随着建筑物能源消耗的不断升高,高精度与高泛化能力的非侵入式负荷监测技术的研究具有重大意义.针对当前负荷分解方法存在的问题,提出了一种基于多尺度特征融合与多任务学习框架的非侵入式负荷监测方法.将实例-批归一化网络与U形网络结合,提取总负荷数据的上下文信息,并利用跨越连接实现对不同尺度的细节特征与全局特征的融合.针对多特征特点,引入高效通道注意力网络,使模型聚焦重要特征.引入多任务学习框架与后处理操作,去除输出的假阳性片段,实现对目标电器的精准识别.将所提模型与几种代表性模型在 UK-DALE(UK domestic appliance-level electricity)数据集与 REDD(reference energy disaggregation data set)上进行对比实验,结果表明,所提模型的性能优于对比模型,具有出色的负荷分解能力与状态识别能力.
Multi-scale Feature Fusion and Multi-task Learning Architecture for Non-intrusive Load Monitoring
As building energy increases,the research of non-intrusive load monitoring(NILM)with high accuracy and generalization ability is of great significance.To address the problems of current NILM methods,a multi-scale feature fusion and multi-task learning architecture for NILM is proposed in this paper.We combine the instance-batch normalization net with UNet to extract the contextual information of the aggregated load data,and skip connections are used to achieve the fusion of detail features and global features at different scales.An efficient channel attention network is introduced for multiple features to make the model focus on important features.This paper introduces multi-task learning and post-processing operations to remove the false-positive segments,which realize the more accurate recognition of the target appliance.In this paper,the proposed model and several typical models are compared and experimented on the UK-DALE dataset and REDD dataset.The experimental results show that the proposed model outperforms the compared models and has outstanding load disaggregation and state recognition performance.
国家科技期刊平台
NSTL
万方数据
维普
