Electrical energy is the very foundation of the modern society. While the depletion of fossil fuels and unreliability of renewables has painted a hazy picture of our future energy supply, the need for efficient utilization of electrical energy and its conservation is greater than ever. Energy management and conservation measures not only generate immediate results but they also ensure smooth operation of the overall system. In addition to the economic benefits, the burden on the transmission and distribution system is reduced.
Building energy disaggregation at appliance level can serve as an important tool for energy planning and conservation schemes. In order to find the energy consumption of each individual load inside a building, there are two options:
- Intrusive load monitoring: Install an energy meter with each appliance to record its power consumption.
- Nonintrusive load monitoring (NILM): Install a single energy meter at the service main connected to the building and estimate the energy consumed by each individual load by the changes observed in the composite voltage and current signals.
Intrusive load monitoring requires a complicated hardware setup and unnecessary intrusion on the consumer’s property. NILM on the other hand, is a sophisticated and advanced technique that requires minimal hardware with no intrusion. A single set of voltage and current sensors are installed on the main cable outside the building or voltage and current readings are taken from the main energy meter already installed. These current and voltage signals, representing the total consumption of the building, are then analyzed using sophisticated techniques to identify individual appliances inside the building and their energy consumption.
How NILM works
The basic methodology of NILM is illustrated in the figure below. The "load" in the block diagram represents the combined load of the whole building. The first stage is the data acquisition stage, where the composite current and voltage signals are obtained. Then, certain algorithms are applied to extract useful information from these signals that can be fed to a classifier. This is the feature extraction stage. Once the features are extracted, a suitable machine learning based classification algorithm is used to classify the appliances inside the building. Supervised learning techniques are most commonly used and they will require a preset database of appliances.
Figure 1: Block diagram representing basic NILM methodology.
Feature extraction and classification stages present the main challenge in developing an effective NILM system. Research is still ongoing in this field, and engineering researchers are slowly moving towards the development of an effective and accurate model.
NILM algorithms can be divided into two categories:
1. Event based
2. Non-event Based
Event based methods focus on changes in parameters of appliances when they change states. Appliances can be classified based on their operating states, typically categorized as:
- On/off appliance: Appliances which have only two states of operation, i.e. either they are on or they are off.
- Continuously variable loads:, with no fixed pattern of states like dimmer lights.
- Finite state machines (FSMs): multistate appliances like washing machine.
- Permanent devices are devices that are on for a long period of time like a DSL router.
Event based methods involve event identification and then keeping track of appliance operation times. The energy consumption is estimated by appliance operation duration. Non-event based algorithms disaggregate loads from the composite signal using various machine learning, optimization and probabilistic techniques. Event based approach is the more popular approach.
Feature extraction
The pioneer research on non-intrusive load monitoring was done by Hart in 1980s. His research was based on steady state power analysis due to its easy availability and extraction but his work discounted the nonlinear loads and hence was not completely effective. Moreover, performance of this model further decreases as the number of appliances increases, especially if many appliances have similar power characteristics.
Researchers have since come up with a possible solution to this drawback by analyzing the signals at higher frequencies and capturing the transient characteristics. Distinct physical characteristics of different appliances cause them each to exhibit unique transient features. To use this property in NILM, proper mechanism is required for appropriate detection of these transients. Short duration of transients calls for sophisticated hardware and software tools.
To put it in simple terms: whenever an appliance changes states, it generates a distinct transient in the composite current signal that can be captured by high frequency sampling. The transient is also present in the voltage signal but it usually not as prominent as it is in a current signal. Once this transient is captured, techniques like wavelet analysis and short-term Fourier transform can be used to extract useful features. Once these features are extracted, they can be thought of as unique "appliance signatures." A classifier is then trained to recognize these signatures. Feature extraction techniques are not limited to the ones mentioned in this article, there are a number of techniques out there and discussing all of them will take a long time so we are leaving things here at the basic level.
Classification
Appliance identification is done by employing machine learning techniques which can be classified as supervised or non-supervised. Supervised learning techniques, as apparent by the name, require labeled data for training whereas unsupervised techniques do not. Since labeled data is available or can easily be generated in case of NILM, supervised learning is the popular choice. The choice of classification method depends largely on the feature extraction algorithm used. Popular classification tools include, artificial neural networks, support vector machines, trees, K-nearest neighbor algorithm and other similar tools.
Final thoughts
NILM is a powerful technique that can be used for a number of energy management and conservation schemes. From smart home energy management systems to selective load-shedding schemes, the possibilities are endless. The research on NILM has come a long way and there are a few companies out there offering closely related services, but large-scale commercialization is still in the pipeline. Researchers from all across the globe are improving the algorithms, and effective and inexpensive NILM services will likely be available all over the world.
