基于一维卷积神经网络的非侵入工业负荷事件检测方法

Non-intrusive industrial load event detection method based on one-dimensional convolutional neural network

针对非侵入式工业负荷事件检测中准确率较低和漏检率较大的问题,提出了一种基于一维卷积神经网络(1D-CNN)的非侵入工业负荷事件检测方法。所提方法在1D-CNN模型中引入Inception-V2模型构建一维Inception-V2卷积神经网络(1D-Inception-V2-CNN)模型,利用多种长度的滑动窗和对应的卷积核实现对数据的读取和压缩,利用1D-Inception-V2-CNN模型对压缩后的数据进行检测和分类,并通过自适应循环检测方法更新网络模型的检测样本库,最终实现对工业用户负荷数据的全面检测。在对实际工业用户的事件检测实验中,所提检测方法的准确率和Fscore分别达到了96.32%和95.42%,与LeNet一维卷积神经网络、二维卷积神经网络和滑动窗累积和算法相比均有明显的提升。实验结果表明,所提方法能够有效地提高工业事件检测的准确率,同时减小事件漏检率。

For low accuracy and high missed detection rate of non-intrusive industrial load event detection,a nonintrusive industrial load event detection method based on One-Dimensional Convolutional Neural Network(1D-CNN)was proposed.In the proposed method,the Inception-V2 model was introduced into the 1D-CNN model to build the OneDimensional Inception-V2 Convolutional Neural Network(1D-Inception-V2-CNN)model.Multiple sliding windows of various lengths and the corresponding convolution kernels were used for reading and compressing the input data.1DInception-V2-CNN model was used for detecting and classifying the compressed data.And an adaptive loop detection method was designed to update the detection sample database of the network model.Finally,the industrial user load data was comprehensively detected.In the event detection experiment of actual industrial users,the accuracy and F-score of the detection method reached 96.32%and 95.42%,which were obviously improved compared with LeNet-one-dimensional convolution neural network,two-dimensional convolution neural network and sliding window cumulative sum algorithm.The experimental results show that the proposed method can effectively improve the accuracy of industrial event detection and reduce the missed detection rate of event.