1. Introduction
Electricity usage profiling is essential for understanding and improving household energy consumption patterns [
1]. By identifying individual appliance-level energy usage patterns, homeowners can make informed decisions on how to manage their energy use, reduce their carbon footprint, and save money on energy bills [
2]. The identification of appliance consumption has been successfully applied to improve the householders’ quality of life in many different scenarios, such as scheduling the use of large consumption appliances [
3], detecting appliance malfunctions [
4], or providing early preventive maintenance [
5], among many others.
Two main groups of approaches have been previously studied to monitor each appliance’s load: intrusive and non-intrusive. Intrusive load monitoring relies on installing additional sensors at the plug level per appliance cost, being more accurate at the expense of a higher price due to the high number of plug-level sensors that need to be manufactured, installed, and maintained. On the other hand, non-intrusive load monitoring (NILM) only relies on the aggregated load measured at the user connection point with their energy distributor. As such, NILM approaches use algorithms and machine learning models to disaggregate the appliance-level load from the aggregated load, leading to a less accurate but more cost-effective approach. Since the NILM problem was first formulated in the mid-1980s [
6,
7], many researchers have proposed different alternatives to address this challenge. These proposals can be categorized into four different groups depending on the strategies used to disaggregate the energy load.
State-based approaches, such as Hidden Markov Models (HMMs) [
8,
9,
10], used to be the state of the art in NILM as there was a clear relationship between each appliance state and the hidden states of the model. HMMs are probabilistic methods that require providing (or learning) a finite set of states, the probabilities of transitioning from one state to another, and the probability of producing an output from the hidden state. The major limitation of this type of approach is its high computational cost, making it too expensive and slow for real-life applications.
Dictionary-based approaches (sparse coding) [
11,
12,
13,
14] aim to find data representation based on the linear combination of a dictionary and a representation where the difference between the aggregated time series and the linear combination is minimized, each element of the dictionary represents a different appliance, and the representation is sparse enough. This minimization problem is NP-Hard and several methods can be used to solve it, such as K-SVD or LASSO. These approaches can provide good results for supervised scenarios and can be fast depending on the formulation and algorithms used to solve the optimization problem. However, they have some major limitations in unsupervised scenarios since the number of appliances for the dictionary must be provided previously.
Neural network-based approaches, with all the advancements made in deep learning over the past decade, have become the state of the art for supervised NILM. These approaches are notoriously slow for training and require large amounts of data but can provide fast and accurate disaggregation once trained. Several different neural network architectures have recently been developed for this purpose, such as the use of U-Net [
15], combinations of convolutional neural networks and Long Short-Term Memory [
16], and generative adversarial networks [
17].
Lastly, event-based approaches [
18,
19,
20] detect the use of an appliance by detecting events where appliances have been switched on or off or changed to a state with considerably different power consumption, usually by edge detection. Once the events are identified, rising and falling power edges are generally matched and some features of each event are extracted (power, duration, etc.). Then, a classification (supervised) or clustering (unsupervised) algorithm is used to map each event to an appliance. These approaches are generally fast due to the dimensionality reduction provided by the event extraction. However, they can only detect appliances with consistent energy consumption in each operational state. The algorithm proposed in this paper is of this kind. A notable methodology of this kind that has seen some success, even in unsupervised scenarios, is the use of Graph Signal Processing (GSP) [
21,
22,
23]. In GSP, a graph is constructed, with each node representing a rising/falling edge of the original time series. Then, the mathematical properties of the graph representation and a weighted adjacency matrix are exploited to convert the problem at hand (clustering, classification) into an unconstrained quadratic optimization problem that minimizes the total graph variation.
Although numerous approaches to NILM have been suggested, the majority of them face limitations that hinder practical deployment in real-life applications. A predominant proportion of these NILM algorithms depend on supervised methods, necessitating energy companies to acquire and install multiple dedicated sensors (one for each appliance) in every customer’s household. Despite offering highly accurate disaggregation results, this approach compromises the intended non-intrusiveness of NILM algorithms and imposes significant economic burdens due to the substantial costs associated with the installation of these devices. Unsupervised approaches, which eliminate the need to install sensors, have barely been studied, with just a few algorithms proposed for this task. However, even these algorithms have some major limitations for real-life applications. For example, in [
21], even though the disaggregation is conducted in an unsupervised manner, several hyperparameters must be tuned manually in order to do so, making it unfeasible to deploy it on a large scale. Furthermore, the disaggregated signals are not mapped to their corresponding appliance, requiring, according to the authors, an additional step comparing each event with a signature database, which defeats the unsupervised purpose of the algorithm. Another example of this situation is found in [
24], where the energy is disaggregated according to an energy consumption survey in Central Europe and a probabilistic HMM framework of household activities. Not only may we be concerned with whether the results of the survey are truly transferable to other regions but this approach requires some supervised information that is not provided to the energy distributor, such as the number of occupants, their age, and the nominal power of the appliances.
As such, the work proposed in this paper presents a new algorithm for low-rate unsupervised NILM that provides the following main contributions to the field:
It is the first unsupervised algorithm that can be deployed in any residential household without any additional supervised information;
We propose a novel event detection algorithm capable of recognizing some instances in which rising/falling edges overlap;
The NILM algorithm provides its disaggregation through knowledge of the common use of appliances and how they work, making it easy to understand but limiting the number of appliances it can detect.
The rest of the document is structured as follows. The proposed methodology is detailed in
Section 2.
Section 3 provides an analysis of the results obtained. And finally, the conclusions of our work are gathered in
Section 4