Elsevier

Applied Energy

Volume 238, 15 March 2019, Pages 796-805
Applied Energy

Can non-intrusive load monitoring be used for identifying an appliance’s anomalous behaviour?

https://doi.org/10.1016/j.apenergy.2019.01.061Get rights and content

Highlights

  • Ruled-based algorithm to detect anomalies in compressor based appliances.

  • Evaluation of the viability of NILM for anomaly detection.

  • Experiments using publicly available four NILM algorithms and three datasets.

  • Analysis of correlation between NILM accuracy and resulting anomaly detection.

  • Discussion of further steps to facilitate effective anomaly detection using NILM.

Abstract

Identification of faulty appliance behaviour in real time can signal energy wastage and the need for appliance servicing or replacement leading to energy savings. The problem of appliance fault or anomaly detection has been tackled vastly in relation to submetering, which is not scalable since it requires separate meters for each appliance. At the same time, for applications such as energy feedback, Non-intrusive load monitoring (NILM) has been recognised as a scalable and practical alternative to submetering. However, the usability of NILM for anomaly detection has not yet been investigated. Since the goal of NILM is to provide energy consumption estimate, it is unclear if the signal fidelity of appliance signatures generated by state-of-the-art NILM is sufficient to enable accurate appliance fault detection. In this paper, we attempt to determine whether appliance signatures detected by NILM can be used directly for anomaly detection. This is carried out by proposing an anomaly detection algorithm which performs well for submetering data and evaluate its ability to identify the same faulty behaviour of appliances but with NILM-generated appliance power traces. Our results on a dataset of six residential homes using four state-of-the-art NILM algorithms show that, on average, NILM traces are not as robust to identification of faulty behaviour as compared to using submetered data. We discuss in detail observations pertaining to the reconstructed appliance signatures following NILM and their fidelity with respect to noise-free submetered data.

Introduction

In buildings, electrical appliance’s faulty behaviour can happen either due to a fault in any appliance part or user negligence, e.g., refrigerant loss in a refrigerator or keeping the refrigerator door open. An instance of faulty behaviour can result in higher energy consumption than its normal behaviour and/or can lead to permanent damage of the appliance. Mostly, such faulty instances are intermittent; identifying them promptly improves appliance maintenance and lifespan, and results in energy savings. In this paper, hereafter, we call an appliance showing faulty behaviour as “anomalous appliance” and the anomalous instance as an “anomaly”.

Identifying faulty behaviour of appliances in buildings has traditionally used submetered data, i.e., measuring energy consumption at appliance level individually, as in [1], [2], [3]. However, the number of submeters or individual appliance monitors increases with the number of appliances or loads, and therefore anomaly detection based on submetering is not a scalable solution, especially in modern households with over 40 electric appliances.1

On the other hand, Non-intrusive load monitoring (NILM) estimates the individual consumption of an appliance within a building from the aggregate meter reading obtained from a smart meter, measuring total household electricity consumption at each sampling point; effectively eliminating the need for submetering. The effectiveness of NILM has been demonstrated in providing appliance [4], [5] and activity-based feedback [6] to consumers, utilities, and policy makers (see [7], [8], [9], [10], [11] for recent reviews).

NILM research has received an increased boost since 2010, primarily due to the roll-out of smart meters worldwide [12] and evidence that the appliance-level feedback to consumers can result in energy savings of up to 15% [13]. Many algorithms have been proposed to improve the disaggregation performance of NILM [14]. High disaggregation accuracy has been reported in the literature (in some cases, around 90% [15], [16], [17]), and presently, more than 30 companies are offering NILM-based solutions [18], e.g., EnerTalk (https://www.enertalk.com/product) from Encorded and SPEED (https://bit.ly/2NLP1Yu) from Enetics provide appliance-level consumption details to households from a bespoke smart meter fitted at the mains. There are other somewhat meter-agnostic offerings too that work on smart meter data from national roll-outs. However, in NILM literature and industrial offerings, the algorithms are not always tested at scale on real, noisy datasets typical of smart meter actual measurements from buildings and households. Furthermore, many NILM solutions are limited to disaggregating few appliances accurately, use multiple features (e.g., active and reactive power, voltage, current) and sampled measurements at 1 Hz that are generally not available from national smart meter deployments, and offer either good classification accuracy (i.e., which appliance was running) or good consumption estimation (i.e., how much the detected appliance consumed in watts) accuracy. Current EU and national law and smart meter deployments do not make data available remotely (e.g., utility) at rates higher than 15–60 min. Furthermore, the feature available is mostly restricted to active power. However, as per the UK Smart Meter specifications [19] and other home energy management providers on the market (as discussed above) with bespoke higher resolution smart meters, the data available to the customer or data owner within the Home Area Network (HAN) is at higher granularity, e.g, 1–60 s, and therefore NILM can provide useful energy feedback directly to the customer.

While the ability of NILM in removing the need of submetered data for itemized billing is well recognized, so far NILM has not been tested for detection of appliance’s faulty behaviour in buildings. To ensure accurate appliance anomaly detection, it is not sufficient to produce an accurate energy consumption estimate, but also to reconstruct with high fidelity the appliance load signature. In this paper, we assess the accuracy of reconstructed appliance load signatures using state-of-the-art NILM methods and therefore the possible impact of NILM on anomaly detection, that depends on these load signatures being replicated accurately. That is, we evaluate whether NILM-generated power traces can be used directly in identifying anomalous appliances.

To identify faulty appliances from a single smart meter, first we use four publicly available, well-established and popular NILM techniques of [20], [21], [22], [16] to obtain disaggregated appliance power traces, and then attempt anomaly detection on these appliance power traces. Given the exploratory nature of the work, and to gain deep insights, we focus our study on the anomaly detection of two major energy consuming appliances in residential buildings, i.e., Air Conditioner (AC) and refrigerator. Typically, an AC runs for limited hours of a day, but often consumes significantly high amount of energy. On the other hand, a refrigerator remains operational 24 × 7, which causes it to consume energy (usually around 7% of the total energy consumption [23]) continuously.

Anomaly detection is performed using a new rule-based proposed algorithm, which we term UNUM2 that first learns the appliance’s ON–OFF cycle frequency and duration during normal operation and then monitors the appliance’s consumption and flags an anomaly whenever a deviation is found.

Our study consists of two steps: (i) Perform energy disaggregation using existing techniques to get NILM data (i.e., appliance-level traces); (ii) Apply the proposed UNUM on both NILM data and submetered appliance data, where testing on submetered data provides the baseline performance of UNUM.

We use energy consumption data of six homes from three different publicly available datasets (REDDs [24], iAWE [25], Dataport [26]) to perform experiments. These datasets provide both aggregate smart meter measurements at 1 min (Dataport) and 1 s (REDD, iAWE) sampling rates, and submetered data at the same rates (which is used purely for baseline performance evaluation).

Contributions of this paper are summarised as:

  • 1.

    A rule-based UNUM algorithm is proposed for detecting anomalies, which uses appliance-level power traces of an AC or a refrigerator.

  • 2.

    An in-depth methodological evaluation of the viability of NILM power traces is provided through careful insertion of well-established AC and refrigerator anomalies and through multiple metrics of assessment, to determine the correlation between NILM accuracy and resulting anomaly detection based on NILM power traces. The generated annotated appliance anomaly dataset is made publicly available.

  • 3.

    Anomaly detection is performed directly on NILM-generated power traces obtained from the smart meter aggregate measurements instead of circuit-level measurements or appliance submetering.

  • 4.

    Robust, methodological evidence is provided via four NILM algorithms and three datasets for experiments. Using publicly available NILM techniques and datasets allows reproducibility of presented results.

  • 5.

    We discuss further steps needed to facilitate effective anomaly detection using NILM-outputs, i.e., appliance-level power traces obtained from NILM.

The remaining paper is organized as follows: Section 2 discuss the related work in the anomaly detection domain. Section 3 discusses the proposed anomaly detection algorithms. Section 4 explains the dataset, baseline algorithms and the evaluation metrics used. Section 5 mentions the results obtained. Section 6 discusses results obtained and Section 7 concludes the paper.

Section snippets

Related work

Related work can be broadly divided into two groups: work on anomaly detection and on NILM.

Anomaly detection: Anomaly detection in energy domain has become a popular research topic with the introduction of smart meters (aggregate load measurements), circuit-level and plug monitors (latter two providing submetering data), which enable logging and analysis of power consumption data. Therefore anomaly detection approaches target either aggregate smart-meter or at submetered load level energy

Methodology

In this paper, we focus our analysis on AC and refrigerator, which are common household appliances. They are both compressor-based and high energy consuming appliances, with the primary contributor to their energy consumption being their compressor. Any fault in the compressor itself or in any other part affecting the compressor gets reflected in the power consumption trace of the appliance. Fig. 1(a) shows the normal functioning of such appliances where each cycle consists of ON and OFF

Dataset

We use energy consumption data of six homes from three different publicly available datasets (four from Dataport, one from iAWE and one from REDD [24], [25], [26]) for the evaluation. Other publicly available datasets (ECO [54], DRED [55], Smart [56], GREEND [57], REFIT [58], UK-DALE [59], AMPds [60], Dataport [26], REDD [24], PLAID [61], tracebase [62]) either do not have both AC and refrigerator or do not have data of considerable duration required for the experiments. Only one home in REDD

Results

In this section, first, we report disaggregation performance of various existing NILM techniques. Then, with UNUM, we show how effective NILM data is for anomaly detection as compared to submetered appliance data.

Table 3 reports ANE for different appliances of six homes using CO, FHMM, LBM, and SSHMM separately. Appliance mapping of these homes is given in Table 4. In Table 3, few entries are >1 meaning that the disaggregation technique predicted an appliance consumed more energy in total (sum)

Discussion

In this section, we discuss our findings through key research questions.

1. How do we know which NILM technique will perform better for anomaly detection without using UNUM?

Our experiments show that a good number of anomalous instances can be identified correctly if the ANE for an appliance is <0.1 as reported in Table 3. Overall, ANE for AC is lowest as compared to remaining appliances, and the top row of Fig. 4 shows that AC anomalies can be detected with a precision of 0.7 and recall of 0.5,

Conclusion & future work

Submetering, i.e., using separate energy monitors for each appliance, to detect appliance-specific faulty behaviour is neither a scalable nor practical solution. Instead, NILM or non-intrusive load disaggregation using only as input, smart meter data, which has shown substantial progress in accurately estimating appliance level energy consumption, seems a good alternative to submetering for identifying faulty appliance behaviour at scale. In order to determine whether the reconstructed

Acknowledgement

We thank Dr. Batra and Zhong for validating the implementations of FHMM and LBM algorithms, respectively. We would like to acknowledge the support provided by ITRA project, funded by DEITy, Government of India, under a grant with Ref. No. ITRA/15(57)/Mobile/HumanSense/01. This work was supported in part by the U.K. Engineering and Physical Sciences Research Council under Grant EP/R512898/1 EPSRC Global Challenges Research Fund Institutional Sponsorship Award 2017 (GCRF)/R171051-102 ENACT

References (63)

  • T. Ganu et al.

    Socketwatch: an autonomous appliance monitoring system

  • W. Pereira et al.

    Unsupervised detection of unusual behaviors from smart home energy data

  • Christensen D, Earle L, Sparn B. Nilm applications for the energy-efficient home. Tech. Rep., Golden, CO (United...
  • Barker S, Kalra S, Irwin D, Shenoy P. Nilm redux: The case for emphasizing applications over accuracy. In: NILM-2014...
  • J. Froehlich et al.

    Disaggregated end-use energy sensing for the smart grid

    IEEE Pervasive Comput

    (2011)
  • K. He et al.

    Non-intrusive load disaggregation using graph signal processing

    IEEE Trans Smart Grid

    (2018)
  • B. Zhao et al.

    Improving event-based non-intrusive load monitoring using graph signal processing

    IEEE Access

    (2018)
  • Parson O. Overview of the nilm field, http://blog.oliverparson.co.uk/2015/03/overview-of-nilm-field.html;...
  • K.C. Armel et al.

    Is disaggregation the holy grail of energy efficiency? the case of electricity

    Energy Policy

    (2013)
  • Faustine A, Mvungi NH, Kaijage S, Michael K. A survey on non-intrusive load monitoring methodies and techniques for...
  • Encorded. Enertalk NILM fusebook, https://www.enertalk.com/product;...
  • S. Makonin et al.

    Exploiting hmm sparsity to perform online real-time nonintrusive load monitoring

    IEEE Trans Smart Grid

    (2016)
  • J. Kelly et al.

    Neural nilm: Deep neural networks applied to energy disaggregation

  • Community N. Companies offering nilm products and services, http://wiki.nilm.eu/companies.html;...
  • UK: Department of Energy & Climate Change. Smart metering equipment technical specifications: Second version (SMETS2)....
  • G.W. Hart

    Nonintrusive appliance load monitoring

    Proc IEEE

    (1992)
  • Kolter JZ, Jaakkola TS. Approximate inference in additive factorial hmms with application to energy disaggregation. In:...
  • Zhong M, Goddard N, Sutton C. Latent bayesian melding for integrating individual and population models. In: Advances in...
  • EIA. International energy outlook, energy information administration (EIA) (DOE/EIA-0484...
  • Kolter JZ, Johnson MJ. Redd: A public data set for energy disaggregation research. In: Workshop on data mining...
  • N. Batra et al.

    It’s different: Insights into home energy consumption in india

  • Cited by (88)

    • A mixed-integer programming approach for industrial non-intrusive load monitoring

      2023, Applied Energy
      Citation Excerpt :

      For the power generation side, disaggregated electricity signals can also be helpful to load forecasting and identification of load flexibility. For equipment manufacturers, it is possible to realize the predictive maintenance of equipment by analyzing individual devices’ electricity signals since faulty devices may have disordered energy signals that can be recognized from disaggregated data [3–6]. Due to the huge difference in power consumption scale and types of electrical equipment among different electricity users, NILM is generally divided into three types: NILM for residential, commercial, and industrial settings.

    • Anomaly detection on household appliances based on variational autoencoders

      2022, Sustainable Energy, Grids and Networks
      Citation Excerpt :

      In the test phase, samples that deviates more than 20% from their assigned operation mode are considered anomalous for that device. Rashid et al. in [16] tried to find anomalies in household appliances by extracting their power consumption with non-intrusive load monitoring (NILM) techniques. To find the anomalies, they applied some thresholds to the energy consumption and duration of the device’s activation based on their expected normal power profile.

    View all citing articles on Scopus
    View full text