计算机科学
GSM演进的增强数据速率
传输(计算)
学习迁移
传递函数
边缘检测
负荷管理
人工智能
实时计算
无监督学习
电子工程
功率(物理)
工程类
控制工程
信号处理
状态监测
最大功率转移定理
作者
Taoyu Lu,Y Li,Ruiyang Yao,Yi Wang
标识
DOI:10.1109/tsg.2026.3699466
摘要
Non-intrusive load monitoring (NILM) is crucial for residential real-time energy management as it provides appliance-level consumption insights. Deploying NILM at the edge enables immediate feedback and preserves user privacy by processing data locally. While deep learning has significantly advanced NILM, its direct deployment on the edge faces major hurdles. The deep learning models are often too large to be applied to resource-constrained hardware. In addition, their performance may degrade substantially for a new household due to cross-household domain shifts. To overcome these two obstacles, this paper presents a novel lightweight transfer learning framework consisting of two main phases: dynamic network compression (DNC) and lightweight cross-domain transfer (LCT). DNC enables the progressive shrinking training of an elastic supernet in the cloud, which can be flexibly compressed into compact subnets tailored to the memory constraints of various edge devices without retraining. LCT enables efficient fine-tuning of deployed subnets at the edge in an unsupervised manner, adapting them to locally unlabeled cross-domain data by only updating lightweight residual modules. Experiments on a hardware testbed demonstrate that the proposed framework significantly outperforms existing methods in both disaggregation accuracy and memory efficiency, demonstrating its capability to deliver high-performance NILM on various edge devices.
科研通智能强力驱动
Strongly Powered by AbleSci AI