WO2014090969A2

WO2014090969A2 - A system for robust load disaggregation

Info

Publication number: WO2014090969A2
Application number: PCT/EP2013/076458
Authority: WO
Inventors: Alex SINTONI; Antonio Ruzzelli; Anthony Schoofs
Original assignee: University College Dublin - National University Of Ireland, Dublin
Priority date: 2012-12-14
Filing date: 2013-12-12
Publication date: 2014-06-19
Also published as: WO2014090969A3

Abstract

An electrical power data processing system has an interface to an electricity power sensor (11) and a processor (12) adapted to perform pre-processing of said sensor data to provide data (13) for load disaggregation. The processor (12) detects data failure and disables load disaggregation on data loss, incomplete data sets, corrupt data, inconsistent data, and unsynchronised data. Also, it automatically detects data loss for a parameter by detecting lack of time synchronisation between parameters of a given electrical reading in response to an event. The event may be a power event having a concurrent variation of a number of electrical parameters. The processor is determines if a current sensor data periodicity pattern falls outside periodicity rules.

Description

^;tA system for robust load disaggregation"

Introduction The invention relates to load disaggregation systems for monitoring of energy consumption.

Conventional methods for acquiring equipment-level insights on electricity spending require installation of recording meters on individual pieces of equipment. Software load disaggregation techniques have been developed to tackle the cost constraints associated to such sub-metering approaches. It is a procedure by which to monitor with a single meter an electrical circuit that contains a number of appliances, which switch on and off independently [Hart 1992]. It has been demonstrated that analysing the load's current and voltage waveforms captured at the meter allows recognition of individual loads [LEEB 95, NIALMS 97]. The power signal pattern recognition consists of processing power characteristics of a given load in order to recognise equipment unique power footprints and disaggregate the load accordingly.

Hart [Hart 1992] laid the basics of power signal pattern recognition with the Non-intrusive Appliance Load Monitoring (NALM) work. Power measurements collected from recording meters are partitioned into segments (steady and changing periods), to characterise the power signal in successive steps and events. Features extracted from the segments are eventually matched to known equipment power footprints, enabling the classification of equipment powered on at a given time and individualised energy computation. The prior art 5-step methodology is depicted in Fig. 1. Event-based load disaggregation systems developed after NALM have similar methodology, innovating at either procedural step to improve a specific feature of the system.

When powered on, each piece of equipment will exhibit a unique set of steady and transient power footprints, which will all together form the equipment power profile [Lee 2004]. While most innovation efforts focus on the improvement and development of new pattern recognition algorithms, the success of load disaggregation with such pattern recognition algorithms highly depends on the quality of equipment profiles, the completeness of data collected at the meters, and the complexity of equipment power patterns and produced power variations.

The main threats to robust load disaggregation are: - Undetected data failure between data collection and data processing: the load disaggregation engine processes compromised data and outputs incorrect equipment power activity reports.

- Processing of power variations unfit for the disaggregation engine: the load disaggregation engine attempts feature classification on noise data (power variations that cannot be used by the load disaggregation system) and outputs incorrect equipment power activity reports.

- Misconfigured equipment profiles: the load disaggregation engine matches power segments with compromised power profiles and outputs incorrect equipment power activity reports.

- Processing of power profiles of equipment known to be off or no longer in use: the load disaggregation engine incorrectly finds matches with equipment known to be off.

- Data overload preventing real-time disaggregation: the load disaggregation engine receives as input too many data points and cannot process them fast enough for real-time and continuous analysis.

An objective of the invention is to control and pre-process data collected to improve robustness in load disaggregation, including: the ability to operate at higher accuracy despite abnormalities and unexpected deviations in signal input, system calibration, signal patterns, software calculation, data size, output classification, or of any other process involved in load disaggregation.

REFERENCES [Hart 1992] G.W. Hart, Non-intrusive Appliance Load Monitoring, IEEE, vol. 80, No 12, 1870- 1891 , 1992

[NIALMS 97] Electric Power Research Institute, Nonintrusive Appliance Load Monitoring System (NIALMS) - Beta-Test Results, TR-108419, Final report, September 1997

[LEEB 95] S. B. Leeb, S. R. Shaw, J. L. Jr. Kirtley, Transient Event Detection in Spectral Envelope Estimates for Nonintrusive Load Monitoring, IEEE Transactions on Power Delivery, vol.10, no.3, pp.1200-1210, July 1995 [Norford] L. Norford, S. Leeb, D. Luo, and S. Shaw, Advanced Electrical Load Monitoring: A Wealth of Information at Low Cost, MIT technical report

[Kolter 12] J. Zico Kolter, T. Jaakkola, Approximate Inference in Additive Factorial HMMs with Application to Energy Disaggregation, Proceedings of the 15th International Conference on Artificial Intelligence and Statistics (AISTATS), La Palma, Canary Islands, 2012

[Lee 04] W. Lee, G. Fung, H. Lam, F. Chan, and M. Lucente, Exploration on Load Signatures, International Conference on Electrical Engineering (ICEE'04), 2004

[EPRI] Electric Power Research Institute, 2012 Research Portfolio: End-Use Energy Efficiency and Demand Response - Program 170, 2012

[Sintoni 2011] Sintoni, A., Schoofs, A., Doherty, A. Smeaton, A.F., O'Hare, G.M.P., Ruzzelli, A.G., Generating Power Footprints without Appliance Interaction: an Enabler for Privacy Intrusion, In 1st IEEE Workshop on Holistic Building Intelligence through Sensing Systems, June 29, 201 1

[Schoofs 2010a] A. Schoofs, A. Guerrieri, D.T. Delaney, G.M.P. O'Hare, A.G. Ruzzelli, Annot: Automated Electricity Data Annotation Using Wireless Sensor Networks, In Seventh Annual IEEE Communications Society Conference on Sensor, Mesh, and Ad Hoc Communications and Networks (SECON'10), 2010

[LIANG 2010] J. Liang, S.K.K. Ng, G. Kendall and J.W.M. Cheng, "Load signature study part 1 : basic concept, structure and methodology," IEEE Trans. Power. Delivery, vol.25, no.2, April 2010

[WO 2009] WO 2009/103998 A2 (SENTEC LTD [GB]; ROSEWELL NEIL ALEXANDER [GB]; STEINER HANS JOACHIM [GB]) 27 August 2009

[US 2011] US 2011/095608 Al (JONSSON KARL S [US] et AL) 28 April 201 1

[Schoofs 2010b] A. Schoofs, A. Sintoni, A. Ruzzelli, G.M.P. O'Hare, NETBEM: Business Equipment Energy Monitoring through Network Auditing. 2nd ACM Workshop On Embedded Sensing Systems For Energy-Efficiency In Buildings (BuildSys), held in conjunction with ACM SenSys, November 2010

[Schoofs 201 1] A. Schoofs, A. Ruzzelli, G.M.P. O'Hare, VLAN Auditing for Preliminary Assessment of After Hours Networked Equipment Electricity Wastage, Elsevier Energy, Volume 36, pp. 6910-6921, 2011

[Schoofs 201 lb] A. Schoofs, G.M.P. O'Hare, A.G. Ruzzelli, Debugging Low-Power and Lossy Networks: A Survey, IEEE Communications Surveys & Tutorials, Volume 13, Issue 1, 2011

Summary of the Invention

According to the invention, there is provided an electrical power data robustness system comprising an interface to an electricity power sensor and a processor adapted to perform processing of sensor data to provide robust data for load disaggregation.

In one embodiment, the processor is adapted to detect data failure on data loss, and/or on corrupt sensor data, and/or on inconsistent sensor data, and/or on unsynchronised sensor data. In one embodiment, the processor is adapted to automatically detect data loss for a parameter by detecting lack of time synchronisation between parameters of a given electrical reading in response to an event. In one embodiment, the event is a power event having a concurrent variation of a number of electrical parameters. In one embodiment, the processor is adapted to determine if a current sensor data periodicity pattern falls outside periodicity rules. In one embodiment, the processor is adapted to, in a learning phase, detect data loss by analysing sampling patterns in the sensor data to automatically detect said data sampling periodicity rules.

In one embodiment, the processor is adapted to, during said learning phase, measure for a specific set of parameters a time period between successive data points when the signal exhibits a steady-state mode, and to separately measure data sampling frequency after each power event as well as the number of data points collected at that frequency and the time period that frequency is maintained. In one embodiment, the processor is adapted to perform computation to determine if a parameter has not been updated, thereby causing data loss, and to perform a computation to update a parameter that has not been updated or received to repair data non-synchronisation and loss. In one embodiment, the processor is adapted to perform cross computation of parameter values to automatically detect lack of sensor data synchronisation.

In one embodiment, the processor is adapted to determine if Apparent, Active and Reactive Power values are correctly linked.

In one embodiment, the processor is adapted to determine if Active Power is correctly linked to RMS Voltage and RMS Current. In one embodiment, the processor is adapted to automatically detect if data for a particular parameter has not been received by analysing data identifiers. In one embodiment, the system is adapted to analyse data identifiers by counting number of unique parameter identifiers received for a pre-configured data set.

In one embodiment, the processor is adapted to perform active power value analysis to automatically detect if a current transformer has incorrect clamping, or is damaged, or is undersized , or is incorrectly wired.

In one embodiment, the processor is adapted to monitor the current values on all current transformers to detect long periods when a given current transformer returns zero values which indicate that it does not operate correctly . In one embodiment, the processor is adapted to differentiate a current transformer problem from lack of power activity by analysing current measurements over three phases and, by detecting when only one or two phases exhibit zero current values, to determine the longest period a current transformer sensor has returned zero value currents. In one embodiment, the processor is adapted to perform a voltage value analysis to automatically detect if a sensor providing the sensor data is incorrectly configured.

In one embodiment, the processor is adapted to monitor in real time the result of the following rules:

if no data point for parameter is received after / then at least one data point is deemed lost, and after a power event, if no data point for parameter / is received after / then at least one data point is deemed lost, and

if less than Ni points are received within Ni*ti period after a power event then at least one data point is deemed lost.

In one embodiment, the processor is adapted to perform data filtering to remove chaotic data variations.

In one embodiment, the processor is adapted to perform timestamp and power value analysis to perform automatic separation of steady state, peak, and transient power steps.

In one embodiment, the processor is adapted to store received data points until it has been established whether they should be filtered out or passed to a load disaggregation system, wherein for each new set of data the processor attempts to detect a new power event relating to significant changes in the values of a defined set of electrical parameters in respect to a reference power state, indicating either a change of a steady power state or the next stage of an ongoing power variation. Preferably, the processor is adapted to determine whether previous data sets are currently stored and not yet processed, wherein if data sets are recorded this indicates that the system has not yet established the nature of an ongoing power transition, wherein absence of data sets indicates the beginning of a power transition, and wherein in both cases new data sets are added to the record and are not released until the nature of the transition has been established.

In one embodiment, the processor is adapted to determine that a transition is finished when no new power event has occurred on N successive data sets or after a time period / or a combination of both. In one embodiment, the processor is adapted to analyse first the last data set collected, wherein if the last data set is the only data set recorded then the power transition is classified as transition between two steady states and data is forwarded without change to a load disaggregation engine, wherein if the number of data sets recorded exceeds two, the processor compares the last data set to a reference power state, wherein if the difference between the reference power state and last data set exceeds a threshold the last data set is forwarded to the load disaggregation engine. In one embodiment, the processor is adapted to automatically transform peak power steps into steady state power steps. embodiment, the processor is adapted to perform the steps of:

(a) collecting peak data.

(b) clustering similar peaks together,

(c) for a period t, measure time interval between successive similar peaks to learn the periodicity of power peaks,

(d) after the period t, monitor time interval between successive similar peaks to determine periods of appliance activity, and

(e) transform peak periodic power activity into steady state continuous power activity during periods when appliances are active.

In one embodiment, the processor is adapted to perform the step (e) by transforming the power peaks into a steady and continuous power activity, which reflects the duration of an appliance's power activity and a total power consumption that equals that of the power peak, wherein if the appliance power activity is detected for a period / made of N successive power peaks, the total power consumed Ptotal expressed in kWh by all the power peaks is calculated as the sum of the N individual power peaks' power consumption Ppeak(i):

Ptotal -∑(Ppeak(i)) with wherein the processor determines a power step of duration t matching the timestamps of the peak activity period, and the amplitude Amp of the power step is calculated by the equation below, in which Amp is expressed in kW and / in hours:

Amp = Ptotal / T

In one embodiment, the processor is adapted to automatically transform transient power steps into steady state power steps.

In one embodiment, the processor is adapted to perform said transformation by collecting transient data, and transforming transient power activity into steady state continuous power activity reflecting period of activity or energy consumed.

In one embodiment, if a transient power activity is detected for a period between tl and tl+t, the processor calculates the amplitude of the power variation Amp between tl and tl+t as being indicative of the pre- and post-transient steady power states, and wherein the processor is adapted to transform transient power data into a power step of amplitude Amp and of duration t starting at tl.

In one embodiment, the processor is adapted to calculate power consumption Piramient of transient power activity and power consumption Pstep of a power step of amplitude Amp and of duration t starting at tl, and to next transform the transient power data into a power step of amplitude Amp and of duration Temp = t*Ptransient/Pstep starting at t2-Temp.

In one embodiment, the processor is adapted to perform data profiling including detection of appliance power states according to power values and steady state transitions.

In one embodiment, the processor is adapted to perform data profiling including detecting incorrect appliances according to start and end power values and step-up and step-down power values.

In one embodiment, the processor is adapted to automatically determine from an external system if an appliance if disabled and to accordingly disable power profiles for such an appliance.

In one embodiment, the processor is adapted to interface with a building management system to perform dynamic access to scheduling reports of timers for appliances.

In one embodiment, the processor is adapted to apply rules as to how to link auxiliary equipment time schedules to measured electrical readings in said sensor data. In one embodiment, the processor is adapted to process data concerning equipment network connectivity state to enable or disable pieces of equipment from a list of loads to disaggregate.

In one embodiment, the processor is adapted to output changes observed on auxiliary networks so that a disaggregation engine with which it communicates can use this information to verify its output accuracy.

In one embodiment, the processor is adapted to output to a disaggregation engine data indicating the number of appliances powered on at a given time based on a building management system configuration.

In one embodiment, the processor is adapted to perform data reduction by generating transition points. In one embodiment, the processor is adapted to perform the steps of filtering out data for unnecessary parameters, and reducing number of data points by creating event transition points. In one embodiment, the processor is adapted to provide an interface for configuration of sensors providing data. In one embodiment, the processor is adapted to provide an output signal to disable disaggregation of certain appliances according to quality of data associated with said appliances.

In another aspect, the invention provides an energy monitoring system comprising a robustness system as defined above in any embodiment and a disaggregation engine linked with the robustness system and adapted to perform appliance disaggregation and provide per-appliance energy output data.

Detailed Description of the Invention

The invention will be more clearly understood from the following description of some embodiments thereof, given by way of example only with reference to the accompanying drawings in which :-

Fig. 1 outlines a five-step disaggregation process of the prior art; Fig. 2(a) is a high level block diagram showing the functional components of a robustness system of the invention;

Fig. 2(b) is a high-level block diagram illustrating operation of the robustness systems in conjunction with state-of-the-art disaggregation systems;

Fig. 3 is a plot showing the effect of non-synchronisation of data points in a 3 -phase system;

Fig. 4 is a plot showing artificial power variations created in case of data failure;

Fig. 5 is a plot showing the effect of appliances' concurrent power activity over the steady and unsteady classification of appliances;

Fig. 6 shows peak power variations that may arise in a power signal; Fig. 7 shows transient power variations that may arise in a power signal; Fig. 8 diagrammatically shows data filtering; Fig. 9 is a flow diagram showing data filtering steps;

Fig. 10 is a plot showing power peak recognition;

Figs. 11 is a plot illustrating aspects of appliance signature profiling, in which N power states are not recorded;

Fig. 12 is a flow diagram showing recording of power states;

Fig. 13 is a plot showing incorrect power step recorded as a power profile;

Fig. 14 is a flow diagram showing how the system ensures that the measured power characteristics are those of a particular appliance;

Fig. 15 is a sample display showing use of network audits for enabling/disabling power profiles in a load disaggregation procedure;

Fig. 16 is a diagram illustrating how data acquisition frequency can affect data processing; and Fig. 17 is a diagram illustrating data reduction via virtualisation of data points on power event transitions

Description of the Embodiments The invention provides a disaggregation robustness system enabling fault-prone application of load disaggregation for monitoring changes in voltage and current into an electrical circuit that powers a number of electrical appliances, and for estimating the number and nature of such loads over time. The system includes software which implements advantageous functionality and for the purposes of clarity these are referred to below as Techniques 1 to 5. These are labelled in Fig, 2(a) across the top. A system of any embodiment may have any one or a combination of some or all of these software functions. For example a system of the invention may have software functions for data failure detection to disable load disaggregation ("Technique 1") and not any of the others.

Additionally, the invention also provides a disaggregation system incorporating both disaggregation functionality as shown across the bottom of Fig. 2 and means to implement some or all of the robustness Techniques 1 to 5. For the purposes of clarity the following description describes the invention as being a robustness system which is linked with a separate disaggregation system (which could be supplied by a third party). A system of any embodiment may be deployed between a data collection system and a load disaggregation system, integrated within a load disaggregation system or combine both, as illustrated in Figure 2(b). The embodiment on the left has electrical circuits 10 being monitored by a meter 11 which feeds a system 12 of the invention, which in turn feeds a load disaggregation system 13. As shown on the right side of Fig. 2(b) the meters 1 1 feed a disaggregation system 15 incorporating a robustness sub-system.

The robustness system incorporates input and output mechanisms to allow configuration of the system and communication with external systems. The input and output mechanisms include amongst others general-purpose input/output pins (GPIOs), LEDs, and software interfaces. Fig. 2c illustrate the use of interfaces within this invention. The system of this invention can receive commands for configuring the system to the data collection and load disaggregation specifics. For instance, the load disaggregation system may interface the robustness processor to indicate when it is in training, profiling or monitoring mode, which may trigger different actions by the robustness processor. Interface to auxiliary data collection networks also exists to provide context to the robustness processor for processing data, as discussed in Technique 5 of this invention. The robustness processor utilises its output interfaces to control the operation of the profiling procedure, and to report status updates over the data robustness. The nature and number of input and output interfaces can vary and Fig. 2c does not include all possible options.

In summary the robustness system of the invention achieves some or all of the following in various embodiments.

Detection of data loss: « automatic detection of parameters not being updated via comparison of the parameters' timestamp records, and/or

_* automatic detection of parameters not being updated via a series of checks on data

collection periodicity and number of points collected, and/or

· automatic generation of parameters not updated via computation from other parameters.

Detection of data non-synchronisation:

• automatic detection of parameters not collected in synchrony with other parameters via analysis of parameters cross computation results.

Detection of data corruption: « automatic detection of parameters not received via parameter id analysis, and/or

» automatic detection of parameters incorrectly updated via parameter value analysis, and/or

• automatic detection of current transformer incorrect clamping via Active Power value analysis, and/or

· automatic detection of broken current transformer via Active Power value analysis, and/or

* automatic detection of undersized current transformer via plateau detection, and/or

* automatic detection of broken current transformer via Voltage value analysis, and/or

■» automatic detection of incorrectly wired current transformer and meter via Voltage value analysis, and/or

• automatic detection of incorrectly configured meter via Voltage value analysis.

Data filtering

<· automatic separation of steady state, peak and transient power steps via timestamp and power value analysis, and/or

« automatic detection of peak power steps' periodicity, and/or

• automatic transformation of peak power steps into steady state power steps, and/or

_* automatic transformation of transient power steps into steady state power steps.

Data profiling ^β automatic capture of all equipment power states via check on power levels values and detection of steady-state transitions, and/or

• automatic detection of incorrect equipment profiling via check on start/end power levels values and checks on step up/step down power values. Data classification β Automatic enabling/disabling of power profiles for load disaggregation for equipment known to be off via auxiliary data acquisition systems.

Data reduction

• automatic data reduction pre-disaggregation via transformation of data points into

transition points.

In more detail, the following are the sub-systems 1 to 5, referred to as ' echniques":

Technique 1 : Detection of data failure between data collection and data processing to disable load disaggregation on compromised data sets.

Technique 2: Filtering out of noise, peak and transient variations from original power measurement to produce clean peak, transient, and steady-state data sets for application of custom load disaggregation engines on each data set.

Technique 3: Detection of incorrect capture of power characteristics and naming during equipment profiling.

Technique 4: Prevention of impossible load disaggregation output via interfacing with pre-existing auxiliary data acquisition systems.

Technique 5: Data reduction pre load disaggregation to prevent data overload and secure real-time disaggregation of future power measurements. This may involve data reduction "post" load disaggregation or "pre" load disaggregation.

In energy monitoring systems such as load disaggregation, the components handling data capture and data communication to the machine running the disaggregation engine must ensure consistent data flow without data loss or data alteration. The robustness of the disaggregation engine is indeed greatly affected if data packets are missing or incorrectly time-stamped, and power measurements are incorrect or inaccurate.

Fig. 3 shows the type of readings that can be measured with a meter from a 3 -phase power supply, illustrating (a) A power step produced by a 3 -phase equipment, (b) A power step produced by a single-phase equipment, and (c) A power step produced by two single-phase equipment (e.g. two separate equipment or two components of a single equipment). This illustration includes meter-reading events (see black dots) to depict the issue of missing or unsynchronised data sets.

Matching power events to known equipment power profiles allows power activity reports for individual equipment to be produced. State-of-the-art software load disaggregation engines analyse the variations of successive values of a set of electrical parameters to recognise major power events, e.g. variations that go beyond a set of thresholds as appropriate for the load disaggregation algorithm. One key element to identifying major power events is the provision of synchronised and complete data for the three power phases. This is an issue for 3-phase circuits more common to commercial and industrial settings. The time synchronisation issue does not exist for one-phase circuits more common to domestic environments. Data loss directly affects the accuracy of disaggregation engines because variations in the power signal are analysed with the wrong data points. Say one data point is lost, up/down variations will be calculated with the last received data point which may not be within the same range as the data point not received, see Fig. 4. The worst-case scenario occurs when lost data contains a power event, which would highly affect the disaggregation accuracy. Data loss should therefore be detected and used to disable load disaggregation for the associated period of time.

Similarly, the lack of time synchronisation between various parameters of a given electrical reading may result in inaccurate load disaggregation. A major power event may indeed be defined as a concurrent variation of a number of electrical parameters, which may not happen if variations occur at slightly different times due to non-correlated time-stamping. Fig. 3 depicts such an issue with Phase B PI value having T2 as power step timestamp and Phase A and C PI values having Tl as power step timestamp. A load disaggregation engine may at Tl analyse variations only over two phases and therefore miscompute the real 3-phase power event for two 1 -phase power events. Such a synchronisation issue often relates to bad design in the data collection system, for example where readings of electrical parameters over various phases are not identically time- stamped, but instead time-stamped independently as they are read. Another source of synchronisation issues relates to the introduction of low-power wireless communication for simplified meter reading. Wired metering solutions are already being upgraded to wireless thanks to the integration of low-cost wireless embedded modules, e.g. advanced metering infrastructures with low-power ZigBee and 6LoWPAN/IP-based communication. As a result, wireless meters are increasingly deployed, each meter communicating over low-power multi-hop wireless medium its readings to a network gateway, which in turn communicates all readings to the load disaggregation processor via a dedicated system infrastructure. A low-power wireless medium affects data time synchronisation because network packets are designed to be short in size in order to prevent power-consuming communication for the low-power nodes. Subsequently, it is common that electrical parameters measured on a meter are transmitted in different network packets from the meter to the network gateway, resulting in timestamp differences if time stamping occurs at the network gateway. Low-power communication media are also known for being a great source of data failure, meaning that some data packets may be lost or sent again with a different timestamp, see [Schoofs 201 lb]. Another data failure relates to the provision of incomplete data sets from the source of electrical readings. Load disaggregation algorithms typically make use of a selected range of electrical parameters for identifying unique load patterns. Data for such a range of electrical parameters may however not be contained within all data sets, if the source of electrical systems cannot measure the given parameters, e.g. a recording meter with limited functionality will not return harmonics data, if measurements are not collected from the meter by the data collection software, or if measurements are not communicated to the load disaggregation system. Instead, values set to NULL or '0' may be fed to the load disaggregation engine, which will on such occasion either generate errors or compute incomplete data sets and output incorrect results if the missing data remains undetected.

Failures related to data corruption are similarly important, as load disaggregation is guaranteed to fail if input data is incorrect. We identify the following issues that can happen at various stages of electrical recording: • Clamping of current transformers in incorrect direction

For correct operation it is essential that the current transformers (CTs) are connected the right way round and that the phase relationships are maintained between the meter and the CTs. Some recording meters support user interfaces that can be used to help determine incorrect connection of voltage or cuixent inputs at setup. Many units do not support such a feature and risks of negative current values and wrong power factor measurements may end up undetected.

• Current transformer ratio is not configured correctly at the recording meter or data collection software

Current transformers are marked with the ratio between the maximum primary current and the maximum secondary current. For example a '300:1 A' current transformer produces a 1A output signal when 300 A is flowing through the load. The CT output signal is received by the meter as input signal and is processed to reconstruct the current flowing through the load. For that process the CT ratio between primary and secondary currents must be configured within the meter. When the CT ratio is incorrectly set the produced power measurements are erroneous.

• Current transformers with different primary or secondary currents are wired to the same meter channel

Channel is the name generally given to the grouping of three meter input terminals used to receive signals from CTs clamped to the three circuits of a 3-phase load. Channels can also be used to receive signals from three independent CTs each clamped to the circuit of a single-phase load. Recording meters can have multiple channels to monitor a combination of single-phase and 3-phase appliances. Current transformers are marked with the ratio between the maximum primary current and the maximum secondary current, which must be configured within the meter as described above. Such a CT current ratio configuration is generally only possible at the channel level, meaning that all CTs connected to a given meter channel must have same maximum primary and secondary currents. When CTs with different ratios are wired to a same channel the produced electrical measurements are erroneous. · Current transformer is not properly sized for the current to be measured

On new construction, the size the CT is typically set to handle about 80% of the circuit breaker capacity, e.g. a 800 amp CT is used if the circuit is served by a 1000 amp breaker. For older buildings it is good practice to add 20 to 30% to the past peak demand. Correct CT sizing guarantees accurate current measurements. The accuracy of current measurements however decreases with low line current and when CT is oversized in respect to line current. Similarly, the conductor size and distance is important. Long secondaiy conductor runs with undersized cable can result in poor accuracy. · Current transformer is broken or not closed properly

Split-core CTs will deliver inaccurate current measurements if the clip-on part of the CT is not properly clipped. Similarly, the secondary circuit of a CT must never be opened when current is flowing in the primary circuit. The voltage in the primary winding can immediately reach several thousand volts and cause serious damage to the transformer. Such damage may not be immediately obvious, but will lead to incorrect current measurements.

• Meter and CT are not wired properly for electrical system (star/delta)

The three windings of a generator, a transformer or electrical loads can be connected either in star or in delta. The setup for recording power measurements on star and delta configurations is directly affected. Positioning of CTs will different, for instance three CTs are needed to monitor current variations on 3-phase 4-wire star circuits, whereas two CTs are needed to monitor current variations on 3-phase 3-wire delta circuits. When electrical configuration is not properly documented it is possible to unknowingly deploy CTs incorrectly and collect erroneous current readings.

• Meter is not configured correctly for electrical system (star/delta)

Depending on the star/delta configuration, the relationship between voltages, currents and their phase relationship differs. The setup for recording power measurements on star and delta configurations is different, for instance meters will measure line-to-line voltage in delta and phase voltage in star. Meters generally provide configuration options to indicate whether star or delta electrical configuration is being metered. Misconfiguration will however source incorrect power measurements to the load disaggregation system.

To summarise, data loss, data non-synchronisation and data corruption at the source should be detected prior to providing data points as input to the disaggregation engine in order to let the system know of data failure. In such cases, data failure should be dealt with when possible and disaggregation of unreliable data sets should be temporarily disabled to increase robustness of the produced output. In one embodiment the system includes a module for detection of data loss. Electrical data acquisition is generally event-based, combining periodic reporting during periods of no power activity and dynamic reporting when power events are detected. Event-based reporting can be implemented at the recording meter or at the data collection software. Electrical readings are continuously sampled at set frequency. Thresholds and conditions are set to detect power events on any combination of electrical parameter and reporting is adjusted accordingly. Event-based data reporting allows accurate power step/footprint characterization with fine-grained data sampling when needed for load classification, such as for capturing transient state power variations when an appliance is switched on/off, and providing less information when load disaggregation does not need to perform classification, such as low-frequency sampling for minor variations.

The system of the invention monitors whether all electrical parameters have been updated after a new reading. The system analyses the timestamp of successive data points received from the data collection system, and only releases a set of electrical parameters for load disaggregation when all electrical parameters have been gathered with the same timestamp. For instance, if the timestamp of all electrical parameters have been updated except that of Parameter N the system can identify data loss for Parameter N and the set of parameters is discarded. The system outputs an error status signal to indicate data loss. In such a scenario that the missing value can be calculated from other correctly updated parameter values the system automatically calculates the missing parameter value to compensate the loss of a data point. For instance, Apparent, Active and Reactive Power values are linked to each other by the ApparentP^A2 = ActiveP^A2 + ReactiveP^A2 equation, and Apparent Power, RMS Voltage and RMS Current are linked by ApparentP = RMSV x RMSC.

The system of the invention also keeps track of data points successively collected and assesses what point should come next. Any abnormality is interpreted as data loss under certain conditions. The system of the invention may first analyse the data sampling patterns to automatically detect the data sampling periodicity and rules specific to each data collection setup. For that procedure, the system measures for either all or a specific set of parameters the time period between successive data points when the signal exhibits a steady-state mode, and separately measures the data sampling frequency after each power event as well as the number of data points collected at that frequency and the time period that frequency is maintained. From this learning phase the system creates a set of data sampling rules. Alternatively, the set of data sampling rules can be programmed in the system of this invention. The set of data acquisition rules are checked in real-time for automated detection of data loss.

The number and type of rules may vary from system to system. By default the system aims to validate three rules into data sampling. The data collection system will typically enforce the time period Ti as the data sample acquisition reporting period as set for parameter i when no major power variations occur, Ni as the number of data points reported at higher frequency after detection of a power event for parameter i, and ti as the higher-frequency reporting period set for the Ni data points. For example, a given electrical data collection system may generate new electrical readings every 60 seconds, and every time the active power variation exceeds 10% three additional readings are produced with a one-second interval between two consecutive ones.

Verifying that rules are verified is a prime indicator of data loss. The system of the invention continuously monitors the result of the following three rules:

Rule 1 : If no data point for parameter is received after ti then at least one data point is deemed lost.

Rule 2: After a power event, if no data point for parameter i is received after ti then at least one data point is deemed lost.

Rule 3 : If less than Ni points are received within Ni*ti period after a power event then at least one data point is deemed lost.

The system of the invention uses its output mechanism to indicate a data loss error.

The system of the invention allows the use of all parameters that influence the way data is sampled for the definition of new rules. The complexity of the rules definition allows various degrees of data loss analysis, ranging from simple knowledge that data has been lost to comprehensive analysis of which data points have been lost. The system can for instance output a variety of data loss error codes for each parameter. In one embodiment the system includes a module by which to detect unsynchronised data values. Some electrical parameters produced by data collection systems are not directly measured by a recording meter but are instead calculated using a variety of electrical parameters measured by a recording meter. For example, Apparent, Active and Reactive Power values are linked by a simple equation. Apparent Power is also linked to RMS Voltage and RMS Current by another simple equation. The system of the invention runs all of such equations using last records of electrical parameters to verify if the equations return correct results. Incorrect results will indicate a data point not updated, which relates to non-synchronisation or possibly data loss. Data corruption happens at various levels of data collection and a number of modules are used in the system of this invention. In one embodiment the system includes a module by which to detect incomplete data sets, i.e. some electrical parameters are not measured by the data collection system. The system of the invention analyses the electrical parameter identifiers to count the number of electrical parameters provided by the electrical system. The system counts the number of parameters by analysing the number of unique parameter identifiers received. An incomplete data set error status is issued by the system if the figure differs from the initial system configuration or changes over time. Same error status is issued if values for a given parameter is monitored by the system as being equal to NULL and if values for a given parameter are always constant, for instance always returning 'Ο', and not aligning variations with other dependent parameter variations.

In one embodiment the system includes a module by which to detect that a current transformer has been clamped to a circuit in the opposite direction. The system of the invention continuously monitor the Current and Active Power values received from the data collection system and detects negative current and active power values which indicate inverse clamping of current transformer over circuit. A CT clamping error status is issued by the system of this invention.

In one embodiment the system includes a module by which to detect that the ratio of the CTs used for measuring current in a circuit have been incorrectly configured at the meter. In another embodiment the system includes a module by which to detect that CTs with different maximum primary and/or secondary currents are connected to the same channel of a meter. For both embodiments, the system measures the power consumed over time on all phases of the power signal and keeps periodic consumption records such as monthly records as well as day and night records. The system offers an input interface by which the consumption day. night and total units over a given time period can be provided, on which the system runs a validation check. The system checks whether the measured consumption falls within the range of the provided data, meaning that CTs are correctly configured, or whether the measured consumption is equal to the units provided as input but divided or multiplied by an integer ratio. For instance, measuring a consumption equal to half of the input consumption indicates a CT ratio misconfigiiration by a factor two. An error on the meter configuration will translate in a same division factor identified on the 3 phases of a given channel. An error linked to the use of CTs with different current ratios on a same channel will translate in a division factor identified on 1 or 2 phases of a given channel only. The system runs a validation check on the three phases of a channel to differentiate CT ratio misconfiguration from incorrect use of CTs with different ratios on the channel.

In one embodiment the system includes a module by which to detect that the CTs deployed are sized too small for the current to be measured. The system of the invention tracks saturation levels in the Current measurements, as a form of plateau detection which indicates that current actually goes higher than the measurements but cannot be measured with the CT. Signs of current saturation are discovered when the values returned by the current parameter reach a maximum value over longer than a period T and/or more than N times.

In one embodiment the system includes a module by which to detect that some of the CTs deployed are either broken or not closed correctly. The system of the invention continuously monitor the current values on all CTs and is set to detect long periods when a given CT returns zero values which indicate that the CT does not operate correctly. The system differentiates a CT problem from simple lack of power activity by analysing current measurements over the three phases and by detecting when only one or two phases exhibit zero current values. Over time the system keeps track for each phase of the longest period a CT has returned zero value currents. When that period exceeds a set period the system issues a CT broken error status message.

In one embodiment the system includes a module by which to detect the type of electrical configuration being metered. The system of the invention analyses the power, current, and voltage values of each phase of the electrical readings. The system of the invention use as reference the known voltage of international electrical systems as recalled below. Electrical configuration Line-Neutral Line-Line

1 -Phase, 2-Wire 120 V with neutral 120 -

1 -Phase, 2-Wire 230 V with neutral 230 -

1 -Phase, 2-Wire 208 V (No neutral) - 208

1 -Phase, 2-Wire 240 V (No neutral) - 240

1-Phase, 3-Wire 120/240 V 120 240

3-Phase, 3-Wire 208 V Delta (No neutral) - 208

3-Phase, 3-Wire 230 V Delta (No neutral) - 230

3-Phase, 3-Wire 400 V Delta (No neutral) - 400

3-Phase, 3-Wire 480 V Delta (No neutral) - 480

3-Phase, 3-Wire 600 V Delta (No neutral) - 600

3-Phase, 4-Wire 208Y/120 V Star 120 208

3-Phase, 4- Wire 400Y/230 V Star 230 400

3-Phase, 4-Wire 415Y/240 V Star 230 415

3-Phase, 4-Wire 480Y/277 V Star 277 480

3-Phase, 4-Wire 600Y/347 V Star 347 600

3-Phase 4-Wire Delta 120/208/240 120,

240

Wild Phase 208

3-Phase 4-Wire Delta 240/41 /480 240,

480

Wild Phase 415

The RMS voltage value of voltage is calculated by dividing the peak voltage value by 2. The system of the invention analyses RMS voltage measurements from the data collection system and multiply them by 2 to identify the voltage specifications of the electrical system. The systems runs the voltage against the table of international voltages to shortlist the potential electrical configurations. In a second step, the system of the invention detects whether RMS current and RMS voltage have values different than zero on the tlu-ee phases AND Active Power has values on only two of the three phases. The latter observation indicates metering of a Delta electrical connection and allows the system of the invention to disambiguate star from delta systems in the case of 3 -phase supply. The system finally either issues a meter configuration status message via its output interface or compares the generated shortlist of potential electrical configuration with an internal configuration set up via the system's input mechanism. A mismatch identified by the system means that either a meter or CTs are incorrectly deployed or that there exists a misconfiguration of the meter for the type of electrical installation. TECHNIQUE 2: Filtering out of noise, peak and transient variations from original power measurement to produce clean peak, transient, and steady-state data sets for application of custom load disaggregation engines on each data set. Chaotic data variations are defined in the system of this invention as noise, peak and transient variations that are not recognisable by a software disaggregation engine. They often appear in environments where a large number of equipment are powered through a single distribution board and operate with short-duration activities. Examples of such environments are domestic households, office rooms, areas heated with inverter-driven equipment dynamically controlled via sensors such as temperature and occupancy sensors. Chaotic variations are also observed with machines made of various electrical components all powering on and off over the course of operation (e.g. dishwasher). Machines with long transient states from off to on states (e.g. booster pumps) and machines operated through variable speed drives generally exhibit continuously varying power patterns.

Event-based load disaggregation systems mainly rely on the existence of power state changes between steady states, which are computed through a variety of algorithmic approaches. Chaotic power variations created by peculiar sets of both profiled and non-profiled appliances result in noise for load disaggregation engines, in that recognizable steady state switches of other appliances end up altered and not usable anymore if power events overlap each other. Furthermore, power levels after a power event are examined to detect new steady states, generally assigned when the power level after a power event remains steady for longer than a time period T configured in the load disaggregation system. Chaotic power variations will affect the duration in which a power level is steady after a power event and an appliance steady state may be missed due to parallel chaotic power activity. Fig. 5 depicts this issue, where the chaotic power activity of a secondary appliance B prevents the recognition of appliance A steady state.

Chaotic variations must be detected, filtered out and separately processed to prevent degradation of event-based load disaggregation. Noise signals should be managed prior to disaggregation in order to optimize the engine's robustness.

The invention provides a pre-processing stage to filter-out power variations that prevent robust load disaggregation. Only data suitable for event-based load disaggregation will be provided as input to an event-based load disaggregation engine. Filtering out chaotic power variations from steady-state power variations fiirthermore may enable application of other types of load disaggregation engines customised for transient and peak data sets. Indeed, one type of load disaggregation engine cannot process all power variations and the system of the invention provides suitable data to each type of load disaggregation engines.

The system of this invention defines power peaks as data variations that fall within the pattern of a series of consecutive points which, from a stable steady-state, vary and go back to the same steady-state without generating a new steady-state (no steady pattern occurs during the variation). Examples of such peaks are given in Fig. 6. Power peaks are typical to equipment with short-duration power activity and equipment with internal electrical components switching on for short periods, for example hot water boilers, cup warmers and professional coffee machines trigger a resistance every few minutes to keep water and cups warm at all times.

Transient state data variations differ from power peaks in that they fall within the pattern of a series of consecutive points which, from an initial steady-state vary over multiple power states and end up in a steady-state different from the initial steady state. Examples of such transient state data are shown in Fig. 7.

The following discusses two methodologies for solving the issue of steady state-altering power variations:

Methodology 1 : Peak/transient filtering for steady-state smoothing

Methodology 2: Transformation of power peaks and transient peaks into steady-state power variations Methodology 1 : Steady-state smoothing

Steady-state smoothing is defined as a technique which smoothes the original power signal so that it produces a new signal made only of power variations between steady states and two new data streams reflecting the filtering out of peaks and transient data. The example in Fig. 8 depicts how power variations containing power peaks and transient data are processed to produce three data sets: steady-state data, peak data and transient data.

As per the description of peak and transient data in Fig. 6 and Fig. 7, data filtering follows the procedure given in Fig. 9. It produces three independent and possibly complementary data sets that respectively contain steady-state-only data to be used for typical event-based load disaggregation, transient data to be used for transient state load disaggregation and Peak data to be used for load disaggregation based on peak monitoring.

The system of the invention stores received data points until it has been established whether they should be filtered out or passed to the load disaggregation system. The objective is to detect and separate peak and transient power variations from steady state transitions. For each new set of data the system tries to detect a new power event. A power event relates to significant changes in the values of a defined set of electrical parameters in respect to a reference power state. The reference power state is generally the current power state defined by lastly received data set. The detection of a power event either indicates the change of a steady power state or the next stage of an ongoing power variation. The system makes the distinction by analysing whether previous data sets are currently stored and not yet processed. If data sets are recorded this indicates that the system has not yet established the nature of the ongoing power transition. No data sets record indicates that this is the beginning of a power transition. In both cases new data sets are added to the record and are not released until the nature of the transition has been established.

The system knows that a transition is finished when no new power event has occurred on N successive data sets or after a time period T or a combination of both. The system analyses all power events data sets collected to filter out transient and peak data. The system runs a procedure by which the last data set collected is first analysed. If the last data set is the only data set recorded then the power transition is classified as transition between two steady states and data is forwarded as is to load disaggregation engine. If the number of data sets recorded exceeds two, the system compares the last data set to the reference power state, which embeds the power values of the pre-variations power state. If the difference between reference power state and last data set exceeds a threshold defined by the system the last data set is forwarded to the load disaggregation engine and all previous data sets are either discarded to forwarded to a transient load disaggregation engine. If the difference between reference power state and last data set is within a threshold defined by the system the last data set is forwarded to the load disaggregation engine and all previous data sets are either discarded to forwarded to a peak load disaggregation engine.

Methodology 2: Transformation of power peaks and transient peaks into steady-state power variations Separation of peak and transient data from steady-state data enables transformation of peak and transient data into steady state data, so that they can be monitored by event-based load disaggregation engines monitoring transitions between steady power states. Peak and transient data variations are (similar to steady state variations) unique to a given appliance.

The system of the invention transforms peak data into data suitable for steady-state load disaggregation engines in a five-step procedure.

1. Collect peak data

2. Cluster similar peaks together

3. For a period T, measure time interval between successive similar peaks to learn the periodicity of power peaks

4. After the period T, monitor time interval between successive similar peaks to determine periods of appliance activity

5. Transform peak periodic power activity into steady state continuous power activity during periods when appliances are active.

Step 1 is performed as explained in methodology 1. Step 2 uses state-of-the-art clustering techniques to match similar power peaks. For Step 3 the system investigates with state-of-the-art time series analysis techniques the periodicity of each power peak. When a periodicity exists and is measured by the system, the system starts monitoring the periods of power activity in Step 4. Identification of on and off periods is based on the presence of power peaks within expected time period, as depicted in Fig. 10. During monitoring, if a new peak appears within the pre- discovered time period after the previous peak then the appliance is set as powered on. If no peaks are retrieved then the appliance is set as switched off until a new peak is detected. Peaks are characterised with their power characteristics, periodicity and duration, which allows peaks to be unique to individual appliances. Such a technique applies to all equipment which generate periodic peak patterns, typical to resistance type of equipment such as hot water boilers, coffee machines and cup warmers. Step 5 is a procedure by which the system transforms the power peaks into a steady and continuous power activity, which reflects the duration of the appliance power activity and a total power consumption that equals that of the power peak. If the appliance power activity is detected for a period T made of N successive power peaks, the total power consumed Ptotal expressed in kWh by all the power peaks is calculated as the sum of the N individual power peaks' power consumption Ppeak(i): Ptotal =∑(Ppeak(i)) with i = [IN]

The system finally produces a power step of duration T matching the timestamps of the peak activity period. The amplitude Amp of the power step is calculated by the system by the equation below, with Amp expressed in kW and T in hours:

Amp = P 'total / T The system of the invention transforms transient data into data suitable for steady-state load disaggregation engines in a two-step procedure.

1. Collect transient data

2. Transform transient power activity into steady state continuous power activity reflecting period of activity or energy consumed.

Step 1 is performed as explained in methodology 1. Step 2 is a procedure by which the system transforms the power transients into a steady and continuous power activity, which can either reflect the exact duration and start/end times of the appliance power activity (equal to the power activity in transient AND steady state), or to reflect the exact total power consumption of the appliance (equal to the sum of power consumption in transient AND in steady state). If a transient power activity is detected for a period T between Tl and Tl+T, meaning that the power levels are in steady state before Tl and after Tl+T, the system calculates the amplitude 'Amp' of the power variation between Tl and Tl+T indicative of the pre- and post-transient steady power states. If the goal is to reflect the exact period of power activity of the appliance while allowing its recognition by the load disaggregation system, the system of the invention transforms the transient power data into a power step of amplitude Amp and of duration T starting at Tl. This guarantees accurate recognition of the appliance period of power activity by load disaggregation at time Tl, but on the other hand produces an inaccurate energy computation as the new power consumption may differ from the real transient power consumption. The system of the invention can output an error code indicating a transient consumption transform.

The system of the invention can also transform the power transient to let the load disaggregation algorithm compute the exact power consumed by the appliance. The system first calculates the real power consumption Ptransient of the transient power activity. The system transforms Ptransient into a power step of amplitude Amp and of duration Ttr = Ptransient/Amp starting at T2-Ttr. This calculation guarantees accurate computation of the appliance power consumption by the load disaggregation engine, but on the other hand produces an inaccurate start of power activity at T2- Ptransient/Amp. The system of the invention can output an error code indicating a transient time transform.

Technique 3 addresses a number of problems that can happen during the procedure of profiling electrical equipment. The procedure of profiling equipment can be either fully automated, semi- automated with manual input, manually operated or fully automated via use of auxiliary systems: - Fully automated setup - The software engine analyses current and voltage variations and clusters similar power variations into signatures [Hart 1992, Sintoni 2011]. In a second step, naming of clustered signatures is performed utilising heuristics and intelligent matching to known and peculiar equipment footprints. For example, the specific on/off periodic pattern of fridges can be used for naming a fridge without requiring human input.

- Semi-automated setup - The software engine analyses current and voltage variations and clusters similar power variations into signatures. Next, whenever one of these unnamed but known power variations appears, the engine asks the user/technician to name or pick from a list which equipment just switched on. The engine can therefore over time associate names to all clustered signatures.

- Manual setup - The software engine is configured to capture the power signal. The appliance is switched on and time is allowed for the software engine to record data embedding the power characteristics of the appliance just switched on. The appliance name is entered during the procedure. The process is repeated for each appliance to be monitored.

- Fully automated setup with auxiliary system - Auxiliary devices (e.g. sensor nodes, current switches, portable sub-meters, etc) are temporarily deployed to capture the true equipment operating state and power footprints [Schoofs 2010a]. Such equipment power reports are forwarded back to the software engine to automatically generate accurate power profiles. Within this invention a profiler is defined as the computer program managing record of power measurements and creation of power profiles. Problem 1 (Fig. I ll: Not all power states of electrical equipment are profiled

A variety of appliances such as dishwashers and industrial machines have finite state machine (FSM) models, meaning that for complete disaggregation each state of the FSM must be profiled as belonging to a given piece of equipment; profiling appliances for a set duration may only allow disaggregation of the initial operation states. Often, the profiler cannot be configured with prior knowledge on equipment operation state and will not know that the piece of equipment switches between multiple power states and only the initial step will typically be profiled or assigned to that equipment. Risks of incorrect or incomplete load disaggregation will therefore happen when the non-profiled power states will not be recognised with the power measurement. When profiling is manually operated, the system of the invention involves human intervention to indicate the nature of the equipment to be profiled, for example single-state or N-state appliance, and that profiling has started. Such intervention will happen through the input mechanism of the system of this invention. When single-state profiling mode is selected, the system of the invention monitors power variations and steady states, and indicates via its output mechanism that profiling can be stopped when a power transition followed by a steady state has been identified. This will allow complete control over profiling of appliances that have or do not have transient start. The output mechanism can be a light alert to the technician conducting the profiling procedure or pre- defined up/down signal on a physical line.

For default cases when the equipment type is unknown and when equipment type is of a multi power state type, the full equipment operation cycle is indicated by the system of this invention to let the load disaggregation profiling engine record all steady power states that belong to that appliance. The appliance is switched on and left powered on for its typical time duration in normal use or until it automatically stops its cycle. That way all power states exhibited by the appliance will be captured by the system. The system's output will indicate that profiling can be stopped when the power level state returns to its pre-profilmg power state meaning that the appliance cycle or operation is complete. The procedure is summarized in Fig. 12, Problem 2 (Fig. 13): The power characteristics recorded are not that of the equipment profiled The profiler will often not be capable of differentiating whether a power step variation captured during a profiling procedure belongs to the equipment being profiled or to another equipment. Indeed, it may happen that a secondary piece of equipment switches on while the profiler computes a given footprint, therefore creating a power variation that will not be that of to the equipment being profiled. Risks of generating an incorrect profile for a given piece of equipment will in turn affect the capability of the load disaggregation to recognise that equipment. Fig. 13 shows an example of an incorrect power step being recorded as a power profile.

The system implements the methodology presented in Fig. 14. Whenever the profiling procedure starts the system makes a record of the START power state before the appliance is switched on or off. When a power event is detected the system makes a record of the power step A and waits for another power event that indicates the switch of the appliance power state and the end of the profiling period. When the second power event occurs the system makes a record of the power step B and of the END power state.

The system then implements a validation step, which consists of comparing:

(1 ) the delta variations of the power-down and power-up events,

(2) the start and end power values before and after the switch on/switch off stages.

The objective of this validation step is to identify the contribution of other equipment that would not have been identified by the profiler. A major difference in the two events power variations calculated as B - A will indicate that an appliance may have been switched on or switched off at the same time the profiled appliance was switched on or off, or that the two power events have been made by two different appliances. A difference in START / END power values would indicate a switch on or switch off event independent from the appliance profiled and which would have occurred during the profiling period. In both cases the system of the invention indicates that profiling step will need to be either canceled or conducted again.

TECHNIQUE 4: Prevention of impossible load disaggregation output via interfacing with preexisting auxiliary data acquisition systems. Often, specific combinations of equipment power activity are impossible whether a range of equipment are programmed to always operate concurrently or complementarity, or whether equipment operation is scheduled during time periods via programmable timers. Similarly, singular equipment power activity may be impossible, whether equipment is known to be broken/disabled or whether it is known from another source of data that the piece of equipment is off. The system of the invention incorporates mechanisms to gather such information from auxiliary data acquisition systems and provides output signals that can be used to prevent inclusion of a piece of equipment for load disaggregation. Similarly, the system of the invention can be configured to provide an output signal used to enable a piece of equipment for load disaggregation, or to inform the number of appliances that should be recognised during load disaggregation engine. Letting a load disaggregation system know of possible and impossible breakdown options (when such an information is available) reduces the risk of the system delivering down incorrect power activity. Validation from data networks:

The system implements a technique with validation from data networks. With Internet connectivity being integrated into the wide range of electronic devices, e.g. net-TV or smart washing machine, most if not all equipment will soon be accessible and controllable via Internet. Networked equipment exhibits network activity on data networks whenever they are powered on and active [Schools 2010b]. Networking indeed requires acquisition of an IP address to communicate on the network and such IP address acquisition can be monitored and used to track networked equipment connectivity and therefore power activity [Schoofs 201 1]. The system monitors the network presence of a networked equipment to confirm or invalidate its inclusion in the list of appliances that can be processed by the load disaggregation engine. Attaching a set of rules to network activity can be used by the load disaggregation engine to filter out those appliances that are not powered on. For instance, Fig. 15 shows how equipment network connectivity state can be used to enable/disable pieces of equipment from the list of loads to disaggregate. The use of rules as to how to interpret and link network connectivity to power activity is depicted with the example of Appliance M being considered as powered off as soon as network probes are not answered. Other rules may be set depending on the equipment and network audit frequency. In some occasion the system of the invention will interface with the load disaggregation system and will incorporate intelligence to match names of equipment monitored on the data network with equipment listed in the library of power profiles. In other cases, the system of the invention will solely output the number of machines detected as powered on the data network, so that this figure is continuously updated over time and can be used by the load disaggregation engine to match detected changes of power activity with similar changes in the number of equipment powered on the data network. For instance, if the load disaggregation engine detects a computer being turned off the system of this invention should at the same time report that one machine disappeared from the data network. A change of connectivity state on a data network should match a change in load disaggregation output, and the robustness processor outputs changes observed on auxiliary networks and the load disaggregation engine can use this information to verify its output accuracy.

Validation from building management systems:

Some equipment is controlled via timers, e.g. hotel washroom air evacuation system that stalls at 8am and stops at 10pm. Dynamic access to scheduling reports of such timers provides knowledge whether a given appliance should be powered at a certain time. Interaction with building management systems is therefore presented in this application as an innovative way of producing rules to disable a number of appliances from the disaggregation and increase the robustness of the disaggregation engine. The methodology is similar to that using network connectivity. The system queries equipment time schedules to enable/disable them from the list of loads to disaggregate based on the current time and whether the appliance is scheduled to be powered on or off. Alternatively the system may return the number of appliances powered on at a given time based on the building management system configuration. The use of rules as to how to link equipment time schedules to measured power activity will depend on various factors including the accuracy at which such time schedules have been produced (e.g. divergence may exist between sheet reports and actual timer settings), whether the time schedules are enforced (e.g. time schedules may not be in effect at that precise moment), and the risk of non- synchronisation between the system and the time schedules clock (a few minutes time difference between the two clocks may lead to equipment being enabled/disabled incorrectly).

TECHNIQUE 5: Data reduction pre load disaggregation to prevent data overload and secure real-time disaggregation of future power measurements. The amount of data collected by individual meters is of critical importance for the robustness of load disaggregation systems, in that it affects on one hand the capability of such systems to monitor all power events if data is not sampled fast enough, and on the other hand the capability of such systems to deliver real-time power disaggregation if data is sampled too often and creates too large data sets to process, see Fig. 16. Sampling frequency is therefore the parameter to control and data storage policy the parameter to optimize.

Many techniques exist to dynamically adjust data acquisition and prevent data overload [LIANG 2010, WO 201 1 , US 2011], typically combining low-frequency periodic data sampling and high- frequency data sampling when power events are detected. Data is produced periodically whenever changes in electrical parameter values are below set thresholds, e.g. apparent power phase A value stays within 1 OOW of its previous value. However, when a variation for a defined set of parameters is above a set threshold, e.g. apparent power increases greater than 100W on phase A, data collection is triggered.

Although dynamic data acquisition frequency reduces greatly the amount of data stored for processing, such a technique must be combined with another round of data reduction pre load disaggregation to filter out that data not of interest for load disaggregation. In various embodiments the functionality to implement Technique 5 follows a 2-step procedure, as follows with each independent from each other:

i. Filter out data for unnecessary parameters, i.e. data initially collected parameters not important for load disaggregation

ii. Reduce number of data points by creating event transition points as illustrated in Fig. 17.

For step 1 the system of the invention processes data received from the data collection system and processes each data set against a filter of allowed parameters for load disaggregation. Concretely, all data of no use for load disaggregation such as voltage measurements are discarded.

For step 2 when a new steady state is discovered two new points are created. The first point is created to represent the last data point captured in the previous steady state and the second point is created to represent the first point captured in the new steady state. When no transient states exists between the two steady states the timestamp of the two new points are separated by at least the minimum time resolution of the system. All other points are discarded. Fig.17 illustrates the transformation of data points into event transition points.

Large data sets are reduced to foster load disaggregation capability to process data points in real- time. Real-time load disaggregation is key to robustness as it allows real-time user interaction and feedback for profiling in manual and semi-automated mode and real-time correlation of power events with external events.

The invention is not limited to the embodiments described but may be varied in construction and detail. For example the invention may be applied to monitoring of flow of other things such as gas or water, for which there is consumption by various different devices.

Claims

An electrical power data robustness system comprising an interface to an electricity power sensor and a processor adapted to perform processing of sensor data to provide robust data for load disaggregation.

A system as claimed in claim 1, wherein the processor is adapted to detect data failure on data loss, and/or on corrupt sensor data, and/or on inconsistent sensor data, and/or on unsynchronised sensor data.

A system as claimed in claim 2, wherein the processor is adapted to automatically detect data loss for a parameter by detecting lack of time synchronisation between parameters of a given electrical reading in response to an event.

A system as claimed in claim 3, wherein the event is a power event having a concurrent variation of a number of electrical parameters.

A system as claimed in any of claims 1 to 4, wherein the processor is adapted to determine if a current sensor data periodicity pattern falls outside periodicity rules.

A system as claimed in claim 5, wherein the processor is adapted to, in a learning phase, detect data loss by analysing sampling patterns in the sensor data to automatically detect said data sampling periodicity rules.

A system as claimed in claim 6, wherein the processor is adapted to, during said learning phase, measure for a specific set of parameters a time period between successive data points when the signal exhibits a steady-state mode, and to separately measure data sampling frequency after each power event as well as the number of data points collected at that frequency and the time period that frequency is maintained.

A system as claimed in any of claims 2 to 7, wherein the processor is adapted to perform computation to determine if a parameter has not been updated, thereby causing data loss, and to perform a computation to update a parameter that has not been updated or received to repair data non-synchronisation and loss.

9. A system as claimed in any preceding claim, wherein the processor is adapted to perform cross computation of parameter values to automatically detect lack of sensor data synchronisation.

10. A system as claimed in claim 9, wherein the processor is adapted to determine if

Apparent, Active and Reactive Power values are correctly linked.

11. A system as claimed in claims 9 or 10, wherein the processor is adapted to determine if Active Power is correctly linked to RMS Voltage and RMS Current.

12. A system as claimed in any preceding claim, wherein the processor is adapted to automatically detect if data for a particular parameter has not been received by analysing data identifiers.

13. A system as claimed in claim 12, wherein the system is adapted to analyse data identifiers by counting number of unique parameter identifiers received for a pre- configured data set.

14. A system as claimed in any preceding claim, wherein the processor is adapted to perform active power value analysis to automatically detect if a current transformer has incorrect clamping, or is damaged, or is undersized , or is incorrectly wired.

15. A system as claimed in claim 14, wherein the processor is adapted to monitor the current values on all current transformers to detect long periods when a given current transformer returns zero values which indicate that it does not operate correctly.

16. A system as claimed in claim 15, wherein the processor is adapted to differentiate a current transformer problem from lack of power activity by analysing current measurements over three phases and, by detecting when only one or two phases exhibit zero current values, to determine the longest period a current transformer sensor has returned zero value currents.

17. A system as claimed in any preceding claim, wherein the processor is adapted to perform a voltage value analysis to automatically detect if a sensor providing the sensor data is incorrectly configured.

18. A system as claimed in any preceding claim, wherein the processor is adapted to monitor in real time the result of the following rules:

if no data point for parameter is received after // then at least one data point is deemed lost, and

after a power event, if no data point for parameter i is received after ti then at least one data point is deemed lost, and

19. A system as claimed in any preceding claim, wherein the processor is adapted to perform data filtering to remove chaotic data variations.

20. A system as claimed in claim 19, wherein the processor is adapted to perform timestamp and power value analysis to perform automatic separation of steady state, peak, and transient power steps.

21. A system as claimed in claims 19 or 20, wherein the processor is adapted to store received data points until it has been established whether they should be filtered out or passed to a load disaggregation system, wherein for each new set of data the processor attempts to detect a new power event relating to significant changes in the values of a defined set of electrical parameters in respect to a reference power state, indicating either a change of a steady power state or the next stage of an ongoing power variation.

22. A system as claimed in claim 21 , wherein the processor is adapted to determine whether previous data sets are currently stored and not yet processed, wherein if data sets are recorded this indicates that the system has not yet established the nature of an on-going power transition, wherein absence of data sets indicates the beginning of a power transition, and wherein in both cases new data sets are added to the record and are not released until the nature of the transition has been established.

23. A system as claimed in claim 22, wherein the processor is adapted to determine that a transition is finished when no new power event has occurred on N successive data sets or after a time period t or a combination of both.

24. A system as claimed in claim 23, wherein the processor is adapted to analyse first the last data set collected, wherein if the last data set is the only data set recorded then the power transition is classified as transition between two steady states and data is forwarded without change to a load disaggregation engine, wherein if the number of data sets recorded exceeds two, the processor compares the last data set to a reference power state, wherein if the difference between the reference power state and last data set exceeds a threshold the last data set is forwarded to the load disaggregation engine.

25. A system as claimed in any of claims 19 to 24, wherein the processor is adapted to automatically transform peak power steps into steady state power steps.

26. A system as claimed in claim 25, wherein the processor is adapted to perform the steps of:

(f) collecting peak data,

(g) clustering similar peaks together,

(h) for a period t, measure time interval between successive similar peaks to learn the periodicity of power peaks,

(i) after the period t, monitor time interval between successive similar peaks to determine periods of appliance activity, and

(j) transform peak periodic power activity into steady state continuous power activity during periods when appliances are active.

27. A system as claimed in claim 26, wherein the processor is adapted to perform the step (e) by transforming the power peaks into a steady and continuous power activity, which reflects the duration of an appliance's power activity and a total power consumption that equals that of the power peak, wherein if the appliance power activity is detected for a period t made of N successive power peaks, the total power consumed Ptotal expressed in kWh by all the power peaks is calculated as the sum of the N individual power peaks' power consumption Ppeak(i):

Ptotal =∑(Ppeak(i)) with i = [1,NJ wherein the processor determines a power step of duration t matching the timestamps of the peak activity period, and the amplitude Amp of the power step is calculated by the equation below, in which Amp is expressed in kW and l in hours: Amp = Ptotal / T

28. A system as claimed in any of claims 19 to 27, wherein the processor is adapted to automatically transform transient power steps into steady state power steps.

29. A system as claimed in claim 28, wherein the processor is adapted to perform said transformation by collecting transient data, and transforming transient power activity into steady state continuous power activity reflecting period of activity or energy consumed.

30. A system as claimed in claim 29, wherein if a transient power activity is detected for a period / between tl and tl+t, the processor calculates the amplitude of the power variation Amp between tl and tl+t as being indicative of the pre- and post-transient steady power states, and wherein the processor is adapted to transform transient power data into a power step of amplitude Amp and of duration / starting at tl

31. A system as claimed in claims 28, 29, or 30 wherein the processor is adapted to calculate power consumption P transient of transient power activity and power consumption Pstep of a power step of amplitude Amp and of duration t starting at tl, and to next transform the transient power data into a power step of amplitude Amp and of duration Temp = t*Ptransient/Pstep starting at t2-Temp.

32. A system as claimed in any preceding claim, wherein the processor is adapted to perform data profiling including detection of appliance power states according to power values and steady state transitions.

33. A system as claimed in any preceding claim, wherein the processor is adapted to perform data profiling including detecting incorrect appliances according to start and end power values and step-up and step-down power values.

34. A system as claimed in any preceding claim, wherein the processor is adapted to automatically determine from an external system if an appliance if disabled and to accordingly disable power profiles for such an appliance.

35. A system as claimed in claim 34, wherein the processor is adapted to interface with a building management system to perform dynamic access to scheduling reports of timers for appliances.

36. A system as claimed in claim 35, wherein the processor is adapted to apply rules as to how to link auxiliary equipment time schedules to measured electrical readings in said sensor data..

37. A system as claimed in claim 36, wherein the processor is adapted to process data concerning equipment network connectivity state to enable or disable pieces of equipment from a list of loads to disaggregate.

38. A system as claimed in claims 36 or 37, wherein the processor is adapted to output changes observed on auxiliary networks so that a disaggregation engine with which it communicates can use this information to verify its output accuracy.

39. A system as claimed in any preceding claim, wherein the processor is adapted to output to a disaggregation engine data indicating the number of appliances powered on at a given time based on a building management system configuration.

40. A system as claimed in any preceding claim, wherein the processor is adapted to perform data reduction by generating transition points.

41. A system as claimed in claim 40, wherein the processor is adapted to perform the steps of filtering out data for unnecessary parameters, and reducing number of data points by creating event transition points.

42. A system as claimed in any preceding claim, wherein the processor is adapted to provide an interface for configuration of sensors providing data.

43. A system as claimed in any preceding claim, wherein the processor is adapted to provide an output signal to disable disaggregation of certain appliances according to quality of data associated with said appliances.

44. An energy monitoring system comprising a robustness system as claimed in any preceding claim and a disaggregation engine linked with the robustness system and adapted to perform appliance disaggregation and provide per-appliance energy output data.