Author: Stephen O'Keeffe, B Eng Marine and Plant Engineering, B Sc Hons, Process Plant Technology, M Eng Sc Sustainable Energy, is an asset integrity engineer at Brookfield Renewable Energy Group, and winner of the €1,500 EirGrid award for his master’s thesis on sustainable energy

Introduction to VEA concepts

Virtual Energy Auditing (VEA) is the practice of disaggregating interval meter consumption data over a time series in order to identify building specific consumption patterns or further to identify individual appliance consumption through a concept known as Non-Intrusive Load Monitoring (NILM). As the name suggests the practice is non-intrusive and is designed to be performed remotely or autonomously without the need for complicated installation or require the same level of ‘feet on the ground’ man hours previously associated with conventional energy audits. The methodologies used are abundant and they are employed through a spectrum of sampling frequencies anywhere from one hour data samples to the MHz range. Each offering different capabilities to infer where energy is being lost or consumed. To varying degrees of success this form of audit can be performed on any metered variable to deduce specific areas of excess consumption and even identify the actual appliances consuming this energy. The degree to which an aggregate electrical signal can be disassembled is dependent on the level of data resolution that can be obtained by the type of meter used and its sampling frequency. For this reason all NILM concepts can be broadly categorised based on their sampling frequency requirements. The majority of NILM concepts can be categorised under four distinct categories: cork1a Each category offers different advantages and also limitations as the disaggregation technique is not always suited to the electrical behaviour of the equipment being monitored. However despite certain limitations many VEA concepts have achieved substantial energy savings through rapid energy assessments and subsequent tackling of worst offending consumers.

Benefits of virtual audits

This form of demand side energy monitoring and the wealth of data produced has numerous advantages for the end use customer, energy supplier and energy research agencies. This data can be used to better model future energy demand, maintain greater supply integrity and allow the building owner more transparency as to how energy is used within the building. To support these functions many companies now provide cheap rapid online auditing services which utilise this data and put it into practice as an effective means of identifying equipment energy ‘hogs’ or ineffective building management controls. Accurate knowledge of appliance energy consumption can be used to support incentive programs to promote ‘off peak’ energy consumption and even offer the end user an accurate means of appliance benchmarking and cost benefit analysis when purchasing new appliances.

Research focus

The focus of this article is exclusively on the consumption of electrical energy and the identification of specific appliance consumption. The disaggregation concept which will be discussed is based on magnitude changes in real and reactive power consumption. This concept is typically referred to as ‘Transitions in Steady State Power’ a concept which was first introduced by George W Hart of MIT and is often favoured over other methodologies due to its relatively low cost and straightforward approach. The underpinning fundamental principle relies on the fact that many appliances have dissimilar electrical characteristics which will in turn generate different signatures upon activation/deactivation. For instance, an electrical element such as an element used in a kettle is purely resistive in nature (i.e. it does not have inductive or capacitive properties). This means that the power drawn by the element will be entirely ‘real’ power with no reactive component. This is due to the fact that both the current and voltage wave forms are in phase with each other. Conversely a motor would have both real and reactive components due to the coil winding resistance and the inductive nature of the device. When plotted on a 2-D plane, Real power (Watts) vs. Reactive power (VAR), the points are immediately distinguishable from each other. The resulting signatures are often quite distinct from each other, however due to signature overlapping which often occurs the concept is not well suited to multiple identical devices or large numbers of devices in the monitored scope. A sample device signature from a three phase motor is represented in Fig.1 showing the device consumption versus time and also the observed signature plot resulting from On/Off events. In this study the magnitude change in overall consumption resulting from device switching is referred to as ‘Delta PQ’ and the kW and kVAR change is denoted dP and dQ respectively. The dP and dQ are used for brevity and the P & Q labels are consistent with electrical engineering convention when representing real and reactive powers. The Delta PQ signature plot is constructed from multiple dP and dQ readings obtained during device fingerprinting and is plotted for visual representation on a 2-D plane in Fig.1. The approach is then to use this signature along with feature recognition algorithms to classify new unknown events as On/Off cycling of various appliances. If the magnitude change and runtime of the device is known then its power consumption can be calculated to a reasonably high accuracy. Although in theory this could be performed by a permanently installed smart meter, in this study a portable energy meter was used which was clamped to the main incomer via current transformers. To address the shortfall in publicly available test data regarding the efficacy of power based NILM in the field multiple studies were performed. The capabilities of the power based approach and how accurate it could be when applied to different equipment under various test conditions was assessed. Three building studies were put to the test;
  1. UCC’s Electrical engineering building (large scale)
Primarily consists of electrical & electronic laboratories, computer labs, lecture rooms and offices with a total floor area 2791 m2.  The electrical signal was sampled at the building main meter (Point of common coupling) therefore the study encompasses the entire building electrical energy consumption. Purpose of study: Uncontrolled tests. To emphasize the complexity of achieving individual appliance recognition using steady state power methods on a building of this size.
  1. UCC’s Cavanagh pharmacy building distribution board (medium scale)
The pharmacy building consists of biochemical & pharmacology laboratories, lecture theatres and offices with a total floor area of 5600 m2. In this case the electrical signal was sampled at a plant room sub distribution board which serves 27 items (primarily fans and pumps operated by three phase motors). Purpose of study: Controlled tests. To assess the performance of the NILM algorithm on a scale where it is frequently employed in commercial energy monitoring applications.
  1. Household appliance study (small scale)
Purpose of study: Controlled tests. To assess the viability of successfully applying this form of power based NILM at a household level and also to examine the variation in power signatures resulting from appliances with dissimilar electrical characteristics.


For this method the initial phase of the virtual audit consists of a period of device fingerprinting. This is the only intrusive step in the virtual audit. With the meter attached each device is cycled on and off a number of times. For many appliances this will result in a repeatable dP & dQ consumption values from the baseline consumption which can then be represented as target groupings on the DeltaPQ plot. Once sufficient values have been obtained, the algorithm can then classify new unknown switching events based on the signatures stored in the signature library. VEA process:
  1. An event is identified using a ‘Threshold exceeded’ logic. In this case when the change in kW consumption exceeds 50W (0.05 kW) between two consecutive data points. 50 Watts was chosen as the steady state variation in power draw did not typically exceed this value on the electrical line under test thus resulting in less false events triggered.
  2. Log time of occurrence of threshold exceeded.
  3. Wait a predetermined window time for the device to run up to full steady value before the signature is extracted. (This is set during the training phase to suit the range of equipment being monitored)aacork
  4. Record value of overall kW and kVAR at this time.
  5. Subtract the value of kW and kVAR prior to the event occurring from the recorded value in step four to give dP and dQ.
Event classification As signature overlap is possible the registered event must be assigned to its group according to ‘closest fit’ or ‘nearest neighbour’ algorithms. In order to increase the identification performance of the algorithm when dealing with unknown events that arise in the region of several proximate groups a distance based nearest neighbour logic was adopted. K-Nearest Neighbour (K-NN) was identified as a suitable candidate for this purpose. See Fig.2, when an unclassified event falls within these overlapping groups this approach will result in increasing the probability of correctly classifying that event. The process is a follows:
  1. The unclassified event is logged through the event detection logic previously outlined and the subsequent dP & dQ values are recorded.
  2. A range check is performed to make sure that the event falls within range of a suitable group or a number of groups for classification otherwise it is discarded as it is highly unlikely that it belongs to any of the devices being monitored.
  3. All other dP/dQ values obtained during the training phase when device fingerprinting are then subtracted from the new event dP/dQ and the differences are evaluated according to a Euclidean distance metric.
  4. The shortest distance from the unclassified event to all known signatures is now known.
  5. The six closest signatures are then selected to determine the classification of the new event. This is known as the K number (k = 6). These values then ‘vote’ to classify the new event.
  6. When the K6 are identified they are normalised to the largest distance in the group of six. This normalises all distances to an index between 0 and 1. Those falling on or below 0.5 are allocated two votes and those above 0.5 are allocated one vote. This serves to weight the decision towards the closer signature values.
  7. The device with greatest number of votes classifies the new event as belonging to its own group.
For this study K=6 was adopted as anything greater would have been likely to include distant signatures. Due to operational restrictions on live plant typically only 8-10 signature values were extracted for each device in the training phase of this study. Ideally at least 50 signatures in each grouping would be desired. The spread of each group would be much more clearly defined and the group density far greater. As the nearest neighbour solution is sensitive to the value of K then greater values of K will result in higher classification accuracy. With three or four proximate groups of 50 values each then a K value of up to 30 may be used. With the addition of a greater number training points the effect of outliners on the KNN-algorithm will be negligible.

Blind testing

To ensure unprejudiced test results a number of blind tests were performed by a third party. The knowledge of which devices were switched under test was withheld from the author until such time as the third party individual was able to verify the algorithm generated switching event record returned by the author. These tests were also conducted under increasingly difficult testing regimes to simulate equipment running in an overlapping and close switching manner. This allowed the performance of the algorithm to be assessed under the types of conditions which could be expected from general operation.

Conclusion and results

The medium scale tests conducted on the pharmacy buildings sub distribution board did confirm that this form of steady state power method could be employed to success on select groups of equipment where the number of possible devices is finite and manageable and where simultaneous switching is typically infrequent. Identification accuracies of 90-100 per cent were achieved under conditions expected from typical operation despite the fact that many of the devices used in the study were identical three phase motors. This accuracy is expected to approach 100 per cent when larger K values are used during classification. The test results showed that the number of devices which could be monitored increased when the individual signature groups were distinct and decreased when signatures were alike. The identification accuracy also increased or decreased according to the same rule. However it was also evident from the other studies that there are still considerable hurdles to overcome until an all inclusive VEA concept is developed. It was evaluated from the large scale tests performed on the Electrical Engineering building that this form of virtual audit was not sufficient in detecting single device switching at this level. Due to large quantities of equipment operating on the line the consumption data frequently fluctuated in a manner where it was not possible to isolate single switching events. It is also apparent that steady state power methods alone are unlikely to be a comprehensive solution to household disaggregation. One reason is the fact that there are a large number of mobile or semi-permanent appliances which can be used. The algorithm would have to either be trained to all new devices or alternatively employ autonomous machine learning to classify new appliances. The household typically includes appliance types which have certain electrical characteristics or operate in a behaviour which is not well-suited to power methods of this sort. Overlapping of device signatures is again more likely to be a concern in a household environment. Although primary consumers may be identified through more complicated steady state power based algorithms, in the long run high frequency sampling methods may prove more valuable in the household as they are typically better suited to low power and variable consumption devices. One factor above all which resulted in poor device recognition was the close switching of appliances. One solution to which is to decrease the sample window time. This is not always possible however as signature values are expected to be extracted at a steady state of consumption. The issue of close switching can be resolved using a method often referred to as ‘Integer programming’. The collective change in kW and kVAR values are they themselves disaggregated based on the most probable combination of devices which could make up this aggregate change. By combining this approach with the one outlined in this article many close switching events could be disentangled and classified separately according to the standard protocol. When it comes to discussing VEA performance it is also worth noting that the question of ultimate accuracy can be misleading as it pertains to the specific application in question which is dependent on the boundaries in which the applied VEA concept operates. For example, very few VEA concepts are suited to detection of variable power consuming devices such as power tools. If all the equipment being monitored operated in this manner then identification accuracy would be 0 per cent. Alternatively if the equipment being monitored was a kettle, toaster, electric shower and a fridge then the accuracy could be 100 per cent. The signature classification approach used in this study is constricted to equipment types which exhibit constant consumption patterns of repeatable magnitude. In many cases this is sufficient to identify most primary equipment in the monitoring scope however devices which exhibit cyclical behaviour consistent with finite state machines (such as a washing machine) may be harder to effectively identify without the inclusion of complex pattern recognition algorithms particularly when overlapping switching events occur during their operation. The performance of this particular approach is subject to the extent of signature separation, sampling frequency and consumption behaviour of the equipment in the monitoring scope. The results of this study indicate that techniques of this sort have the capacity to formulate worthy energy monitoring solutions where primary energy users can be identified to an acceptable level of accuracy particularly when more complex algorithms are employed. There is however no size that fits all and further development is necessary before a complete unrestricted VEA concept is to be realised. [caption id="attachment_24077" align="aligncenter" width="300"]aacor1 Fig.1 – Right: Delta PQ signatures observed from switching of a three phase motor operating a hood extraction fan (Designated FN1002). Left: Time vs. kW & kVAR[/caption] [caption id="attachment_24079" align="aligncenter" width="300"]aacor2 Fig.2 – Classification using K – Nearest Neighbour. FN1005 & FN1006 are identical three phase motor operated hood extraction fans. PU2001 & PU4010 are similar sized motor driven DHW pumps[/caption] The research for this paper was supported by Intelligent Efficiency Research Group (IERG) and its director, Dominic O’Sullivan