A Comprehensive Review on the NILM Algorithms for Energy Disaggregation
Akriti Verma, Adnan Anwar, M. A. Parvez Mahmud, Mohiuddin Ahmed, Abbas Kouzani
HHighlights
A Comprehensive Review on the NILM Algorithms for Energy Disaggregation
Akriti Verma,Adnan Anwar• This paper presents an up to date outline of NILM framework and its related strategies and methods for the energydisaggregation problem.• The paper presents an experimental overview of the application of NILM-API, which was released with nilmtk-contrib,on three publicly available data sets, draws conclusions and highlights on future research directions.• A detailed overview on the energy disaggregation problem is presented. Here, we have shown the advantages of theNILM API through which algorithmic comparisons can be defined with relatively little model knowledge. a r X i v : . [ ee ss . SP ] F e b Comprehensive Review on the NILM Algorithms for EnergyDisaggregation
Akriti Verma a , Adnan Anwar b , ∗ a School of Information Technology, Deakin University, Geelong, Australia ([email protected]) b Strategic Centre for Cyber Security Research and Innovation (CSRI), Deakin University, Australia ([email protected])
A B S T R A C T
The housing structures have changed with urbanization and the growth due to the construction of high-rise buildings all around the world requires end-use appliance energy conservation and managementin real time. This shift also came along with smart-meters which enabled the estimation of appliance-specific power consumption from the building’s aggregate power consumption reading. Non-intrusiveload monitoring (NILM) or energy disaggregation is aimed at separating the household energy mea-sured at the aggregate level into constituent appliances. Over the years, signal processing and machinelearning algorithms have been combined to achieve this. Incredible research and publications havebeen conducted on energy disaggregation, non-intrusive load monitoring, home energy managementand appliance classification. There exists an API, NILMTK, a reproducible benchmark algorithm forthe same. Many other approaches to perform energy disaggregation have been adapted such as deepneural network architectures and big data approach for household energy dis aggregation. This pa-per provides a survey of the effective NILM system frameworks and review the performance of thebenchmark algorithms in a comprehensive manner. This paper also summarizes the wide applicationscope and the effectiveness of the algorithmic performance on three publicly available data sets.
1. Introduction
There has been immense research on developing techno-logical solutions in order to address the energy requirementswhich are rising exponentially and thereby making the chal-lenge of energy conservation harder day by day. The increas-ing energy demands not only affects a country’s economy butalso comes with significant negative implications on the en-vironment. Hence, the only effective way to conserve energynow is to encourage its efficient usage. A fine-grained mon-itoring of energy consumption and communicating the sameto the consumers can help in the noteworthy reduction of en-ergy wastage [33], [9]. The problem was originally studiedby Hart [13] in the early 1980s and due to the availabilityof larger data sets in recent years, smart meter roll outs, andamidst climate change concerns, there has been a renewedinterest among researchers in this field. The main objectiveof energy disaggregation is to offer detailed energy sensingand to provide information on the breakdown of the energyconsumed, which would moreover enable the automated en-ergy management systems to depict appliances with a highrate of energy consumption, allowing them to devise energyconservation strategies such as re-scheduling of high powerdemanding operations for the off-peak times[40]. Further-more, companies would be able to develop a better under-standing of the relationship between appliances and their us-age patterns. There are two primary approaches to energy disaggregation, specifically Intrusive Load Monitoring (ILM) ⋆ This document is the results of the research project partly funded bythe School of IT, Deakin University ∗ Corresponding author [email protected] (A. Anwar) (A.Anwar)
ORCID (s): (A. Anwar) and Non-Intrusive Load Monitoring (NILM). Intrusive loadmonitoring consists of measuring the electricity consump-tion of one or a few appliances using a low-end meteringdevice, particularly requiring one or more than one sensorper appliance, whereas NILM just requires only a single me-ter per house or a building that is to be monitored. Non-intrusive load monitoring (NILM) or energy disaggregationis aimed at separating the household energy measured at theaggregate level into constituent appliances. The presence ofnilmtk-contrib [5], an open-source, reproducible state-of-the-art energy disaggregation implementation has unfoldedthe means for comparisons of the different algorithms exe-cuting energy disaggregation. It has also enabled researchersto assess the generality of NILM approaches as it can be ap-plied to multiple data sets accessible online. The versatil-ity of NILM API makes experimentation in this field easyby lowering the entry barriers and making the implementa-tion generic irrespective of the datsets and algorithms whichhelped to progress research in this area.
There has been a good number of research and publi-cations in energy disaggregation, non-intrusive load mon-itoring, home energy management and appliance classifi-cation. In the recent past, machine learning strategies forNILM have attracted a lot of recognition due to breakthroughsin research disciplines such as computer vision. NILM al-gorithms are trained and tested on energy consumption datasets. Such data sets include aggregate-level energy readingsfrom smart meters as well as appliance-level energy readingsfrom measurement equipment such as smart plugs. In thecourse of the years, a vast number of publicly-available datasets have been released. With NILMTK, an open-sourcetoolkit was designed specifically to enable the comparison
Akriti et al.:
Preprint submitted to Elsevier
Page 1 of 11 eview on the NILM Algorithms
Table 1
Articles included in systematic literature reviewSubject Purpose Discussiononframe-work DiscussiononNILMTK-API Relateddatasetrefer-ences Algorithmimple-menta-tion NILMTK-APIimple-menta-tion Empiricalanaly-sis Futurere-searchprospectsNon-Intrusive LoadMonitoring SystemFramework and LoadDisaggregation Algo-rithms: A Survey[34] presents a generalNILM frameworkand reviews publiclyavailable data sets. Yes Yes Yes No No No YesA Survey on Non-Intrusive Load Moni-toring Methodies andTechniques for En-ergy DisaggregationProblem[11] an overview of theNILM system and its as-sociated methods andtechniques for energydisaggregation problemfollowed by the reviewof the state-of-the-artNILM algorithms. Yes Yes Yes No No No NoBuilding power con-sumption datasets:Survey, taxonomy andfuture directions[17] survey, study and visu-alize the numerical andmethodological natureof building energy con-sumption datasets. Yes No Yes No No No YesOn performance eval-uation and machinelearning approachesin non-intrusive loadmonitoring[23] aims to determine theaccuracy as well asthe generalisation abil-ities of existing NILMalgorithms on the datasets REDD, UK-DALE,and Dataport. Yes Yes Yes No No No YesProspects of ApplianceLevel Load Monitoringin Off-the-Shelf EnergyMonitors: A TechnicalReview[12] indicates a trend to-wards the incorpora-tion of appliance-levele-monitoring and loaddis aggregation, alongwith requirements toimplement load disag-gregation in the nextgeneration e-monitors. Yes No Yes No No No YesAn Overview of Non-Intrusive Load Monitor-ing: Approaches, Busi-ness Applications, andChallenges[39] survey of NILM sys-tem framework andadvanced load disag-gregation algorithms,reviews load signa-ture models, presentsexisting datasets andperformance metrics. Yes Yes Yes Yes No Yes YesLiterature Re-view of PowerDisaggregation[18] reviews the currentstate of the algorithmsand systems of NILM. Yes No No No No No YesLoad DisaggregationTechnologies: RealWorld and LaboratoryPerformance[27] reviews recent fieldstudies and labora-tory tests of NILMtechnologies. Yes No No Yes No Yes Yes
Proposed Comprehen-sive review on energydisaggregation reviews of the state-of-the-art techniques,implements NILMTK,presents analysis anddraws conclusions Yes Yes Yes Yes Yes Yes Yes
Akriti et al.:
Preprint submitted to Elsevier
Page 2 of 11eview on the NILM Algorithms of energy disaggregation algorithms in a reproducible man-ner [5]. Performance evaluation and comparison of NILMalgorithms remain open research challenges for several rea-sons [23] [30] [16]. Table-1 is a tabular representation ofthe literature review compiled from similar survey paperswhich contains synthesised information about the purposeof research and their major findings in this field. It can beobserved that the prospects of NILM application are both,wide and versatile. There have been multiple attempts todevise the generalisation abilities of existing NILM algo-rithms in order to incorporate the methodology to appliance-level monitoring and disaggregation. It is also challengingto collect and store data at the required sample rates in or-der to make data sets available for these algorithms. Thispaper provides a concise and clear overview of the NILMframework and the NILMTK-API by implementing the APIon three publicly available data sets and comparing the ob-served results.The main contributions of this survey article are as fol-lows:1. This paper presents an up to date outline of NILMframework and its related strategies and methods forthe energy disaggregation problem.2. The paper presents an experimental overview of theapplication of NILM-API, which was released withnilmtk-contrib, on three publicly available data sets,draws conclusions and highlights on future researchdirections.3. A detailed overview on the energy disaggregation prob-lem is presented. Here, we have shown the advantagesof the NILM API through which algorithmic compar-isons can be defined with relatively little model knowl-edge.
The remainder of this paper is structured as follows. First,we provide a brief background on the load monitoring ap-proaches which is followed by a formal introduction to theenergy disaggregation problem and a discussion about theavailable implementations and their general framework. Wethen introduce the State-of-the-art Disaggregation Techniques,the NILMTK-contrib repository, and provide an overviewof the supported algorithms. Next, we describe the bench-mark data sets, including the features taken into considera-tion for the experiment API. We then demonstrate the valueof this implementation through an empirical comparison, be-fore summarising our analysis.
2. Background on Load MonitoringApproaches
Load monitoring and identification is a tool for evalu-ating the electrical energy usage and operating condition ofindividual appliances, based on the analysis of the compos-ite load measured from the main power metre in the house.They can supply the customer and the utility with informa-tion such as the type of load, the detail of the electricity consumption and the operating conditions of the appliances.Load monitoring is essential for energy management solu-tions as it gives us statistical insights on appliance energyconsumption and their patterns which can be used for loadscheduling and optimal energy utilisation.
It consists of measuring the electricity consumption ofone or a few appliances using a low-end metering device.The term intrusive, here, means that the meter is located inthe habitation, typically close to the appliance that is mon-itored. The Intrusive load monitoring system may well bea standard metering system that measures the energy con-sumption of an appliance by connecting power meters toeach appliance within the household. Therefore, it requiresentering the house, thus the system is remarked as intrusive.It provides accurate results, however, imposing high costsand an elaborate installation which usually requires wiringand data storage units for the households concerned [1]. In-trusive load monitoring techniques could also be direct or in-direct monitoring techniques [35]. Direct monitoring tech-niques which are called physical intrusive signatures mea-sure the electrical characteristics of each appliance’s powerdemand. The physical intrusive signatures are generated bya tiny low device attached to the power cord of an appliancefor measuring the energy consumption by the appliances.Whenever the appliance is switched on, the device sends asignal to the data collector indicating the operating state ofthe appliances. The power drawn by the appliance is oftencalculated by measuring the electromagnetic field generatedby the flow of current through the wire. This technique pro-vides accurate measurement, but it is not cost effective [35].
It consists of measuring electricity consumption using asmart meter, typically placed at the meter panel. Relying ona single point of measure it is also called one-sensor meter-ing. The qualification of non-intrusive means that no extraequipment is installed in the house. With NILM, the ap-pliance signatures are superposed and, for comprehendingthe contribution of single appliances, they have to be sep-arated. This operation is called dis aggregation of the to-tal electricity consumption. Non-intrusive load monitoringis a convenient means of determining the energy consump-tion and therefore the state of operation of individual appli-ance supported analysis of the aggregate load measured bythe main electric meter in a building. NILM is a process ofanalysing changes within the voltage and current going intoa building and deducing what appliances are utilized in thebuilding also as their individual energy consumption [1].It’s called non-intrusive because it does not require intrud-ing into the house or consumer premises when measuringthe energy consumption of various appliances. Smart me-ters with NILM technology are utilized by companies to sur-vey the particular uses of electrical power in different homes.NILM is taken into account for the cost-effective alternativeto intrusive monitoring techniques. The thought of analysingthe power flow to see household appliances and report on
Akriti et al.:
Preprint submitted to Elsevier
Page 3 of 11eview on the NILM Algorithms their operating condition started when George W. Hart wascollecting and analysing load data as a part of a residentialphoto-voltaic system [13]. The basic monitoring principleis to acknowledge a step change in active and reactive powerwithin the total load produced by altering the operating stateof the various customer’s appliances. A NILM is installedtemporarily to analyse the characteristics of the applianceswhich may be used in suggesting ways of reducing consump-tion and costs.
3. The Energy Disaggregation Problem
The housing structures have changed with urbanisationand immersed the development of high-rise buildings all roundthe world which needs end-use appliance energy conserva-tion in real time. This shift also came together with smart-meters which enabled the estimation of appliance-specificpower consumption from the building’s aggregate power con-sumption reading. Almost two decades back, Hart proposeda method for the dis aggregation of electrical loads by exam-ining only the appliance specific power consumption signa-tures within the aggregated load data. This method is con-sidered to be non-intrusive as it avoids any equipment in-stalled inside the customer’s property. The aggregated datais acquired from the main electrical panel outside the build-ing or the residence. The goal is to partition the whole-housebuilding data into its major constituents. Non-Intrusive LoadMonitoring (NILM) is an attractive method for energy disaggregation, as it can discern devices from the aggregateddata acquired from a single point of measurement.
Non-Intrusive Load Monitoring (NILM) is an attractivemethod for energy disaggregation, as it can discern devicesfrom the aggregated data acquired from a single point ofmeasurement.As mentioned in the nilm toolkit[5][4], the model onwhich NILM works is as follows: For an observed time se-ries of aggregate measurements 𝑌 = ( 𝑌 , 𝑌 , ..., 𝑌 𝑇 ) , where 𝑌 𝑡 ∈ 𝑅 + represents the energy or power (active) measuredin Watt-hours or Watts by an electricity meter at time 𝑡 . Thissignal is assumed to be the aggregation of energy consumedby the component appliances in a building. In the following,we assume there are 𝐼 appliances, and for each appliance,the energy signal is represented as 𝑋 𝑖 = ( 𝑥 𝑖 , 𝑥 𝑖 , ..., 𝑥 𝑖𝑇 ) where 𝑥 𝑖𝑡 ∈ 𝑅 + . The main readings can then be repre-sented as the summation of the appliance signals and 𝐸 𝑡 where 𝐸 𝑡 is an error term. The aim of the energy disag-gregation problem, i.e., NILM, is to recover the unknownsignals 𝑋 𝑖 given only the observed aggregate measurements 𝑌 . Figure-1 demonstrates one such scenario where a high-rise building H contains multiple apartments which are allequipped with a smart meter (SMX, where X is the apart-ment number) and the building itself has a smart-meter ag-gregator (SM0) that is connected to the electricity service.This building can be considered as a meter-group. If theelectricity consumption data for this building is collectedand sampled at a given rate, then applying NILM to this me- Power TimeEnergy dis aggregation
SM3SM2SM1SM0SM4
Utility H Figure 1:
Energy dis aggregation
NILM
Bayesian Procedures Big data applicationsSoft ComputingNeural NetworksSupervisedLSTM UN SupervisedData analyticSimilarity modelsEvent DetectionFCM clustering Fuzzy logic
Figure 2:
State-of-the-art Energy dis aggregation techniques ter group will give us the appliance-wise energy consump-tion patterns for this group.
4. State-of-the-art Dis aggregation Techniques
The energy disaggregation problem was introduced byGeorge Hart in the early 1980s[13]. Owing to the efforts to-wards energy conservation and emission reduction, there hasbeen extensive research in this field which also got boosteddue to smart meter roll outs across different countries. Thereare more than 10 publicly available data sets from across dif-ferent geographies to support significant research activities.The growing interest in this field has led researchers to tryand implement various algorithms for solving the energy disaggregation problem, involving neural networks, big data,soft-computing and statistical methods as shown in Figure-2.
Akriti et al.:
Preprint submitted to Elsevier
Page 4 of 11eview on the NILM Algorithms
Many methods used probabilistic procedures to explic-itly model the energy consumption of an appliance usinga hidden Markov model (HMM) [25], [29], [38], [37].There are both supervised and unsupervised approaches tothe problem including signal processing methods that useappliance’s features for the purpose of disaggregation [14],[19]. Another class of methods utilizes soft-computing tech-niques to solve NILM by employing fuzzy clustering [31],[26], and [32]. Additionally, there is also a set of algorithmsbased on factorization procedures leveraging the low-rankstructure of energy consumption to perform energy disag-gregation [3],[8], and [24]. The open-source API for non-intrusive load monitoring toolkit (NILMTK), was releasedin 2019 in order to facilitate easy comparison of NILM al-gorithms in a reproducible fashion and was meant to be thesource library for energy dis aggregation, data set parsers,and reference benchmark algorithm implementations. Thenilmtk-contrib [6] lessens the barrier to entry for algorithmdevelopers and simplifies the definition of algorithm com-parison experiments by rewriting the disaggregation API andimplementing a new experiment API, along with a numberof functional improvements to the toolkit’s installation pro-cess, data set parsers, and documentation.
Neural networks are being successfully applied to a vari-ety of load scenarios and have achieved better scores in termsof accuracy as well as producing generalization for unseenhouses [22], [15]. Recurrent neural networks and LongShort Term Memory are the most popular neural networkalgorithms. There also exists an RNN based approach forNILM on small power office equipment [2].
Autoencoders are simple neural architectures which areclosely related to principle component analysis and if the ac-tivation function used within the autoencoder is linear withineach layer, the latent variables present at the bottleneck (thesmallest layer in the network) directly corresponds to theprincipal components from PCA. They are unsupervised ma-chine learning algorithms that project the data from a higherdimension to a lower dimension using non-linear transfor-mation while trying to preserve the important features ofthe data and removing the non-essential parts. The encoderfunction of the network compresses or down samples the in-put into a fewer number of bits and maps the original data toa latent space, which is present at the bottleneck whereas thedecoder function tries to reconstruct the input using only theencoding of the input and maps the latent space at the bottle-neck to the output. The denoising autoencoder is a specificdeep neural network architecture designed to extract a par-ticular component from noisy input. They work by addingsome white noise to the data prior to training but comparethe error to the original output when training thereby forcingthe network to not become overfit to arbitrary noise presentin the data. Hence, it subtracts the noise in order to producethe underlying meaningful data. Well-known applications ofa DAE include removing grain from images and reverb from speech signals. It has been used for NILM by consideringthe mains signal to be a noisy representation of the appli-ance power signal, where the mains reading is assumed tobe the sum of the power consumption of the target applianceand noise. Since DAE denoises on a per-appliance basis, itneeds multiple trained models to dis aggregate a group ofappliances. Besides, the DAE here receives a window ofthe mains readings of fixed length and outputs the inferredappliance consumption for the same time window. The ar-chitecture of the network remains as that was proposed innilmtk-contrib.
Recurrent Neural Networks are used to process sequencessuch as a time series prediction or natural language process-ing. It takes one element at a time while retaining the mem-ory of previously encountered states, working on the princi-ple of saving the output of a particular layer (or state) andfeeding this back to the input in order to predict the outputof the layer. Therefore, at each element the model consid-ers not only the current inputs but also what it remembersfrom the preceding elements. This enables the network tolearn long-term dependencies from a series of events whichmeans that the model can take the entire context into consid-eration while making a prediction. An RNN contains a layerof memory cells and the network is in the form of a chainof repeating modules of a neural network. RNNs, however,suffer from the problem of vanishing gradients and explod-ing gradients. The gradients carry information used in theRNN, and when the gradient becomes too small, the param-eter updates become insignificant whereas a gradient explo-sion occurs when large error gradients accumulate, resultingin very large updates to the neural network model weighingduring the training process. These challenges can be solvedusing LSTMs (Long Short Term Memory), that are widelyused memory cells which maintain a cell state as well as acarry for ensuring that the signal is not lost during the pro-cessing of a sequence. It also has a chain-like structure butwith multiple communicating neural network layers whichdecide the amount of data that is required to be retained, thesignificance of the data to be remembered and the part of thememory cell that impacts the output at the given time-step.Moreover, RNNs are a type of neural network that allows forconnections between neurons of the same layer which makesRNNs well suited for sequential data, as in NILM. The em-ployed RNN model receives a sequence of mains readingsand outputs a single value of power consumption of the targetappliance. The network also utilizes long short-term mem-ory (LSTM) units to overcome the vanishing gradient prob-lem by storing values in their built-in memory cell. The ar-chitecture of the network remains as that was proposed innilmtk-contrib.
The sequence to sequence learning model is a deep learn-ing concept which is used to convert from one sequence toanother. It contains an encoder RNN to understand the input
Akriti et al.:
Preprint submitted to Elsevier
Page 5 of 11eview on the NILM Algorithms sequence and a decoder RNN to decode the thought vectorthereby constructing an output sequence. An encoder net-work condenses an input sequence into a vector, and a de-coder network unfolds that vector into a new sequence. Thecommon denominator between encoder and decoder archi-tectures in the sequence-to-sequence model are RecurrentNeural Networks. In the encoder, a new word of the in-put sentence is fed at each step of recurrence which is ex-ploited by the next state in the sequence in the subsequentstep. The decoder starts by receiving as input the final stategenerated by the encoder, that ideally contains all the infor-mation stored inside the input sentence and thereafter theRNN carries it to each successive step which aims to pre-dict the output by using a discrete probability distributionand the loss function.. Here, the main idea is to learn a re-gression map from the mains sequence to the correspondingtarget appliance sequence. The seq2seq model is defined bythe regression 𝑥 𝑡 ∶ 𝑡 + 𝑊 ∕1 = 𝑓 ( 𝑌 𝑡 ∶ 𝑡 + 𝑊 ∕1 , 𝑄 ) + 𝐸 𝑡 where 𝑓 is a neural network. The architecture of the networkand other hyperparameters remain as they were proposed innilmtk-contrib [5]. Sequence to point learning (seq2point) models the in-put of the network as a mains window, and the output as themidpoint element of the corresponding window of the tar-get appliance. The intuition behind this method is that themidpoint of the target appliance should have a strong corre-lation with the mains information before and after that timepoint. Seq2point learning could be viewed as a non-linearregression. The architecture of the network and other hyper-parameters remain as they were proposed in nilmtk-contrib.
Gated Recurrent Unit (GRU) is a new generation of Neu-ral Networks that attempts to reduce the computational de-mand while maintaining the same performance by replac-ing the LSTM units by light weight Gated Recurrent Units(GRU) and optimising the recurrent layer sizes to reduce re-dundancy as well as minimising the risk of a vanishing gra-dient. They are a solution to short-term memory with in-ternal mechanisms called gates that can regulate the flow ofinformation. These gates learn which part of the data is im-portant and pass relevant information down the long chainof sequences. GRUs are similar to LSTM except that theydo not have the cell state and use the hidden state to transferinformation. It also only has two gates, a reset gate and up-date gate. The reset gate determines how much of the pastknowledge to forget whereas the update gate decides what in-formation to throw away and what new information to add.The working of GRU continues such that when the reset gateis near to zero, the hidden state is constrained to disregard thepast hidden state and is reset with the current input. Thispermits the hidden state to discard any information that’sfound to be insignificant within the future. This result per-mits a more compact representation. Whereas the upgradegate controls how much information from the past hidden state will be exchanged to the current hidden state. The ac-tuation of the GRU at a specific time may be a straight addi-tion between the past actuation and the candidate activation,where an update door chooses how much the unit overhaulsits activation or content.The Online GRU model receives thelast available mains readings 𝑌 𝑡 ∶ 𝑡 + 𝑊 ∕1 as input and usesthem to calculate the power consumption 𝑥 𝑗 ( 𝑡 + 𝑊 ∕1) of asingle appliance 𝑗 , for the last time point. The architecture ofthe network and other hyperparameters remain as they wereproposed in nilmtk-contrib. A good deal of big-data approaches have popped for en-ergy disaggregation with the application of data analyticsin smart meters [36] such as Neighbourhood NILM whichworks on the intuition that ‘similar’ homes have ‘similar’ en-ergy consumption on a per-appliance basis [7] and a scalablethree-level learning framework for smart cities that matchesthe hierarchical nature of big data generated by smart citieswith a goal of providing different levels of knowledge ab-stractions [28].
Feature extraction and pattern recognition tasks for non-intrusive residential electrical consumption using fuzzy logichave proven to be feasible. There are soft computing tech-niques to identify the behavior of each of the devices fromaggregated consumption records [31]. A fuzzy classifierwith the Fuzzy C-Means (FCM) clustering and optimizationalgorithms to identify the energizing and de-energizing sta-tuses of each appliance, has also been proposed [26].
Bayesian algorithms are being used for event detectionand identification of the individual contribution of appliances.Findings for event detection method based on cepstrum smooth-ing [10] and LSTM models [20] along with a modular wayto address multi-dimensionality issues that arise when thenumber of appliances increase [21] have been devised.
5. Datasets and Algorithms
The Reference Energy Disaggregation Data Set (REDD),is a freely available data set containing detailed power usageinformation from several homes, which is aimed at further-ing research on energy dis aggregation. REDD is the firstpublic energy dataset released by MIT in 2011. REDD con-tains high and low-frequency readings from 6 households inthe USA recorded for short period (between a few weeks anda few months). This data set is widely used for the evaluationof NILM algorithms.
UK-DALE is an open-access data set from the UK record-ing Domestic Appliance-Level Electricity at a sample rate of
Akriti et al.:
Preprint submitted to Elsevier
Page 6 of 11eview on the NILM Algorithms
Figure 3:
Predicted energy consumption: IAWE 𝑘𝐻𝑧 for the whole-house and at 𝐻𝑧 for individual ap-pliances. The data set contains 16 kHz current and voltageaggregate meter readings and 6 second sub-metered powerdata from individual appliances across 3 UK homes, as wellas 1 second aggregate and 6 second sub-metered power datafor 2 additional homes. An update to the data set was re-leased in August 2015 which has expanded the data availablefor house 1 to 2.5 years. The Indraprastha Institute of Information Technology re-cently released the iAWE data set, which contains aggregateand sub-metered electricity and gas data from 33 householdsensors at 1 second resolution. The data set covers 73 daysof a single house in Delhi, India.
For the purpose of this experiment, five neural networkbased algorithms have been utilised, namely: DAE, RNN,Sequence-to-Sequence, Sequence-to-Point and OnlineGRU,all of which have been described in detail in section 4.1 ofthis article.
The tests were run on a machine with GeForce GTX1660 Ti/PCIe/SSE2 GPUs with Intel® Core™ i7-9750H CPU@ 2.60GHz × 12 and 16 GB RAM. The sample period was60 seconds. All neural algorithms were trained for 5 epochswith a batch size of 32.
For each data set, the network was trained for a periodof 30 days based on the data and tested on the subsequent20 days using the nilmtk-contrib API to predict the energy
Akriti et al.:
Preprint submitted to Elsevier
Page 7 of 11eview on the NILM Algorithms
Table 2
MAE against Ground Truth for Iawe
MAE- Iawe WindowGRU RNN DAE Seq2Point Seq2Seq washing machine 66.726196 65.665489 65.275597 65.736549 65.298805fridge 53.850956 68.835884 51.917274 36.775997 54.116047computer 24.112394 24.148933 24.46452 24.64801 24.25812air conditioner 169.634262 139.516449 139.706543 120.483185 141.557159television 1.347255 1.353503 1.347071 1.475708 1.449378
Figure 4:
Predicted energy consumption: REDD consumed per appliance and report the mean absolute er-ror(MAE) in each case. The outputs from the tests on thethree datasets under consideration do not collectively revealthe best performer out of the five algorithms used. More-over, there is no clear winner in each case except UK-dalewhere Sequence-to-Sequence leads. This could be due to thedifferences in the data used. Mean absolute error measures the average magnitude of the errors in a set of predictions,without considering their direction. It’s the average over thetest sample of the absolute differences between predictionand actual observation where all individual differences haveequal weight. It expresses the average model prediction er-ror in units of the variable of interest. The metric rangesfrom 0 to ∞ and is indifferent to the direction of errors. It is Akriti et al.:
Preprint submitted to Elsevier
Page 8 of 11eview on the NILM Algorithms
Table 3
MAE against Ground Truth for REDD
MAE- Redd WindowGRU RNN DAE Seq2Point Seq2Seq washing machine 49.6175 60.661507 49.541073 47.514114 52.302326fridge 59.622471 68.384491 63.540127 55.803658 63.520336light 30.128523 32.435764 31.222496 31.245703 31.557386sockets 0.870423 0.888707 0.939906 0.827897 0.883221microwave 28.598331 28.599768 34.998123 21.999201 24.803219
Figure 5:
Predicted energy consumption: UKdale a negatively-oriented score, which means lower values arebetter. It effectively describes the magnitude of residualsand does not indicate the under performance or over per-formance of the model. Each residual contributes propor-tionally to the total amount of error, meaning that larger er-rors will contribute linearly to the overall error. It is notice-able from Table-2 that there is no clear winner in the case of iAWE and the mean absolute errors are very high whichcould be because the data set is based on a single housingfacility. The mean absolute errors remain similar in cases ofwashing machine, computer and television with the reportederrors being the minimum for television. The case of air con-ditioner reports the maximum observed error from all threescenarios which is greater than 100 for all five algorithms
Akriti et al.:
Preprint submitted to Elsevier
Page 9 of 11eview on the NILM Algorithms
Table 4
MAE against Ground Truth for UKdale
MAE-UKdale WindowGRU RNN DAE Seq2Point Seq2Seq washing machine 50.239201 55.136917 53.854832 49.03191 43.448345fridge 44.169342 44.513477 45.257683 44.960304 44.8424light 33.446453 34.21154 35.531647 33.719086 35.57534dish washer 49.290272 47.874489 39.505829 49.656605 41.864983microwave 21.169703 16.606073 25.805212 20.247181 19.278296 and varies largely for every algorithm. The performance ofSequence-to-Sequence and DAE remains similar across allfive appliances with DAE performing the best for the televi-sion and Sequence-to-Sequence performing the best for thewashing machine. It can also be observed from Figure-3, thedifference in the load cycles of different appliances.It is evident from Table-3 that none of the algorithms inconsideration perform the best overall in the case of REDDalthough Sequence-to-Point leads in four of the five mostused appliances. This could be due to the dissimilar usagetrends of the appliances that the graphs clearly demonstratein Figure-4, hence, signifying the energy consumption pat-terns of the appliances. The mean absolute errors are thelowest in this case making it a preferred data set to base ourconclusions on.The mean absolute errors in for the sockets remain lessthan 1 which is the minimum error reported by an algorithmin all three scenarios. The errors show maximum variationin the case of the washing machine followed by the fridge.There is little variation in the reported values of mean ab-solute error for the sockets which is again followed by thelight. The performance of WindowGRU and RNN is com-parable across light, sockets and microwave but differ largelyin cases of the washing machine and fridge.Table-4 shows that Sequence-to-Sequence performs thebest for most cases of UK-Dale and the mean absolute errors,although high, are similar for nearly all appliances. The pre-dictions from the graph indicate how some appliances haveclearly defined usage patterns (as seen in Figure-5) whereasothers are used on a daily basis. The mean absolute errorstays the lowest in the case of microwave with 16.60 beingthe minimum as reported by RNN. It can also be noticed thatthe mean absolute error for the fridge remains similar acrossall five algorithms followed by little variation in the case oflight and shows the maximum variation for the washing ma-chine.Different appliances use different quantities of energyand, thus, compared to a high energy consumption load, theerrors measured relative to low energy use might be less sig-nificant. Furthermore, some end devices operate less fre-quently than others, so if a statistically substantial numberof run times is not recorded, metric results may not be in-dicative of efficiency. In view of these problems, it is recom-mended that every metric assessment reflects some levelingof the use of energy or other basis for appropriate compar-isons of results across different end uses which coule be a fixed energy usage per end use, a fixed number of real eventsper end use, or a fixed period that is capable of capturing avariety of conditions.
6. Conclusion and future work
In this paper, we demonstrated how nilmtk-contrib pro-vides an interface to energy disaggregation problems andalso the API through which algorithmic comparisons can bedefined with relatively little model knowledge, thus enablingempirical evaluations to be easily generated. Therefore, in-creasing the rate of progress within the field and supportingthe progress of research in NILM.Our experimental results carried out on three publiclyavailable data sets namely IAWE, REDD and UKdale donot indicate the presence of any patterns in the output. Thiscould be a result of the large differences in the data sets. Al-though, it suggests that the API can be used for data sets fromdifferent geographies.With the need for energy and resources rising each day,careful and efficient utilisation is the only way, to conserveit and energy dis-aggregation can be an essential elementin the conservation of energy since it elaborates the energyusage tendencies. The trends observed in the energy-usagepatterns from a household can be used for the purpose ofsecurity as an anomaly in them might represent a sign ofappliance failure or illegal use of supplied electricity. Theappliance usage patterns can also be used to calculate andcontrol the amount of carbon emissions.
References [1] Aladesanmi, E., Folly, K., 2015. Overview of non-intrusive load mon-itoring and identification techniques. IFAC-PapersOnLine 48, 415–420.[2] Astal, M..T.E., Kalloub, M., Abu-Hudrouss, A.M., Frey, G., 2020.Office appliances identification and monitoring using deep leaningbased energy disaggregation for smart buildings, in: IECON 2020The 46th Annual Conference of the IEEE Industrial Electronics So-ciety, pp. 1986–1991. doi: .[3] Batra, N., Jia, Y., 2018. Transferring decomposed tensors for scal-able energy breakdown across regions, in: The Thirty-Second AAAIConference on Artificial Intelligence (AAAI-18).[4] Batra, N., Kelly, J., Parson, O., Dutta, H., Knottenbelt, W., Rogers,A., Singh, A., Srivastava, M., 2014. Nilmtk: An open source toolkitfor non-intrusive load monitoring, in: Proceedings of the 5th Interna-tional Conference on Future Energy Systems, Association for Com-puting Machinery, New York, NY, USA. p. 265–276. URL: https://doi.org/10.1145/2602044.2602051 , doi: . Akriti et al.:
Preprint submitted to Elsevier
Page 10 of 11eview on the NILM Algorithms [5] Batra, N., Kukunuri, R., Pandey, A., Malakar, R., Kumar, R., Krys-talakos, O., Zhong, M., Meira, P., Parson, O., 2019a. Towards re-producible state-of-the-art energy disaggregation, in: Proceedingsof the 6th ACM International Conference on Systems for Energy-Efficient Buildings, Cities, and Transportation, Association for Com-puting Machinery, New York, NY, USA. p. 193–202. URL: https://doi.org/10.1145/3360322.3360844 , doi: .[6] Batra, N., Kukunuri, R., Pandey, A., Malakar, R., Kumar, R., Krysta-lakos, O., Zhong, M., Meira, P., Parson, O., 2019b. Towards repro-ducible state-of-the-art energy disaggregation, in: Proceedings of the6th ACM International Conference on Systems for Energy-EfficientBuildings, Cities, and Transportation, pp. 193–202.[7] Batra, N., Singh, A., Whitehouse, K., 2015. Neighbourhood nilm: Abig-data approach to household energy disaggregation. arXiv preprintarXiv:1511.02900 .[8] Batra, N., Wang, H., Singh, A., Whitehouse, K., 2017. Matrix factori-sation for scalable energy breakdown, in: Proceedings of the... AAAIConference on Artificial Intelligence.[9] Darby, S., et al., 2006. The effectiveness of feedback on energyconsumption. A Review for DEFRA of the Literature on Metering,Billing and direct Displays 486, 26.[10] De Baets, L., Ruyssinck, J., Develder, C., Dhaene, T., Deschrijver, D.,2017. On the bayesian optimization and robustness of event detectionmethods in nilm. Energy and Buildings 145, 57–66.[11] Faustine, A., Mvungi, N.H., Kaijage, S., Michael, K., 2017. A sur-vey on non-intrusive load monitoring methodies and techniques forenergy disaggregation problem. arXiv preprint arXiv:1703.00785 .[12] Haq, A.U., Jacobsen, H.A., 2018. Prospects of appliance-level loadmonitoring in off-the-shelf energy monitors: A technical review. En-ergies 11, 189.[13] Hart, G.W., 1992. Nonintrusive appliance load monitoring. Proceed-ings of the IEEE 80, 1870–1891. doi: .[14] He, K., Stankovic, L., Liao, J., Stankovic, V., 2018. Non-intrusiveload disaggregation using graph signal processing. IEEE Transactionson Smart Grid 9, 1739–1747. doi: .[15] He, W., Chai, Y., 2016. An empirical study on energy disaggregationvia deep learning, in: 2016 2nd International Conference on ArtificialIntelligence and Industrial Engineering (AIIE 2016), Atlantis Press.pp. 338–342.[16] Herrero, J.R., Murciego, Á.L., Barriuso, A.L., de La Iglesia, D.H.,González, G.V., Rodríguez, J.M.C., Carreira, R., 2017. Non intrusiveload monitoring (nilm): A state of the art, in: International Confer-ence on Practical Applications of Agents and Multi-Agent Systems,Springer. pp. 125–138.[17] Himeur, Y., Alsalemi, A., Bensaali, F., Amira, A., 2020. Buildingpower consumption datasets: Survey, taxonomy and future directions.Energy and Buildings 227, 110404. URL: , doi: https://doi.org/10.1016/j.enbuild.2020.110404 .[18] Jiang, L., Li, J., Luo, S., Jin, J., West, S., 2011. Literature review ofpower disaggregation, in: Proceedings of 2011 International Confer-ence on Modelling, Identification and Control, IEEE. pp. 38–42.[19] Jin, Y., Tebekaemi, E., Berges, M., Soibelman, L., 2011. A time-frequency approach for event detection in non-intrusive load moni-toring, in: Signal Processing, Sensor Fusion, and Target RecognitionXX, International Society for Optics and Photonics. p. 80501U.[20] Jones, R., Klemenjak, C., Makonin, S., Bajić, I.V., 2020. Stop: Ex-ploring bayesian surprise to better train nilm, in: Proceedings of the5th International Workshop on Non-Intrusive Load Monitoring, pp.39–43.[21] Kaselimi, M., Doulamis, N., Doulamis, A., Voulodimos, A., Pro-topapadakis, E., 2019. Bayesian-optimized bidirectional lstm regres-sion model for non-intrusive load monitoring, in: ICASSP 2019-2019IEEE International Conference on Acoustics, Speech and Signal Pro-cessing (ICASSP), IEEE. pp. 2747–2751.[22] Kelly, J., Knottenbelt, W., 2015. Neural nilm: Deep neural networksapplied to energy disaggregation, in: Proceedings of the 2nd ACM in-ternational conference on embedded systems for energy-efficient built environments, pp. 55–64.[23] Klemenjak, C., 2018. On performance evaluation and machine learn-ing approaches in non-intrusive load monitoring. Energy Informatics1, 391–395.[24] Kolter, J.Z., Batra, S., Ng, A.Y., 2010. Energy disaggregation via dis-criminative sparse coding, in: Advances in Neural Information Pro-cessing Systems, pp. 1153–1161.[25] Kolter, J.Z., Johnson, M.J., 2011. Redd: A public data set for energydisaggregation research, in: Workshop on data mining applications insustainability (SIGKDD), San Diego, CA, pp. 59–62.[26] Lin, Y.H., Tsai, M.S., Chen, C.S., 2011. Applications of fuzzy classi-fication with fuzzy c-means clustering and optimization strategies forload identification in nilm systems, in: 2011 IEEE international con-ference on fuzzy systems (FUZZ-IEEE 2011), IEEE. pp. 859–866.[27] Mayhorn, E.T., Sullivan, G.P., Petersen, J.M., Butner, R.S., Johnson,E.M., 2016. Load disaggregation technologies: real world and labo-ratory performance. Pacific Northwest National Laboratory (PNNL),Richland, WA (US), Tech. Rep. PNNL-SA-116560 .[28] Mohammadi, M., Al-Fuqaha, A., 2018. Enabling cognitive smartcities using big data and machine learning: Approaches and chal-lenges. IEEE Communications Magazine 56, 94–101.[29] Parson, O., Ghosh, S., Weal, M., Rogers, A., Prior, N.I.L.M.U., 2012.Models of general appliance types.[30] Pereira, L., Nunes, N., 2018. Performance evaluation in non-intrusiveload monitoring: Datasets, metrics, and tools—a review. Wiley Inter-disciplinary Reviews: data mining and knowledge discovery 8, e1265.[31] Puente, C., Palacios, R., González-Arechavala, Y., Sánchez-Úbeda,E.F., 2020. Non-intrusive load monitoring (nilm) for energy disag-gregation using soft computing techniques. Energies 13, 3117.[32] Shenavar, P., Farjah, E., 2007. Novel embedded real-time nilm forelectric loads disaggregating and diagnostic, in: EUROCON 2007-The International Conference on" Computer as a Tool", IEEE. pp.1555–1560.[33] Shorrock, L., Utley, J., et al., 2003. Domestic energy fact file 2003.Citeseer.[34] Sun, M., Nakoty, F.M., Liu, Q., Liu, X., Yang, Y., Shen, T., 2019.Non-intrusive load monitoring system framework and load disaggre-gation algorithms: A survey, in: 2019 International Conference onAdvanced Mechatronic Systems (ICAMechS), pp. 284–288. doi: .[35] Zeifman, M., Roth, K., Stefan, J., 2013. Automatic recognition of ma-jor end-uses in disaggregation of home energy display data, in: 2013IEEE International Conference on Consumer Electronics (ICCE), pp.104–105. doi: .[36] Zhang, Y., Huang, T., Bompard, E.F., 2018. Big data analytics insmart grids: a review. Energy informatics 1, 1–24.[37] Zhong, M., Goddard, N., Sutton, C., 2014. Signal aggregate con-straints in additive factorial hmms, with application to energy disag-gregation, in: Advances in Neural Information Processing Systems,pp. 3590–3598.[38] Zhong, M., Goddard, N., Sutton, C., 2015. Latent bayesian meld-ing for integrating individual and population models, in: Advances inneural information processing systems, pp. 3618–3626.[39] Zhuang, M., Shahidehpour, M., Li, Z., 2018. An overview of non-intrusive load monitoring: Approaches, business applications, andchallenges, in: 2018 International Conference on Power System Tech-nology (POWERCON), IEEE. pp. 4291–4299.[40] Zoha, A., Gluhak, A., Imran, M.A., Rajasegarar, S., 2012. Non-intrusive load monitoring approaches for disaggregated energy sens-ing: A survey. Sensors 12, 16838–16866.
Akriti et al.: