Yehui Li

Yehui Li

Auto-generated profile, based on publicly available information

Yehui Li

Department of Electrical and Electronic Engineering, The University of Hong Kong, Hong Kong, SAR, China

Research Focus & Metrics

Academic Career Length

6Years (2021-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
17 October 2024 in IEEE Transactions on Smart Grid
Verified
Journal Article
Closed access
YLYehui Li
Yehui Li
RYRuiyang Yao
Ruiyang Yao
DQDalin Qin
Dalin Qin
YWYi Wang
Yi Wang
Non-intrusive load monitoring (NILM) is a critical technology for disaggregating appliance-specific energy usage by only observing household-level power consumption. If NILM can be performed on end devices (such as smart meters), it can facilitate electricity demand identification and electricity behavior perception for real-time demand-side energy management. However, implementing high-performance NILM models on end devices presents an unresolved issue, encompassing two primary challenges: hardware resource constraints and data resource paucity on end devices. To this end, this paper proposes a lightweight federated learning approach for on-device NILM by combining neural architecture search (NAS) and federated learning. Firstly, a memory-efficient NAS approach is investigated to determine a personalized model within the resource constraints of end devices. Secondly, a federated mutual learning approach is designed to orchestrate the cooperation of distributed end devices with heterogeneous personalized models in a privacy-preserving manner. Case studies on two real-world datasets verify that the proposed method for appliance-level power disaggregation outperforms conventional methods in accuracy and efficiency.

30 November 2021 in IEEE Transactions on Instrumentation and Measurement
Verified
Journal Article
Closed access
YLYehui Li
Yehui Li
CYChangli Yu
Changli Yu
BXBingsen Xue
Bingsen Xue
Hongpeng ZhangHongpeng Zhang
Hongpeng Zhang
XZXingming Zhang
Xingming Zhang
The detection of metal particles in oil is of great significance to the fault diagnosis of mechanical equipment. Real domain detection methods cannot simultaneously identify the sizes and materials of particles. Because a metal particle in an alternating magnetic field induces a complex domain signal whose strength relates to the particle properties, this paper introduces a double lock-in amplifier circuit (DLAC) into a triple-coil inductive particle sensor to detect the complex domain signal under noise interference conditions. Furthermore, various sizes of red copper, brass, aluminum, iron, and type 403 steel particles are chosen as the experimental materials. For the first time, the complex domain signals induced by the five kinds of particles are detected. Hence, in the complex plane, a method is proposed to identify the size and material of a particle based on the amplitude and phase of these signals. With the high sensitivity and strong antinoise ability of the DLAC, this system can detect 25 μm iron particles and 70 μm copper particles. This study presents a new strategy for the online detection of particles in oil.

15 August 2025 in IEEE Transactions on Smart Grid
Verified
Journal Article
Closed access
YLYehui Li
Yehui Li
Xianhao ChenXianhao Chen
Xianhao Chen
YWYi Wang
Yi Wang
With the growth of energy demand and the popularization of distributed energy resources, home energy management (HEM) has emerged as a crucial technology to improve energy efficiency and reduce electricity costs. HEM systems at the edge can provide more efficient and personalized energy services for households through real-time intelligent analysis. However, distributed edge devices may not have adequate storage space to support computation-intensive decision-making algorithms. To address this problem, this paper proposes EdgeHEM, an edge reinforcement learning framework for HEM that considers the memory constraints of edge devices. Specifically, a dynamic sparse learning strategy with topology evolution is explored to overcome memory limitations on the network scale. Furthermore, a compressed federated learning approach with gradient approximation is developed to leverage the transitions cached in the memory of multiple edge devices. Extensive experiments are conducted on the established hardware testbed to demonstrate the efficiency and practicality of EdgeHEM using real-world datasets.

17 February 2025 in IEEE Transactions on Smart Grid
Verified
Journal Article
Closed access
XCXueyuan Cui
Xueyuan Cui
BLBoyuan Liu
Boyuan Liu
YLYehui Li
Yehui Li
YWYi Wang
Yi Wang
Model predictive control (MPC) has been widely adopted for indoor temperature control and building energy management. There are two steps in traditional MPC: 1) modeling thermal dynamics as the state space function to represent the temperature variation influenced by thermostatically controlled loads (TCLs); 2) formulating an optimization problem for optimal scheduling of TCLs within the control horizon. However, such a “model-then-control” strategy could result in biased control because of the unaligned modeling error and control cost, i.e., minimization of model errors may not necessarily lead to minimal costs against actual thermal dynamics in buildings. To tackle this problem, we advocate for a “smart model-then-control” (SMC) strategy that incorporates thermal dynamics modeling into the temperature control task. In particular, instead of using mean squared errors (MSE), we adopt the control objective as the task-specific loss function to guide the model training. We further formulate an Input Convex Neural Network (ICNN)-based surrogate loss function, which is differentiable and convex for effective training. In this way, the objectives of both model training and temperature control in MPC are well-aligned to obtain cost-effective decisions. We validate the performance of the SMC strategy in single-zone and multi-zone buildings. The simulation results show that it can reduce control costs by 5.97% and 2.10% respectively when compared with traditional MPC.

2 June 2025 in IEEE Transactions on Smart Grid
Verified
Journal Article
Closed access
YLYehui Li
Yehui Li
HDHaoyuan Deng
Haoyuan Deng
YWYi Wang
Yi Wang
Local energy communities establish a platform for prosumers and consumers to share surplus generation via peer-to-peer (P2P) energy trading. Traditional optimization techniques for energy management problems necessitate precise models and accurate predictions. In contrast, reinforcement learning has gained widespread attention as it can handle system uncertainties without relying on models. However, multi-agent reinforcement learning (MARL) approaches for P2P energy trading involve a large volume of communications and computations, hindering the practical deployment of decision-making algorithms on edge devices whose communication and computation resources are very limited. To this end, this paper proposes an efficient and privacy-preserving MARL approach for edge computing-enabled P2P energy trading. Specifically, we design an information interaction framework that shares representations rather than sensitive observations to protect the privacy of various agents. In addition, we develop an event-triggered communication approach that controls the frequency of interactions through a gating mechanism to reduce communication overhead among agents. To save the memory footprint of each agent, we investigate an evolutionary training method that updates the networks using perturbation without backward gradient computation. Case studies on a real-world dataset demonstrate that our proposed method yields significant efficiency improvements while maintaining high decision-making performance.

10 September 2025 in IEEE Transactions on Smart Grid
Verified
Journal Article
Closed access
YLYehui Li
Yehui Li
DQDalin Qin
Dalin Qin
TLTaoyu Lu
Taoyu Lu
YWYi Wang
Yi Wang
Increasing numbers of distributed photovoltaic (PV) systems are integrated into power grids. Accurate distributed PV power forecasting is pivotal in guiding individual users to manage their electricity usage. The forecasting model is often deployed on low-cost local devices of distributed systems to reduce communication costs and protect data privacy. However, the offline trained models may not perform satisfactorily on these systems due to the distribution shift caused by spatial and temporal context variations, which calls for online model updating to tackle such problems. Nevertheless, the memory consumption required for on-device model training is prohibitive for edge devices with limited memory capacity. To enable on-device forecasting model updates, we propose a tiny online learning framework to tackle the distribution shift problem in a lightweight manner. First, we introduce a tiny model adaptation strategy that freezes dense layers while updating memory-efficient components online. Second, we develop a tiny distribution shift detection mechanism that compares residual distribution discrepancies without catching historical data to identify shifting samples. Third, we present a tiny parameter importance evaluation approach that determines critical parameters to guide dynamic sparse training. Extensive experiments conducted on a hardware testbed show the effectiveness of our framework in improving forecasting accuracy (up to 14.5%) and training efficiency (13.1 $\times $ ) compared to vanilla online learning methods.