WO2015059272A1 - Improved non-intrusive appliance load monitoring method and device - Google Patents

Abstract

The current invention concerns a NIALM method for monitoring the electrical appliances in a system comprising the steps of: detecting events in a measured electrical signal comprising power consumption information of said system, said events representing state transitions of the appliances in the system; characterizing said events by an event signature, taking into account differences between steady states before and after the event and/or taking into account transient states located between steady states before and after the event; clustering events into a set of clusters on the basis of their signatures; identifying components on the basis of said set of clusters; identifying appliances from the components identified in the previous step;optionally providing appliances energy consumption, characterized in that said clustering comprises an initial clustering of said events into an initial set of clusters on the basis of a first clustering criterion, and a subsequent reclustering of at least one of said initial clusters on the basis of a second clustering criterion different from the first clustering criterion.

Description

Improved non-intrusive appliance load monitoring method and

device

Technical field

The invention pertains to the technical field of automatic appliance detection, in particular Non-Intrusive Appliance Load Monitoring (NIALM), which refers to the automated detection of the state of appliances, e.g. household appliances, industrial appliances, a company's appliances, etc., from a total energy consumption signal. More in particular, the present invention relates to obtaining information at the component level of electrical appliances in a system, e.g. a household or a company, from a measurement of the total electricity consumption of said system. This information at the component level may include but is not limited to: the presence, the state and/or the energy consumption. The system may also be comprised of a single or a few appliances, whereby the methods, devices and/or systems of the present invention can be used for condition monitoring of one or more appliances. Background

WO 2012/160062 Al discloses a method for detecting transitions in a measured signal e.g. a discrete time signal and current waveform, which is induced by elements i.e. components of appliances, of a physical system i.e. a house. The method involves generating a residual signal, i.e. transition likelihood signal, from a measured signal, i.e. a discrete time signal, where the residual signal is provided with high amplitude when transitions occur and with low amplitude in other cases. Rules concluding that transitions occur when the residual signal is larger than a threshold stable value, are provided, where the stable value is automatically defined from local values of the residual signal and defined as a function of local background noise. The measured signal is filtered before generating residual signal. The method enables defining a time index corresponding to end of transient state to maximize a distance between transient states and a straight line passing through the transient state, thus improving separation between transient and steady states, and hence detecting transitions in the measured signal, and monitoring automatic-setup non-intrusive appliance load for identifying appliances energy consumption in a reliable manner. This document also discloses an automatic-setup non-intrusive appliance load monitoring method (50) for identifying appliances energy consumption and comprising the steps of:

• detecting transitions in a measured signal;

• characterizing differences between steady states before and after these transitions; • characterizing transient states located between steady states;

• identifying components;

• identifying appliances from the components identified in the previous step;

• providing appliances energy consumption. The step of identifying components is hereby performed by selecting spectral features or features describing a current waveform for the steady states, with the possibility of the chosen features being grouped together in well-separated clusters for different components.

Hereby, a first step of component identification consists in clustering and classifying the components according to their nature (motor, resistor, heater, television are examples of different natures). A second step classifies the components within their nature cluster; this is the component classification itself.

Identifying a nature of a component is a classification problem with a predefined number of classes. Therefore, a K-means algorithm is preferably chosen. A K- means method is a clustering technique that partitions n observations into K clusters, where K is a predefined value. It classifies the components according to their nature; there will be as much classes as defined component natures. Three examples of different natures are resistors, motors and electronic devices. Not all features are relevant when looking only at a component nature. Preferably, when searching a nature of a component, one can consider THD (total harmonic distortion), or ratio Q/P (Q = reactive power and P = active power).

The second classification step is far different because the number of clusters to identify is generally large and unknown; no prior information is generally available about a number of appliances contained in a house, so no prior information is available concerning the identification of components within their nature. Hence, a DBScan (Density Based Spatial Clustering for Applications with Noise) classification algorithm is preferably used. Looking at the steady states, the components will differ for instance by a magnitude of their consumption from others within a component nature cluster. As far as the transient state features are concerned, two nearly identical but different components could have different transient shapes. Note hereby, that although the number of clusters does not have to be specified up front when using the DBSCAN algorithm, the number of clusters which is found by the DBSCAN algorithm has been observed not always to correspond with the actual number of electrical components in the system. Although the above methodology, in particular regarding the second step above, may work in a controlled set-up, it has been found that problems (that are sometimes important) remain in real-life application of the above clustering and component identification methods. More in particular, the fact that the number of clusters to identify is unknown for any household, company, industry or other environment, poses severe problems for present state of the art clustering techniques within the field of NIALM.

WO2009/103998 discloses a method of inference of appliance usage from a point measurement on a supply line, said supply line being common to multiple appliances and/or components of appliances comprising the steps of:

• obtaining data from said measurement point;

• sampling power and reactive power at intervals substantially throughout periods of operation of said appliances or components of appliances corresponding to appliances or components of appliances being in ON and/or OFF modes of use;

• identifying characteristics of events by assessing power and reactive power change during an event; and by

• assessing one or more additional characteristics derivable from said power and reactive power to characterize an appliance;

· grouping events and/or cycles of events into clusters of similar characteristics; and

• inferring appliance usage based on said grouping.

This document does not seem to disclose how to separate clusters of events in an automated method, independent of the system is being monitored. This document discloses the possibility of using a look-up table in which ranges of clusters' properties are provided that correspond to particular appliances and appliance components. Obviously, such a look-up table technique will depend highly on the quality and continuous updating of the look-up table, and will allow identification of a limited number of electrical components only. Furthermore, this document discloses a step- wise cluster identification on the basis of the distance of events, whereby events within a pre-set distance are assumed to belong to the same cluster, after which the events belonging to the cluster with the greatest number of events are removed from the data, and the next cluster is identified.

Note that methods such as the one described in WO2009/103998, use a pre-set maximal inter-event distance, on the basis of which events are assigned to a cluster, whereby clusters identification is made at least partially on the basis of a minimal number of events assigned to that cluster. Both steps introduce a certain degree of arbitrariness into the method which results both in uncertainties with respect to the correctness, as well as in errors for the resulting identified clusters as well.

Prior art techniques for identifying electrical components within a system from a measured electrical power consumption signal, have been noticed to require extra input with respect to the appliances comprised in the system, either by simply providing the method with characteristic signals of each single appliance up front, or by providing an extensive look-up table of characteristic signals of appliances. Such techniques could be deemed impractical for systems in which the number or type of appliances changes over time, such as a common household or a company, and could be deemed unworkable as it requires constant updating of a look-up table with characteristic signals from all possible appliances.

Prior art techniques for identifying electrical components within a system from a measured electrical power consumption signal, have been seen to be not completely reliable with respect to the grouping of events extracted from the measured signal into clusters. Furthermore, prior art techniques do not seem to allow fully computing or estimating the uncertainty or error which arises from the clustering of the events.

The present invention provides a NIALM method which allows improved component detection, in particular in the case no prior information on the components or appliances in the system is known. In such a case, the number of components (N) is a priori unknown and an automated determination of this number of components is required. Hereby, each correctly detected event can be associated to one of the N components. The present invention hereto provides an improved clustering technique for events, which in cooperation with the other steps of the method, allows better identifying electrical components.

More in particular, techniques proposed in the literature use prior information about parametric finite state machine (FSM) models, parameters being automatically tuned to fit to the observations. As reported by Johnson and Willsky, "Bayesian Nonparametric Hidden Semi-Markov Models", arXiv preprint, 2012, this information should be learned from the data. Since information can be extracted from the aggregate consumption as e.g. detailed in this document, one could wonder if this information can be straightforwardly computed to obtain component state sequences. Given the assumption that appliances are either on/off loads or systems of such loads, the ideal case for the evaluation of the state sequence is as follows: first, events correspond to on/off switching of components and there are no fictitious events and, second, for a given component, signatures of both on and off transitions belong to one single cluster. Further, no events related to other components share that cluster. In this ideal case, components are directly represented by clusters, their operations correspond to time intervals between on and off events and the evaluation of their state sequence simply consists to pair successive on and off events that are within identical clusters. The aggregate consumption is then modeled by combining the individual sequences.

Actual situations differ from this ideal case in several ways: SI :

Some on/off switching are missing from the event sequence and there are fictitious events.

Why: Because there are false (positive and negative) detections in the segmentation step, or because there is noise or non-linearities.

So what: Neither complementary events to missing ones nor fictitious detections should be paired with other events. S2 :

Events of a single component are classified into different clusters.

Why: On and off event signatures might differ because of nonideal segmentation of turn-on transients or variable consumption of loads during their operations, because of noise or because of non-linearities.

- So what: On and off events of different clusters should be paired to obtain component state sequences.

S3 :

Events corresponding to different components belong to a single cluster.

Why: There are components with similar properties w. r.t. the clustering parameters (or completely similar components).

So what: Events of that single cluster must be distributed into different component state sequences.

S4 :

There are events whose signatures are combinations of others

- Why: Some on/off switching are simultaneous switch events.

So what: These events should be paired to several events distributed in different clusters. The problem consists to find the underlying electrical components responsible for the observed aggregate consumption. Only then could states sequences be correctly evaluated and cope with the deviations from the ideal case (SI to S4).

Ideally, one cluster corresponds to one electric component and inversely. In practice, several clusters might correspond to a single component, e.g. if turn-on and turn-off features differ, and several components might correspond to a single cluster, e.g. a household or a company which own many similar electrical components. Therefore, prior art methods using clustering do not automatically lead to component detection.

The present invention provides a NIALM method which allows an improved component detection, in particular in case a system comprises one or more components having different turn-on and turn-off features and thus giving rise to separate clusters, or in case of a system comprising many similar electrical components which gives rise to events which are attributed to one cluster, or in any of the above four cases SI to S4. The present invention hereto provides an improved component detection technique, which in cooperation with the other steps of the NIALM method, allows to better identify electrical components on the basis of clusters of events.

Summary of the invention

The present invention provides a non-intrusive appliance load monitoring (NIALM) method for monitoring (preferably electrical) components of (preferably electrical) appliances in a system. The NIALM method comprises the steps of: detecting events in an electrical signal (that is preferably a measured electrical signal), said electrical signal comprising power consumption information of said system, said events representing state transitions of the appliances in the system;

- characterizing each event by an event signature, by taking into account differences between steady states before and after each event and/or by taking into account transient states located between steady states before and after each event;

clustering events into a set of clusters on the basis of their event signatures; - identifying components on the basis of said set of clusters;

optionally identifying appliances from the components identified in the previous step;

optionally providing appliances' energy consumption or providing the components' energy consumption. Hereby, said clustering comprises an initial clustering of said events into an initial set of clusters on a basis of a first clustering criterion, and a subsequent reclustering of at least one of said initial clusters on the basis of a second clustering criterion different from the first clustering criterion. Preferably said first clustering criterion is computed by taking into account event signatures from subtantially all events detected in said electrical signal, and said second clustering criterion is computed by taking into account event signatures of essentially only events in said one initial cluster which is to be reel uste red.

Preferably, said electrical signal is a measured electrical signal. The NIALM method can be used to identify (preferably electrical) components in a system on a basis of a (preferably measured) electrical signal comprising power consumption information of said system. The identification of the (preferably electrical) components allows a further identification of (preferably electrical) appliances, which is particularly useful in systems where the number and/or nature of appliances is not known a priori, or can change in time. Therefore, in a preferred embodiment, the method comprises a step of identifying appliances from the components on the basis of the identified (preferably electrical) components.

The NIALM method can also be used for condition monitoring of an appliance. Typically, but not in a limitative way, condition monitoring of an appliance can be performed to check if or when the appliance is functioning normally or abnormally, and/or to give an alarm if a malfunction is observed. The appliance which is being monitored using the NIALM method of the present invention can be identified using a method as described in this document, or by another method.

In a preferred embodiment, the method comprises providing energy consumption of appliances or of components. Energy consumption of appliances or of components can be given in a form of a report to a user of an appliance or of the system, or to the electricity provider to detect malfunctions or to optimize energy efficiency either for a single system or for a number of systems.

The present invention further provides for a recursive clustering method for clustering events obtained from a (preferably measured) electrical signal comprising power consumption information of a system, said events being at least partially characterized by an event signature, said recursive clustering method comprising : an initial step of clustering events of an initial event set into a set of clusters using a clustering criterion for deciding whether two events belong to the same cluster and/or whether an event belongs to an existing cluster, said clustering criterion being computed on a basis of event signatures from substantially all events from said initial event set; and

a recursive step of reclustering events belonging to a cluster into a set of sub- clusters using an updated clustering criterion for deciding whether two events belong to the same cluster and/or whether an event belongs to an existing cluster, said updated clustering criterion being computed on a basis of event signatures substantially only of events from said cluster, thereby obtaining one, two or more sub-clusters, preferably whereby if two or more sub-clusters are obtained, these two or more sub-clusters are reclustered using the present recursive step.

The present invention also provides for a method for clustering events detected in a (preferably measured) electrical signal comprising electrical power consumption information of a system, each of said events being at least partially characterized by an event signature comprising a set of event features, the method comprising the steps of:

(a) defining principal components of said event features for said events;

(b) computing a clustering criterion on the basis of the principal components of substantially all of said events;

(c) clustering said events into a set of clusters; and being characterized in that it comprises the steps of:

(d) sub-clustering at least one cluster of said set of clusters by:

(i) redefining principal components of the event features for events belonging to said one cluster, thereby substantially excluding event features for events not belonging to said one cluster;

(ii) recomputing the clustering criterion for said one cluster on the basis of the redefined principal components of substantially all of the events belonging to said one cluster;

(iii) clustering said events belonging to said one cluster in one, two or more sub-clusters;

(e) optionally, if two or more sub-clusters are obtained after step (d)(iii), performing step (d) taking at least one of said sub-clusters from step (d)(iii) as a cluster to be sub-clustered.

Hereby, step (d) is preferably performed recursively for each cluster of said set of clusters and/or for all sub-clusters obtained by performing step (d) until no further sub-clustering in two or more sub-clusters is obtained. Typically, the events relate to on-off switching of components of appliances in the system . Events can be classified by their signatures, which refer to a set of properties or features which characterize transitions in the measured signal. These properties or features can be computed from a current waveform extracted from the measured signal, and preferably from a delta waveform extracted from the measured signal. More preferably, features are computed on the basis of a frequency spectrum of the delta waveform, for example features could relate to harmonic magnitudes or phases. Preferred features which are computed to characterize events are : active power (P), reactive power (Q), fundamental frequency (Hi), odd harmonics preferably of order 3 to 13 and preferably normalized to the amplitude of the fundamental frequency com ponent of the current, continuous component and/or second order harmonics which a re preferably normalized to the amplitude of the fundamental frequency com ponent of the current, total harmonic distortion (THD), ratio | Q/P| , maximal I_M and minima l I_m values of the current. As a non-limiting, but preferred example, a signature of an event k, preferably a steady-state event, could be written as X_ss(k) = [P, Q, Hi, H₀/i, H₂/i, H^odd(_3→i3₎/i, THD, | Q/P| , I_M, I_m], wherein H_n/i refers to the n'th Harmonic normalized to the magnitude of the fundamental frequency component of the current.

In a preferred embodiment, and most preferably for transient state events, a signature of an event comprises a time series of a quantity derivable from the electrica l signa l, preferably a quantity being the active power and/or the current. Typically said time series of sa id quantity comprises a set of va lues for sa id number ordered chronologica lly.

When clustering events, a decision criterion, or clustering criterion, needs to be applied to decide whether or not an event belongs to a cluster, or two events belong to the same cluster. Such a criterion preferably com prises computing an inter-event distance to compare events, said inter-event distance being compared to one or more delimiting va lues, which are parameters of the clustering a lgorithm .

In the present invention it should be noted that in the reclustering or recursive step, whereby a parent cluster is being reclustered or sub-clustered into child clusters or sub-clusters, the clustering criterion is updated su bstantia lly only on the basis of the events which have been identified previously as being part of the pa rent cluster, i.e. events which could have been part of the initial event set, but were not identified as being part of the parent cluster, are not taken into account when updating the clustering criterion for reclustering or subclustering the parent cluster. This allows a reclustering of the parent cluster which is specifica lly ada pted for that parent cluster, which allows zooming in on the parent cluster and use a clustering criterion which is especially selective for events in that parent cluster. This concept makes the clustering method more robust, and more applicable for all types of measured signals coming from all types of systems, and in particular for systems comprising appliances which are not frequently or regularly used, or for systems in which new appliances are added or old ones are removed. Furthermore, it has been observed that methods of the present invention provide an improved clustering for systems comprising small and/or similar loads.

Therefore, in an embodiment, events are compared by computing an inter-event distance. This distance could preferably, and in particular for comparing events characterized by steady-state features, be a Euclidean distance or a Mahalanobis distance, or any combination thereof. This distance could preferably, and in particular for comparing events characterized by transient-state features, be a Minkowski distance, a Dynamic Time Warping (DTW) distance, a Longest Common Subsequence (LCSS) distance, or any combination thereof. Also preferred are distances which are at least partly in a one-to-one correspondence with the above-mentioned distances or any combination thereof.

In a preferred embodiment of the methods for clustering events as disclosed herein, events are characterized by steady-state features or transient-state events. In the method for clustering wherein the clustering criterion is computed on the basis of the principal components of events, preferably events which are characterized by steady- state features are clustered as these events have been noticed to comprise principal components which can be identified unambiguously in a principal component analysis, e.g. by means of looking at those linear functions of the event features which maximize the variance, e.g. as further explained in I. T. Jolliffe, "Principal Component Analysis", Vol. 30, Springer Series in Statistics, Springer, 2^nd ed., 2002.

In a preferred embodiment, said clustering and/or reclustering and/or subclustering is performed using a density-based algorithm, such as a DBSCAN algorithm, an OPTICS algorithm and/or a DBCLASD algorithm, said algorithm based on a density defined by the amount of events which can be found in an e-neighborhood of a point of the cluster, said e-neighborhood of a point p defined as

N_£(p) = {q e X\d(p, q)≤ e}, where X is the signature space, preferably a feature space, d() represents a distance function between points and/or events, and e is an inter-event distance delimiting the neighborhood, whereby two kinds of points are defined : core points which have a density \N_e(p) \≥ MinPts, MinPts being a number higher than 1, and preferably lower than 20, and border points which belong to an e-neighborhood of a core point without being a core point itself, whereby two points are defined as density-reachable if there exists a sequence of points such that each one belongs to the an e-neighborhood of its predecessor, the latter being a core point, whereby two points are defined as density- connected if they are density-reachable from a common point, whereby said clustering criterion comprises evaluating if an event is density-connected to a point and/or an event in a cluster.

The applicant would like to note that descriptions of neighborhoods which are in correspondence, in particular in a bijectional correspondence, with the description given here above can also be used.

In a more preferred embodiment, said clustering criterion is computed and/or recomputed by computing and/or recomputing e and/or MinPts, preferably e.

Preferably, said events are characterized by steady-state features and/or transient- state features.

The inventors have found that prior art techniques seem to rely on external input, e.g. maximal inter-event distance or minimal number of events, in order to determine which event belongs to which cluster or which cluster can be deemed complete. Although such prior art techniques allow identification of relatively simple components at a reasonable level, these techniques seem to fail when applied to more intricate systems, such as households or companies, which comprise electrical appliances with rather complicated electrical components.

This is at least partly due to a degree of arbitrariness which is introduced in the method, in particular in a step related to clustering of events, prior to identification of components on the basis of the computed clusters. Although the results for the component identification can be improved on a case-by-case basis, i.e. largely depending on the type of system which is being monitored, it is clear that the prior art does not disclose a method which can be applied to any type of system, or the systems which change over time, e.g. a household or a factory. The inventors have also found that the current invention results in a clustering of events which is clearly better than prior art techniques, and thus leads to better component identification and appliance load monitoring. The present invention further allows to quantify or objectivate any error which is made during the clustering. The information on the error allows to improve the techniques further or to identify irregularities in the system. In a further aspect, the present invention also provides a component detection method comprising the steps of: obtaining clusters of events obtained from a (preferably measured) electrical signal comprising power consumption information of a system, said events comprising on-events and off-events, whereby each of said events comprises a cluster ID which allows to identify the cluster to which the event belongs, preferably whereby said clusters of events are obtained using a method according to the present invention;

pairing on-events and off-events from said measured electrical signal, thereby obtaining a set of paired events, each paired event comprising paired cluster ID information representing the cluster ID of the on-event and the cluster ID of the off-event in said paired event;

detecting (preferably electrical) components by identifying recurrent paired cluster ID information within said set of paired events. In a preferred embodiment, said step of pairing on-events and off-events comprises extremizing a cost function which depends on variations in energy, in power, in current, in energy fit errors, in power fit errors, in current fit errors or in any combination thereof, preferably in a linear combination of at least two of said variations. More preferably, this cost function depends on variations in power and/or in power fit errors.

The present invention further concerns the use of any of the clustering methods and/or the component detection method in a non-intrusive appliance load monitoring method and/or for condition monitoring of an appliance.

The component detection method of the present invention is particularly suited for a system wherein electrical appliances are on/off components, or consist of such on/off components, and, their consumption can be modeled by constant power segments during steady states. For these systems, any on-event should be followed by an off- event with similar power step (with opposite sign). "Explained power steps" are the ones for which complementary power steps can be identified following the idea that: if a component has been turned on at time t_on, it should be turned off at a time t_off > t_on. Further, constant power draw is assumed and we search for the pair (t_on, t_off) minimizing fit-errors.

Note that the component detection method of the present invention allows not relying on prior models of the system and/or its appliances and is robust to clustering errors. The present invention also concerns a processing unit arranged for performing a NIALM method, any clustering method and/or a component dtetction method according to the present invention. Preferably said processing unit is arranged for performing a recursive clustering method or a method for clustering according to the present invention, and a component detection method according to the present invention.

The present invention further concerns a device, preferably a computer-mountable and/or meter-mountable device, comprising instructions for executing a NIALM method, any clustering method and/or a component detection method according to the present invention. Herein, 'meter-mountable' device refers to a device which can be linked to an electrical meter, such as an electrical meter which can be found in a residence, a business or any system comprising an electrically powered device, and which measures the consumption, and/or optionally the generation, of electrical energy. Preferably said processing unit is arranged for performing a recursive clustering method or a method for clustering according to the present invention, and a component detection method according to the present invention.

In a preferred method, said device is a separate device, i.e. which can be releasably linked to a computer or a meter or a system comprising electrical appliances. In another preferred method, said device is integrated in a meter, such as a smart meter, or in a circuit breaker of said system.

The present invention further concerns a NIALM system comprising a client device and a server device, whereby said client device and server device are linkable, and optionally are linked, and whereby said client device and server device are configured to together execute a NIALM method according to the present invention, a recursive clustering method according to the present invention, a method according to the present invention, and/or a component detection method according to the present invention. This system allows performing the separate steps of the methods according to the present invention on the client device or on the server device or on both. In particular certain steps of the methods of the present invention can be performed on the client device while the remaining steps of the methods can be performed on the server device, i.e. the client device can be configured to perform a first subset of the steps of the methods of the present invention, while the server device can be configure to perform a second subset of the remaining steps of the methods of the present invention. In a preferred embodiment, said client device is configured to obtain a measured electrical signal comprising power consumption information of a system. In an embodiment, the client device is configured to perform the following steps: detecting events in a (preferably measured) electrical signal comprising power consumption information of said system, said events representing state transitions of the appliances in the system;

- transferring information representing said detected events to said server device; and the server device is configured to perform the following steps: obtaining said information representing said detected events from said client device;

- characterizing said events by an event signature, by taking into account differences between steady states before and after each event and/or by taking into account transient states located between steady states before and after each event;

clustering events into a set of clusters on the basis of their event signatures; - identifying components on a basis of said set of clusters;

optionally identifying appliances from the components identified in the previous step;

optionally providing appliances' energy consumption or providing the components' energy consumption. In an embodiment, the client device is configured to perform the following steps: detecting events in a (preferably measured) electrical signal comprising power consumption information of said system, said events representing state transitions of the appliances in the system;

characterizing said events by an event signature, by taking into account differences between steady states before and after the event and/or taking into account transient states located between steady states before and after the event;

transferring information representing said event signatures to said server device; and the server device is configured to perform the following steps: obtaining said information representing said event signatures from said client device;

clustering events into a set of clusters on the basis of their signatures;

identifying components on the basis of said set of clusters; optionally identifying appliances from the components identified in the previous step;

optionally providing appliances' energy consumption or providing the components' energy consumption. In an embodiment, the client device is configured to perform the following steps: detecting events in a measured electrical signal comprising power consumption information of said system, said events representing state transitions of the appliances in the system;

characterizing said events by an event signature, taking into account differences between steady states before and after the event and/or taking into account transient states located between steady states before and after the event;

clustering events into a set of clusters on the basis of their signatures;

transferring information representing said set of clusters to said server device; and the server device is configured to perform the following steps: obtaining said information representing said set of clusters from said client device;

identifying components on the basis of said set of clusters;

optionally identifying appliances from the components identified in the previous step;

optionally providing appliances' energy consumption or providing the components' energy consumption.

In the NIALM system's embodiments given above, said clustering comprises an initial clustering of said events into an initial set of clusters on a basis of a first clustering criterion, and a subsequent reclustering of at least one of said initial clusters on a basis of a second clustering criterion different from the first clustering criterion.

The use of a NIALM system with a client device and a server device provides a number of advantages, including, but not limited to: the possibility to optimize and update the algorithms in an easy way, in particular for the algorithms of the server device, and in particular the clustering algorithm,

the possibility to collect data centrally in order to assess and improve the efficiency of the used algorithms, the possibility to perform an analysis of the aggregate energy or power consumption of a number of systems.

Therefore, the present invention also concerns a database comprising information representing identified components in a system, preferably of a multitude of systems, said information obtained using a NIALM method according to a method disclosed in this document. Preferably the database also comprises information representing identified appliances and/or information representing energy consumption of the appliances and/or components in said system or said multitude of systems.

Description of figures Figure 1 shows three methods to evaluate similarities between time series: the Minkowski distance, the dynamic time warping (DTW) and the least common subsequence (LCSS).

Figure 2 shows how the amount of outliers mainly decreases with k, the amount of nearest neighbors in the k-dist criterion. Figure 3 shows that when increasing k, the decrease of outlier points is mainly located in the origin of the P-Q plane but small clusters are then considered as outliers.

Figure 4 illustrates substructures which are better emphasized if principal components are evaluated for specific data subsets. Figure 5 shows a modeled aggregate power based on paired events. The evaluation is independent from the classification results.

Figure 6 shows that, once pairs have been assigned to components, the considered power steps derive from the clusters.

Figure 7 shows sequences of on-off cycles which are generated for nine appliances such that their aggregate consumption is not cyclic over the three-hour period.

Figure 8 shows nine appliances which draw specific current waveforms.

Figure 9 shows that the microwave is not a two state-device. Moreover, it exhibits cycling operations depending on the average power asked by the user.

Figure 10 shows events and their features - events corresponding to: the kettle, the vacuum cleaner and the microwave oven.

Figure 11 shows events and their features, zooming in at the origin of the P-Q plane. Figure 12 shows that some turn-on transients systematically lead to several detections.

Figure 13 shows power variations of the ventilator which are significant compared to the power it consumes. Outliers risk to be associated with clusters of such a variable and low power device.

Figure 14 shows that clusters which are obtained after a first clustering iteration do not correctly reveal the complete data structure.

Figure 15 shows that an iterative clustering procedure according to the present invention efficiently captures the data substructure. Figure 16 shows a classification according to the highest membership value based on the Mahalanobis distance.

Figure 17 shows that there is almost no confusion between clusters of transient patterns. However, no clusters are discovered for the kettle, the ventilator and the vacuum cleaner. Figure 18 shows that component ID are assigned to recurrent pairs of nonoutliers. As a result, equivalent component ID are assigned to on and off event clusters.

Figure 19 shows first, the modeled aggregate power is evaluated with all paired events; and second, electrical components are discovered from recurrent pairs of nonoutliers. Finally, state sequences are evaluated based on classified event pairs. Figure 20 shows paired events are assigned to components according to their cluster identifiers.

Figure 21 shows state sequences evaluated from the event pairs assigned to pure component. Results from the state sequence evaluation are given in the top plots and the measurement in the bottom ones. The drawings of the figures are neither drawn to scale nor proportioned. Generally, identical components are denoted by the same reference numerals in the figures.

Detailed description of the invention

NIALM techniques are either pattern recognition-based or optimization-based approaches, also referred to as event-based or non event-based. In the former case, samples correspond to state changes of appliances, referred to as 'events'. These methods are also referred to as event-based NIALM. Signatures are used to associate events to appliances. State sequences of appliances then result from classification algorithms. In supervised approaches, changes of appliance states are matched one- by-one to known signatures. Where learning is unsupervised, the recurrence of signatures is exploited to recognize appliances. The present invention concerns an unsupervised, event-based NIALM method according to claim 1. The present invention also concerns methods which optimize the NIALM method, including a recursive clustering method for recursively clustering similar events according to claim 5, a method for clustering similar events taking into account principal components of the event's signature according to claim 9, a component detection method according to claim 13, as well as the use of any of these methods or their combination in a NIALM method or for condition monitoring of an appliance. The present invention also concerns a processing unit arranged for executing any of these methods and a device comprising instructions for carrying out these methods, as well as a NIALM system comprising a client device and a server device as specified in claims 19 and 20 and a database according to claim 21.

Unless otherwise defined, all terms used in disclosing the invention, including technical and scientific terms, have the meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. By means of further guidance, term definitions are included to better appreciate the teaching of the present invention. As used herein, the following terms have the following meanings:

"A", "an", and "the" as used herein refers to both singular and plural referents unless the context clearly dictates otherwise. By way of example, "a compartment" refers to one or more than one compartment.

"About" as used herein referring to a measurable value such as a parameter, an amount, a temporal duration, and the like, is meant to encompass variations of +/- 20% or less, preferably +/-10% or less, more preferably +/-5% or less, even more preferably +/-1% or less, and still more preferably +/-0.1% or less of and from the specified value, in so far such variations are appropriate to perform in the disclosed invention. However, it is to be understood that the value to which the modifier "about" refers is itself also specifically disclosed.

"Comprise", "comprising", and "comprises" and "comprised of" as used herein are synonymous with "include", "including", "includes" or "contain", "containing", "contains" and are inclusive or open-ended terms that specifies the presence of what follows e.g. component and do not exclude or preclude the presence of additional, non-recited components, features, element, members, steps, known in the art or disclosed therein.

The recitation of numerical ranges by endpoints includes all numbers and fractions subsumed within that range, as well as the recited endpoints. The expression "% by weight", "weight percent", "%wt" or "wt%", here and throughout the description unless otherwise defined, refers to the relative weight of the respective component based on the overall weight of the formulation.

As used herein, the DBSCAN algorithm refers to density-based spatial clustering of applications with noise. In DBSCAN, a distance parameter, i.e. an inter-event distance parameter, defines the neighborhood of a point, i.e. an event. A zone is considered as dense if a sufficiently large number of points, i.e. events, is found within the given neighborhood. DBSCAN hereby requires only two parameters and no range of values must be specified for the number of clusters to investigate. Instead, the number of clusters is an output of the algorithm. We refer to M. Ester, H.-P. Kriegel, J. Sander and X. Xu, "A Density-Based Algorithm for Discovering Clusters in Large Spatial Databases with Noise", Knowledge Discovery and Data Mining 96: 226-231, 1996 for more detailed information about this algorithm.

As used herein, the OPTICS algorithm refers to an algorithm for ordering points to identify the clustering structure. We refer to M. Ankerst, M. M. Breunig, h-P Kriegel and J. Sander, "OPTICS: ordering points to identify the clustering structure", ACM SIGMOD Record 28 :49-60, 1999, for more information.

As used herein, the DBCLASD algorithm refers to a distribution-based clustering algorithm for mining in large spatial databases. We refer to X. Xu, M. Ester, H. P. Kriegel and J. Sander, "A distribution-based clustering algorithm for mining in large spatial databases", International Conference on Data Engineering, Vol. 1, pp. 324-331, 1998, for further information.

When clustering events, a decision criterion needs to be applied on whether or not an event belongs to a cluster, or two events belong to the same cluster. Such a criterion preferably comprises computing an inter-event distance to compare events. Therefore, in an embodiment, events are compared by computing an inter-event distance. This distance could preferably, and in particular for comparing events characterized by steady-state features, be a Euclidean distance or a Mahalanobis distance, or any combination thereof. This distance could preferably, and in particular for comparing events characterized by transient-state features, be a Minkowski distance, a Dynamic Time Warping (DTW) distance, a Longest Common Subsequence (LCSS) distance, or any combination thereof. Also preferred are distances which are at least partly in a one-to-one correspondence with the above-mentioned distances or any combination thereof.

As used herein, the term "Euclidean distance", "Mahalanobis distance", "Minkowski distance", "Dynamic Time Warping (DTW) distance" and "Longest Common Subsequence (LCSS) distance" refer to specific types of distances between events, which are discussed in what follows.

Euclidean and Minkowski distance

wherein p=2 for the Euclidean distance, and p= l, 3, 4 or more for a Minkowski distance. Herein, X and Y refer to event signatures characterized by a set of N features X, and y, resp.

Preferably, these features are normalized, e.g. with respect to the minimum and maximum values of the features from the events which are taken into account to define or compute the clustering criterion, according to

wherein x, refers to the i'th feature in the signature, and x_i;n refers to the normalized feature value. Mahalanobis distance:

Let X_c (n_c x p) be a set of n_c p-feature signatures grouped within a cluster. The variance-covariance matrix of cluster X_c is

Let further x be a signature whose distance w.r.t. X_c is to be evaluated and X_c the vector of mean feature values of X_c. The Mahalanobis distance is defined as follows:

This distance is evaluated in the original feature space, i.e. in the features described in Chapter 6. The membership value is simply expressed as the inverse of the point-to- cluster distance. The Mahalanobis distance better deals with the dispersion of clusters along different directions than e.g. the Euclidean. Therefore, the Mahalanobis distance is better suited to identify the cluster to which outlier events belong.

Outliers are isolated points in the feature space. They do not systematically exhibit extreme feature values. Instead, the distance to their nearest neighbors is high compared to the one of non outlier points. Consequently, the distance of a point to its k nearest neighbors is an image of its likelihood to be an outlier. We refer to such a distance as k-dist. Outliers are then points with the highest k-dist values.

As an example for deciding which events are outlier events, a threshold on k-dist values could be defined to decide whether or not points are outliers. In an embodiment, a maximum nonoutlier (MNO) value of the k-dist distribution could be used : it is defined as the highest observed value below the upper quartile increased by 1.5 times the interquartile range. Following this definition, the threshold value on the k-dist is obtained with the following formula :

€ out Hers max [k ^~™~ dist(p) < Q s + l.T5(Q¾— ¾s)

p where Q₇₅ and Q₂5 are the upper and lower quartiles of the k-dist distribution. Herein, k can preferably take any value between 1 and 20, preferably between 4 and 15, more preferably 10.

Dynamic time warping (DTW) distance and least common subsequence (LCSS) distance: The principle of the dynamic time warping (DTW) and the least common subsequence (LCSS) are illustrated in figure 1, which shows three methods to evaluate distances, in particular distances between time series: the Minkowski distance, the dynamic time warping (DTW) and the least common subsequence (LCSS).

DTW allows to deal with sequences with different lengths and different localization in the window. LCSS further allows to skip samples considered as outliers in the pattern. When sequences that we compare are isolated patterns, the globally optimal alignment obtained with DTW is a suitable way to measure the similarity. Considering that different transients have been separated by the segmentation and that turn-on transients follow a continuous process on the time axis (no step in time), the DTW is chosen over the LCSS method. Further, the DTW has the advantage that the entire sequences must be matched which avoids that spurious transient states be matched to subparts of true transients. In other words, DTW performs matching with time shift analysis robust to positive false detections. The average distortion is used instead of the total distortion in order to normalize for different segment lengths. The objective of DTW is to compare two sequences X = (xi, .., x_N) and Y = (yi, .., y_M) of length N and M. A warping path is a sequence p = (pi, .., p_L) with pi = (πι,ιτΐι) satisfying the three conditions:

Boundary condition : Pi = (1, 1) and p_L = (N,M)

Monotonicity condition : n_! < n₂ < ... < n_L and mi < m₂ < ... < m_L

- Step size condition : p,₊₁ - p, e {(1, 0), (0, 0), (0, 1)} for I 6 [ 1,L - 1]

Provided a distance metric d(x,, y,) between two elements x, and y, (e.g. Euclidean distance), the total cost of a warping path p is i

e_F(X, 1' ) = \_^ t f\ .<·,„ . y. if )

1=1 and the optimal warping path p* is the one minimizing the total cost among all possible paths. The distance DTW(X, Y) is defined as the total cost associated to the optimal path p* :

DTW(X, Y) = c_p*(X, Y)

= min {c_p(X, Y) | p is a warping path}

Finally, the average DTW distance is DTW(X,Y)/L where L is the length of the optimal path p*.

The present invention will be now described in more details, referring to examples that are not limitative. Examples

Example 1 : Feature extraction on the basis of principa l components

Electric signatures have been evaluated because they characterize the electric consumption of components, not because they allow to classify components. Features might be correlated a nd less features could then be used to represent the data structure. Also, some features might be better than others to reveal similarities and differences between signatures. Considering non relevant features would add similarities between events.

In this example, we present the principal component analysis (PCA) as a tool to extract relevant features for clustering purpose. Because PCA is based on data scattering analysis, outliers might severely impact the results. They should consequently be removed from the data before that PCA be performed. Their detection is the subject of Section 2. Then, in Section 3, we expose how the features are extracted from the steady state signatures in our NIALM procedure. 1. Principal component analysis

We propose a concise description of the principal component analysis based on the work by I. T. Jolliffe, "Principal Component Ana lysis", Vol. 30, Springer Series in Statistics, Springer, 2^nd ed ., 2002.

"The central idea of principal component analysis (PCA) is to reduce the dimensionality of a data set consisting of a large number of interrelated variables, while retaining as much as possible of the variation present in the data set. This is achieved by transforming to a new set of varia bles, the principa l components (PCs), which are uncorrelated, and which are ordered so that the first few retain most of the variation present in all of the original variables". Suppose that x is a vector of p random varia bles, and that the variances of the p random varia bles. The first step is to look for a linear function α Ί x of the elements of x having maximum variance, where Oi is a vector of p constants an, a₁₂, _lp, and ' denotes transpose, so that

Next, look for a linear function a'₂x, uncorrected with αΊχ and having maximum variance. The procedure is repeated so that at the k'th stage a linear function a x is found that has maximum variance subject to being uncorrelated with α Ίχ, α '₂χ, ..., a'_k-ix. The k'th derived varia ble, a'_kx is the k'th PC. Detailed equations for the principal component analysis can be found in the work by Jolliffe.

The number of features will be smaller if the initial ones are not independent. Jolliffe suggested that a subset of the principal components be chosen such that their cumulative percentage of tota l variation be between 70% and 90%. Let o² _j be the variance of the data along principa l component PC,, the cumulative percentage of total variation corresponding to the PC k is :

We set that value to 85% in our implementation . The principa l components (PC) taken into account are then the set of first PCs covering 85% of the variance. In addition to adding relevant variations, considering more PCs will, on the one hand, add distinctions between points of a same cluster and, on the other hand, add similarities between points of different clusters.

2. Detection of outlier points

Outliers might impact the result of the principal component analysis. They are mainly related to errors from the underlying data generation, e.g . they can be related to errors insignal segmentation leading to outliers. Herein, the errors can be due to noise or non-linearities, or due to transients which can be very long and hence could be seen as two or more different events. They are also rare events resulting from simultaneous on/off switching of components. They should be removed from the data set before searching for the principa l components. Outliers a re isolated points in the feature space. They do not systematica lly exhibit extreme feature values. Instead, the distance to their nearest neighbors is high com pared to the one of non outlier points. Consequently, the distance of a point to its k nearest neighbors is an image of its likelihood to be an outlier. We refer to such a distance as k-dist. Outliers are then points with the highest k-dist values. Signatures consist of features, with very different magnitudes. In order to give similar impact to each feature in the k-dist evaluation, data are normalized . The min-max normalization can be appropriate to detect the outliers and the reduced data are then : ΪΙΧ1ΙΧ 5¾

-Vt fc [l. jj

Each feature x_i;n ranges between 0 and 1 and the k-dist of events is the Euclidean distance between their normalized feature vectors and the one of their kth nearest neighbor. A threshold on k-dist values must be defined to decide whether or not points are outliers. X. M. Lopez et al., "Clustering methods applied in the detection of Ki67 hot- spots in whole tumor slide images: an efficient way to characterize heterogeneous tissue-baked biomarkers", Cytometry, Part A: the journal of the International Society for Analytical Cytology, Vol. 81(9) : 765-775, September 2012, reported on the use of the maximum nonoutlier (MNO) value of the k-dist distribution : it is defined as the highest observed value below the upper quartile increased by 1.5 times the interquartile range. Following this definition, the threshold value on the k-dist is obtained with the following formula : toutliers = max [fc - Mst (p) < Q75 + 1.75(^75 ^~ Q s)]

P where Q₇₅ and Q₂₅ are the upper and lower quartiles of the k-dist distribution. Another approach to detect outliers is proposed by Jolliffe. It consists to look at the extreme values of the last PCs.

A value must be chosen for k, the number of neighbors taken into account in the kdist. Figure 2 illustrates the relative evolution of the number of outliers according to the value of k for five different data subsets. We see that increasing the number of neighbors considered in the k-dist decreases the number of outliers. Variations are mainly located in the origin of the P - Q plane which corresponds to false positive detections and small loads whose signature evaluation is impacted by fluctuations of the mixture. However, increasing k also leads to considering points of small clusters as outliers. These two aspects are illustrated in figure 3.

Detecting loads consuming low power compared to the aggregate consumption is the biggest challenge of NIALM. Consequently, it is desired that the false positive detections be correctly identified as outliers. This will facilitate the detection of points which cluster around the origin of the P - Q plane and are related to signatures of small loads. Taking this reasoning into account and the choice of DBSCAN parameters presented below, k is set to 10. 3. Non outlier PCA

Outliers add non relevant scattering of the data. Obviously, that impacts the evaluation of the principal components which are based on variance-covariance evaluation. The principal component analysis should consequently be evaluated solely with the non outlier points.

Furthermore, because PCA selects features in order to maximize the variance of the data, variables with higher amplitudes will be given higher weight and will be systematically considered in the first PCs. For instance, the active power ranges from some dozens of watts up to some kilowatts whereas normalized harmonic amplitudes range between 0 and 1. Data should then be reduced before the PCA be applied in order to give the same weight to the different features.

Our non outlier PCA works as follows:

Ignore outliers and,

a. Normalize the data

Outliers have been identified. It is then possible to reduce their impact on data normalization. The min-max normalization defined above is applied to the data, but the outliers are ignored from the search for the minimum and maximum values. This avoids that data be compressed because of extreme feature values of outliers.

where the exponent NO stands for non outliers. X _n = (x i,n , x NO p,n b. Evaluate the principal component transformation a = PCA(X ^O _n) where PCA() refers to the calculations of the linear combinations O i to a_p explained in Section 1.

Then, for the entire data set

c. normalize with the transformation of step a,

With X_n = (x_1; nr d. project the data into the new features with the transformation matrix obtained in b.

XPC = ^ft %n

PCA has been applied to the normalized dataset whose normalized active and reactive power are shown in the upper plot of figure 4. The two first PC are plotted in the leftmost below plot. The PCA is also applied to the same data set limited to events with active power first higher and then lower than 500W (respectively in the middle and right-most plots). For each of these three cases, the weights assigned to the features in the first PC are given in Table 1 and the percentage of total variance of the three first PCs in Table 2.

Table 1 : Weights of the features in the PCI for the data in figure 4

Table 2: Cumulative percentage of total variance Different comment can be formulated based on this illustrative example.

In the three cases, three features are given quasi null weight | Q/P| , H_0n, H_2n. This means either that they are heavily correlated with other features ( | Q/P| ) or that there is no variations along these features (H_0n and H_2n).

When the entire data set is considered, the first PC is the combination of most features. But for both data subsets, only some features have relatively high weights: the PCI of the > 500W' subset involves the fundamental components whereas the one of the < 500W subset is mainly limited the harmonic content and the reactive power. This could have been expected.

Whereas for the > 500W data subset the PCI is sufficient to explain the variance of the data, PC2 is needed in the case of the entire data set to satisfy the 85% criterion. For the < 500W subset, the three first PCs must be used.

It follows that adapting the extracted features to the considered data subset will improve the clustering.

Example 2 : Component discovery

The method to discover instances of electrical components from clusters and the event time sequence can be illustrated by the following example.

First, on and off events are paired such that the fit-error is minimized. Second, cluster identifiers of paired events yield links between clusters which reflect electrical components. Third, assign event pairs to components to derive state sequences.

1. Pairing events by minimizing energy errors and error variations Considering two-state devices, if a component is turned off, it has previously been turned on. Following this, off-events should be paired to on-events to represent occurrences of unknown components. Constant power loads being assumed, the aggregate power is straightforwardly obtained for any sample k as the power of all pairs with k_on < k < k_off . If y_k is the power measured at sample k and AP_k is a vector whose nonzero entries correspond to the signed power step values of paired events, the modeled aggregate power at sample k, y_k, can be written as

y₀ stands to account for loads that would have been switched on before the first sample of the observation window.

The ideal set of event pairs is the one for which the modeled aggregate power is the closest to the measured one. At each time, the power of the unknown components considered as active should sum up to the measured aggregate power. However, segmentation errors or simultaneous switch events lead to impossibility to truly fit the measured power simply by pairing events. Their impact on subsequent decisions should be minimized. This can be achieved by penalizing the bestfit solution with the total variation of the fit-error. We formalize this in the next paragraphs.

In the ideal case, all on/off switching have been detected and all events are correctly paired. In that case, differences between the modeled power (y) and the measured one (y) derive from non constant power consumptions. Among different pairing possibilities, the optimal one is the one which minimizes the difference between y and y. But to deal with missed events, a gap should be allowed and its variations should then be minimized. The fit-error and the variations of this error are respectively defined as:

where y_k is the median value of the power samples within steady state k. Note that also energy fit-errors and variations of energy fit-errors can be defined analogously for different time periods. This corresponds to modeling the observed aggregate power as constant segments. The use of median values instead of mean values allows the method to be more robust to segmentation errors: if there is a non detected power step within a steady state (i.e. there are actually two steady states), the longer one determines the constant power segment.

Pairing on-events and off-events can be done by extremizing a cost function which depends on variations in energy, in power, in current, in energy fit errors, in power fit errors, in current fit errors or in any combination thereof, preferably in a linear combination of at least two of said variations, more preferably variations in power and in power fit errors.

In a preferred embodiment, the cost function is given by:

wherein preferably λ= 1 and β= 1.

For the remainder of this example, we define the cost function :

where k varies on a considered window as explained in the next subsection. The optimal set of event pairs is the one minimizing this cost function. Summation limits are not explicited since no optimization window needs to be defined yet at this stage in the procedure.

The use of the energy error instead of the power error (first term of cost function equation) could be investigated. But variations of power fit errors (second term) should be kept since variations in modeled energy are not relevant. If the first term were to be replaced by energy errors, weights between both terms should be carefully studied.

1.1. Searching for locally optimal pairs

An approach to find the optimal solution consists to test all combinations of event pairs. Of course, this is neither possible nor judicious. We propose a windowed approach to find locally optimal solutions. The use of optimization programming techniques is discussed in the next section.

A local window is considered, the length of which is a parameter (we set the initial value to W = 10 minutes). If events candidate to be paired to an on-event are investigated, the window starts right before the studied on-event. Similarly, it ends right after studied off-events. In this local window, each candidate k' is tested as follows in three steps: a. the amplitude of the turn-off power step is assigned to AP_k and AP_k with appropriate signs,

b. the modeled aggregate power is evaluated with the equation given above where y₀ is the power of the first sample in the window,

c. the cost function l^~ _k< is evaluated with summations limited to the local window.

Since off-event signatures are not impacted by turn-on transients, they are more reliable. This is why they are considered instead of on-event signatures in Step a. Step b allows to hide deviations from ideal fit due to previous erroneous pairing since is systematically initialized to the first sample in the local window. In step c, the minimum cost l^~ _k< obtained with the different candidates in the window is compared to the costs of two other cases: • ηο pair with event k' (Γ_ΝΡ) : AP_k is kept equal to zero, considering that event k is a fictitious transition

• 'candidate for event k is out of window' (r₀w) : AP_k is set to the observed power step at event k without assigning any complementary power step.

The next step is conditioned by the case that minimized the cost function :

• r_k< is minimum

Events k and k' are paired. If the candidate k' was previously paired to another event, the previous pair is canceled. The entries of ΔΡ are updated accordingly.

• r_NP is minimum

The event is not paired. It probably results from a false positive detection or its partner has not been detected.

• r₀w is minimum

If the 'candidate out of window' solution minimizes the cost function, the local window size is enlarged and new candidates are investigated in the new time interval.

This process is repeated for all events in a global window. An example of modeled aggregate power is shown in figure 5.

2. From clusters to electrical components

Events have been paired and each event has been assigned to a cluster. We extract pairs which are constituted by non-outlier points, since these points are likely to be true events.

If events associated to different clusters are often paired, these two clusters are considered to contain events from a single component. In such a case, an instance of electrical component is recorded and its electrical properties derive from the electric signatures of the clusters. Pairs of nonoutlier events with identical cluster ID lead to a direct association of a cluster to a component.

To summarize, a component can be defined as a recurrent link between clusters required to appropriately model the aggregate power.

3. Evaluation of component state sequences Instances of electrical components have been learned from pairs of nonoutlier points. Since all events have been assigned to a cluster owing to the point-to-cluster Mahalanobis distances, a first approximation for state sequences of the discovered components consists of pairs of events assigned to the corresponding clusters. An example is given in figure 6.

However, the discovered components could also be used as input for NIALM methods with prior information. Example 3 : Results of laboratory testing

Algorithms were developed considering both laboratory and field data. However, to evaluate performances at the different stages of the NIALM procedure, a reliable ground truth is required. The latter is only available with laboratory data. Algorithms have been developed with the aim of providing parameter settings which are not tuned to specific data. Consequently, a new data set was generated in the laboratory (ELDA) specifically for this performance evaluation.

In Section 1 below, the data set is presented. The results are given in Section 2 and discussed in Section 3.

1. Data set Measurements have been performed in the laboratory developed for this research (ELDA). In addition to measuring the voltage and the aggregate current, each appliance is individually monitored. The appliance level consumption is consequently known.

To obtain the ground truth about times of appliance state changes, the power consumed in each outlet is thresholded. However, this works only for two-state appliances since, with multi component appliances, once a device is in operation additional state changes cannot be detected by simply thresholding the power draw. For example, if the heating resistor of a washing machine is in operation, state changes of the drum cannot be detected by thresholding the power since the latter varies between e.g. 2000W and 2200W. Consequently, mainly two-state devices were considered.

Nine devices were cyclically turned on and off during a 3-hour period : a kitchen mixer, a hair dryer in low power mode, a kettle, a 50W incandescent light bulb, a ventilator, a vacuum cleaner, a 200W halogen light bulb, a 10W economic light and a microwave oven. All these devices but the microwave are two-state appliances. The power they consume during one on-off cycle is shown in figure 7 and their current waveforms are shown in figure 8. Cycles have been generated such that repetitive simultaneous state changes are avoided. However, simultaneous on/off switching might still occur. Cyclic behavior of appliances has no impact on the results since times and durations of use are not taken into account. Furthermore, even if appliances are cyclically operated, the aggregate consumption is not cyclic and various load combinations are obtained.

Some occurrences of the microwave oven are shown in figure 9. This appliance has more than two states: it consists of a light bulb, a motor and a magnetron. Furthermore, it can be noticed that, according to the average power asked by the user, the magnetron is turned on and off during its operations. Consequently events were manually added to the ground truth to account for these state changes.

2. Performance evaluation

In this section, results obtained at the different stages of the consumption analysis procedure are assessed and illustrated. Results will also be analyzed with the aim to identify suggestions for future work. 2.1. Event detection

The power trace was median filtered with L_M = 50 and the cusum detector parameters were set to v_m = 20 and N_d = 50. The rate of correct detections and the amount of missed detections can be evaluated for each individual load. The figures are reported in Table 3. We observe a global rate of 91%.

Detection rate obtained with v_m = 20 and N_d = 50. Most missed detections comes from low power loads. 2.2. Feature evaluation

The detected events are plotted in the ΔΡ-AQ plan in figure 10 and figure 11. Each couple color-symbol represents one appliance, as indicated by the ground truth. In the following paragraphs, we comment on what can be observed from these plots. Turn-on events with slow time decay lead to several detections. In the studied data set, this impacts the segmentation of turn-on transients corresponding to the mixer and the vacuum cleaner, as shown in figure 12. This leads to different steady state feature values for turn-on and turn-off events. This can be seen with the green point cluster in figure 10 for the vacuum cleaner and with the red point cluster in figure 11 for the mixer. The events corresponding to both these loads form substructures. Furthermore, fictitious second events limit the search for the knee in the power trace which prevent from properly capturing the transient dynamics.

Events corresponding to the microwave exhibit a more complex scattering. Based on the power cycles shown in figure 9, two observations can be made. First, events corresponding to its light lead to points in the cluster around the origin of the P-Q plane (see figure 10). Second, turn-on transients result systematically in two detected events, as illustrated in figure 12. The second ones form the cluster right under 1000W in figure 10. They are not labeled in the ground truth and are then considered as false positive detections. It can be observed in figure 11 that most fictitious detections lead to points located between clusters of low power loads. Some also lie in clusters formed by events of these loads (in the P-Q plan). Moreover, the relative impact of variations in the mixture on the feature evaluation is more important for these low power loads. In addition, the ventilator shows relatively important power variations with respect to the power it consumes. This is illustrated in figure 13. This will probably harden the identification task.

Note finally that, when transients states are defined such that they contain two switch events, the assigned ground truth label is the one of the firstly actuated appliance. This explains why red crosses indicating events of the economic light lie with events of the 200W light bulb in figure 11.

2.3. Clustering the signatures

Clusters obtained with a first run of DBSCAN are shown in figure 14 and the ones resulting from the iterative procedure are shown in figure 15. Note that these classification results are only based on steady-state features since no mean to obtain a single data partitioning from SS and TS clusters has been implemented. Clusters of transient patterns are however shown at the end of this section. Three observations should be done:

• Low power loads are well identified thanks to the adaptive feature extraction and definition of the o-neighborhood in the iterative clustering procedure. This is observed by comparing figure 14 and figure 15.

• Second events of microwave turn-on transients are also divided in two clusters.

This is not desired but could be expected since these events clearly form two clusters in the P-Q plot.

• However, it can be observed that the cluster of on-events of the vacuum cleaner is not divided although distinct clusters appear in the P-Q plot. This means that they are close enough thanks to other features.

The link between clusters and monitored appliances is given in Table 4. All figures but the ones in the first row are numbers of events. The table is structured as follows:

• Each column corresponds to one cluster: the first row gives the cluster identifier (ci_D) corresponding to figure 15 and events are distributed in rows according to the ground truth. Outliers are reported in the last column along with nondetected events.

• Rows correspond to appliances as defined by the ground truth. The last row corresponds to events for which no identifier has been assigned in the ground truth. They are mostly false positive detections, but second events of microwave turn-on transients also lie in this category.

Table 4: There is almost no confusion between clusters resulting from the iterative clustering procedure.

Differences in the total number of events in Table 4 (and following) compared to Table 3 are due to simultaneous switch events. Since, in these cases, one event appears in the result instead of two, it is only counted once and is considered to belong to the appliance firstly turned on.

We analyze further the results given in Table 4.

• Single clusters are identified for the hair dryer (ci_D = 3), the kettle (ci_D = 1), the 50W light bulb (Ci_D = 14), the 200W light bulb (d_D = 10) and the economic light (do = 11). The low confusion between events of the economic light and the ventilator observed at this stage is an interesting result. This would probably not be achievable if only P and Q were considered.

• Two clusters are identified for events corresponding to the mixer: one for on- events (Ci_D = 9) and one for off-events (Ci_D = 12). The situation is similar for the vacuum cleaner (ci_D = 4 and Ci_D = 5). It is expected that the time sequence analysis reveal that these clusters belong to a single component.

• The power consumed by the ventilator is low and shows variation of the order of 20%. This is illustrated in figure 13. Consequently, outliers are associated to its cluster (ci_D = 2).

• The situation is more complex for the microwave. Its turn-off events form a single cluster (ci_D = 8). On the contrary, its turn-on transients are systematically segmented in two events: the first ones form one cluster (ci_D = 6) while the second ones are grouped in two clusters (ci_D = 7 and Ci_D = 15). · Except for clusters 2 and 13, all events appearing in entries others than the main one result from simultaneous switch events. In these cases, only the appliance firstly turned on was reported in the ground truth.

These results show that clusters are appropriate to learn electrical component signatures since there is little confusion between components: columns contain manly one single entry.

Membership values are evaluated with the Mahalanobis distance in the initial feature space. The classification corresponding to the minimum distance is shown in figure 16 and the corresponding figures are given in Table 5. In this table, figures in the rightmost column correspond to non detected events. The most important observation is that outliers are mostly assigned to one cluster, the one constituted by events of the ventilator (d_D = 2). This suggest that a method to limit the impact of outliers be investigated in future research. For instance, it could be useful that a 'bin' cluster be defined or that a threshold on membership values be used not to associate spurious events to clusters. § 12 3 1 1 2 4 -5 10 6 8 11 13 15 '

Mixer 46 35 - - 2 17 - - - 1 - - 1 3 8

Hair dryer - 92 - - 5 2 - 2 - - - 3

Kettle 96 - - 5 2

SOW light bulb - - - 83 21 - 1 2 - 4

Ventilator - - 1 - 88 - - 2 2 2 - 1 15

Vacuum cleaner - - - - - - 49 51 1 - 1

20OW light bulb - - 14 1 - 127 2 1 2 4 microwave oven - - - 13 10 - 1 77 70 - 1 8 eco light - - 2 - 5.1 - 1 3 2 1 118 7 - 64

/ - - 257 2 1 1 1 3 21. 44

Table 5 : Outliers are associated to clusters with the Mahalanobis distance evaluated in the initial feature space. Spurious events are mostly assigned to the cluster

corresponding to the ventilator. Note that microwave oven events assigned to clusters 2 and 14 correspond to the light bulb belonging to this appliance.

Regarding transient states, patterns corresponding to the identified clusters are given in figure 17. No cluster are found for the kettle, the vacuum cleaner and the ventilator. This could be expected for the kettle and the ventilator since their turn-on transients are only step-ups. This is however surprising for the vacuum cleaner; this means its transient patterns are too scattered to be clustered. There are two clusters for the microwave oven corresponding to the two parts of its turn-on transients. Transients of the mixer are divided into two distinct clusters.

Again, the discovered clusters are appropriate to learn component signatures since there is almost no confusion between components.

2.4. Discovering the components

Events are paired following the local window approach of the best-fit technique exposed earlier in this document. The modeled aggregate power resulting from paired events is shown in figure 19. Components are defined as recurrent associations of cluster IDs within nonoutiier event pairs. The resulting component identifiers (compi_D) are shown in figure 18 and given in Table 6. 9 4 5 1 12 £ 6 10 7 8 '1 11 /

Mix. - _ - - - _ - 1 - _ _ 94

H, J f ΘΓ - 44 10 _ - _ _ _ _ _ 50

Kettle _ - - 72 - - - _ _ - - _ _*i 1

SOW light - - _ - IS - _ _ _ 2 _ - 91

Ventilator - - - 1 - 14 - _ 2 1 9 - 82

Vac, clean. - - _ - _ - _ _ - _ -

200W light - - 10 - _ - - 54 _ - _ - 87 in. oven - - - - - - - - 18 34 _ - 125 eco light. - - _ 1 - - 1 - - 1 1 20 214

/ - _ _ _ _ 2 _ - 1:9 36 5 _

Ta ble 6 : All a ppliances have a pure or quasi pure group of event pairs that can be used to learn component signatures.

The following observations can be formulated : · All appliances have a pure or quasi pure group of event pairs that can be used to learn component signatures.

• There is one component (comp_ID = 5) composed of events from the hair dryer and the 200Wlight bulb, whereas these two components are not confused regarding the signature classification (Ta ble 5) . This suggests that the membership values be used in some ways in the pairing procedure. On the contrary, the confusion between the ventilator and the economic light (comp_ID = 3) was already present in the classification.

• Since second events of microwave turn-on transients were divided in two clusters, pairs remain divided in two component identifier. Future works could address the analysis of their time sequences to detect that they belong to one single electrical component. Actually, they should not exhibit simulta neous on states and should have similar duration of use distribution .

• First events of microwave turn-on transients are successfully rejected . In future works, additiona l analysis could detect that such a rejected cluster derive from the segmentation of single transients. For instance, this could be based on the following considerations : events of D = 6 (figure 15) are always followed by events of D = 7/15 and the sum of their linea r features lies in cluster Ci_D = 8 which exhibits subsequent off-events.

2.5. Energy disaggregation

Although the proposed tool focuses on discovering the set of components, it is interesting to look at how component state sequences are evaluated and how aggregated energy can be segmented into the one of individual components. Since a membership value has been assigned to all events. The method simply consists to assign component identifier to event pairs according to their cluster identifiers, the resulting classification is given in figure 20 and Table 8. The state sequence of each pure component is evaluated and shown in figure 21. The state sequence of the ventilator is based on comp_ID = 2 although it is not a pure component. Furthermore, both compi_D = 7 and compi_D = 8 are manually associated to the microwave oven. Regarding the economic light, it can be seen in figure 21 that, although paired events correspond to events of this light, pairs do not properly link successive on and off events. This also occurs in a lesser way for the ventilator.

The energy corresponding to the state sequences is contrasted with the one obtained from the submetering in Table 7. The low percentages of energy recovery suggest that more robust techniques be implemented to exploit the discovered component signatures and better evaluate the state sequences.

Table 7: Energy assigned to individual components (in Wh). The link between the components and the devices is manually defined. T (True) stands for energy consumed by the appliance and properly assigned to a component state sequence. F (False) stands for energy assigned to a state sequence whereas the corresponding appliance exhibits no consumption. 9 4 5 1 12 2 6 10 7 8 3 11 /

Mix... 36 1 2 1 - 73

II. Dryer - 60 20 - - 3 - - - - - - 21

Kettle - - - 87 _ - _ - _ - _ _ 16

SOW light - _ _ - 26 4 _ - _ 2 1 - 78

Ventilator - - _ 1 - 42 - 2 2 1 14 - 47

Vac. clean. - - - _ _ 82 - _ - _ 20

200V light - - 19 - - _ 1 77 - 1 - - 53 in. oven - - 1 - 2 3 _ - 19 42 _ - 113 eco light - - _ 2 _ 16 1 1 - 1 50 37 143

/ - - _ - - 36 _ - 20 43 22 1

Table 8: Paired events are assigned to components according to their cluster identifiers.

3. Discussion The results presented in this example showed that it is possible to discover the electrical components responsible for the aggregate consumption, and this, without prior information. Component state sequences have been evaluated with the present algorithm, mainly for illustrative purpose.

Claims

Claims

1. Non-intrusive appliance load monitoring (NIALM) method for monitoring components of appliances in a system and comprising the steps of: detecting events in an electrical signal comprising power consumption information of said system, said events representing state transitions of the appliances in the system;

characterizing each event by an event signature, by taking into account differences between steady states before and after each event and/or by taking into account transient states located between steady states before and after each event;

clustering events into a set of clusters on the basis of their event signatures; identifying components on the basis of said set of clusters, characterized in that said clustering step comprises an initial clustering of said events into an initial set of clusters on a basis of a first clustering criterion, and a subsequent reclustering of at least one of said initial clusters on the basis of a second clustering criterion different from the first clustering criterion.

2. NIALM method according to claim 1, comprising the step of:

identifying appliances from the identified components.

3. NIALM method according to any of the claims 1 or 2, comprising the step of:

- providing energy consumption of identified appliances and/or of said identified components.

4. NIALM method according to any of the previous claims, characterized in that said first clustering criterion is computed by taking into account event signatures from substantially all events detected in said electrical signal, and in that said second clustering criterion is computed by taking into account event signatures of substantially only events in said one initial cluster which is to be reclustered.

5. Recursive clustering method for clustering events obtained from an electrical signal comprising power consumption information of a system, each event being at least partially characterized by an event signature, said recursive clustering method comprising : an initial step of clustering events of an initial event set into a set of clusters using a clustering criterion for deciding whether two events belong to the same cluster and/or whether an event belongs to an existing cluster, said clustering criterion being computed on a basis of event signatures from substantially all events from said initial event set; and

a recursive step of reclustering events belonging to a cluster into a set of sub- clusters using an updated clustering criterion for deciding whether two events belong to the same cluster and/or whether an event belongs to an existing cluster, said updated clustering criterion being computed on a basis of event signatures substantially only of events from said cluster, thereby obtaining one, two or more sub-clusters, preferably whereby if two or more sub-clusters are obtained, these two or more sub-clusters are reclustered using the present recursive step.

6. Recursive clustering method according to claim 5, characterized in that clustering and/or reclustering is performed using a density-based algorithm, such as a DBSCAN algorithm, an OPTICS algorithm and/or a DBCLASD algorithm, said density- based algorithm being based on a density defined by the amount of events which can be found in an e-neighborhood of a point of the cluster, said e-neighborhood of a point p being defined as N_e(p) = {q e X\d(p, q)≤ e], where X is the signature or feature space, d() represents a distance function between points and/or events, and e is an inter- event distance delimiting the neighborhood, whereby two kinds of points are defined : core points which have a density \N_e(p) \≥MinPts, MinPts being a number higher than 1, and preferably lower than 20, and border points which belong to an e-neighborhood of a core point without being a core point itself, whereby two points are defined as density-reachable if there exists a sequence of points such that each one belongs to the an e-neighborhood of its predecessor, the latter being a core point, whereby two points are defined as density-connected if they are density-reachable from a common point, whereby said clustering criterion comprises evaluating if an event is density- connected to a point and/or an event in a cluster.

7. Recursive clustering method according to claim 6, characterized in that said clustering criterion is computed and/or recomputed by computing and/or recomputing e and/or MinPts, preferably e.

8. Recursive clustering method according to any of claims 5 to 7, characterized in that said events are characterized by steady-state features and/or transient-state features.

9. Method for clustering events detected in an electrical signal comprising power consumption information of a system, each of said events being at least partially characterized by an event signature comprising a set of event features, the method comprising the steps of: (a) defining principal components of said event features for said events;

(b) com puting a clustering criterion on the basis of the principa l components of substantia lly a ll of sa id events;

(c) clustering sa id events into a set of clusters; characterized in that it comprises the steps of:

(d) sub-clustering at least one cluster of said set of clusters by:

(i) redefining principa l components of the event features for events belonging to said one cluster, thereby substantially excluding event features for events not belonging to sa id one cluster;

(ii) recomputing the clustering criterion for said one cluster on the basis of the redefined principa l components of substa ntially all of the events belonging to sa id one cluster;

(iii) clustering sa id events belonging to sa id one cluster in one, two or more sub-clusters;

(e) optiona lly, if two or more sub-clusters are obtained after step (d)(iii), performing step (d) by taking at least one of said sub-clusters from step (d)(iii) as a cluster to be sub-clustered.

10. Method according to claim 9 characterized in that step (d) is performed recursively for each cluster of said set of clusters and/or for all sub-clusters obtained by performing step (d) until no further sub-clustering in two or more sub-clusters is obta ined.

11. Method according to claims 9 or 10, characterized in that said events are characterized by steady-state features.

12. Method according to claims 9 to 11, whereby step (c) and/or step (d)(iii) is performed using a density-based algorithm, such as a DBSCAN algorithm, an OPTICS algorithm and/or a DBCLASD algorithm, said density-based a lgorithm being based on a density defined by the amount of events which can be found in an e-neighborhood of a point of the cluster, sa id e-neighborhood of a point p being defined as N_e(p) = {q e x\d(p, q)≤ e], where X is the signature or feature space, d() represents a distance function between points and/or events, and e is an inter-event distance delimiting the neighborhood, whereby two kinds of points a re defined : core points which have a density \N_e(p) \≥ MinPts, MinPts being a number higher than 1, and preferably lower than 20, and border points which belong to an e-neighborhood of a core point without being a core point itself, whereby two points are defined as density-reachable if there exists a sequence of points such that each one belongs to the an e-neighborhood of its predecessor, the latter being a core point, whereby two points are defined as density- connected if they are density-reachable from a common point, whereby said clustering criterion comprises evaluating if an event is density-connected to a point and/or an event in a cluster, preferably whereby said clustering criterion is computed and/or recomputed by computing and/or recomputing e and/or MinPts, preferably e.

13. Component detection method comprising the steps of: obtaining clusters of events obtained from an electrical signal comprising power consumption information of a system, said events comprising on-events and off-events, whereby each of said events comprises a cluster ID which allows to identify the cluster to which the event belongs, preferably whereby said clusters of events are obtained using a method according to any of the claims 3 to 10;

pairing on-events and off-events from said electrical signal, thereby obtaining a set of paired events, each paired event comprising paired cluster ID information representing the cluster ID of the on-event and the cluster ID of the off-event in said paired event;

detecting components by identifying recurrent paired cluster ID information within said set of paired events.

14. Component detection method according to claim 13, characterized in that said step of pairing on-events and off-events comprises extremizing a cost function which depends on variations in energy, in power, in current, in energy fit errors, in power fit errors, in current fit errors or in any combination thereof, preferably in a linear combination of at least two of said variations, more preferably variations in power and in power fit errors.

15. Use of a method according to any of claims 5 to 14 in a non-intrusive appliance load monitoring method.

16. Use of a method according to any of claims 1 to 14 for condition monitoring of an appliance.

17. Processing unit arranged for performing a method according to any of the claims 1 to 14, preferably said processing unit being arranged for performing a recursive clustering method according to any of claims 5 to 8 or a method according to any of claims 9 to 12, or a component detection method according to claim 13 or 14.

18. Device, preferably a computer-mountable or meter-mountable device, comprising instructions for executing a method according to any of claims 1 to 14, preferably said device being arranged for performing a recursive clustering method according to any of claims 5 to 8 or a method according to any of claims 9 to 12, or a component detection method according to claim 13 or 14.

19. NIALM system comprising a client device and a server device, whereby said 5 client device and server device are linkable, and optionally are linked, and whereby said client device and server device are configured to together execute a NIALM method according to any of the claims 1 to 4, a recursive clustering method according to any of the claims 5 to 8, a method according to any of the claims 9 to 12, and/or a component detection method according to claim 13 or 14.

10 20. NIALM system according to claim 19, whereby said client device is configured to obtain a measured electrical signal comprising power consumption information of a system.

21. Database comprising information representing identified components in a system, preferably of a multitude of systems, said information being obtained by using 15 a NIALM method according to any of the claims 1 to 4, preferably whereby the database comprises information representing identified appliances and/or information representing energy consumption of the appliances and/or of the components in said system or said multitude of systems.