Multivariate event detection methods for non-intrusive load monitoring in smart homes and residential buildings
Introduction
In the current context of energy transition, it is essential to better monitor the domestic consumption. Feedback on electricity consumption appears to be a major tool to save energy through more conscious user behaviour [1]. NILM has emerged as a promising solution to provide a breakdown of the residential energy consumption without instrumenting HEAs [2], [3], [4]. The appliances identification is made through the analysis of the current and voltage signals using a single measuring device connected to the house electrical panel [5]. This information is relayed back to the consumers, who can therefore make an informed decision about energy savings. Throughout literature, a large number of NILM approaches are reported [2], [6]. Most of the methods can be classified into supervised or unsupervised methods. The former rely on classification algorithms that require a sufficient amount of labeled data for a training process [6], [7]. On the opposite, the latter methods are based on clustering procedures which do not need a labeled training dataset [6]. Finally, NILM approaches can be categorized into non event-based and event-based approaches [8]. Non event-based methods disaggregate the power consumption using techniques such as the Hidden Markov models (HMM) [9], [10] which identify repetitive HEAs’ patterns in the load curve based on HEAs previous states. The event-based approaches assume that each event in the total household power consumption is a response to a HEA state change [8]. Fig. 1 shows an overview of event-based NILM approaches. In this research work, a specific interest is given to HEAs’ event detection which is an elementary step in the NILM pipeline [11]. According to the existing literature on detection algorithms for event-based NILM methods, two main approaches are considered: the heuristic and the probabilistic methods [12]. Algorithms under the first category analyze a time series of data, looking for changes above a given threshold. In [13], the events are detected in the total active power signal by computing the absolute value of the difference between two consecutive samples, which is then compared to a pre-defined threshold. M.N. Meziane et al. in [14], propose an event detector called High Accuracy NILM Detector (HAND) that tracks the standard deviation variation of the current signal envelope using a sliding window. A threshold separates the events characterized by high standard deviations from the steady states defined by much smaller standard deviations. They obtained a probability of detection of 96.7% on simulated data.
Algorithms in the second category of probabilistic methods comprise three entities: A stochastic process under observation, a change point at which the statistical properties of the process undergo a change, and a decision maker that observes and detects the change of the process statistical properties [15], [16]. In the NILM context, three algorithms are mainly used to detect changes in active power signals: the Generalized Likelihood Ratio (GLR) test, the Chi-squared Goodness Of-Fit (χ2 GOF) and the CUSUM detectors. Authors in [17], [18] use the GLR approach to test if two consecutive time frames share a common distribution by deriving a decision function from the log-probability distribution ratio before and after a potential change of the mean value. In [18], [19], a χ2 test statistic is used to assess if two neighboring windows share a common distribution. If this is not the case, an event is then detected. Researchers in [20], [21] apply a CUSUM algorithm for the detection of both the beginning and the end of a HEA transient-state active power signal. Zhu et al. [21] reached a probability of detection of 90% by applying their approach on real data comprising 200 events of eight different HEAs switched on and off.
The common thread in all these research works is that the detection is done in a univariate context by only considering the active power signal, whereas prior literature on NILM focuses on identifying an effective features’ set that defines a unique HEA signature [22], [23], [24]. Fig. 2 illustrates three power time series related to the same scenario of several HEAs switched on and off using our own acquisition system [25], and three power time series derived from the Building-Level fully-labeled Dataset for Electricity Disaggregation (BLUED) [26]. It can be observed that changes are present in the active power time series, the reactive power time series and their 4th and 5th order harmonic. The information provided by these time series could be used together to improve the detection robustness. To the best of our knowledge, there is no existing literature on multivariate change detectors designed for NILM event-based methods. Consequently, the goal of this paper is to demonstrate the benefit brought by using a multivariate approach for change detection with the most relevant features. Reference [27] presents our preliminary results related to the application of univariate change detectors on active power signals. In this paper, the pool of studied change detectors is enlarged and extended to the multivariate case by considering several power features such as the active power, the reactive power and their respective harmonics. Four algorithms suited for mean change detection are considered in both the univariate and the multivariate cases. These change detectors are the Bayesian Information Criterion (BIC), the Cumulative Sum (CUSUM), the Hotelling T2 test (equivalent to the GLR statistic test for normal distributions) and an updated version of the Effective Residual algorithm described in [27].
This paper is organized as follows. In Section 2, the four change detectors are presented. Their performances are assessed through Monte Carlo experiments using detection performance evaluation metrics, considering both univariate and multivariate cases. In Section 3, the proposed Feature Selection Algorithm for Detection Purposes (FSADP) is introduced. Section 4 focuses on the experimental results when applying the FSADP on real-world power signals corresponding to a scenario of HEAs switched on and off. Firstly, power signals are obtained from the BLUED dataset current and voltage measurements [26], and secondly, power signals are computed from current and voltage measurements using our own acquisition system. The subsequent discussion addresses the selection of the power features that are best suited for the detection task. Finally, the main conclusions and contributions of the current work are provided in Section 5.
Section snippets
Hypothesis testing framework
In this subsection, the mathematical formulation of the detection problem is introduced. Let xm ∈ lRp be a multidimensional time series such as where xm,j is the value of feature at time instant m. Let with be the matrix (w being an even integer) of the last samples of xm until the current time instant n. Since a HEA transient state is not necessarily an abrupt change, we use the “Window with Margins”
Motivations
In NILM approaches, the features are extracted attributes from current and voltage measurements. In Fig. 2, it can be well observed that an event induces a common change for different electrical features at different scales. In other words, some features might be informative and useful whereas others are not. FS is a process commonly used for classification tasks to reduce overfitting and to avoid the curse of dimensionality problem [39], [40]. The selected subset contains the minimal number of
Application to NILM
In this section, the Hotelling T2 detector, which showed the best performances, is combined to the proposed FSADP and applied to real data streams in order to find the most relevant features for the detection task. We first introduce the power features computed using current and voltage measurements. Then we present the publicly available BLUED dataset dedicated to NILM event-based approaches and a new proposed dataset built using our acquisition system. Both datasets contain several events of
Conclusion
This paper explores and extends to the multivariate case four methods for detecting HEAs state changes in a NILM context. Furthermore, a new labeled dataset is introduced and made available for the sake of reproducible research. We also propose a new method for properly selecting power features that improve detection performances over the existing univariate methods based on active power signal only. Our study further supports prior NILM studies that also emphasized the need of an effective set
Declaration of Competing Interest
All authors of the submitted manuscript entitled “Multivariate Abrupt Change Detection for Non Intrusive Load Monitoring in Residential Buildings” confirm that they have contributed significantly for this research work, revised it critically and approved the manuscript. We also confirm that it has not been submitted earlier in any journal and is not being considered for publication elsewhere.
As scientists, we certify that the paper has been submitted with full responsibility and research
References (45)
- et al.
Smart homes, home energy management systems and real-time feedback: lessons for influencing household energy consumption from a swedish field study
Energy Build.
(2018) - et al.
Using pattern recognition approach to disaggregate the total electricity consumption in a house into the major end-uses
Energy Build.
(1999) - et al.
Non-intrusive load monitoring by using active and reactive power in additive factorial hidden markov models
Appl. Energy
(2017) - et al.
Non-intrusive electrical load monitoring in commercial buildings based on steady-state and transient load-detection algorithms
Energy Build.
(1996) - et al.
Comparative study of event detection methods for non-intrusive appliance load monitoring
Energy Procedia
(2014) - et al.
On the bayesian optimization and robustness of event detection methods in NILM
Energy Build.
(2017) - et al.
A novel CUSUM-based approach for event detection in smart metering
IOP Conf. Series: Mater. Sci. Eng.
(2018) - et al.
Comprehensive feature selection for appliance classification in NILM
Energy Build.
(2017) Fundamentals of Statistical Signal Processing: Estimation Theory
(1997)- et al.
Feature selection and its use in big data: challenges, methods, and trends
IEEE Access
(2019)
A supervised event-based non-intrusive load monitoring for non-linear appliances
Sustainability
Performance evaluation in non-intrusive load monitoring: datasets, metrics, and tools: a review
Wiley Interdiscip. Rev.: Data Min. Knowl. Discov.
A generative model for non-intrusive load monitoring in commercial buildings
Energy Build.
Model for electric load profiles with high time resolution for german households
Energy Build.
NILM Techniques for intelligent home energy management and ambient assisted living: a review
Energies
Machine learning approaches for non-intrusive load monitoring: from qualitative to quantitative comparation
Artif. Intell. Rev.
Review of non-intrusive load appliance monitoring
Ieee 3rd advanced information technology, electronic and automation control conf. (iaeac)
Time series distance-based methods for non-intrusive load monitoring in residential buildings
Energy Build.
Event detection for non intrusive load monitoring
Iecon 2012 - 38th annual conf. on ieee industrial electronics society
Leveraging smart meter data to recognize home appliances
Ieee int. conf. on pervasive computing and communications
High accuracy event detection for non-intrusive load monitoring
Ieee int. conf. on acoustics, speech and signal processing (icassp)
Detection of Abrupt Changes: Theory and Application
Cited by (31)
A low complexity binary-weighted energy disaggregation framework for residential electricity consumption
2023, Energy and BuildingsNon-intrusive condition monitoring based on event detection and functional data clustering
2023, Engineering Applications of Artificial IntelligenceRobust event detection for residential load disaggregation
2023, Applied EnergyElectric energy disaggregation via non-intrusive load monitoring: A state-of-the-art systematic review
2022, Electric Power Systems ResearchCitation Excerpt :The event detection module of NILM looks for events in the aggregate signal using either an expert heuristics model, a probabilistic model, or a matched filter model [34]. Few examples of these models include a change in amplitude, cumulative sum, generalized likelihood ratio, goodness-of-fit [35], Dynamic Time Warping, clustering [36], multivariate event detection [37], and maximum and minimum points based [38] event detectors. Following confirmation of an event, the signature extraction module comes into action.
Incorporating coincidental water data into non-intrusive load monitoring
2022, Sustainable Energy, Grids and NetworksCitation Excerpt :A fundamental step in any event-based NILM method is detecting the events accurately, where an event in general refers to a mode transition of any appliance. Due to the presence of noise and fluctuations in the power signal, it is essential to establish a reliable event detection procedure [39]. In a majority of research articles on event-based NILM, an event is detected based on the difference of two consecutive samples considering a predefined threshold.