This invention relates to systems and methods used to monitor an electrical system in an industrial, commercial or domestic environment. More particularly, it relates to systems and methods of benefit to electricity suppliers or aggregators, and also to consumers, for monitoring sub-sets of the electrical load behind an electrical meter.

Electricity supplies to a client consumer typically comprise of one or more main feeds to the client location, where the feed first goes to a meter, and from there it splits off into various circuits as required, with each circuit generally being used to supply electricity for particular loads, or types of load. For example, in a domestic environment, there may be different circuits for lighting and sockets (with separate circuits for upstairs and downstairs, or different regions of a property), outside supply, etc. Other circuits may exist for individual items that have a large current draw, such as cookers, showers, or electrical heaters. In a commercial environment such as a large store, there may be circuits for e.g. lighting, warehouse conveyors, tills, heating, and air conditioning. Similarly, in an industrial environment, there may be circuits for various items that consume significant amounts of electricity, such as motors, mills, crucibles, etc., as well as the usual lighting and socket circuits.

For industrial and commercial clients, there is currently a system in place known as

Demand Side Response (DSR), or, synonymously, Demand Side Flexibility (DSF). This is a system whereby the electricity supplier or an electricity aggregator may request that a client switches off an electrical load for a period, or delays switching one on, in times of peak load on the system. The loads that typically fall under DSR control will tend to be those that are not time-critical. For example, a large freezer unit may be switched off for a few minutes, which will have no great effect on the freezer functionality. Similarly, a request may be made to delay switching on a significant electrical load like a heater for a crucible. These interventions have the effect of smoothing out, or balancing the load on the electricity grid by lowering peak demand. Such balancing can provide significant savings for the electricity generating companies, as they do not need such an extensive network of power stations at the ready to cope with these, often short-lived, demands.

On the client side, the implementation of the requests to switch off or delay the switch-on of electrical loads may be done either manually (with a person controlling appropriate switchgear), or automatically, with the supplier or aggregator having the ability to switch on or off the electrical loads. Where the control is done manually, then there may be disputes between the supplier and the client as to what load was switched off, or when a given load was switched off. Such data may not be immediately apparent from the electricity meter, due to random switch-on of other systems etc., masking the switching process of the desired load. Where the control is done automatically, both the client and supplier need to be sure that a load in question is switched appropriately.

There is thus a need for electricity companies to be confident that switching requests are complied with both in the time and duration of execution and the level of reduction or shift in electric load during the time period. A solution is to have smart meters attached to all loads subject to DSR. This however can be expensive to retrofit to a given installation, as smart- meters are relatively expensive.

There is thus a desire at the supply side to have reliable information as to what electrical loads are in operation at the client side, both for DSR type applications, and also more generally, for disaggregation of client electricity usage for various purposes.

US Patent 4,858,141 discloses a technique for non-intrusive load monitoring (NILM), which comprises taking measurements from an electricity meter at regular, frequent intervals, and pairing increases in power draw with subsequent equal sized decreases in power draw, and calculating an energy consumption of an electrical load based upon the paired readings. NILM can be very useful where the client load characteristics are relatively simple.

However it suffers from various drawbacks. Primarily, it requires a high sample rate - typically of at least 1 second intervals, and preferably shorter - of the load measurements to reduce the chances of it missing a switching event. For example, if multiple switching operations take place between meter readings, then it will get confused as to which event is associated with which electrical load. There is also general uncertainty in the process as to how many different loads there may be, and so electricity suppliers tend not to rely too much on NILM techniques for reliable disaggregation, particularly for DSR, and so it is unsuitable for DSR billing settlement. It tends not to be particularly useful for situations more complex than a domestic environment.

Embodiments of the invention have the object of addressing one or more shortcomings of the background art, and/or to provide an alternative thereto. According to a first aspect of the present invention there is provided an electrical load monitoring system comprising of:

a) a primary energy meter, attached to an electrical supply line that supplies a plurality of independently switchable electrical loads within an electrical installation, the meter being able to measure at least one power parameter of the electrical loads;

b) a plurality of switch monitors, each adapted to monitor the status, either on or off, of an associated electrical supply switch associated with an independently switchable load, c) a hub, in communication with the plurality of switch monitors, arranged to record at intervals i) the status of the switches as determined by the switch monitors, and simultaneously, ii) a measurement of the at least one power parameter from the primary energy meter; where (i) and (ii) together comprise a measurement event;

wherein the hub further comprises a processor and memory, and is arranged to store a plurality of measurement events, and to apply regression to the stored

measurement events to estimate the current drawn by the electrical loads behind each monitored switch.

The invention provides a way of disaggregating the loads behind the primary meter, which is typically a fiscal meter - i.e. the meter that is used by an electricity supplier for billing purposes - with a defined accuracy, that in many cases, and with appropriate setup and commissioning, is good enough for DSR billing settlement and other purposes. Thus, the output from the regression provides an estimate, at the time of the latest measurement interval, of power being used by the electrical load behind each monitored switch. The processor is thus arranged to calculate a disambiguated power drawn by loads behind the monitored switches at a given instant with reference to the regression output and the switch state at the given instant.

Note that, in a typical installation, the primary meter and any sub-meters present are able to measure voltage and current, and hence power drawn by the loads they supply. Some meters may be arranged just to measure current flow, from which power may be inferred from knowledge of the system voltage. Either way, the power is often the term ultimately required by users (either a plant operator or an electricity supplier), and so such terms will be herein referred to as power parameters, and may comprise of the power or the current, and may further comprise of the voltage, the power factor, or other such parameters that may be useful in determining the power draw. The power may be multiplied by, or integrated over time to produce an energy reading, which may then be used for billing purposes. Often, the voltage supplied to loads is assumed to be the same throughout a client installation, and so, where this is the case, it only needs to be measured at one location (typically the primary meter). Advantageously, the meter will be arranged to measure at least the current drawn, as then such readings will be of direct use in, for example, subtracting off currents from complex loads before applying the regression.

When being used for DSR purposes, for example when a load is being controlled using DSR automatically by an electricity supplier, embodiments of the invention allow the supplier to switch off a load, and based upon the regression output and switch state, to be sure, even if the overall load current of the client installation increases following the switch- off, that the load in question was switched off, and the increased current draw was caused by other loads.

An installation of an embodiment of the invention within a client’s electrical network will preferably have all electrical loads on the client side of the fiscal meter monitored by means of being able to record whether each is, at a given measurement instance, switched on or off, through the use of monitored switches. In some instances that may not be possible, and there may be, for example, behind one or more monitored switches, more than one electrical load, each having its own particular load characteristics. These multiple loads sitting behind a switch may be in unknown states at any given instant. The status (e.g. whether switched off or on) of multiple loads behind a monitored switch cannot be ascertained using the regression analysis of the invention, and any switching of these multiple unmonitored loads can appear as measurement“noise”, which reduces the accuracy of the estimations (from the regression) of the monitored loads. A further source of measurement noise or error is where there are electrical loads that do not sit behind a monitored switch, in which case their changing power requirements again will tend to reduce the accuracy of the regression estimations of the monitored loads.

It is advantageous to ameliorate these situations, which may be done, in various embodiments, in different ways. For example, in an embodiment of the invention where there are a plurality of loads that are not individually monitored by means of a monitored switch, then a sub-meter may be positioned to record the power taken by an (otherwise) unmonitored load, the sub-meter then being read at a measurement instance, and the power drawn from the load components it is monitoring being subtracted from the total power (as read by the fiscal meter) before applying the regression to estimate the power being drawn by the loads sitting behind the monitored switches. This may be useful for example where there are lots of small loads, such as in a lighting system, where individual lights are controllable by users, and where it would be cost prohibitive to monitor each one individually. Thus, in some embodiments at least one load is monitored by a sub-meter, and wherein a reading recorded by the sub-meter is subtracted from the reading recorded by the primary meter before applying the regression.

Another approach for disaggregating a plurality of loads behind a monitored switch is to carry out a local NILM process (as described above) on those loads, where those loads are not monitored by, for example, individual sub-meters or monitored switches. In such cases, a meter (either the primary meter, or, if available, a sub-meter that is monitoring a power parameter drawn by the multiple loads) is polled at a suitable frequency (preferably at least every second), and the data so gathered used in an NILM process to infer individual load power draws. So, advantageously, in some embodiments, at least two electrical loads are monitored by a meter, and a Non-lntrusive Load Monitoring process is carried out on the loads to disambiguate such loads behind the meter. This will provide an estimate, subject to the various potential errors involved in NILM as mentioned above, of the loads behind this meter. This estimate can feed into the regression process, by finding NILM events that are not time correlated with the monitored switches and subtracting these from the total meter reading before applying the regression.

During the commissioning of an embodiment of the invention in a client installation, a training cycle will be used before regression analysis can be performed. This comprises recording a plurality of measurement events to build up a training dataset that is then used for the regression analysis, where each measurement event comprises of a measurement of one or more power parameters. A measurement event may also advantageously comprise the time of the event. When a system according to the present invention is installed in a client facility, then this training cycle may be entered as part of a

commissioning process. The training cycle involves recording a plurality of measurement events, preferably taken during normal operation of the system, and further preferably at different times of day, and different power draw periods, such as at weekends and during working days, with different switch settings of the monitored switches. By“normal operation”, it is meant that the client facility should be operating in a manner that is consistent with its operation when not undergoing a training cycle. However, the measurements events captured during the training cycle should capture each monitored switch in both of their on and off positions, preferably more than once, such as more than 5 times, 10 times, 30 times or more than 100 times. This should preferably be done until sufficient sample data points are gathered to perform the regression process, to a defined accuracy. The number of training points required will depend upon the complexity of the electrical load installation, and hence the number of monitored switches, as would be appreciated by a person of ordinary skill in the relevant art. It will, in typical environments normally number in the hundreds or thousands, with more data being required where there are more monitored switches present. It will be understood that the number of

measurements taken will affect the accuracy of the regression process. Too few measurements will tend to give larger errors in predicting the current or power drawn by a given load. It will also be appreciated that, once a sufficient number of measurements have been taken, the accuracy will approach a working limit, whereby further training

measurements do not have an impact upon the accuracy of the estimation process.

As the training dataset grows (and up until the threshold where sufficient readings have been taken) it becomes easier to disambiguate power draw contributions from individual electrical loads using the linear regression techniques employed with the present invention. An embodiment may be arranged to run a regression at various times during the recording of the training dataset, with estimations of current drawn by a given load being compared with actual power measurements of the same load. This will give an indication of the error in the regression process, and also whether the error is decreasing with an increasing training dataset. Thus, it provides a means for determining when a training dataset is of a sufficient size.

The training cycle therefore comprises recording at least one power parameter, including power drawn at the primary meter and any sub-meters (either directly, or by measurement of current, and preferably also voltage (which may otherwise be inferred from other knowledge of the system voltage) and the status of the monitored switches at intervals. The measurement intervals may typically be every 30 seconds, 1 minute, 5 minutes or 10 minutes, unless a particular meter is being used in NILM monitoring, in which case the measurement interval will preferably be at least ever second. Note that the measurement interval need not be fixed, but may vary, sometimes considerably, between measurements.

The gathering of information making up the dataset may continue beyond the duration during which the original training cycle is active. Thus, the dataset may be expanded or replaced with live data, comprising measurements taken from the electricity meter along with the status of the monitored switches during subsequent operation of the client facility, and the expanded or renewed dataset used in inferring which electrical loads are currently drawing power.

Thus, the processor is arranged to process the data gathered, comprising at least one power parameter measured at intervals, along with simultaneously monitored switch status, using linear regression and knowledge of the current monitored switch status, to associate power drawn with a particular electrical load.

The association uses regression to find weighting factors (e.g. currents) for each switch that represent the current drawn from each switch during the training cycle. Then, during a subsequent operational phase, these weights are applied to the switches to establish the energy usage.

Advantageously, at least one switch monitor may be arranged to monitor the status of its associated switch by means of a current transformer. This provides a simple technique for measuring whether any current is flowing to the loads associated with the switch, and so is indicative of the status of the switch.

Alternatively, or as well, at least one switch monitor may be arranged to monitor the status of its associated switch by means of, for example, making a measurement of resistance or voltage across the switch, or voltage across the load. As will be understood, if there are zero volts across the load between a switched live to neutral then it is in an OFF state, equivalently if there is very low resistance across the switch it is ON assuming it is in a normally functioning environment.

There may be different types of switch monitor for different switches, as governed, for example, by what may be already present in existing systems.

Other ways of monitoring the status of monitored switches may be employed, such as determining the physical position of a switching lever etc. Of course, where remotely controlled relay switches are used, then the hub will be able to read the status of the switch from whatever system is controlling the switches. Advantageously therefore, the hub may be integrated into any such systems so as to be able to read the switch status. Advantageously, one or more of the switch monitors may include a wireless communication means for transmitting information to the hub. In environments where this is possible, it saves on installing additional wiring systems, which may be expensive or inconvenient to do.

In some environments it may be more appropriate to use wired connections between the monitored switches and the hub. The wired connection may comprise a separate, dedicated data network, or may be integrated along with other existing networks. In some embodiments the electrical cabling itself (i.e. that cable carrying the current for the monitored loads) may be used to additionally carry switch state data. Such environments may comprise, for example, those where a wireless link cannot be made, due to blockage or significant attenuation of the wireless signal, or where safety concerns do not allow EM fields to be generated.

The hub may advantageously be a computer system. It may comprise, for example, a processor and memory, along with other standard functions of a computer system such as input/output capabilities. It may conveniently comprise, for example, a small form factor computer such as a Cubox or Beaglebone or Axiomtek or the like, having hardware to allow it to communicate with the primary meter, and other components of particular embodiments the invention, such as the monitored switches and any sub-meters that may be present.

Advantageously the hub is further arranged to communicate to external (to the client installation) systems, such as the billing system of the aggregator or energy supplier, or the DSR/DSF dispatch system of the aggregator, energy supplier or Distribution Network Operator (DNO), or a centralised energy monitoring system. Such communication may conveniently be done using existing WAN links, so providing internet access, but any suitable system may be employed.

The regression performed by embodiments of the invention is preferably a linear regression, but the invention should not be limited to such, and thus covers other such forms of estimation, based upon a training set. Preferably a least squares linear regression is performed. The regression will typically be linear however non-linear regression may also be used. Examples of non-linear regressors include, but are not limited to: softmax; support vector machines; (deep) feed-forward neural networks; (deep) recurrent neural networks, etc. According to a second aspect of the invention there is provided an electrical load monitoring system comprising of a software program product arranged to run on a computer, the computer being arranged to communicate with a primary energy meter in a client electrical installation and a plurality of switch monitors, wherein

a) the primary energy meter is attached to an electrical supply line that supplies a plurality of independently switchable electrical loads within an electrical installation, the meter being able to measure at least one power parameter of the electrical loads;

b) the plurality of switch monitors are each adapted to monitor the status, either on or off, of an associated electrical supply switch associated with an independently switchable load,

c) the software is arranged to instruct the computer to record at intervals i) the status of the switches as determined by the switch monitors, and simultaneously, ii) a measurement of the at least one electrical parameter from the primary energy meter; where (i) and (ii) together comprise a measurement event;

wherein the computer further comprises a processor and memory, the software being arranged to store a plurality of measurement events, and to apply regression to the stored measurement events to estimate the current drawn by the electrical loads behind each monitored switch.

The system of the second aspect may be arranged to function, when coupled to the primary energy meter and switch monitors, in a manner as described in relation to the first aspect of the invention.

According to a third aspect of the invention there is provided a method of monitoring a plurality of switched electrical loads located behind a primary energy meter, wherein the electrical loads have monitoring means adapted to measure the state of their associated switches, comprising the steps of:

a) measuring a power parameter from the energy meter, and substantially simultaneously measuring the state of switches supplying the electrical loads, and storing said power parameter and switch states as a measurement event;

b) repeating step (a) to build up a plurality of measurement events with the switches in different positions.

c) applying regression to the measurement events to estimate the current drawn by the electrical loads behind each switch.

The method may be used to carry out other aspects of the invention as explained herein. It will be appreciated the various aspects, and embodiments as described herein may be taken independently or in any suitable combination, unless such features are incompatible. In particular, features described in relation to one aspect may be applied to another aspect as appropriate, and claims may be drafted accordingly. Features described in relation to software may be carried out in firmware or hardware as appropriate, and features described in relation to hardware may be carried out in software or firmware where appropriate.

According to a further aspect of the invention there is provided a non-transitory computer- readable medium storing instructions that, when executed by one or more processors, are configured to cause the one or more processors to perform operations comprising the method as presented herein.

Embodiments of the invention are further described hereinafter, by way of example only, with reference to the accompanying drawings, in which:

Figure 1 diagrammatically illustrates a simplified view of a client installation, at which is installed an embodiment of the present invention;

Figure 2 diagrammatically illustrates a similar installation to that shown in Figure 1 , but wherein two further unmonitored electrical loads are present; and

Figure 3 shows three graphs that give an indication, using simulations, of how the accuracy of an embodiment of the invention varies with the number of measurement events used in its training data.

Shown in Figure 1 is a highly simplified representation of a client installation, such as a factory, with hardware elements of the invention located therein. Although it is described as being (in this example) a factory, it may, in other embodiments, equally be a commercial unit, a domestic unit, an agricultural unit, or any other suitable consumer of electricity.

Factory unit 10 has various items of equipment or electrical loads 11-17 installed therein. These all require supplies of electricity. Some, such as items 11-15 may be single, large, high power items, whilst others, such as 16 and 17 may comprise of a plurality of smaller individual elements, that generally work together, to provide their function, such as a lighting function (comprising of a large number of individually switchable lights) or an IT function (comprising of several computers, monitors or the like, independently switchable by their users). Also shown are four monitored isolation switches 18-21 , which may be used to switch on and off electrical loads supplied by them. Thus, switch 18 can be used to control power delivery to electrical load 11 via cable 25; switch 19 can be used to control power delivery to load 12 via cable 26, switch 20 can be used to control power delivery to loads 13 and 14 via cable 27, and switch 21 can be used to control power deliver to load 15 through cable 28. Load 14 is monitored by current transformer (CT) 29, that is able to measure current flowing thereto at an accuracy sufficient to provide a measurement for fiscal purposes. Note that electrical loads 16 and 17 are not switched (or at least not switched by a monitored switch). Instead, they are monitored by a sub-meter 30. The factory is supplied with electricity from grid equipment 35, and the electricity goes through primary electricity meter 36, from where it is fed to the electrical loads mentioned above through cable 37.

A hub 40 is connected to the electricity meter, and is able to poll the meter 36 for a power parameter measurement as desired. In this embodiment both voltage and current are measured, but other embodiments may also include other parameters such as power factor. The hub also has a connection to the monitored switches 18-21 , sub-meter 30 and CT 29 as indicated functionally by connection 42. This may be, in practice a wired connection or a wireless connection, or a combination of both. The hub is able to request any given switch or switches to provide its or their present status (e.g. either on or off) back to the hub. Thus, the hub is able, at arbitrary time intervals and instances, to measure total power draw at a given time, and also to measure the status of the monitored switches at the given time instance.

The hub comprises of interface hardware for communicating with the various switches and the electricity meter, and also comprises a processor and memory store. The processor may be a microprocessor, and conveniently in some embodiments may comprise of a small computer such as a Cubox from SolidRun Ltd., or a Beaglebone, or an Axiomtek ICO-120. Although shown in Figure 1 as a separate item from that of meter 36, it may, in some embodiments, be integrated with the meter 36.

When the system as described above has been installed and is operational, before it can provide information relating the current or power draw of each monitored load, it is necessary for a training cycle to be instituted. During this training cycle the client installation 10 will operate as normal, with the loads 11-17 being operated as required to perform the various tasks required by the client. This involves switching the various monitored switches to operate the equipment, resulting in different power consumption and current draws over the operational period. The system records at intervals, typically of between every 30 seconds to every 5 minutes the current switch state - i.e. the status (on or off) of each monitored switch, the electrical current drawn by the CT 29, and the power or current drawn by sub-meter 30 and primary meter 36. This information is stored in a table for subsequent use by a linear regression algorithm. The table is built up over time, until it is judged that it holds sufficient data for accurate estimation of current draw by a given load behind a monitored switch.

When taking data for the table, the system measures the voltage and current drawn by the loads 16 and 17, as shown by sub-meter 30. As this current is therefore known with some accuracy, it is subtracted from the current measured by primary meter 36, so that the difference current is that drawn by the loads behind monitored switches 18-21 and CT 29. The current drawn by the load being monitored by CT 29 can also be subtracted before the regression analysis takes place. The primary meter 36 has no idea as to the individual current drawn by any individual load behind the monitored switches, but the invention is able to provide an estimate of this, as described below.

Once the training data has been gathered, then the system is able to start estimating current or power associated with each of the monitored loads.

A linear regression analysis provides an output comprising of a series of weights, which are representative of the electrical currents being drawn by the various monitored loads. Thus, it may, when given the total power being drawn at a given instant, and the switch state at that instant (e.g. that switches 18, 20, 21 are on, and switch 19 is off, with CT 29 also showing no current being drawn), provide an estimation that the current drawn by the load behind switch 18 is x amps, the current drawn by the load 13 behind switch 20 is y amps, and the current drawn by the load behind switch 21 is z amps. This information is provided by the linear regression without direct measurement of the power or current taken by those individual loads.

Tables 1 and 2 show a highly simplified set of (simulated) sample training data to illustrate the principle more clearly. This set of training data assumes perfect measurement of the total current drawn, and hence no error is present. Each row records the switch status of two devices, Device A and Device B (with a“1” representing the switch being on, and a“0” representing the switch being off), along with the total current drawn (“Total”) as measured by a (perfect, error free) primary meter, and thus is a measurement event.

Device A Device B Total

1 0 30

0 1 20

1 0 30

1 1 50

1 1 50

0 1 20

0 0 0

0 1 20

1 1 50

1 1 50

0 0 0

1 1 50

0 1 20

0 1 20

1 1 50

Table 1 : Simplified two-device regression training sample data.

The training data is provided to the linear regression“Linest” function of Microsoft Excel, and Table 2 shows the results of the regression analysis. The regression output produces the second“Estimate” row, whereas the first row,“Actual” data is provided merely for comparison purposes, and is not known to the algorithm itself. The“Actual” row therefore represents the actual (as simulated) current consumption of each device. It can be seen that the regression estimate matches perfectly with actual consumption, and hence the error between the two, as shown in the third row, is zero. An energy provider could thus obtain an (in this case perfect) estimation of the current consumption of each monitored device using embodiments of the invention. Device A Device B

Actual 30 20

Estimate 30 20

Error 0 0

Table 2: Results of using Microsoft Excel linear regression Linest function on the data from Table 1

In order to perform the linear regression for example with Microsoft Excel the total current draw is treated as the single known Y and the switch states are the known Xs. The output of the linest function includes a weight for each input X value, and this weight is the estimated current draw for the load behind the associated switch. Although Microsoft Excel is used here, for convenience, it is likely that an installation of the invention would use a regression program that is written specifically for the task. This may conveniently use standard (e.g. open-source) libraries such as the Java Apache Commons Math Library, or Google ^(RTM) Tensorflow, or some other such library, or alternatively may be coded specially for the task.

It will be appreciated that the above example is trivial, but is provided as a simple example for clarity.

Tables 3 and 4 show a slightly more realistic simulation example, whereby there are now five devices, and wherein there is measurement error, as may be caused by use of a real- world meter, and by small non-metered loads being present in a system. Devices A-E again represent loads, whereby a“1” in the appropriate column indicates the monitored switch is on, and a“0” indicates it is off. The“Total” column represents the actual current drawn by the devices which are switched on (which is not known to the system of course). The“Error” column is a randomly generated inaccuracy value of between -2% and +2%, representing the measurement errors. The“Measured” column represents the

measurement as made from a primary meter, and thus has the errors present in the measurements. This (rather than that in the“Total” column) forms part of the data used in the regression process.

The total current as measured at a primary meter in each row is thus the sum of the (unknown prior to the operation of the invention) currents being drawn by those devices that are switched on, plus a contribution from the random inaccuracy percentage value. The error is estimated as a percentage of the current load. For example for measurement at peak load of around 100, we add an a contribution of up to +/- 2 watts in this case.

Device A Device B Device C Device D Device E Total Error % Measured

1 0 0 1 1 100.00 0.4 99.58

1 0 1 1 1 110.00 0.3 109.63

1 1 1 0 1 110.00 1.4 111.54

0 0 1 0 1 60.00 1.0 59.42

1 0 1 1 1 110.00 1.1 108.77

0 0 1 0 1 60.00 1.4 60.84

1 1 0 1 0 70.00 1.9 71.33

1 0 0 1 0 50.00 1.8 49.10

0 1 0 0 0 20.00 1.7 19.66

0 1 1 0 0 30.00 1.6 29.53

1 1 1 0 1 110.00 1.8 111.93

0 0 0 1 1 70.00 0.5 70.34

1 1 1 1 0 80.00 2.0 81.59

0 1 1 1 1 100.00 1.2 101.17

1 1 1 1 0 80.00 1.0 80.81

1 0 0 0 1 80.00 1.5 78.81

1 0 0 0 0 30.00 0.6 30.17

1 0 1 0 1 90.00 0.9 90.80

1 1 1 0 0 60.00 1.0 59.42

0 0 0 0 1 50.00 1.6 49.20

Table 3

It can be seen that there are 20 measurement events in this sample of training data. The regression analysis, as performed again by the Linest function, gives the results in Table 4. It can be seen for example that Device A, has an actual current draw of 30 amps, but the estimate as predicted by the regression is 30.28 amps, leading to an error of -0.9%.

Device A Device B Device C Device D Device E

Actual 30 20 10 20 50

Estimate 30.28 21.38 10.11 20.15 50.70

Table 4

The actual (for the purposes of the simulation) current drawn by each device when switched on is shown in the“Actual” row of Table 4, and again this is not explicitly known by the system.

It can be seen in Table 4 that there are some more significant errors here, of up to 6.9% in this particular example, with the estimation of the current draw of Device B against its actual current draw. This error can be improved by increasing the number of measurement events making up the training set. Table 5 shows a similar system to that of Tables 3-4, but instead using 200 measurement events rather than the 20 used previously. It can be seen that this brings the estimation error down significantly.

Device A Device B Device C Device D Device E

Actual 30 20 10 20 50

Estimate 30.21 19.97 9.87 20.11 50.31

Error -0.7% 0.1% 1.3% -0.6% -0.6%

Table 5

Clearly, most real client installations will be more complex than that shown in the

simulation, but it can be seen that provided the errors due to unmonitored current drains can be kept within reasonable levels, a good estimation of the current (and hence power) levels drawn by particular loads can be achieved. The results may be used for any required purpose, either by the energy suppliers, distribution network operators or by the clients themselves, e.g. for DSR purposes.

Figure 2 shows a system similar to that of Figure 1 , but with two additional equipment loads 50, 52 present. These loads are unmonitored by any monitored switches, sub-meters or CTs, and so their power draw cannot be ascertained by just using the regression process described in relation to Figure 1. Left unaccounted for, their power draws would effectively add a greater degree of error to the regression output otherwise computed.

One way to disambiguate the power drawn by them is to use NILM. Such a process is suitable where the loads exhibit discrete on-off behaviours (such as, for example, the switching on or off of a motor or lighting circuit), and are less suitable where the load tends to vary (e.g. a dimmable lighting load or variable speed motor). The meter 36 is polled here at 1 second intervals, to catch any transitions in the measured power parameter. Given knowledge of the monitored switch positions, any transitions that occur at the same time as switching events from switches 18 to 21 can be ignored during the NILM processing.

Remaining transitions must therefore be due to loads 50 and 52, and by tracing the power parameter changes the current drawn by each may be distinguished. The NILM estimations of current drawn by loads 50 and 52 may then be subtracted from current readings recorded by meter 36 before the regression is applied, to improve the accuracy thereof.

If, because of the load type, NILM is not suitable, then the loads should be placed behind a monitored switch, or behind a sub-meter whereby the currents drawn by them can be subtracted before the regression is applied.

Figure 3 shows how the accuracy of an embodiment of the invention can vary with the size of the training data set. Six graphs are shown, in three pairs, where each graph has the number of measurement events (observations) on the x-axis, and each pair shows the mean and standard deviation of error percentage, as viewed across all the devices in the network, on the y-axis. The three pairs show these figures for 10, 100, and 1000 devices. The loads are assumed to switch on and off at random times, whereas in practice it is likely that there will be more correlation between loads and their switching times. The error is the percentage error of the estimated device power usage as compared to the actual power usage for the device. Each simulation is repeated 10 times, and the result of each repetition plotted, so giving the spread in the y-axis.

Figure 3a and 3b show graphs representing a simulation of a client system having ten independent electrical loads, each with a monitored switch. Figure 3a shows mean error whilst Figure 3b shows standard deviation. It can be seen that as the training data set increases in size the percentage error converges towards zero for both the mean error and standard deviation of the error. Both the mean and standard deviation go below 1% by about 500 measurements.

Figure 3c and 3d show similar graphs, but this time where there are a simulated 100 switched devices in a monitored system. Here, a mean error value of 0.1 % with standard deviation of 1% is reached at around 20,000 measurement events with convergence towards zero as the observations increase.

Figure 3e and 3f again show similar graphs but this time with 1000 switchable client loads being simulated. The graphs again converge towards a percentage error of zero. After 100,000 observations the mean error value has reached about +/- 0.1 % but the standard deviation of the error is still around 6%. Although the invention has been described in relation to current measurements being obtained by the regression process, it may also be used, as appropriate, to obtain power or energy measurements. Also, the data making up the measurement events may comprise not only current draw information, but as well as or instead of, any other pertinent information such as power, energy, power factor, voltage etc., as would be understood by a normally skilled person.

The functions described herein as provided by individual components could, where appropriate, be provided by a combination of components instead. Similarly, functions described as provided by a combination of components could, where appropriate, be provided by a single component.

Throughout the description and claims of this specification, the singular encompasses the plural unless the context otherwise requires. In particular, where the indefinite article is used, the specification is to be understood as contemplating plurality as well as singularity, unless the context requires otherwise.

Features, integers, characteristics or compounds, described in conjunction with a particular aspect, embodiment or example of the invention are to be understood to be applicable to any other aspect, embodiment or example described herein unless incompatible therewith. All of the features disclosed in this specification (including any accompanying claims, abstract and drawings), and/or all of the steps of any method or process so disclosed, may be combined in any combination, except combinations where at least some of such features and/or steps are mutually exclusive. The invention is not restricted to the details of any foregoing embodiments. The invention extends to any novel one, or any novel combination, of the features disclosed in this specification (including any accompanying claims, abstract and drawings), or to any novel one, or any novel combination, of the steps of any method or process so disclosed.