WO2011128883A2

WO2011128883A2 - An energy monitoring system

Info

Publication number: WO2011128883A2
Application number: PCT/IE2011/000024
Authority: WO
Inventors: Antonio Ruzzelli; Gregory O'hare; Anthony Schoofs
Original assignee: University College Dublin - National University Of Ireland, Dublin
Priority date: 2010-04-15
Filing date: 2011-04-15
Publication date: 2011-10-20
Also published as: WO2011128883A3

Abstract

An energy consumption monitoring system comprises n electrical sensor (2), a processor (3) adapted to recognise appliances according to inputs from the sensor, and a user interface. The processor generates, in a training phase, an appliance energy profile for each of a plurality of appliances, and uses artificial intelligence functions to identify usage of a particular appliance according to sensor outputs and the appliance profiles. At least some profiles identify the associated appliance as one or a combination of resistive, inductive, or capacitive components, and at least some profiles identify predicted variation patterns for said components, and include parameter values for real power, reactive power, and power factor.

Description

"An Energy Monitoring System¹

INTRODUCTION Field of the Invention

The invention relates to monitoring of electrical energy in locations such as homes or offices. Prior Art Discussion

It is well known that more care in utilisation of energy-consuming appliances can result in considerable reductions in energy requirements.

A system to monitor energy usage by appliances has been described in US6816078 (US2003/0093390). This describes a sensor and an estimating means which learns based on a training set in order to estimate individual electricity consumption of appliances in operation. Fundamental and higher harmonics of total load current and their phase differences are monitored. US6906617 describes an arrangement to determine instantaneous power consumption of an appliance to determine its status, according to a comparison with a characteristic power consumption pattern.

US2010/0280978 (published after our priority date) describes a system to collect waveform data for multiple appliances in order to generate power consumption data on a per-appliance basis.

WO2009081407 describes a process of detecting an electrical signature of an appliance and subsequently inferring statuses of appliances. Actions may be taken such as reducing power consumption to below a threshold.

The invention is directed towards providing an improved energy monitoring system with an ability to more accurately monitor usage by individual appliances. Summary of the Invention

According to the invention, there is provided an energy consumption monitoring system comprising:

an electrical sensor,

a processor adapted to recognise appliances according to inputs from the sensor, and a user interface,

wherein the processor is adapted to:

generate, in a training phase, an appliance energy profile for each of a plurality of appliances, and

use artificial intelligence functions to identify usage of a particular appliance according to sensor outputs and the appliance profiles.

In one embodiment, the processor is adapted to guide user inputs to assist with generating the profiles.

In one embodiment, at least some profiles identify the associated appliance as one or a combination of resistive, inductive, or capacitive components. In one embodiment, at least some profiles identify predicted variation patterns for said components.

In one embodiment, at least some profiles include parameter values for real power, reactive power, and power factor.

In one embodiment, at least some profiles additionally include parameter values for peak current, RMS current, peak voltage, RMS voltage, and sampling frequency.

In one embodiment, the artificial intelligence functions includes a neural network.

In one embodiment, the neural network comprises an activation function in the form of a sigmoid function for training according to the profiles. In one embodiment, the processor is adapted to use a combination of sensory inputs to process energy readings enabling signature acquisition, system setup automation, system calibration, system monitoring, system control, and system optimization. In one embodiment, the processor is adapted to generate a user interface for profiling appliances, and to augment reliability of user inputs during system profiling

In one embodiment, the processor is adapted to autonomously refine the system. In one embodiment, the processor is adapted to implement a control loop that feeds back appliance monitoring accuracy data to refine parameters of the artificial intelligence functions.

In one embodiment, the processor is adapted to control the capture of power data including appliance start and stop activity to autonomously generate appliances signatures.

In one embodiment, the processor is adapted to recalibrate an external appliance load monitoring system.

In one embodiment, the processor is adapted to perform online profiling and processing of appliance data.

In one embodiment, the processor is adapted to compute an appliance activity based on the sensor inputs, in which filters are associated with appliances, and to return a positive appliance activity when the combination of expected inputs is verified.

In one embodiment, the processor is adapted to make appliance state information available on a communication medium, thereby avoiding need for software support on a host system, and reducing the provision of annotated data.

In one embodiment, the processor is adapted to operate as an online or offline energy recommender system, which breaks out energy costs to per-appliance costs, and recommends appliances' replacements to reduce users' costs.

In one embodiment, the processor is adapted to group energy users into groups of users. In one embodiment, the processor is adapted to group users having similar energy usage profiles, for example with the objective of making the groups of users appear as a single entity for energy providers, and allowing specific group energy tariffs.

In one embodiment, the processor is adapted to perform the steps of:

during training, as appliances are activated and deactivated, generating signature vectors with data for a plurality of appliance electrical energy consumption parameters including apparent power, real power, and power factor, and storing the vectors in a training set; adopting input neurons and matching them with hidden neurons, which perform activation adequately for the inputted parameters used to generate the signature vectors: implementing a threshold function;

tune weights within the neural network; and

in real time as appliances are used, sampling said electrical energy consumption parameters and feeding them as inputs to the trained neural network, which calculates the combination of appliances closest to the input values and provides a sequence of appliance identifiers as an output.

According to another aspect, the invention provides a computer program product comprising a computer usable medium having a computer readable program code adapted to perform processor operations of a system as defined above in any embodiment when executing on a digital processor. DETAILED DESCRIPTION OF THE INVENTION

Brief Description of the Drawings

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 is a diagram illustrating context of an energy monitoring system of the invention; Fig. 2 is a diagram showing the architecture of a processing core of the system;

Fig. 3 is a plot illustrating power as a function of time for operation of a number of appliances;

Figs. 4(a), (b), and (c) are plots illustrating parameters for operation of the processing core;

Fig. 5 is an example of sensory information captured next to a kettle;

Fig. 6 is a plot illustrating the relationship between reactive and active power; Fig. 7 is a set of plots showing active power signatures for four appliances; Fig. 8 is diagram showing generation of appliance activity states;

Fig. 9 is a diagram showing example vectors generated during training; Fig. 10 is a diagram illustrating hining of weights in the neural network; and

Fig. 11 is a diagram showing sampling and appliance recognition in real time. Description of the Embodiments Referring to Fig. 1 an energy monitoring system 1 comprises a non-intrusive sensor 2 and a processing core 3. In hardware terms the sensor 2 comprises a current transformer (CT), a microprocessor and related circuitry, a radio unit, a power unit (either a battery or connector to the mains), and a receiver appliance connected to the processor. The processor may comprise a local and/or a remote processor, such as a laptop or a PC or a server. The sensor can be any of known sensors for measuring the various electrical parameters outlined in the description below.

In functional terms the processing core comprises:

- An "appliance profiling" layer that guides user profiling of electrical appliances within a premises and that generates a database of unique appliance signatures. - An "activity recognition" layer that utilises the unique signatures and trains a machine learning technique, such as a recursive weighted distance between combination of signatures and the signal, an artificial neural network (ANN), or a Bayesian classifier, that is then employed to recognize appliance activities in real time.

- A unique signature state information database used to store and retrieve profiled appliances for broad reuse and faster new profiling of similar appliances.

- An "automated data annotation" layer that automatically annotates energy data to facilitate system training and validation. The following are advantageous aspects of operation of the processing core, and which are explained in more detail below:

- automating energy monitoring,

- facilitating the system's setup, termed profiling,

- generating and managing unique appliances' signatures,

- recognising appliances based on recursive weighted distance between the signature and the signal from the meter, and

- an appliance descriptor database based on unique signatures.

Appliance Profiling

The appliance profiling layer guides user inputs to generate a unique appliance signature for each particular appliance. This is a one-off procedure per appliance, which can be either guided by the user or system-automated.

Used-guided profiling: The system 1 requires the user to turn-on the appliance, wait for a few minutes while the system captures data to profile the appliance, and finally turn-off the appliance. In case the appliance has multiple states such as stand-by or cycles, the layer requires those states to also be profiled.

Automated profiling: The system recognises a certain threshold variation of one or more parameters and triggers automatically the profiling phase.

The appliance profiling layer identifies key parameters provided by the electricity meter used for the generation of a unique appliance signature. These parameters can be transmitted to the processing unit either periodically or the transmission is triggered by variation of a threshold, namely throshold delta reporting. A train of those parameters is captured and stored into an appliance signature database.

According to each appliance's internal circuit, an appliance can be classified as resistive, inductive, or capacitive type. Inductors and capacitors affect the power consumption by shifting the alternate current with respect to the alternate voltage.

The following are the parameters forming a unique signature:

Active power (Ac) (also known as "real power"), when an appliance changes state. Capacitors delay the current with respect to the voltage while the opposite happens for inductors. Considering that the power is the multiplication of voltage and current, if voltage and current are shifted, the power transferred to the appliance is less.

Resistance (which does not affect the current/voltage shift).

Variation of reactive power (Rp), when an appliance changes state, to discriminate between appliances with similar electricity consumption.

In case the meter does not provide the reactive power parameter, then the power factor (Pf) and apparent power (Ap) constituents are captured. The reactive power is then calculated by the following: Rp = sqrt{(Ap)-(Rp)} = sqrt {(Ap) -(Pf-Ap)}.

Peak current (Pc) to further discriminate between appliances with a dissimilar current peak following a change of activity states (e.g. start-up, steady state, standby, power down). Pc relates to the appliance circuit specifics, as it represents the maximum amount of energy the appliance allows before reacting.

RMS current (Crms) that provides consumption information independently from the voltage given by the energy provider.

Peak voltage (Pv) and RMS voltage (Vrms) relate to the specific voltage provided when the signature is made. These are useful to reinforce the machine learning technique about the voltage provided at the moment of profiling. The system 1 captures a train of the above parameters for a time period per appliance state change. Overall, the system identifies 6 inner appliance specifics to generate a unique appliance signature, which is the base to discriminate between multiple appliances activity.

Additional factors captured when profiling appliances are the signature length and the meter sampling frequency. These parameters are key to translate signatures from dissimilar types of energy meters into a standard signature. In fact, when profiling appliances, the system 1 may generate signatures of dissimilar duration in order to capture diverse appliance power modes. Signature length and sampling frequency allow the profiling layer to avoid inconsistencies between signatures generated with meters at dissimilar sampling frequencies by translating signatures into a standard frequency before storing it in the relational database.

Activity Recognition

Method 1: ANN

Following the profiling phase, the generated signatures are used to train the ANN for the recognition of appliances. The basic element of the ANN is a neuron, which can be represented as a simple succession of mathematical operations, such as weight balancing, sum and an activation function. Each input of a neuron is balanced by a different weight and is then aggregated into an activation function that can be as simple as a step function, or a more complex function such as hyperbolic tangent.

The ANN consists of interconnected neurons, as shown in Fig. 2. It provides a balance between complexity and response time. The first layer consists of input neurons i.e., neurons with one or more inputs connected to external or internal data. The second layer consists of hidden neurons that have inputs connected to the outputs of the first layer and are not in direct relation with inputs and outputs of the ANN. The third layer consists of output neurons that have inputs connected to the outputs of the hidden neurons. Output neurons represent the direct outputs of the ANN. The connections between the layers and neurons can vary. For example, the input layer can be connected to the output of the ANN in order to provide a feedback informing the new input state about the previous output. The system 1 uses a similar feedback mechanism to improve the accuracy by allowing the user to notify the system of an incorrect guess. We now describe the ANN learning process. At the beginning of the learning phase, all weights are random. Weight coefficients of neurons are modified using the data from a training data set, based on combination of signatures. The data and corresponding weight modifications are propagated through the hidden layer and then the output layer. During the training phase, the output signals generated by the ANN are then compared against the desired values, namely the target as given in the training data set. The difference between the target and the output layer of the network is called error delta Δ. This error is then back propagated to the hidden neurons and the input neurons and the weight of each neuron may be modified accordingly. The calculation used to tune the weights is as follows:

Wn=Wo-(A*Lr)

where Wn is the new Weight, Wo is the old weight and Lr is the learning rate set to a constant value to avoid inconsistencies due to rapid weight changes. The back propagation ends when all the layer weights are adjusted and the ANN is the system 1 for another wave of data. The more sophisticated the training data set, the finer the weights will be tuned, to recognize any combination of appliance activation. To improve system usability, the system 1 implements an automatic learning program (ALP) that allows autonomous training of the system. The ALP uses the generated signatures and creates a training data set with all possible combinations of appliance activity, which is then used to tune the neuron weights autonomously.

It is important to determine the correct number of neurons and the configuration of the network. In general, a large number of hidden neurons may result in long training times and a system that may perform extremely well on the training data set but cannot handle unseen data. In contrast, an inadequate number of neurons causes an inability to handle complex combination of appliances and poor results. Although the literature offers several programs to find the optimal number of neurons, we opted for tuning the ANN empirically. A number of trials allowed the identification a saturation point after which the network showed little further improvement. The system 1 adopts 6 input neurons matched with 6 hidden neurons, which performs adequately for the 6 parameters used to generate the signatures.

Finally, an important aspect of the ANN is an activation function, which shapes the output of the neurons. Considering that the output of the network should correspond to one of the profiled appliances, after several empirical trials we adopted a general sigmoid function, which is close to a threshold function with smooth angles to allow some neuron uncertainty in case of errors within the order of 5% or more, to account for small variations of power delivered by the energy provider.

Advantages of the ANN to perform appliance recognition include:

Ability to handle any type of data.

Unnecessary prior understanding of appliance behaviour.

Easy extensibility to higher number of inputs, many types of values or dissimilar kind of data.

Learning process that can be automated for example through additional profiling sensors that can turn on/off appliances remotely.

Ability to learn while running through mechanisms of error feedback from the user; the ability to handle multiple simultaneous appliance states. In contrast, a drawback of the prior ANN approach is the lengthy training process that may take some minutes for profiling of more than 15 appliances, or if some appliances have long signatures e.g. a washing machine with a multi-state signature.

Method 2: Matrix-based probabilistic approach (MP A)

Based on the generated appliance signatures, combinations of all those signatures are calculated and stored in the database under the form of matrices where each column (or row) represents an appliance state. Either periodically or through a parameter delta reporting, the energy monitoring unit sends data to the processing unit, which is then stored in an array. Appliances are identified by calculating a weighted distance between that array of received parameters and the generated matrices.

The advantages of an MP A include: (1) avoiding the appliance recognition training phase; (2) simplified representation and management of appliance signature and recognition; (3) tuneable recognition control based on varaition of weights in the calcualtion of the distance.

Unique Signature State Information Database

An important part of the system 1 is the repository used for storing profiled appliances. Once a new appliance is profiled, the signature, together with some metadata relative to the appliance model and location, are stored locally and duplicated in a remote database. The duplication allows the creation of a common repository of signatures used to train the ANN and share signatures with other users, namely Unique Signature State Information (USSID) database. In fact, providing a common signature repository for multiple sites can progressively reduce the initial profiling phase required by the system 1. The USSID consists of 6 main relational tables. Three main tables, namely captured parameters, physical and environmental relate directly to the signature. As certain types of appliances with either dissimilar models or from various manufacturers are likely to have dissimilar signatures, the table of physical data captures the specifics of the appliance. As the USSID database grows, it is necessary to provide techniques to present to the user an initial standard set of relevant signatures. Through the environmental table the system can provide a common list of appliance signatures based on user location (password protected). It is in fact common to have same appliance models concentrated within the same area or region (e.g. electric showers are very common in Ireland and UK while a certain HVAC model are more common in warmer countries). By using the system 1 interface, the user can then browse the list of appliance models in the area or search for other signatures should the appliance not be in the list. Currently, the USSID system in the system 1 is implemented in an SQL-based relational database. Furthermore, the environmental table provides information of surrounding conditions during measurements as this may affect the signature accuracy. USSED was designed with a broad use in mind such as a large number of signatures generated by contributors. To address multiple contributors for the same signature, the database implements a "contributor" table that includes a confidence rate, which increases according to the reputation of the contributor. We envision that a reputation would increase based on collection of opinions from other users. The USSID system can handle multiple signatures of the same appliance ED according to the "signature property" table. Similar to contributor reputation, the "energy meter" table stores the accuracy of the energy meter, which can be used to tune the appliance recognition algorithm. For example, in the system 1 this would enable testing the meter accuracy and associate accuracy levels to different activation functions.

Signature ID (SID) Device Location

Real Power Temperature

Power Factor Humidity

RMS Current

RMS Voltage

User ID (UID)

Peak Current

Name

Peak Voltage

Confidence Rate

Sampling Rate

Association

Ttmestamp

State: [startup, steady,

shutdown, off]

Meter ID (MID)

Device Type

Accuracy

Type

8¾> - j

Model

Make SID (Primary key)

Power Rating AID

Voltage Rating MID

Frequency Rating UID

Table 1 : Integration of unique signature state information

Automated Data Annotation

The generation of training data used by appliance load monitoring systems to train the machine learning algorithms and to validate the systems' output generally requires user input to annotate what appliance is on at any instance. The manual generation of training data is often replaced by self-generated training data, where combinations of the set of already recorded signatures are produced, as done in the system 1. The results are nevertheless not always accurate since real power patterns may be emergent and different from the simple addition of appliances power signatures. For example, appliances connected to the same line as an appliance drawing high current may suffer voltage fluctuation, thereby generating an overall power signal lower than the one produced from adding each appliance power signature. At a later stage, when the system is ready to perform real-time recognition of appliance load monitoring, its accuracy needs to be verified. Typically, the user verifies the veracity via direct observation, which is not scalable and a tiresome task. Another related use case is for the acquisition of appliances signatures. Human intervention is generally required to control appliances and to indicate when the latter are switched on and switched off. In one embodiment, there is a technique to reduce human supervision by introducing automated data annotation. Appliances signatures can be acquired, real training data can be produced, and the system accuracy can be compared with real appliances' usage, without requiring human intervention. A set of sensor nodes is attached to the appliances or to the outlet to which they are plugged. The system 1 obtains the appliance operating states and marks the data captured at the energy monitor with the appropriate appliances states. Such a technique is not advantageous for long-term appliance activity recognition, as it is intrusive and less scalable. We therefore use it for the system set-up phase, where these issues are less important.

Multiple types of sensory data are used to identify the operating state of an appliance, and include but are not restricted to:

- Temperature

- Light

- Sound

- Vibration

- Humidity

- Magnetic field variations

- Current variations

Once the set-up phase is completed, the extra sensor/actuator nodes are removed and are not needed anymore. They may however be reused at a later stage for system re-calibration and system monitoring.

A. Data Acquisition System

Recent standardization efforts have generated an increasing trend towards the integration of sensor systems in building automation systems, allowing the connection of low-cost sensing and monitoring units and the gathering of energy consumption information in real-time.

Although the system 1 is independent from the communication protocol used by the energy monitor, the unit used for testing transfers data via a ZigBee-based acquisition system to a gateway connected to a local machine, which connects to either a local or remote relational database for storage. The system 1 system resides on the local machine and processes energy data as they arrive from the network. In particular the system 1 is able to firstly generate appliance signatures and then train an ANN to recognize appliance activities in real time. This starts with the appliance profiling phase, a one-off procedure that allows the system 1 to characterize appliances that the user wants to recognize. The profiling will create a set of unique appliance signatures that will then be used for the real-time activity recognition. To keep record of appliance activity times, once an appliance is turned on/off, the system records this into a dedicated table in a remote database.

Appliance Profiling

An advantageous aspect is what parameters will contribute to the generation of a given signature. For example, the real power consumption can discriminate between appliances with dissimilar power consumption but may fail when appliance consumption is similar.

In order to identify the important constituents for a unique appliance signature, we now highlight the main electrical parameters for an appliance working on alternating current (AC). According to its internal circuit, an appliance can be predominantly resistive, inductive, or capacitive. For example a kettle is almost purely resistive while a fan can be predominantly inductive. Inductors and capacitors affect the power consumption by shifting the alternating current with respect to the alternating voltage. In particular, capacitors delay the current with respect to the voltage, while the opposite happens for inductors. Considering that the power is the multiphcation of voltage and current, if voltage and current are shifted, the power transferred to the appliance is less. This effect is captured by the active and reactive power components, which, in mathematical terms, correspond to real and imaginary parts respectively, as shown in Fig. 6. In general, appliances work through the real power (active), while the reactive power (passive) is due to the presence of storage elements in the appliance circuit (inductors or capacitors), does not work at the load and heats wires. Pure resistive appliances show no shift of current and voltage, the reactive part is null and all the power is transferred to the load. In contrast, the larger the current/voltage shift the greater the imaginary component. Reactive and active powers are key parameters to calculate the power factor, which is captured by the energy meter.

Equation 1 reports the relation between the active, reactive and power factor.

S = P +jQPf= P=\S\ (1) where S = Apparent/Complex power, Q = Reactive Power, P = Active Power, Pf = Power Factor, and |S|= real part of the apparent power.

Unique Appliance Signature

Based on the relations between the power components and how they map to appliance types, this section introduces the constituents of a unique appliance signature. The real power is the first important constituent that can discriminate appliances of dissimilar consumption, as shown for example in Fig. 3. To address appliances with similar consumption, the power factor can discriminate between appliances of resistive, capacitive and inductive types. Following, the peak current relates to the appliance circuit specifics, as it represents the maximum amount of energy the appliance allows before reacting. The system 1 collects also RMS current that provides consumption information independently from the voltage given by the energy provider. Finally, peak voltage and RMS voltage relate to the specific voltage provided when the signature is made. Overall, the system identifies 6 constituents to generate a unique appliance signature, which is the base to disCTiminate between multiple appliance activities. Additional factors captured when profiling appliances are the signature length and the meter sampling frequency. These parameters are key to translate signatures from dissimilar types of energy meters into a standard signature. In fact, when profiling appliances, users may generate signatures of dissimilar duration in order to capture diverse appliance power modes. For example, an electric oven presents an initial period of almost constant current draw followed by periodic deactivations when the set temperature is reached, as shown in Fig. 3. Finally, to avoid inconsistencies between signatures generated with meters at dissimilar sampling frequencies, the system 1 implements a simple function that translates signatures into a standard frequency before storing it in the relational database. Figure 4 shows power signatures for four appliances of different lengths and standard sampling frequency of 1 value per minute. In order to account for the aging of an appliance, each unique signature has an associated expiration date. In the case of a signature, which is expired or about to expire, the system can inform the user of the necessity of renewing a signature either manually or automatically. Furthermore, the system can also identify the appliance aging process by looking at parameter variations over time. Should the system identify repetitive anomalous signature variations, it may require a new signature to be generated.

Training and Recognition We now describe the ANN learning process in more detail. At the beginning of the learning phase, all weights are random. Weight coefficients of neurons are modified using the data from a training data set, based on a combination of appliance signatures. The data and corresponding weight modification propagated through the hidden layer and then the output layer. During the training phase, the output signal generated by the ANN is then compared against the desired value, namely the target, as given in the training data set. The difference between the target and the output layer of the network is called "error _". This error is then back propagated to the hidden neurons and the input neurons and the weight of each neuron may be modified accordingly. Equation 2 provides the calculation used to tune the weights.

Wn = Wo - (h *Lr) (2) where Wn is the new Weight, Wo is the old weight and Lr is the learning rate sets to a constant value to avoid inconsistencies due to rapid weight changes.

The back propagation ends when all the layer weights are adjusted and the ANN is ready for another "wave" of data. Naturally, the more sophisticated the training data set, the finer the weights will be tuned to recognize any combination of appliance activation. To improve system usability, the system 1 implements an automatic training program (ALP) that allows autonomous training of the system. ALP uses the generated signatures and creates a training data set with all possible combinations of appliance activity, which is then used to tune the neuron weights autonomously.

It is important to determine the correct number of neurons and the configuration of the network. In general a large number of hidden neurons may result in long training times and a system that may perform extremely well on the training data set but cannot handle unseen data. In contrast, an inadequate number of neurons cause the inability to handle complex combination of appliances and poor results. Although the literature offers several programs to find the optimal number of neurons, we opted for tuning the ANN empirically. A number of trials allowed the identification a saturation point after which the network showed little further improvement. The current version of the system 1 adopts 6 input neurons matched with 6 hidden neurons, which performed adequately for the 6 inputted parameters used to generate the signatures.

Finally, an important aspect of the ANN is the Activation Function, which shapes the output of the neurons. Considering that the output of the network should correspond to one of the profiled appliances, after several empirical trials we adopted a general sigmoid function, which is close to a threshold function with smooth angles to allow some neuron uncertainty in case of errors within the order of 5% or more e.g., to account for small variation of power delivered by the energy provider. The system 1 in this manner generates comprehensive information concerning current energy consumption in a binding, breaking down for the user the consumption pattern, which may be as illustrated in Fig. 3.

The following are some advantages and features of the system of the invention in various embodiments:

- Using electric parameters that constitute a unique appliance signature.

- Using a wireless energy monitoring unit to gather those parameters.

- Storing the signatures in a distributed database.

- Automatic neural network training based on downloading standard appliance signatures from such a database according to geographic location of the user.

- Adding more signatures into the system, therefore, real-time appliance recognition activity that can be used not only for energy consumption estimation but also for several applications such as real-time energy peak levelling, user activity recognition, and faulty appliance detection.

- Automated data annotation to reduce human supervision in the procedure, and to increase system acceptance and utilization from non-technical users.

Tests (Figs 4(a) to 4(c)

Referring to Fig. 4 plots illustrate parameters for some tests.

A user profiles an that shows both similar power consumption and power factor characteristics as other existing signatures. We identified an electric fire as a resistive appliance with similar consumption of microwave and kettle. The electric fire was introduced in the kitchen and its signature was generated. The objective of this experiment was to test whether the system 1 could still distinguish appliances based on the remaining signature parameters. Fig. 4a shows the real appliance activity, which was manually annotated and used for comparison against the output from the system 1. Fig. 4b shows the raw output from the ANN in response to the input energy data. An appliance is considered active if the output from ANN is equal or greater than 1. The output coming from the system is then passed through a filter module, as shown in Fig. 4c, which converts the raw output into actual appliance activity, display the results on the user interface and annotate the activity into a dedicated table in the database. We collected information over 65 minutes while many low consuming un-profiled appliances where also activated occasionally. In contrast, we never activated the electric fire for the entire duration of the experiment to simulate the worst-case scenario of having an unknown high consuming appliance. By doing this, the system could not reinforce its neurons by comparing the heater against other appliances activity and learn the difference. In spite of this adverse condition, the experiment demonstrated how the system guessed correctly 55 times over a total of 65, which corresponded to an accuracy greater than ~84%.

In order to test system scalability, we performed tests regarding the time needed to train the ANN with a various number of appliance signatures. Average training time showed a large dependence on the machine used and the length of signatures. However, we can report with confidence that, for a regular Pentium 4 machine, the training remained below 1 minute for up to 15 appliances while showing a much larger increase when more signatures where introduced. Based on this, a limit of the system 1 is the inadequacy to recognize a great number of appliances in a building. Instead, it should be trained to recognise a subset of appliances activity at a time in the order of 15 or less. Finally, the presented evaluation demonstrated how the system 1 could guess with high level of accuracy appliances that have an almost constant level of energy consumption, while further tests are needed to understand the performance for appliances with dissimilar power and changing power consumption due to power cycles.

With respect to automated data annotation for system training and validation, we performed tests to identify what sensor information can be used to infer appliance activity. An example measurement is given in Fig. 5. The results show that appliance activity can be discriminated using simple sensors. We conducted further tests and achieved over 90% accuracy in detecting a printer and desktop computer activity. Using extra energy monitors during the set-up phase provides 100% accuracy but is more expensive. The integration of annotated data into the system 1 system to automate signature acquisition, and both system training and validation, consists of associating real-time information on appliances' states to the energy data, so that the system 1 can automatically capture electricity signatures, train its neural network, and can compare the neural network output to real appliance activity to validate the system output. The generation of appliance activity states follows the process depicted in Fig. 8. Appliance filters are created for each appliance, and compute the appliance activity output based on the right combination of sensor activity outputs. This processing can be either done at the sensor node or at any other machine. When the processing is done on the sensor node, the appliance state information is sent to a machine (e.g. the one where the system 1 runs), as a form of either an XML file, or a flow of data on the serial line delimited by well-defined packet headers, or written in a database. Making the appliance state information available as a form of a string of bytes on a serial port or any other communication media has the advantage that the annotation system does not require any software support on any machine. Energy monitoring systems just need to read the serial port and get the information (e.g. no need of extra software to write appliances states into an XML file or a database). Appliances filters are either hard-coded in software, or selected by the user, such as with buttons or via a user interface.

Referring to Fig. 9, during training, appliances are activated and deactivated individually and signature vectors are taken and stored in a training set. The system 1 adopts 6 input neurons matched with 6 hidden neurons, which performed adequately for the 6 inputted parameters used to generate the signatures: apparent power, real power, power factor, peak current, rms current, peak voltage. For the Activation Function, the system 1 adopts a general sigmoid function, which is close to a threshold function with smooth angles to allow some neuron uncertainty in case of errors within the order of 5% or more e.g., to account for small variation of power delivered by the energy provider.

As shown in Fig. 10, once a set of appliances is profiled, the training set is then used to tune the weights internal to the neural network. The weights are an internal part of the neural network relative to each neuron.

Referring to Fig. 11, parameters from the meter arrive periodically approximately every 5 seconds and intervals TO, Tl, T2...Each sample arriving from the meter is fed as an input to the trained ANN. The ANN calculates the combination of appliances closest to the input values and provides a sequence of appliance IDs as an output.

The invention is not limited to the embodiments described but may be varied in construction and detail.

Claims

1. An energy consumption monitoring system comprising:

an electrical sensor,

wherein the processor is adapted to:

2. An energy consumption monitoring system as claimed in claim 1, wherein the processor is adapted to guide user inputs to assist with generating the profiles.

3. An energy consumption monitoring system as claimed in claims 1 or 2, wherein at least some profiles identify the associated appliance as one or a combination of resistive, inductive, or capacitive components.

4. An energy consumption monitoring system as claimed in claim 3, wherein at least some profiles identify predicted variation patterns for said components.

5. An energy consumption monitoring system as claimed in claims 3 or 4, wherein at least some profiles include parameter values for real power, reactive power, and power factor.

6. An energy consumption monitoring system as claimed in claim 5, wherein at least some profiles additionally include parameter values for peak current, RMS current, peak voltage, RMS voltage, and sampling frequency.

7. An energy consumption monitoring system as claimed in any preceding claim, wherein the artificial intelligence functions includes a neural network.

8. An energy consumption monitoring system as claimed in claim 7, wherein the neural network comprises an activation function in the form of a sigmoid function for training according to the profiles.

9. An energy consumption monitoring system as claimed in any preceding claim, wherein the processor is adapted to use a combination of sensory inputs to process energy readings enabling signature acquisition, system setup automation, system calibration, system monitoring, system control, and system optimization.

10. An energy consumption monitoring system as claimed in claim 9, wherein the processor is adapted to generate a user interface for profiling appliances, and to augment reliability of user inputs during system profiling

11. An energy consumption monitoring system as claimed in any preceding claim, wherein the processor is adapted to autonomously refine the system.

12. An energy consumption monitoring system as claimed in claim 11, wherein the processor is adapted to implement a control loop that feeds back appliance momtoring accuracy data to refine parameters of the artificial intelligence functions.

13. An energy consumption monitoring system as claimed in any preceding claim, wherein the processor is adapted to control the capture of power data including appliance start and stop activity to autonomously generate appliances signatures.

14. An energy consumption monitoring system as claimed in any preceding claim, wherein the processor is adapted to recalibrate an external appliance load monitoring system.

15. An energy consumption monitoring system as claimed in any preceding claim, wherein the processor is adapted to perform online profiling and processing of appliance data.

16. An energy consumption monitoring system as claimed in any preceding claim, wherein the processor is adapted to compute an appliance activity based on the sensor inputs, in which filters are associated with appliances, and to return a positive appliance activity when the combination of expected inputs is verified.

An energy consumption monitoring system as claimed in any preceding claim, wherein the processor is adapted to make appliance state information available on a communication medium, thereby avoiding need for software support on a host system, and reducing the provision of annotated data.

An energy consumption monitoring system as claimed in any preceding claim, wherein the processor is adapted to operate as an online or offline energy recommender system, which breaks out energy costs to per-appliance costs, and recommends appliances' replacements to reduce users' costs.

An energy consumption monitoring system as claimed in any preceding claim, wherein the processor is adapted to group energy users into groups of users.

An energy consumption monitoring system as claimed in claim 19, wherein the processor is adapted to group users having similar energy usage profiles, for example with the objective of making the groups of users appear as a single entity for energy providers, and allowing specific group energy tariffs.

21. An energy monitoring system as claimed in any of claims 7 to 20, wherein the processor is adapted to perform the steps of:

tune weights within the neural network; and

22. A computer program product comprising a computer usable medium having a computer readable program code adapted to perform processor operations of a system of any preceding claim when executing on a digital processor.