Towards Low-Cost Energy Monitoring

Introduction

Concern over decreasing energy resources, increasing energy demand, carbon dioxide (CO2) emissions, and global climate change has motivated serious efforts to make more efficient use of energy resources. Buildings consume more than 40 percent of overall energy expenditure worldwide¹. A significant portion of this energy is considered as wastage. Effective energy conservation in buildings is critical both to reducing the seemingly ever-increasing energy demand and to mitigating concerns over global climate change. However, as a first step towards energy conservation, greater transparency over the traditional invisibility of energy usage must be engendered. Appliance-level energy usage monitoring is considered is a pre-requisite for energy conservation in buildings (Hutton, 1986). Recent research efforts have produced solutions that range from low-cost to high-cost methodologies. Two major categories of such methodologies are often referred as: distributed monitoring or Intrusive Load Monitoring (ILM), and single point monitoring or Non-Intrusive Load Monitoring (NILM) (Froehlich, 2011).

Non-intrusive load monitoring is considered to be a low-cost and simple technique to acquire individual appliance power consumption information. Metering devices are installed at fuse boxes to obtain energy usage measurements for whole buildings; these measurements are then further analyzed using pattern recognition algorithms to disaggregate appliance specific information. Figure 1 highlights appliance specific information that is present in an aggregated measurement of a household. In the recent past, researchers have presented many variations of the NILM method. However, a common principle is to create appliance specific energy usage signatures by turning them on/off, measure aggregated power usage of the building, and then, based on these known signatures, analyze power changes within the measurements to disaggregate appliance specific information.

Figure 1.

Power usage highlighting step changes corresponding to individual appliance events

NILM techniques presented in literature are mostly based on two major appliance features: Steady State and Transient State. Steady-state analysis requires less expensive hardware in comparison with transient-state analysis. Despite an intensive amount of research in the area, NILM techniques still have limitations. These include: manual annotation of training data sets when using supervised machine learning, limitations when distinguishing appliances with similar load profiles and low-energy consuming appliances, and validity of performance evaluation metrics.

Background

Hart (1992) proposed non-Intrusive Appliance Load Monitoring (NALM) later referred as Non-Intrusive Load Monitoring (NILM) or single-point sensing. According to Hart, NILM monitors aggregated energy load, and analyzes measurements to identify certain appliance signatures so as to obtain information on appliance activities within a monitored environment. For example, consider a simple case where a building contains a microwave oven that consumes 1200 watts when turned on at high state; then a sudden increase or decrease of 1200 watts in total load will indicate that the microwave is turned on or off. Similarly, other appliances with different characteristics or signatures can be identified from an aggregated load measurement. The very first step in NILM is to install a metering device, which is often a smart meter, to measure energy consumption of a building in real-time. This information is then transferred to a PC via gateway. Appliance signatures are learned by turning each appliance on/off during the system-training phase and then the signatures are annotated appropriately. The next step is to observe a step change in total load and utilize appropriate machine learning algorithms to identify and classify appliances based upon learned signatures. Appliances having different load signatures are easy to identify but classifying those having similar signatures is a complex task.