Models

To address the privacy concern in NILM applications, federated learning (FL) could be leveraged for NILM model training and sharing. When applying the FL paradigm in real-world NILM applications, however, we are faced with the challenges of edge resource restriction, edge model personalization, and edge training data scarcity. We design FedNILM, a practical FL paradigm for NILM applications at the edge client. Specifically, FedNILM delivers privacy-preserving and personalized NILM services to large-scale edge clients, by leveraging i) collaborative data aggregation through federated learning, ii) efficient cloud model compression via filter pruning and multi-task learning, and iii) personalized edge model building with unsupervised transfer learning. Our experiments on real-world energy data show that FedNILM can achieve personalized energy disaggregation with the state-of-the-art accuracy, while preserving the user privacy.

Recent investigations have shown that deep neural networks (DNNs) based approaches are promising for the NILM task. Nevertheless, they normally ignore the inherent properties of appliance operations in the network design, potentially leading to implausible results. We are thus motivated to develop the dual Deep Neural Networks (dual-DNN), which aims to i) take advantage of DNNs' learning capability of latent features and ii) empower the DNN architecture with identification ability of universal properties. In the design of dual-DNN, we adopt one subnetwork to measure power ratings of different appliances' operation states, and the other subnetwork to identify the running states of target appliances. The result is then obtained by multiplying these two network outputs and meanwhile considering the multi-state property of household appliances. To enforce the sparsity property in appliance's state operating, we employ median filtering and hard gating mechanisms to the subnetwork for state identification. Compared with the state-of-the-art NILM methods, our dual-DNN approach demonstrates a 22% performance improvement in average on two public benchmark datasets.

As the number of electric vehicles (EVs) increases, large-scale residential EV charging will burden the power grid, posing problems for both planning and operations. Promptly capturing EV charging events can help mitigate this problem. However, most existing grid operators lack dedicated sensors for residential EV monitoring. This motivates NILM as a technique to gain fine-grain EV charging information. We design EVSense, a robust deep neural network (DNN) based model for non-intrusive EV charging detection. We also show how to use federated transfer learning (FTL) to deploy our system on resource-constrained edge devices. This makes EVSense a feasible solution for large-scale EV monitoring. We evaluate EVSense on both real-world and synthetic datasets and find that it can achieve higher precision and more robust charging detection compared to existing learning-based and rule-based approaches. We also find that, due to the use of FTL, EVSense also shows excellent scalability despite diversity in residential load profiles, sampling rates, and seasons.

Occupancy detection can greatly facilitate HVAC and lightning control for building energy saving. Sensor based occupancy detection is usually costly and may suffer from high false positive rates. As such, occupancy detection using load curve data has been proposed. Such type of methods, however, normally relies on tedious and nontrivial model training process.
To overcome this pitfall, we develop a simple, non-intrusive occupancy detection approach that does not require any model training and only uses load curve data and readily available appliance knowledge. The method consists of three main steps: i) the appliances' mode states are firstly decoded via a carefully designed total variation minimization problem; ii) the human actions are recovered with a-priori knowledge of human-activated switching events; iii) the occupancy states are then inferred based on the recovered human actions along with empirical association rules. We evaluate our approach and compare with existing methods with real-world data. The results show that our approach can achieve similar performance to those using supervised machine learning.

In power systems, load curve data is one of the most important datasets that are collected and retained by utilities. The quality of load curve data, however, is hard to guarantee since the data is subject to communication losses, meter malfunctions, and many other impacts.
We propose a new approach to analyzing load curve data. The method adopts a new view, termed portrait, on the load curve data by analyzing the periodic patterns in the data and reorganizing the data for ease of analysis. Furthermore, we introduce algorithms to build the virtual portrait load curve data and demonstrate its application on load curve data cleansing. Compared to existing regression-based methods, our method is much faster and more accurate for both small-scale and large-scale real-world datasets.

The rapidly growing edge computing market, supported by the edge cloud (EC) infrastructure, has imposed significant operating costs on EC operators (ECOs), particularly the energy cost. By incorporating green energy through the installation of solar photovoltaic (PV) panels and wind turbines (WT), ECOs have great potentials in reducing their energy expenses. Additionally, integrating the battery energy storage system (BESS) to store surplus green energy and discharge it during peak power periods can further minimize energy costs under prevailing energy charge and demand charge tariffs. In this study, we assess the cost-saving potential for ECOs in light of current trends in PV and WT installations as well as demand charge electricity billing tariff. Based on our observations and settings within ECs, we mathematically model the energy scheduling problem in distributed green ECs with the objective of minimizing the energy costs. To solve this problem, we propose a multi-agent reinforcement learning based approach, along with a customized invalid action masking method to handle the inherent strong time coupling characteristics. This new method is capable of adapting to time-varying and fluctuating power from green resources and ECs while making real-time decisions regarding BESS discharge/charge operations. Experimental results obtained under realistic environmental conditions using power traces from renewable energy generators and ECs demonstrate that our approach can significantly cut down the ECO's energy cost by up to 66%.