基于Attention机制优化CNN-seq2seq模型的非侵入式负荷监测

Non-intrusive Load Monitoring Based on CNN-seq2seq Model Optimized by Attention Mechanism

非侵入式负荷监测是智能用电的一个关键环节。本文提出了一种基于注意力机制优化的卷积神经网络-序列到序列模型,用于提高非侵入式负荷监测的分解精度。首先,通过K-means++算法对各设备进行聚类操作,并将其转换成对应的独热码;其次,通过卷积神经网络结构提取总负荷数据的特征,并利用序列到序列结构捕捉负荷特征的时序依赖关系;最后,使用注意力机制优化卷积神经网络-序列到序列模型的整体结构。该模型通过将时间序列数据与设备状态编码之间形成一一映射关系,简化了模型结构。借助于能关注负荷重要特征的注意力机制,提升了模型的分解精度,并使用AMPds2开源数据集验证了模型的有效性。

Non-intrusive load monitoring(NILM)is a key link for intelligent power consumption.In this paper,a convolutional neural networks-sequence-to-sequence(CNN-seq2seq)model optimized by the Attention mechanism is proposed to improve the decomposition accuracy of NILM.First,the K-means++algorithm is used to cluster various kinds of equipment and convert them into the corresponding one-hot code.Then,the features of total load data are extracted by the CNN structure,and the time-series dependence of load features is captured by the seq2seq structure.Finally,the Attention mechanism is used to optimize the overall structure of the CNN-seq2seq model.This model simplifies its structure by forming a one-to-one mapping relationship between the time series data and the state coding of equipment.With the Attention mechanism which can focus on the important features of load,the decomposition accuracy of the model is improved,and its effectiveness is verified by using the AMPds2 open-source data set.