基于改进非局部注意力模块的非侵入负荷辩识

Non-intrusive load identification based on improved non-local attention module

针对目前基于深度学习的非侵入式负荷辩识领域中存在的模型复杂度高、参数量大及获取长距离特征间依赖关系的能力弱等问题,提出一种基于注意力机制的轻量级负荷辨识模型.该模型以低时间维度的设备电流信息为输入,通过引入改进非局部注意力模块建模不同时间电流的特征关系,建立轻量级的时间残差卷积神经网络.在公开PLAID(即插设备标识数据集)和WHITED(全球家庭和工业瞬态能源数据集)上的实验表明:在设备识别率分别达到97.32%和99.32%的情况下,模型的计算量低至4×10^(5),且模型的参数量小于5.2×10^(4).

Aiming at the problems of high model complexity,large number of parameters,and weak ability to obtain long-range dependencies between features in the field of deep learning-based non-intrusive load identification,a lightweight load identification model based on attention mechanism was proposed.The model took the device current information of low time dimension as input,and built a lightweight time residual convolutional neural network by introducing an improved non-local attention module to model the feature relationship of currents in different times.Experiments on the public PLAID(plug-level appliance identification data set)and WHITED(worldwide household and industry transient energy data set)show that the model computational complexity is as low as 4×10^(5)and the number of parameters is less than 5.2×10^(4),when the recognition rate of electrical appliances reaches 97.32%and 99.32%,respectively.