GB2488164A

GB2488164A - Identifying electrical appliances and their power consumption from energy data

Info

Publication number: GB2488164A
Application number: GB1102855.2A
Authority: GB
Inventors: Pias Marcelo
Original assignee: GLOBOSENSE Ltd
Current assignee: GLOBOSENSE Ltd
Priority date: 2011-02-18
Filing date: 2011-02-18
Publication date: 2012-08-22
Also published as: GB201102855D0

Abstract

In a system comprising a plurality of different types of powered device or appliance, aggregate energy usage data for the system is obtained from a sensor, and an individual device is identified by comparing the aggregate energy usage data with predetermined energy profiles of known devices. An estimated level of energy consumption attributable to the device can also be inferred. The sensor may be a non-intrusive load monitor e.g. a clamp sensor around a cable or a plug-through socket sensor. The electrical appliance may be identified by training a hidden Markov model classifier on known appliances, and operating the trained model on Fourier transformed power data. The profile may be a power usage signature. Feedback may be provided to a user of the appliance to reduce his energy consumption. The measured power data may be sampled at different rates according to the types of device connected to the system.

Description

TITLE: METHODS AND APPARATUS FO********MONOQ POWER CONSUMPTION

DESCRIPTION

The present invention relates to methods and apparatus for monitoring power consumption and particularly but not exclusively to non-intrusive load monitoring (NILM).

Recent UK Government reports indicate that energy demand is rising, natural resources are limited, and renewable green energy sources are far short of meeting current energy needs. The UK, like many other countries, spends over two-thirds of its total energy on residential households and non-residential sectors (business) excluding only transport.

Energy saving in these sectors is seen as an effective solution, which not only reduces demand, but also lowers carbon emissions.

Quantitative measures for energy savings are both essential and timely for the government and general public. In the UK, the government is committed to raise the green agenda profile with carbon reduction across central Government buildings within the next five years.

While people are usually keen to reduce their energy consumption, their perceived inability to measure and control their energy usage remains an obstacle.. A recent Carbon Trust survey indicated that 85% of people are keen to reduce their carbon footprint at work but are not adequately equipped, supported, or informed as how to do so.

Energy measurement is also important to utilities companies who require information on appliance-level energy usage of households and office spaces for forecasting fifture electricity arid gas demand, and to support customer retention programmes. The central and local government similarly rely on such information to planning expansion of national power grids, and to support long-term strategies including, for instance, the roll-out of.srn art meters in the UK, and preparation for a smart grid' infrastructure.

A range of commercial consumer devices is available that can either monitor the energy consumption for the entire building (aggregate-level energy data) through low-power clamp-on sensors (e.g. current transformer) or at the individual power socket with tow-power plug-through sensor (e.g. shunt resistor) devices (power socket-level data).

However, these systems are either limited in the information they are able to provide or (in the case of plug-through sensors) expensive to implement for a whole building.

U.S. 4,858,141, US 2010/0292961 and US 2010/0289643) disclose methods relying on ON/OFF power states of appliances in order to determine power usage. However, many of today's electric and electronic appliances exhibit more than said power states including (but not limited to) stand-by, power savings and power shutdown.

tn accordance with a first aspect of the present invention, there is provided a method of monitoring power consumption in a system comprising a plurality of different types of powered device, comprising: receiving aggregate energy usage data for the system, the aggregate energy usage data including a plurality of energy usage measurements for the system recorded sequentially over a period of time; determining from the aggregate energy usage data an estimated level of energy consumption attributable to a first type of device of the system by comparing the aggregate energy usage data with a pre-stored energy profile corresponding to the first type of device; and providing an output indicating the estimated energy consumption of the first type of device.

In this way, an automated method may be provided for identifying power consumption attributable to a particular type of powered device in a system. The aggregate energy usage data may relate to any type of energy supply, e.g. electrical or gas supplies.

The system may comprise any group of devices powered by a common power supply, e.g. system of powered devices in a building or other structure.

In one embodiment the energy usage data has a sampling rate selected to allow matching between the aggregate energy usage data and the pre-storcd energy profile. In one embodiment, the energy usage data has a sampling rate above 1 lIz (e.g. above 10 Hz). In this way, the method may be used to detect electronic equipment having a relatively high frequency profile. In another embodiment, the energy usage data has a sampling rate below 1 Hz. In this way, the method may be configured to detect electrical appliances such as fridges and cookers which have a relatively low frequency profile.

hi one embodiment, the determining step comprises determining from the aggregate energy usage data the identity of the first type of device by comparing the energy usage data with the pre-stored energy profile of the first type of device (e.g. prior to determining the estimated level of energy consumption attributable to the first type of device).

In one embodiment, the determining step comprises identifying the presence of a contribution from the first type of device in the aggregate energy usage data by identifying a candidate energy profile for the first type of device (e.g. based on the pre-stored energy profile of the known device). The determining step may further comprise: comparing the candidate energy profile with the pre-stored energy profile of the known device; and determining the degree of resemblance between the candidate and pre-stored profiles. In one embodiment, the output indicating the estimated energy consumption of the first type of device is provided only if the degree of resemblance exceeds a predetermined threshold value.

In one embodiment, the aggregate energy usage data is obtained at a point (e.g. single.point) through which power passes prior to reaching the plurality of devices. In one embodiment, the aggregate energy usage data is obtained at a point near a main power supply to the system (e.g. using a clamp-on sensor).

In one embodiment, the aggregate energy usage data is received over a communications link (e.g. over a communications network or dedicated wireless/wired communications link).

In one embodiment, the output is provided over a communications link (e.g. over a communications network or dedicated wireless/wired communications link).

In one embodiment the pre-stored energy profile is a generic energy profile corresponding to a typical profile for the first type of device.

In another embodiment, the pre-stored energy profile is a specific recorded energy profile for the first type of device. The specific energy profile may be recorded in a learning step in which an energy profile for the first type of device in the system is recorded and labelled.

In one embodiment, the method further comprises the step of filtering the aggregate energy usage data (e.g. prior to the determining step).

In one embodiment, the method further comprises the step of dividing the aggregate energy usage data into segments (e.g. segments of time or segments based on a threshold (e.g. power threshold)).

The determining step may comprise at least one of: time domain analysis; frequency domain analysis; and time-frequency domain analysis.

The energy profile for the first type of device may comprise a power usage signature, In one embodiment the power usage signature corresponds to power usage whilst the device is in operation. in another embodiment, the power usage signature corresponds to patterns of switching between different power usage levels of the device.

In one embodiment the determining step comprises analyzing a sequence of switching events. In one embodiment, one or more of the switching events in the sequence may be switching events associates with other types of device in the system.

In one embodiment, the determining and providing steps occur in substantially real time (e.g. to provide a continuous output).

lit one embodiment, the method further comprising the step of providing feedback on power usage. The feedback may be provided to a user via a communications device (e.g. portable communications device).

In one embodiment, the feedback includes a usage comparison with.a reference usage level.

In one embodiment, the feedback includes an efficiency comparison with a reference efficiency level.

The method may further comprise repeating the determining and providing steps for at least a second device of the system using a pre-stored energy profile corresponding to the second device of the system.

In accordance with a second aspect of the present invention there is provided, a computer program comprising program instructions for causing a computer to perform the method of the first aspect of the invention, The computer program may be embodied on one or more of a record medium; a computer memory; a read-only memory; a carrier signal (e.g. electrical carrier signal).

In accordance with a third aspect of the present invention, there is provided apparatus comprising an appliance recognition device configured to perform the method of the first aspect of the invention.

In one embodiment, the appliance recognition device comprises a memory device for storing instructions for carrying out the method of the fourth aspect of the invention and a processor for processing the instructions stored in the memory device.

In accordance with a fourth aspect of the present invention, there is provided a method of identifying a powered device, comprising: receiving energy usage data associated with the device, the energy usage data including a plurality of energy usage measurements for the device recorded sequentially over a period of time; determining from the energy usage data the identity of the powered device by comparing the energy usage data with a pre-stored energy profile of a known device type; and providing an output indicating the determined device type.

In this way, an automated method may be provided for identifying a device type of powered device. The energy usage data may relate to any type of energy supply, e.g. electrical or gas supplies.

In one embodiment the energy usage data has a sampling rate selected to allow matching between the energy usage data and the pre-stored energy profile. In one embodiment, the energy usage data has a sampling rate above 1 Hz (e.g. above 10 Hz). in this way, the method may be used to detect electronic equipment having a relatively high frequency profile. In another embodiment, the energy usage data has a sampling rate below 1 Hz. In this way, the method may be configured to detect electrical appliances such as fridges and cookers which have a relatively low frequency profile.

In one embodiment, the determining step comprises identifying a candidate energy profile for the device from the energy usage data (e.g. based on the pre-stored energy profile of the known device). The determining step may further comprise: comparing the candidate energy profile with the pre-stored energy profile of the known device; and determining the degree of resemblance between the candidate and pre-stored profiles. In one embodiment, the output indicating the type of device is provided only if the degree of resemblance exceeds a predetermined threshold value.

In one embodiment, the energy usage data is device specific energy usage data (e.g. relating substantially to a single device to be identified). For example, the energy usage data may be obtained at a power supply point where the power measured is the power supplied only to the device (e.g. using a plug-through sensor at a power socket to which the device is exclusively connected).

In another embodiment, the energy usage data is aggregate energy usage data (e.g. relating to a system comprising a plurality of devices powered by a common power supply).

In one embodiment, the aggregate energy usage data is obtained at a point (e.g. single point) through which power passes prior to reaching the plurality of devices, In one embodiment, the aggregate energy usage data is obtained at a point near a main power supply to the system (e.g. using a clamp-on sensor).

In one embodiment, the energy usage data is received over a communications link (e.g. over a communications network or dedicated wireless/wired communications link).

In one embodiment, the output is provided over a communications link (e.g. over a communications network or dedicated wireless/wired communications link). Typically the output will be provided to a visual display of a communications device (e.g. display screen of a desktop computer or portable communications device).

in one embodiment the pre-stored energy profile is a generic energy profile corresponding to a typical profile for the known type of device.

In another embodiment, the pre-stored energy profile is a specific recorded energy profile for the device. The specific energy profile may be recorded in a learning step in which an energy profile for the device is recorded and labelled.

Tn one embodiment, the determining step further comprises the step of filtering the energy usage data (e.g. prior to the comparing step).

In one embodiment, the determining step further comprises dividing the energy usage data into segments (e.g. segments of time or segments based on a threshold (e.g. power threshold)).

The determining step may comprises at least one of time domain analysis; frequency domain analysis; and time-frequency domain analysis.

The energy profile for the known type of device may comprise a power usage signature. in one embodiment the power usage signature corresponds to power usage whilst the device is in operation. In another embodiment, the power usage signature corresponds to patterns of switching between different power usage levels of the device.

In one embodiment, the determining step comprises analyzing a sequence of switching events. In one embodiment, one or more of the switching events in the sequence may be switching events associates with other types of device in the system.

In one embodiment, the method further comprises: determining from the energy usage data an estimated level of energy consumption attributable to the device; and providing an output indicating the estimated energy consumption of the device. In one embodiment, the determining and providing steps occur in substantially real time (e.g. to provide a continuous output).

In one embodiment, the method further comprising the step of providing feedback on power usage. The feedback may be provided to a user via a communications device (e.g. portable communications device).

In one embodiment, the feedback includes a usage comparison with a reference usage level.

In accordance with a fifth aspect of the present invention there is provided, a computer program comprising program instructions for causing a computer to perform the method of the fourth aspect of the invention.

The computer program may be embodied on one or more of: a record medium; a computer memory; a read-only memory; a carrier signal (e.g. electrical carrier signal).

in accordance with a sixth aspect of the present invention, there is provided apparatus comprising an appliance recognition device configured to perform the method of the fourth aspect of the invention.

In one embodiment, the appliance recognition device comprises a memory device for storing instructions for canying out the method of the fourth aspect of the invention and a processor for processing the instructions stored in the memory device.

Embodiments of the present invention will now be described by way of example with reference to the accompanying figures, in which: FIG. IA and FIG. lB illustrate a residential household energy monitoring system.

The house has a range of typical electric and electronic appliances under the user's control; FIG. 2A and FIG.. 2B illustrate a typical non..residential building (e.g. corporate office, manufacturing facility, government building) which has its energy consumption monitored by a energy monitoring system. The diagram shows typical office equipment; FIG. 3 illustrates one embodiment of the said apparatus, which forms a sustainable end-to-end sensing, and feedback loop that is unique to this invention; FIG. 4 is a flow chart of an exemplary operation of the present invention method for automatically recognising appliances from either aggregate-level energy data or socket-level energy data; FIG. 5 is a chart graph of an exemplary identification of appliances for a three-bedroom family house; FIG. 6 is a flow chart of an exemplary operation of the present invention method with more detailed technical description of machine learning components; FIG. 7 illustrates an example of mathematical tools that may be employed in some implementations of the present invention; FIG. 8 is a flow chart of an exemplary operation of the machine learning process involved for building a database and classification model of known appliances from previously collected energy sensor datasets; and FIG. 9 illustrates an example of mathematical tools that may be employed to create a database of known appliances by learning from existing energy datasets.

The present invention seeks to address a key challenge in the energy sector that is carbon emission reduction through more effective utilization of energy sources and energy saving.

*This invention focuses directly on the energy use that is human-related (i.e. is attributed to user behavior and thus under user control).

Bringing energy awareness and ownership to residential householders and stakeholders in non-residential buildings with personal feedback and real-time information for individual appliances and office equipment permits users to reduce their energy consumption with smaIl behavioral changes.

Empowering the user with detailed energy information feedback capabilities of individual appliances brings potentially huge benefits: Fine-grained energy monitoring of individual electric and electronic appliances; o Bring awareness arid ownership of the energy challenge to influence user habits and behaviours in a positive way towards energy savings and carbon reductions; o Develop feedback ioops to connect data about energy use of appliances with business decision-making; o Introduce efficient power monitoring of individual appliances for optimization of equipment operation, for instance, duty cycling of ICT equipment in office spaces; automatically breakdown the aggregate-level energy data into the various types and classes appliances that contribute to the overall energy usage.

The present invention relates generally to energy saving, power consumption monitoring of appliances. The innovation directly relates to users behavioural change which has been shown to contribute to 30% of total energy consumption in buildings and close to 100% of consumption at households. In one described embodiment, a sustainable end-to-end invention is presented that is comprised of a set of machine learning data processing algorithms at its core. These algorithms transform the raw sensor data into meaningful actionable knowledge that can be presented to users, form part of key reports for building managers, or delivered to senior managers (and policy makers) in support of improved

decision making (and policy specification).

While people are usually keen to reduce their energy consumption, their perceived inability to measure and control *their energy usage remains an obstacle. The present invention seeks to address this problem by allowing sources of energy waste (large or small) to be identified and managed by the user with the help of personalized feedback systems. In this way, to the user may be motivated and empowered to take good care of the appliances and equipment that they have control of such as their house's appliances (e.g. fridge, lights, TV sets) and office/workspace equipment (e.g. computer, display, laptop, office light). In one embodiment, the invention comprises a novel apparatus and method that uses machine learning to automatically recognise, identify and classify from raw energy sensor data the type and class of an appliance, its energy consumption and frequency of usage. The apparatus and method can automatically break-down the energy consumption data of an entire building or area of said building into the individual electric and electronic appliances that contribute to the entire energy consumption. The invention may comprise an automated machine learning process for analyzing patterns in the energy data to infer what appliances are being used and their individual energy consumption in residential households and non-residential buildings (e.g. corporate, manufacturing facilities).

One application of the present invention is to automatically recognize appliance types and classes (e.g. kettle, shower pump, fan oven, game console, TV, fridge, washing machine) the processing of aggregate-level electricity consumption data provided by a low-power clamp-on energy sensor device (e.g. based on a current transformer) that is monitoring an entire house, Another application is the automated recognition of the appliance type and class (e..g whether it is a TV set) from the energy data provided by a plug-through energy low-power sensor device (e.g. based on a shunt resistor) that said appliance is connected to.

In one embodiment, the apparatus comprises a sensing component at a location where environmental sensing and data collection takes place, a server component which hosts the invention computations and data analytics, and a feedback system that may comprise one or more interfaces for user engagement and behavioural change. These components are networked using a communication.media such as wireless radio; forming a web of things framework that interconnects all mentioned apparatus components in a transparent manner.

in some implementations of the invention, part or the entire computation and data analytics may be performed in a system component close to the source of the sensor data, for example, in a mobile communication device (e.g. smartphone, iPhone), tablet computers (e.g. Android, iPads).

This invention may employ a low-power networked sensor data collection system that measures energy consumption (electricity or gas) and communicates the sensor measurements wirelessly (e.g. WiFi, ZigBee, 6LoWPAN) or wired (Ethernet or Power Line Communication). The energy-related data (e.g. average and instantaneous real power, apparent power, power factor, instantaneous and RMS voltage and current) may be collected from either clamp-on sensors (aggregate-level) or plug-through socket sensor devices (appliance-level). The system may be capable of adapting its sensors sampling rate (low to high) in order to achieve the data resolution needed for the identification and recognition process. High sampling rates (e.g. 15Hz) are suitable for electronic equipment including (but not limited to) TV sets, video game consoles, desktop and laptop computers.

Lower sampling (<1Hz) can be used for electric appliances, for instance, fridges and cookers, The present invention provides an argument in support of cheap energy measurement at the aggregate points as opposed to expensive individual appliance monitoring at the power sockets. Whenever individual appliance monitoring via plug-through sensor device is necessary the invention method allows for automatic identification of the appliance class at the power socket.

The inferred appliance type/class and its energy consumption have a number of benefits to the users including (hut not limited to): (i) it demonstrates that the time-series energy information per appliance enables users to engage and become aware of their activities throughout the day, and (ii) it allows users to monitor the performance and efficiency of their appliances and make informed decision whether to replace any of them it in the near future; (iii) to make energy (e.g. in kWh or Kg C02) information that is related to users' house and office appliances visible through real-time personalised feedback on mobile phone, email and web portal, (iv) to encourage users to engage in group-based energy saving activities through *a web portal; (v) to support decision-makers in non-residential building to make informed decisions regarding their energy usage policies; (vi) to support the local and central government in planning expansion of their national power grid towards a smart grid'.

The highlighted apparatus forms a sustainable end-to-end sensing and feedback ioop that is unique to this invention. At the centre of this loop lies the user (human) element which represents the focal point of the system for impact and end result. Unlike traditional systems or related inventions, the composition of apparatus components (i) addresses a wide range of stakeholders including but not limited to the building users, building managers, bill payers, and environmental managers, (ii) is self-reinforcing and sustainable such that it forms an ecosystem, referred to as the web of things', in integrating and connecting apparatus components to enhance the invention results beyond the aggregate impact of individua.l components.

The aforementioned ecosystem is supported by a set of rich, innovative, context-oriented, and dynamic information that is intelligently extracted from raw data (as described below).

Detailed description of the drawings

The present invention provides an elegant means of automatically recognising types and classes of appliances and their energy consumption from raw energy sensor data. The apparatus and method use pattern recognition and machine learning to (1) automatically break-down the energy consumption data of an entire building or area of said building into the various types arid classes of electric and electronic appliances that contribute to the entire energy consumption; (2) automatically recognise the type and class of an appliance from the energy data collected via the power socket that said appliance is plugged to.

FIG. IA and FIG 1 B illustrate an energy (e.g. electricity or gas) monitoring system within a residential household. The energy data is collected through a damp-on sensor which is a part of a device ill capable of measuring various parameters for electricity or gas consumption at the utility meter Location. This data is referred as aggregate-level energy data as it contains useful infonnation (for example patterns and signature of power cycles) about all the appliances and equipment that consume energy within the residential building.

Such aggregate-level data is collected at a measurement point afler the main power supply of the house 105. The energy monitoring sensor device 111 comprises a processing unit (e.g. a microcontroller, ARM-based processor), on-board memory (e.g. flash), sensors (e.g. current transformers) and a network communication interface which can be wireless (e.g. WiFi, ZigBee, 6LoWPAN) or wired (Ethernet, Power Line Communication). The device is capable of on-board real-time processing, and communication of raw energy sensor data (raw timestamp sample data) or derived energy-related information such as root-mean- *square (RMS) of voltage, current and power values. The device runs a real-time operating system, for instance, an Embedded Linux or straight embedded C code.

In some implementations, the energy monitoring may also happen at the power sockets by means of plug-through cnergy monitoring devices 110,112. This type of data is referred here as socket-level energy data, In such scenarios, it is also convenient to automatically identify the type of appliances that are connected to plug-through devices 110,112. For example, whenever a user plugs their TV set to said device, a real-time on-device implementation of the present invention may automatically detect whether it is a TV and what characteristics it possesses.

The devices 110,112 are comprised of a processing unit (e.g. a microcontroller, ARM-based processor), on-board memory, sensors (e.g. shunt resistor), actuators (e.g. power switches and relays) and a network communication interface which can be wireless (e.g. WiFi, ZigBee, 6LoWPAN) or wired (Ethernet, Power Line Communication).

Examples of energy measurements, in particular electricity-related, are instantaneous and average real power (in Watts), apparent power (in VA), reactive power (in Var), power factor, instantaneous and RIVIS current (in Amps) and voltage (in Volts), Examples of energy measurements for a gas supply may include instantaneous usage and sector area pressure.

The aggregate-level energy data provided by the clamp-on energy sensor device II 1 presents signature and patterns of usage for all the electric and electronic appliances and equipment in the house. The said data shows a combination of patterns and signatures that may not be easy to distinguish by visual data inspection but can be accurately identified and characterized by the present invention. In one embodiment, the aggregate-level energy data contains patterns for power cycles of electric appliances such as fridge 130, cooker and fan oven 120, washing machine 170, lights 140,141,142; and electronic devices including a desktop computer/display 160, TV 190 and a video game console 180.

The individual socket-level energy data provided by plug-through sensor devices contains power cycle patterns for the individual appliance connected to the plug-through device 110,112. In some implementations, the present invention accurately may recognise and identify that the appliance connected to the plug-through sensor device 110 is a washing machine 170.

The user 150 is at the centre of the monitoring operations with real-time feedback on the energy consumption of the said identified appliances. In one embodiment, the user receives tailored feedback via their mobile phone (e.g. iPhone, Android, smartphones), via SMS text messaging, a tablet computer or via a web portal.

FIG. 2A and FIG. 2B illustrate an energy monitoring system for non-residential buildings such as business workspaces (corporate, government and manufacturing facilities). In one embodiment, the user 250 may be an employee of a company or government department, the building facilities manager, the ICT manager or a decision-making manager. Similarly to the residential household system presented in FIG. 1B, the energy monitoring system for non-residential buildings is capable of collecting aggregate-level energy data for a particular area of a building (e.g. an office's floor or corridor). This may be achieved through clamp-on energy monitoring sensor devices 211 installed at the power circuit boards 205 of the monitored area. In some implementations, the energy consumption of individual office equipment including (but not limited to) a desktop computer 220, a computer display unit 221, mobile phone charger 222 may be monitored through a plug-through sensor device 210. In some other implementations, the existing automated metering and sub-metering infrastructure installed at the building may be used to provide the aggregate-level energy data.

The present invention accurately and automatically recognises the various electric appliances and electronic devices connected to the power sockets at the monitored area (e.g. individual users' desks and/or shared common areas). The invention is capable of processing aggregate-level energy data (e.g. power circuit board monitoring) or individual socket-level energy data. In some scenarios, electronic appliances comprise desktop computers 220, computer display units 221, mobile phone chargers 222, desk lamps 223 and offlce lights 270. In one embodiment, there is a common area where ICT office equipment may be shared among users. Examples of shared ICT equipment include (but not limited to) photocopiers 240, fax machines 230, and printers 260. in some scenarios, electric appliances in shared spaces such as fridge, microwave oven and water cooler in shared kitchens may be monitored via the said system.

The present invention promotes the vision that bringing energy awareness and giving ownership of the energy challenge to the building users should influence user habits and behaviours in a positive way towards energy savings and carbon reductions. The invention brings energy awareness with the energy consumption monitoring of the individual appliances that are under the user's 250 control. The real-time feedback for the recognised appliances empower the user and motivate them to take good care of the appliances and office equipment that they have control of (e.g. their computer, display, laptop, office light, desk fan, personal heater, mobile phone battery chargers, ete). The energy health state' of said appliances can be monitored allowing a user to minimise energy usage with simple yet sustainable actions (e.g. switching off equipment before meetings or lunch break).

FIG. 3 illustrates one embodiment of the said apparatus, which forms a sustainable end-to-end sensing, and feedback loop that is unique to this invention. At the centre of this loop lies the user 305 (human) element which represents the focal point of the system for impact and end result. The web of things' WoT ecosystem is an end-to-end user-centric and privacy-compliant system comprised of web services including real-time energy sensing 310, energy data inference 320 and feedback 330.

The apparatus comprises of a sensing component where environmental sensing and data collection takes place 310, a server component which hosts the inventions computations and data analytics 320, a feedback system that may comprise one or more interfaces for user engagement and behavioural change 330, and a communication media such as wireless radio which interconnects all said apparatus components. In some implementations, the use of TPv6 (e.g. IP61oWPAN) for communication gateways supports the Web of Things (WoT) where secure web services are used to access individual sensor/actuator nodes.

FIG. 4 is a flow chart of an exemplary operation of the present invention machine learning method for automatically recognising appliances from either aggregate-level energy data or individual socket-level energy data. The first step 410 is to collect energy sensor data (e.g. real power, apparent power, power factor, voltage and current). Data collection is achieved through energy monitoring devices 111,110,112, 211, 210. Such energy data may be provided in the form of statistics, for example, mean, standard deviation and root-mean-square (RMS) of power, voltage or current. This data is fed into an inference engine for processing through machine learning algorithms 420. The result of such a processing is a list of the recognised appliances present in the energy data 430. In some implementations, the present invention method will be implemented on-board of energy sensor monitoring devices (e.g. plug-through devices 110,210). This is to support on-appliance real-time display feedback to the user 150, 250. in other implementations, the method may be implemented in a processing server machine that can be located either within the premises of the monitored building, or externally (e.g. in the cloud or in a rented server space).

FIG. 5 is a chart graph of an exemplary identification of appliances related to aggregate-level energy usage for a three-bedroom family house. The energy data was collected via a wireless clamp-on energy monitoring sensor device installed at the house's utility meter box. It shows the matching of automatically recognised house's appliances to the raw energy data on typical weekend day (e.g. Saturday). These results are achievable through the machine learning process exemplified in FIG. 4. The list of recognised appliances include fudge 130, shower pump, kettle, fan oven 120, video game console 180.

FIG.. 6 is a flow chart of an exemplary operation of the present invention method with more detailed technical system components. The collected energy sensor data 610 may be processed on-device for real-time feedback on the appliance, or fransmitted to a processing server located in the building or externally (e.g. in the Cloud), The communication 620 may be done via wireless (e.g. WiFi, ZigBee, 6LoWPAN) or wired link (e.g. Ethernet, Power Line Communication). The appliance recognition and data inference engine 630 is the machine learning processing unit that (a) relies on signal processing and filtering techniques (e.g. low pass and high-pass filters) to smooth the energy data 631; (b) segments the energy data into window of samples 632 suitable for computing statistics or features which present high discriminative properties 633. The output of the energy data feature extraction 633 is a feature data space fed into a machine learning classifier which is built upon a previously created model capable of matching the input (raw energy data) to a known appliance class.

In some implementations, the users may receive personalised feedback on the energy consumed by the recognised appliances. The feedback may be delivered via a mobile phone (e.g. mobile web browsing), tablet-based computers, SMS text messaging or web portal.

In some implementations, the list of recognised appliances may be used to support comparison of energy efficiency of the users' appliances to their neighbours similar appliances. In one embodiment, the user receives information on how their fridge 130 compares to their neighbours' fridges in terms of usage and energy efficiency. To support this, a database of appliances may be created from the list of recognised appliances for a number of residential households) or a number of non-residential buildings.

In some implementations, the recognised appliances may be used to support the users in selecting suitable utility energy tariffs in the so called electronic tariffing and dynamic pricing for a smart grid scenario in which appliances from a househoidibuilding will communicate/interact with appliances from other houses/buildings through smart meters so that they can coordinate and operate at suitable times of day to achieve efficient overall power load for a particular geographical region (neighbourhood, city level). The suppliers will be able to better manage the energy demand/supply. This *creates an economical ecosystem where tariffs are now published electronically by utilities and made available in devices sitting at the users house (e.g. a tablet computer in the kitchen, which is connected to the smart meter). Such tariffs consider power load, appliances in use, and offer prices that may change dynamically throughout the day/week or month to reflect the current demand. in this smart grid scenario, the ability to identify the energy usage per individual appliances helps the users in matching their energy pattern at appliance level with current electronic tariff offers from the utility companies/suppliers.

In one embodiment, the list of recognised appliances may be used to support campaigns of energy awareness, user engagement, and user behaviour change within H residential households and workplaces.

FIG. 7 illustrates an example of mathematical tools that may be used in some implementations of the present invention method illustrated in FIG. 6.

Signal conditioning (filtering) 631: in some implementations, the raw energy sensor (aggregate or socket-level) signal data can be smoothed with a low-pass filter to reduce measurement noise. In one embodiment, a filter design i.s ii 9 low-pass Butterworth filter with cut-off frequency between 30Hz to 40Hz as illustrated in 710, where H is a transfer function of powers of Z, .0 and b are the low-pass filter coefficients.

Data segmentation 632: the smoothed sensor data may be segmented based on fixed time duration intervals, or based on a particular event or threshold (say whenever the real power is above 200 Watts). One example is the segmentation by a fixed time interval of 30 seconds.

}nergy data feature extraction 633: computed energy data features should have high discriminative ability. in some implementations, time and frequency domain features including mean, median, standard deviation and FF1 of a window of energy data samples may be employed. Other types of features may be applied such as time-frequency domain (Wavelet transforms) and application-domain specific features such as statistics of time intervals for known appliances, for instance, the mean and standard deviation of the real power values above.a certain threshold value. Features may be correlated to create new features, for example, by mean.s of cross-relation transforms between them. The feature data space may increase in size and dimensions. Dimensionality reduction techniques may be applied to said feature data space (PCA for Gaussian data, ICA for non-Gaussian data).

In one embodiment of the invention, the initial set of coefficients xO) from a Fast Fourier Transform (FFT) as illustrated in 720 may be used as feature. For example, the first seven FFT coefficients xO) in 720 may be used, where N corresponds to the size of the sampling window (e.g. 60 seconds of energy sensor data) and co,,expff2*pi*i,A1\9.

Appliance classifier.33: this machine learning classifier finds a mapping from the input feature energy data onto a finite number of appliance classes (e.g. kettle, shower pump, fan oven, game console, TV, fridge, washing machine).

Since the energy data shows sequential patterns, in some implementations it may be suitable to use classification methods that explore the correlations between observations that are close in the sequence. In one embodiment, a switch-on event of the cooker follows a switch-on event of the kitchen light. An example of a classification method that explores such correlations is Hidden Markov Models (HMMs) as described in 730. This method [8 finds the state sequence Q fqj q2.. q'} which corresponds to types/classes of appliances having the highest probability of generating the observation sequence 0 = (Oj 02... 07'] (the energy input data), given the trained model Z %z) is the probability of being in state S1 at time t, given 0 and t, which can be computed from 730.

in some implementations, other types of classifiers may be used including but not limited to k-nearest neighbour (k-NN), naïve Bayesian, support vector machine (SVM), Gaussian Mixture Models (GMMs), artificial neural networks (ANN), decision trees, threshold and rule-based classifiers. The selection of a particular classification method is an implementation decision.

In some other implementations, the implementer may have to consider the accuracy and performance of a classifier for real-time implementation cases.

In some implementations, latency is tolerable so processing intensive methods may be used for offline data analysis.

As one of ordinary skill in the art will appreciate, the classifiers may be combined into multi-stage classifiers for reasons such as accuracy and performance improvement.

FIG. 8 is a flow chart of an exemplary operation of the machine learning process to construct a database and classifier model of known appliances as learnt from previously collected energy datasets.

FIG. 9 illustrates an example of mathematical tools that may be employed to create a database of known appliances by means of learning from existing energy datasets. In one embodiment, the HMM model 2 730 is created by learning from previously collected energy datasets which have been labelled accordingly. For example, the output of the process 820 may be a table where each row is a sample of a particular feature (e.g. FF1 720) with an additional column, which is the label of the type, and class of an appliance.

For example, the data representation for a fridge in said feature data table may correspond to a number of rows with the label fridge' in the last column.

The labelling process may be carried out automatically by collecting energy data from a range of appliances within a time period (e.g. 3 months). The data collection may be achieved through a network of plug-through energy sensor devices where a label is manually added to each individual appliance dataset logged into a database. H In some other implementations, the user as part of a calibration and survey procedure may collect the training datasets.

In one embodiment, the classifier model adapts itself (self-learning) based on new energy data provided by the energy monitoring system. This may be used in some implementations as means of improving the accuracy and breadth of the present invention.

In one embodiment, the Hidden Markov Model (HMM) is built from a learning process using energy datasets (training datasets). A method to generate the 1-1MM is a maximum likelihood which seeks to compute f that maximises the likely of the sample of training energy data sequences, = {Qk}k=t, namely, P(zIA). , (i,j) 91.0 is the probability of being in 5e at time t and in Sj at time t+J, given the whole observation 0 and 2.

a,(z) explains the first t observations and ends in state S at time t. Moving on to state S1 with probability a,1, generate the (t+1)st observation, and continue from Sj at time 1+] to generate the rest of the observation sequence. A normalisation is achieved by dividing for all such possible pairs that can be visited at lime t and t+I.

Claims

Claims: I. A method of monitoring power consumption in a system comprising a plurality of different types of powered device, comprising: receiving aggregate energy usage data for the system, the aggregate energy usage data including a plurality of energy usage measurements for the system recorded sequentially over a period of time; determining from the aggregate energy usage data an estimated level of energy consumption attributable to a first type of device of the system by comparin.g the aggregate energy usage data with a pre-stored energy profile corresponding to the first type of device; and providing an output indicating the estimated energy consumption of the first type of device.
2. A method according to claim 1, wherein the determining step comprises iS determining from the aggregate energy usage data the identity of the first type of device by comparing the energy usage data with the pre-stored energy profile of the first type of device.
3. A method according to claim 2, wherein the determining step comprises identifying the presence of a contribution from the first type of device in the aggregate energy usage data by identifying a candidate energy profile for the flrst type of device.
4. A method according to claim 3, wherein the determining step further comprises: comparing the candidate energy profile with the pre-stored energy profile of the known device; and determining the degree of resemblance between the candidate and prc-stored profiles.
5. A method according to claim 4, wherein the output indicating the estimated energy consumption of the first type of device is provided only if the degree of resemblance exceeds a predetermined threshold value.
6. A method according to any of claims 1-5, wherein the aggregate energy usage data is obtained at a point through which power passes prior to reaching the plurality of devices.
7. A method according to claim 6, wherein the aggregate energy usage data is obtained at a point near a main power supply to the system.
8. A method according to any of claims 1-7, wherein the aggregate energy usage data is received over a communications link.
9. A method according to any of claims 1-8, wherein the output is provided over a communications link.
10. A method according to any of claims 1-9, wherein the pre-stored energy profile is a generic energy profile corresponding to a typical profile for the first type of device..
Ii. A method according to any of claims 1-9, wherein the pre-stored energy profile is a specific recorded energy profile for the first type of device.
12. A method according to claim 11, wherein the specific energy profile is recorded in a learning step in which an energy profile for the first type of device in the system is recorded and labelled.
13. A method according to any of claims 1-12, wherein the method further comprises the step of filtering the aggregate energy usage data..
14. A method according to any of claims 1-13, wherein the method further comprises the step of dividing the aggregate energy usage data into segments.
15. A method according to any of claims 1-14, wherein the determining step comprises at least one of: time domain analysis; frequency domain analysis; and time-frequency domain analysis.
16. A method according to any of claims 1-15, wherein the energy profile for the first type of device comprises a power usage signature.
17. A method according to claim 16, wherein the power usage signature corresponds to power usage whilst the device is in operation.
18. A method according to claim 16, wherein the power usage signature corresponds to patterns of switching between different power usage levels of the device.
19. A method according to any of claims 1-18, wherein the determining step comprises analyzing a sequence of switching events.
20. A method according to claim 19, wherein one or more of the switching events in the sequence are switching events associates with other types of device in the system.
21. A method according to any of the preceding claims, wherein the determining and providing steps occur in substantially real time.
22. A method according to any of the preceding claims, wherein the method further comprises the step of providing feedback on power usage.
23. A method according to claim 22, wherein the feedback is provided to a user via a communications device.
24. A method according to claim 22 or claim 23, wherein the feedback includes a usage comparison with a reference usage level.
25. A method according to any of claims 22-24, wherein the feedback includes an efficiency comparison with a reference efficiency level.
26. A method according to any of the preceding claims, wherein the method further comprises repeating the determining and providing steps for at least a second device of the system using a pre-stored energy profile corresponding to the second device of the system.
27. A computer program comprising program instructions for causing a computer to perform the method of any of claims 1-26.
28. Apparatus comprising an appliance recognition device configured to perform the method of any of claims 1-26.
29. A method of identifying a powered device, comprising: receiving energy usage data associated with the device, the energy usage data including a plurality of energy usage measurements for the device recorded sequentially over a period of time; determining from the energy usage data the identity of the powered device by comparing the energy usage data with a pre-stored energy profile of a known device type; and providing an output indicating the determined device type.
30. A method according to claim 29, wherein the determining step comprises identifying a candidate energy profile for the device from the energy usage data.
31. A method according to claim 30, wherein the determining step further comprises: comparing the candidate energy profile with the pre-stored energy profile of the known device; and determining the degree of resemblance between the candidate and pre-stored profiles.
32. A method according to claim 31, wherein the output indicating the type of device is provided only if the degree of resemblance exceeds a predetermined threshold value.
33. A method according to any of claims 29-32, wherein the energy usage data is device specific energy usage data.
34. A method according to claim 33, wherein the energy usage data is obtained at a power supply point where the power measured is the power supplied only to the device.
35. A method according to any of claims 29-33, wherein the energy usage data is aggregate energy usage data, *
36. A method according to claim 35, wherein the aggregate energy usage data is obtained at a point through which power passes prior to reaching the plurality of devices.
37. A method according to claim 36, wherein the aggregate energy usage data is obtained at a point near a main power supply to the system.
38. A method according to any of claims 29-37, wherein the energy usage data is received over a communications link.
39. A method according to any of claims 29-3 8, wherein the output is provided over a communications link.
40. A method according to any of claims 29-39, wherein the pre-stored energy profile is a generic energy profile corresponding to a typical profile for the known type of device.
41. A method according to any of claims 29-39, wherein the pre-stored energy profile is a specific recorded energy profile for the device.
42. A method according to claim 41, wherein the specific energy profile is recorded in a learning step in which an energy profile for the device is recorded and labelled.
43. A method according to any of claims 29-42, wherein the determining step further comprises the step of filtering the energy usage data.
44. A method according to any of claims 29-43, wherein the determining step further comprises dividing the energy usage data into segments.
45. A method according to any of claims 29-44, wherein the determining step comprises at least one of: time domain analysis; frequency domain analysis; and time-frequency domain analysis.
46. A method according to any of claims 29-45, wherein the energy profile for the known type of device comprises a power usage signature.
47. A method according to claim 46, wherein the power usage signature corresponds to power usage whilst the device is in operation.
48. A method according to claim 46, wherein the power usage signature corresponds to patterns of switching between different power usage level.s of the device. H
49. A method according to any of claims 29-48, wherein the determining step comprises analyzing a sequence of switching events.
50. A method according to claim 49, wherein one or more of the switching events in the sequence are switching events associates with other types of device in the system.
51. A method according to any of claims 29-50, wherein the method further comprises: determining from the energy usage data an estimated level of energy consumption attributable to the device; and providing an output indicating the estimated energy consumption of the device.
52. A method according to claim 51, wherein the determining and providing steps occur in substantially real time.53.. A method according to any of claims 29-52, wherein the method further comprises the step of providing feedback on power usage.54, A method according to claim 53, wherein the feedback is provided to a user via a communications device.55. A method according to claim 53 or claim 54, wherein the feedback includes a usage comparison with a reference usage level.56. A method according to any of claims 53-55, wherein the feedback includes an efficiency comparison with a reference efficiency level.57. A computer program comprising program instructions for causing a computer to perform the method of any of claims 29-56.58. Apparatus comprising an appliance recognition device configured to perform the method of any of claims 29-5 6.