Load Fine-grained Identification Based on Color Encoding

在非侵入式负荷识别任务中,仅使用单一负荷特征对设备进行辨识时,存在特征重叠现象,无法满足对设备进行细粒度分类的需求。为此,该文提出一种基于颜色编码的非侵入式负荷细粒度识别方法。首先,采用Fryze功率理论将高频采样电流分解为有功、无功电流,并对高频采样电压与无功电流进行标准化处理以构建二维U-I轨迹图像。然后,利用颜色编码技术对轨迹图像进行处理,在R、G、B三通道中分别融合有功电流、轨迹变化信息及瞬时功率,得到彩色U-I轨迹图像。最后,构建卷积神经网络,对彩色U-I轨迹图像进行特征提取,实现对设备的分类。在此基础上,文中提出了一种自主学习方法,实现对负荷识别模型自主更新。使用PLAID和WHITED数据集对本算法的识别效果及自主学习方法进行测试。结果表明,文中方法使得U-I轨迹所携带的信息量增加,增强了负荷特征的独特性,从而实现对设备的细粒度识别;自主学习方法能够学习新型电器并更新模型,提升了负荷识别模型场景适应能力。

In the task of non-intrusive load identification,only using a single load feature to classify devices may appear overlapping characteristics,which does not meet the requirements for the fine-grained classification of the devices.Therefore,a non-intrusive fine-grained load identification based on color encoding is proposed in this paper.The Fryze power theory is firstly used to separate the high-frequency sampling current into an active and a reactive component,and the high-frequency sampling voltage and the reactive current are standardized to construct the two-dimensional voltagecurrent(U-I) trajectory images.Then the trajectory images are processed by the color encoding technology,and fused with the active current,the trajectory change information and the instantaneous power respectively in the R,G,and B channels to acquire the color U-I trajectory images.Finally,the convolutional neural network is constructed to extract the features of the color U-I trajectory images,realizing the classification of the devices.On this basis,a self-learning method is proposed to realize the autonomous updating of the load identification model.The PLAID and WHITED datasets are used to test the recognition effect of this algorithm and the self-learning method.The experimental results show that the proposed method increases the amount of information carried by the U-I trajectory,enhances the distinguishability of load features,and realizes the fine-grained identification of devices.The self-learning method can learn new electrical appliances and update the model,which improves the scene adaptability of the load identification model.