Load monitoring

Abstract

A method, apparatus, and computer program are provided for load monitoring for use in an industrial facility comprising a plurality of devices. The method comprises : obtaining energy data associated with the industrial facility, wherein the energy data relates to an aggregate energy consumption of the plurality of devices within the industrial facility over a time period; determining load identification data and initial load disaggregation data for each of the plurality of devices using a respective joint machine learning model; and providing an output comprising a final load disaggregation data for the plurality of devices based, at least in part, on the load identification data and the initial load disaggregation data.

Description

[0001] Load Monitoring

[0002] Field

[0003] The present specification relates to load monitoring, particularly to non-intrusive load monitoring in industrial settings.

[0004] Background

[0005] Industrial loads consume large amount of energy and are considered an essential demand-side resource. Industries often have diverse processes, and each process contributes to the overall energy consumption. There remains a need for improvement in non-intrusive load monitoring techniques in industrial settings.

[0006] Summary

[0007] In one aspect, this specification describes a method for load monitoring for use in an industrial facility comprising a plurality of devices, the method comprising: obtaining energy data associated with the industrial facility, wherein the energy data relates to an aggregate energy consumption of the plurality of devices within the industrial facility over a time period; determining load identification data and initial load disaggregation data for each of the plurality of devices using a respective joint machine learning model; providing an output comprising a final load disaggregation data for the plurality of devices based, at least in part, on the load identification data and the initial load disaggregation data.

[0008] In some examples, if the load identification data indicates an off status for one or more of the plurality of devices, the method further includes determining the initial load disaggregation data for a respective one or more of the plurality of devices to be a false positive; and modifying the initial load disaggregation data based at least in part, on the determined false positives, wherein the final load disaggregation data is further based on the modified initial load disaggregation data. In some examples, modifying the initial load disaggregation data comprises discarding load aggregation data corresponding to the determined false positives.

[0009] In some examples, the load identification data indicates operational modes for one or more of the plurality of devices having an on status. Some examples further include mapping the final load disaggregation data for one or more of the plurality of devices on one or more discrete labels corresponding to an operational mode indicated for the respective one or more of the plurality of devices.

[0010] Some examples further include determining a fault condition in one or more of the plurality of devices if number of false positives determined for the respective one or more of the plurality of devices is higher than a threshold number.

[0011] In some examples, each joint machine learning model comprises: a classification model for determining the load identification data corresponding to a respective device of the plurality of devices, and a prediction model for determining the load disaggregation data corresponding to the respective device of the plurality of devices.

[0012] In some examples, the industrial facility comprises a plurality of devices having variable energy consumption modes.

[0013] In another aspect, this specification describes an apparatus configured to perform any method as described above..

[0014] In another aspect, this specification describes computer-readable instructions which, when executed by a computing apparatus, cause the computing apparatus to perform any method as described above..

[0015] In another aspect, this specification describes a computer program comprising instructions for causing an apparatus to perform at least the following: obtaining energy data associated with the industrial facility, wherein the energy data relates to an aggregate energy consumption of the plurality of devices within the industrial facility over a time period; determining load identification data and initial load disaggregation data for each of the plurality of devices using a respective joint machine learning model; providing an output comprising a final load disaggregation data for the plurality of devices based, at least in part, on the load identification data and the initial load disaggregation data.

[0016] In another aspect, this specification describes a computer-readable medium (such as a non-transitory computer-readable medium) comprising program instructions stored thereon for performing at least the following: obtaining energy data associated with the industrial facility, wherein the energy data relates to an aggregate energy consumption of the plurality of devices within the industrial facility over a time period; determining load identification data and initial load disaggregation data for each of the plurality of devices using a respective joint machine learning model; providing an output comprising a final load disaggregation data for the plurality of devices based, at least in part, on the load identification data and the initial load disaggregation data.

[0017] In another aspect, this specification describes an apparatus comprising: at least one processor; and at least one memory including computer program code which, when executed by the at least one processor, causes the apparatus to: obtain energy data associated with the industrial facility, wherein the energy data relates to an aggregate energy consumption of the plurality of devices within the industrial facility over a time period; determine load identification data and initial load disaggregation data for each of the plurality of devices using a respective joint machine learning model; provide an output comprising a final load disaggregation data for the plurality of devices based, at least in part, on the load identification data and the initial load disaggregation data.

[0018] In another aspect, this specification describes an apparatus comprising a first module configured to obtain energy data associated with the industrial facility, wherein the energy data relates to an aggregate energy consumption of the plurality of devices within the industrial facility over a time period; a second module configured to determine load identification data and initial load disaggregation data for each of the plurality of devices using a respective joint machine learning model; and a third module configured to provide an output comprising a final load disaggregation data for the plurality of devices based, at least in part, on the load identification data and the initial load disaggregation data.

[0019] In yet another aspect, this specification describes a method for training a joint machine learning model for one or more of a plurality of devices in an industrial facility comprising a plurality of devices, the method comprising: obtaining annotated load identification data and load disaggregation data, wherein the annotated load identification data comprises information of one or more operational modes of one or more devices in the industrial facility; training a classification model within the joint machine learning model with the annotated load identification data, and corresponding energy data associated with the industrial facility; and training a prediction model within the joint machine learning model with the load disaggregation data and corresponding energy data associated with the industrial facility, wherein the energy data relates to an aggregate energy consumption of the plurality of devices within the industrial facility. Some examples include performing data annotation for obtaining the annotated load identification data, wherein performing data annotation comprises: analysing energy usage data corresponding to one or more of the plurality of devices in the industrial facility in one or more operational modes at an on status of the respective one or more of the plurality of devices; and annotating the energy usage data with discrete labels based on the one or more operational modes of the respective one or more of the plurality of devices.

[0020] In some examples, the data annotation is performed based on information obtained from sensors corresponding to one or more of the plurality of devices.

[0021] Some examples include testing the joint machine learning model, comprising : determining a first evaluation score for the classification model by comparing predicted load identification data with ground truth load identification data based on a first one or more metrics, wherein the predicted load identification data is an output from the classification model, and the ground truth load identification data is based, at least in part, on the annotated load identification data; and / or determining a second evaluation score for the prediction model by comparing the predicted load disaggregation data with ground truth load disaggregation data based on a second one or more metrics, wherein the predicted load disaggregation data is an output from the predicted model, and the ground truth load disaggregation model is based, at least in part, on the load disaggregation data. In some examples, the first one or more metrics comprise one or more of a precision metric and / or a recall metric. In some examples, the second one or more metrics comprise a mean absolute error metric, a mean absolute percentage error metric, and / or a normalized error in assigned energy metric.

[0022] In another aspect, this specification describes an apparatus configured to perform any method for training as described above.

[0023] In another aspect, this specification describes computer-readable instructions which, when executed by a computing apparatus, cause the computing apparatus to perform any method for training as described above.

[0024] In another aspect, this specification describes a computer program for training a joint machine learning model for one or more of a plurality of devices in an industrial facility comprising a plurality of devices, the computer program comprising instructions for causing an apparatus to perform at least the following: obtaining annotated load identification data and load disaggregation data, wherein the annotated load identification data comprises information of one or more operational modes of one or more devices in the industrial facility; training a classification model within the joint machine learning model with the annotated load identification data, and corresponding energy data associated with the industrial facility; and training a prediction model within the joint machine learning model with the load disaggregation data and corresponding energy data associated with the industrial facility, wherein the energy data relates to an aggregate energy consumption of the plurality of devices within the industrial facility.

[0025] In another aspect, this specification describes a computer-readable medium (such as a non-transitory computer-readable medium) comprising program instructions stored thereon for performing at least the following: obtaining annotated load identification data and load disaggregation data, wherein the annotated load identification data comprises information of one or more operational modes of one or more devices in the industrial facility; training a classification model within the joint machine learning model with the annotated load identification data, and corresponding energy data associated with the industrial facility; and training a prediction model within the joint machine learning model with the load disaggregation data and corresponding energy data associated with the industrial facility, wherein the energy data relates to an aggregate energy consumption of the plurality of devices within the industrial facility.

[0026] In another aspect, this specification describes an apparatus comprising: at least one processor; and at least one memory including computer program code which, when executed by the at least one processor, causes the apparatus to: obtain annotated load identification data and load disaggregation data, wherein the annotated load identification data comprises information of one or more operational modes of one or more devices in the industrial facility; train a classification model within the joint machine learning model with the annotated load identification data, and corresponding energy data associated with the industrial facility, and train a prediction model within the joint machine learning model with the load disaggregation data and corresponding energy data associated with the industrial facility, wherein the energy data relates to an aggregate energy consumption of the plurality of devices within the industrial facility.

[0027] In another aspect, this specification describes an apparatus comprising a first module configured to obtain annotated load identification data and load disaggregation data, wherein the annotated load identification data comprises information of one or more operational modes of one or more devices in the industrial facility; a second module configured to train a classification model within the joint machine learning model with the annotated load identification data, and corresponding energy data associated with the industrial facility; and a third module configured to train a prediction model within the joint machine learning model with the load disaggregation data and corresponding energy data associated with the industrial facility, wherein the energy data relates to an aggregate energy consumption of the plurality of devices within the industrial facility.

[0028] In another aspect, this specification describes a system comprising: a first apparatus configured to train a joint machine learning model for one or more of a plurality of devices in an industrial facility comprising a plurality of devices, comprising: obtaining annotated load identification data and load disaggregation data, wherein the annotated load identification data comprises information of one or more operational modes of one or more devices in the industrial facility; training a classification model within the joint machine learning model with the annotated load identification data, and corresponding energy data associated with the industrial facility, and training a prediction model within the joint machine learning model with the load disaggregation data and corresponding energy data associated with the industrial facility, wherein the energy data relates to an aggregate energy consumption of the plurality of devices within the industrial facility; a second apparatus configured to perform load monitoring for use in the industrial facility comprising a plurality of devices, comprising: obtaining energy data associated with the industrial facility, wherein the energy data relates to an aggregate energy consumption of the plurality of devices within the industrial facility over a time period; determining load identification data and initial load disaggregation data for each of the plurality of devices using a respective joint machine learning model; and providing an output comprising final load disaggregation data for the plurality of devices based, at least in part, on the load identification data and the initial load disaggregation data.

[0029] Brief Description of Drawings Example embodiments will now be described, by way of example only, with reference to the following schematic drawings, in which:

[0030] FIGs. 1 and 2 are block diagrams of systems in accordance with example embodiments;

[0031] FIG. 3 is a flowchart of an algorithm in accordance with example embodiments;

[0032] FIG. 4 is a block diagram of a system in accordance with example embodiments;

[0033] FIG. 5 is a flowchart of an algorithm in accordance with example embodiments;

[0034] FIG. 6 is a block diagram of a system in accordance with example embodiments;

[0035] FIG. 7 is a flowchart of an algorithm in accordance with example embodiments;

[0036] FIGs. 8 to 10 are block diagrams of systems in accordance with example embodiments;

[0037] FIG. 11 is a block diagram of components of a system in accordance with example embodiments; and

[0038] FIG. 12 shows an example of tangible media for storing computer-readable code which when run by a computer may perform methods according to example embodiments described above.

[0039] Detailed Description

[0040] The scope of protection sought for various embodiments of the invention is set out by the independent claims. The embodiments and features, if any, described in the specification that do not fall under the scope of the independent claims are to be interpreted as examples useful for understanding various embodiments of the invention.

[0041] In the description and drawings, like reference numerals refer to like elements throughout.

[0042] FIG. 1 is a block diagram of a system, indicated generally by the reference numeral 10, in accordance with example embodiments.

[0043] The system shows an industrial facility 11, which may comprise a plurality of devices, such as a plurality of industrial appliances. Industrial appliances consume large amounts of energy when compared to domestic appliances, industrial appliances may often be considered an essential demand-side resource (e.g., demand-side resources serve resource adequacy needs by reducing electrical load (referred to as 'load'), which reduces the need for additional generation of electricity). Industries often have diverse processes involving one or more appliances, and each process contributes to a facility's overall energy consumption. It is desirable to understand the energy demand of each process and / or each appliance in order to plan efficiently (e.g., production planning), and such an energy demand may be understood by monitoring the energy usage of said appliances.

[0044] Power consumption of loads (e.g., appliances) may be monitored either intrusively or non-intrusively. In intrusive load monitoring (ILM), loads are monitored by installing sensors on individual loads. This approach of load monitoring requires excessive monetary and temporal costs for data acquisition, complex installation and ongoing maintenance. Non-intrusive load monitoring (NILM) is a low-cost approach that extracts the individual energy usage of an appliance from the total energy measurements at regular intervals. Unlike ILM, NILM is cost effective and easy to install as it uses a single energy meter for data identification and sampling. The measurements include power, voltage, and current signals. NILM may provide energy usage information for each appliance, thus enabling an understanding of each process' contribution such that operations may be optimized, and efficiency may be improved. The determination of energy usage per appliance may be referred to as load disaggregation. Load disaggregation may allow targeted load management which enables administrators (e.g., production planning team(s)) to improve on energy conservation and cost savings. The following description may refer to power consumption and energy consumption interchangeably.

[0045] System 10 further shows a database 12 comprising combined energy data from all industrial appliances that are a part of the industrial facility 11. A deep learning algorithm 13 may be used for performing load disaggregation by receiving the combined energy data from the database 12, and providing as an output an estimate of how much energy is being used by each industrial appliance. For example, the deep learning algorithm 13 may indicate, as an output, individual energy usage of a chip press 14a, soldering oven 14b, and vacuum pump 14c. Therefore, the load disaggregation data may indicate energy usage by load type (e.g., type of appliance).

[0046] In example embodiments, the deep learning algorithm 13 may comprise a machine learning model, such as a neural network. A neural network may be trained with inputs, including energy data from the industrial facility and corresponding ground truth load disaggregation data. The neural network may comprise an input layer, one or more hidden layers, and an output layer. At the input layer, energy data from the industrial facility and corresponding ground truth load disaggregation data (e.g., obtained from power meters or the like) may be provided. The hidden layers may comprise a plurality of hidden nodes, where the processing may be performed based on the data received. At the output layer, load disaggregation data may be provided which may attribute energy data to one or more of the devices 14a, 14b, 14c.

[0047] In example embodiments, ground truth data may be obtained by performing data collection from appliances, which may require sensor installation at each of the appliances. The sensor installation may be temporary, such that the sensor data may only be used for obtaining ground-truth data in order to facilitate the training of the deep learning algorithm 13, where the trained deep learning algorithm 13 may later be used for non-intrusive load monitoring (e.g., load disaggregation in the absence of sensor data from individual appliances).

[0048] Traditional load disaggregation methods in NILM may not take into consideration the operational status of each appliance (e.g., as each appliance may not comprise sensors for reporting on / off / low power / high power status), and this may lead to false positives in the determined load disaggregation data. Traditional load disaggregation methods further fail to provide a joint learning approach for load identification and load disaggregation. The example embodiments below are aimed at improving the determination of load disaggregation.

[0049] FIG. 2 is a block diagram of a system, indicated generally by the reference numeral 20, in accordance with example embodiments. The system 20 shows a joint machinelearning (ML) model 21, which receives combined energy data 22 (e.g., from a plurality of devices (e.g., appliances) within an industrial facility) and provides a final load disaggregation data 23. The joint ML model 21 comprises a load identification module 24 and a load disaggregation module 25. The joint ML model 21 may further provide, as an output, load identification data 26. The outputs of the joint ML model 21 are discussed in further detail below.

[0050] An individual joint ML model 21 may be trained for each of the respective appliances in the industrial facility 11. For example, the joint ML model 21 may be trained for a first device with ground truth load identification data for the first device and with ground truth load disaggregation data for the first device. Of course, if there are a plurality of a same / similar devices (e.g., same / similar category, model, energy consumption), the same joint ML model may be used for the plurality of same devices.

[0051] In example embodiments, load identification may be an indication of an operational status of a device. For example, load identification data may indicate whether a device is switched on or off at a given time or given time period. Alternatively, or in addition, load identification data may indicate an operational mode of the device. For example, many devices may have variable energy consumption modes or different operational modes, such as low power mode (e.g., relatively low operational speed, low operational time for a given session, low precision, or the like), high power mode (e.g., relatively high operational speed, high operational time for a given session, high precision, or the like), and / or a plurality of operational modes requiring various levels of power or energy. The load identification data may indicate (e.g., by estimation) which operational mode a device may be operating at a given time instance or time period.

[0052] In example embodiments, load disaggregation data may indicate an estimation (using non-intrusive load monitoring) of how much energy is used by a respective device at a given time period, for example, indicating what percentage of the total energy consumption of the industrial facility can be attributed to the respective device. As there may not be individual sensors and / or power meters used in systems with NILM, it may be desirable for the estimation of load aggregation data to be as accurate as possible.

[0053] In example embodiments, the joint ML model comprises a classification model for determining the load identification data corresponding to a respective device, such that the load identification module 24 may be implemented using the classification model.

[0054] In example embodiments, the joint ML model 21 comprises a prediction model for determining the load disaggregation data corresponding to a respective device, such that the load disaggregation module 25 may be implemented using the prediction model.

[0055] Generally, a classification model may be used to categorize data into predefined groups, or classes. Classification models may be trained on data, then evaluated on test data, and finally used to predict the class of new data based on known classification of ground-truth data. In contrast, generally, prediction models (e.g., such as regression models) may use past examples to predict future data rather than classifying data based on known classification groups. Accuracy of prediction models may rely heavily on amount of ground truth data. Since load disaggregation data for industrial settings may not be readily available due to its large scale, prediction models for load disaggregation that are used for domestic settings (e.g., home settings, or other low-scale settings such as small offices, shops, or the like) may not provide results with high accuracy in industrial settings. The example embodiments below provide techniques for using a single joint ML model 21 that is trained for performing both load identification and load disaggregation, such that the load disaggregation data has higher accuracy due to the inputs from the load identification module.

[0056] FIG. 3 is a flowchart of an algorithm, indicated generally by the reference numeral 30, in accordance with example embodiments. The algorithm 30 provides a method for load monitoring (e.g. NILM) for use in an industrial facility comprising a plurality of devices, wherein the method comprises operations that may be performed in an inference phase using the joint ML model 21.

[0057] The algorithm 30 starts with operation 31, where energy data (e.g., combined energy data 22) associated with an industrial facility (e.g., industrial facility 11) may be obtained. The energy data may relate to an aggregate energy consumption of the plurality of devices within the industrial facility over a time period. The time period may be a predetermined time period and / or a dynamically determined time period (e.g. determined from a user input, user setting, and / or learnt dynamically based on historical information or from an ML model).

[0058] Next, at operation 32, load identification data and initial load disaggregation data for each of the plurality of devices may be determined using a respective joint machine learning (e.g., deep learning) model, such as the joint ML model 21. For example, if the industrial facility has n number of devices (e.g., appliances), a separate joint ML model may be trained for each of the n number of devices. The respective joint ML model may be used for determining the load identification data (e.g., operational status such as ON / OFF state, and / or operational mode such as low / high power mode) for the respective device using the load identification module 24. Further, an initial load disaggregation data (e.g., initial estimation how much of the aggregate energy consumption can be attributed to that device) may be obtained for the respective device.

[0059] At operation 33, an output may be provided such that the output comprises a final load disaggregation data for the plurality of devices based, at least in part, on the load identification data and the initial load disaggregation data. As will be explained in further detail below, the final load disaggregation data may be provided by modifying the initial load disaggregation data (if required) based on the load identification data. FIG. 4 is a block diagram of a system, indicated generally by a reference numeral 40, in accordance with example embodiments. The system 40 comprises a database 41 that may comprise combined or aggregated energy data associated with an industrial facility, where the industrial facility may comprise a plurality of appliances, such as appliance 1, appliance 2, appliance 3,... appliance N. Each appliance may have a respective joint ML model trained with ground truth data corresponding to the respective appliance. For example, joint ML model 42, comprising load identification module 42a and load disaggregation module 42b, may correspond to appliance 1; joint ML model 43, comprising load identification module 43a and load disaggregation module 43b, may correspond to appliance 2; joint ML model 44, comprising load identification module 44a and load disaggregation module 44b, may correspond to appliance 3; and joint ML model 45, comprising load identification module 45a and load disaggregation module 45b, may correspond to appliance N. The energy data from the database 41 may be provided as inputs to each of the joint ML models 42 to 45. Based on said inputs, each joint ML model may provide outputs comprising load identification data and initial load disaggregation data for the respective appliances, and said outputs may be provided to a module 49 that may perform one or more of operations 46 to 48.

[0060] The load identification data from the plurality of joint ML models may be checked at operation 46. In example embodiments, if the load identification data indicates an OFF state (having an OFF status and being in an OFF state are used interchangeably in this specification) for one or more of the plurality of appliances, the initial load disaggregation data for the respective one or more devices may be determined to be a false positive at operation 47. The initial load disaggregation data may then be modified based at least in part, on the determined false positives. The final load disaggregation data (e.g., provided in operation 33) may be based on the modified initial load disaggregation data. Modifying the initial load disaggregation data may comprise discarding load disaggregation data corresponding to the determined false positives. Alternatively, or in addition, if the load identification data indicates an ON state (having an ON status and being in an ON state are used interchangeably in this specification) for one or more of the plurality of appliances, the initial load disaggregation data is considered to be valid, and modification of the initial load disaggregation data for the respective appliance may not be required for obtaining the final load disaggregation data.

[0061] In example embodiments, the load identification data indicates operational modes for one or more of the plurality of devices having an ON status. In example embodiments, the final load disaggregation data may be mapped on different operational modes identified in the load identification data (for example, if load identification data indicates that appliance 1 is in an ON state and is in a high-power operational mode). If the load disaggregation data estimates that over a 24-hour time (e.g. dependent on sequence length) period, the appliance 1 uses lOOkW, this load disaggregation data may be mapped on the high-power operational mode of the appliance 1.

[0062] In example embodiments, a fault condition may be determined for one or more devices of the industrial facility if number of false positives detected (e.g., over a period of time) for the respective one or more devices exceeds a threshold number. For example, if the number of false positives detected with a low power mode of operation of appliance 1 over a certain time period (e.g., one week), exceeds a threshold number , it may be determined that the appliance 1 has a fault condition in the low power mode of operation. This information of a predicted fault condition may be provided to an administrator such that the fault condition may be investigated and / or repaired. The threshold number may be a predetermined threshold number, and / or a dynamically determined threshold number (e.g. determined from a user input, user setting, and / or learnt dynamically based on historical information or from an ML model).

[0063] It will be appreciated that using the joint ML model may reduce occurrence of false positives, and it also can associate predicated disaggregated values to one of the operational modes of an appliance. Load disaggregation values that associate to appliance operational modes could help identify if any faults happened in a specific mode of operation of an appliance. Also, the use of the joint ML model may help to understand energy consumptions of each appliance at various operational modes.

[0064] FIG. 5 is a flowchart of an algorithm, indicated generally by the reference numeral 50, in accordance with example embodiments. The algorithm 50 provides operations for training a joint machine learning model for one or more of a plurality of devices in an industrial facility comprising a plurality of devices. The algorithm 50 comprises operations that may be performed in a training stage of the joint ML model 21. It will be appreciated that the operations 52 and 53 may be performed independently of each other in any order or sequence, or they may be performed substantially simultaneously. In example embodiments, the algorithm 50 for training the joint ML model may be executed at an apparatus, device, module (e.g. software module) that is the same as that at which the algorithm 20 is executed. Alternatively, the algorithm 50 for training the joint ML model may be executed at an apparatus, device, module (e.g. software module) that is different from that at which the algorithm 20 is executed.

[0065] At operation 51, annotated load identification data and load disaggregation data may be obtained, for example, from a database. The annotated load identification data may comprise ground truth energy data corresponding to one or more of a plurality of devices in the industrial facility. The annotated load identification data comprises information of one or more operational modes of one or more devices in the industrial facility. For example, the annotated load identification data may comprise ground truth data with labels, such as ON / OFF status, or labels corresponding to variable operational modes of each device. The process of data annotation is described in further detail below with respect to FIG. 7.

[0066] Next, at operation 52, a classification model within the joint machine learning model may be trained with the annotated load identification data, and corresponding energy data associated with the industrial facility, such that the classification model may learn to classify incoming energy data into labels (e.g. labels corresponding to operational mode or ON / OFF status), as provided by the annotated load identification data.

[0067] Next, at operation 53, a prediction model within the joint machine learning model may be trained with load disaggregation ground truth data, and corresponding energy data associated with the industrial facility, such that the prediction model may learn to predict load disaggregation of one or more devices based on incoming combined energy data associated with the industrial facility.

[0068] In some examples, the energy data relates to an aggregate energy consumption of the plurality of devices within the industrial facility.

[0069] FIG. 6 is a block diagram of a system, indicated generally by the reference numeral 60, in accordance with example embodiments. The system 60 shows the various modules involved in training of a joint ML model, such as the joint ML model 21. Combined energy data may be provided as an input to the data annotation module 61. The data annotation module 61 may then provide annotated data to a training module 62. The training module 62 may provide as an output, an initial trained model to the testing module 63, where the initial trained model may be tested and updated based on the testing. The final output may be a trained joint ML model that is ready for use in an inference process. Inference using the joint ML model may be based on the algorithm 30, as described with reference to FIG. 3. For example, in the inference process, the trained joint ML model may determine load identification data of an appliance (e.g. whether an appliance is in an OFF state or ON state), and may further determine final load disaggregation data corresponding to the appliance (e.g. load disaggregation data may at least partially be dependent on an initial load disaggregation data and the load identification data).

[0070] FIG. 7 is a flowchart of an algorithm, indicated generally by the reference numeral 70, in accordance with example embodiments. The algorithm 70 comprises operations that may be performed for data annotation, for example, at the data annotation module 61, such that said data annotation may be used for training a joint ML model for a first device. At operation 71, energy data for the first device may be obtained from a database or directly from the industrial facility. The energy data may be based on sensor data (ground truth load disaggregation data) from one or more sensors at the first device (e.g., sensors on each device configured to identify ON / OFF status). The sensor data may only be used during the training of the joint ML model, such that the load disaggregation during inference remains a non-intrusive load monitoring where there may not be any sensor data available. The energy data may be pre-processed at operation 72, such that the data is in an appropriate format for being analysed. For example, data pre-processing, as performed at operation 72, may comprise transformations applied to raw data in order to obtain a standardized dataset for provision to the joint ML model. For example, a Random Forest algorithm may not support null values, and therefore a dataset may be pre-processed (standardized) by removing null values before being provided to a Random Forest algorithm. It is appreciated that different types of machine learning modules may require different types of pre-processing or standardization of data.

[0071] At operation 73, the operating and non-operating power ranges of individual devices may be identified. These power or energy ranges may indicate how much power a respective device consumes while it is turned ON (operating power range) in one or more operational modes, or when it is turned OFF (non-operating power range, such as sleep or standby mode, or fully switched off). For example, performing data annotation may comprise analysing energy usage data corresponding to one or more of the plurality of devices in the industrial facility in one or more operational modes at an ON status of the respective one or more of the plurality of devices. At operation 74, information may be obtained from an industrial domain expert (or "domain expert") (e.g., a person who is an expert in the field of load monitoring or load management) that verifies the power ranges identified in operation 73 . Operation 75 may comprise a data exploration step, where the identified power ranges may be further analysed or explored, such that an amount or percentage of data present in different power ranges may be identified.

[0072] Next, at operation 76, based on the information from the domain expert and / or the data exploration step, the energy usage data may be annotated with discrete labels based on the one or more operational modes of the respective one or more of the plurality of devices. The annotation may be based on input from the domain expert, or they may be based on an automated process (e.g., determined historical information, or the like). For example, the energy usage data may be annotated with labels corresponding to different operational modes, such as ON / OFF mode, high-power mode, low-power mode, or the like.

[0073] In example embodiments, the data annotation process is performed on energy data obtained from sensors in each device / appliance to define energy ranges for different modes of operations of these devices / appliances. These energy ranges may be identified through data analysis and from domain expert knowledge. The annotation process is performed from data analysis and domain expert knowledge to discretize continuous appliance energies into appliance operational modes in industrial domain. This process may help to learn appliance operational modes via model training.

[0074] FIG. 8 is a block diagram of a system, indicated generally by the reference numeral 80, in accordance with example embodiments. The algorithm 80 comprises operations that may be performed for training a joint ML model for a first device. Energy data (e.g., training data comprising information of aggregate energy consumption from an industrial facility) may be provided as inputs to a training module 82 of the joint ML model. The training module 82 may further receive annotated load identification data 84 and load disaggregation data 85 as ground truth inputs. The annotated load identification data 84 may comprise outputs from a data annotation module, such as the data annotation module 61 providing annotated data according to the algorithm 70. The load disaggregation data 85 may comprise ground truth data relating to load disaggregation for the first device in the industrial facility (e.g., ground truth data that is obtained over a long time period). The annotated load identification data 84 may be an input to a load identification module 82a of the training module, and the load disaggregation training data 85 may be an input to a load disaggregation module 82b of the training module 82. As an output of the training module, a trained model 83 may be provided.

[0075] In example embodiments, the joint ML model is proposed that trains and tests on this annotated data (representing, for example, different operational modes of appliances)( e.g., a vacuum pump having low vacuum, medium vacuum and high vacuum operating modes). The inference engine (joint ML model 21) obtains the main energy data, then identifies load identification and disaggregation and uses load identification to minimize false positives. It also can associate load disaggregation values to various load identification classes. This process helps to identify more insights of appliances at various load identification classes. The joint ML model may combine multiple input sources of data, but a model can be trained for a given appliance to detect both identification and disaggregation. The joint ML model may be structured such that two different models are not required for an appliance to detect load identification and disaggregation. In this joint ML model, multiple input sources of a data may be processed in order to provide load classification data and / or load disaggregation data for one or more appliances by using a classification model and a regression joint loss function respectively.. The joint ML model may be built independently for each appliance to learn the appliance mode of operations and load disaggregation jointly and simultaneously. The joint mapping function, from the main energy to discrete labels of appliance load identification and continuous values of appliance load disaggregation, may be learnt using a joint loss function between discrete and continuous values. It is understood that, from main energy data of an industrial facility, learning discrete values of individual device operational modes may be more efficient than learning continuous values of load disaggregation used in traditional load disaggregation techniques. This is because, in the example embodiments using load identification in addition to load disaggregation, the learning process in based on a single main energy value where discretization labels (annotated load identification data) having more input data points are being mapped to each class, whereas in traditional load disaggregation techniques, continuous values are being mapped with less input points. The joint ML model may be trained using training data, and the trained joint ML model 83 may be evaluated using testing data (discussed in further detail with respect to FIG. 9).

[0076] FIG. 9 is a block diagram of a system, indicated generally by the reference numeral 90, in accordance with example embodiments. The algorithm 90 comprises operations that may be performed for testing a trained joint ML model for a first device. For example, test data (e.g., test energy data) may be used to evaluate the trained joint ML model via both classification performance metrics (e.g., for the load identification module) and prediction performance metrics (e.g., for the load disaggregation module) on each of the appliances / devices of the industrial facility.

[0077] In example embodiments, main energy test data 91 may be obtained from a database and provided to a trained joint ML model for a first device (e.g., appliance 1). The trained model may then be evaluated in an evaluation module 93. The evaluation module 93 may test a model 94 based on input 96 (comprising annotated load identification test data (ground truth data) and provided to a load identification module 94a), and further based on input 97 (comprising load disaggregation test data (ground truth data)) and provided to a load disaggregation module 94b). The output of the model 94 may be an evaluation score 95 for the trained model 92 for the first device.

[0078] In example embodiments, testing the trained model 92 may comprise determining a first evaluation score for the classification model (for load identification) by comparing predicted load identification data (output of module 94a) with ground truth load identification data (input 96) based on a first one or more metrics. For example, the predicted load identification data is an output from the classification model, and the ground truth load identification data is based, at least in part, on the annotated load identification data.

[0079] In example embodiments, testing the trained model 92 may further comprise determining a second evaluation score for a prediction model (for load disaggregation) by comparing predicted load disaggregation data (output of module 94b) with ground truth load disaggregation data (input 97) based on a second one or more metrics. For example, the predicted load disaggregation data is an output from the predicted model, and the ground truth load disaggregation model is based, at least in part, on the load disaggregation data.

[0080] In example embodiments, the first one or more metrics comprise at least one of a precision metric and a recall metric. In example embodiments, the second one or more metrics comprise at least one of a mean absolute error metric, a mean absolute percentage error metric, and a normalized error in assigned energy metric.

[0081] FIG. 10 is a block diagram of a system, indicated generally by a reference numeral 100, in accordance with example embodiments. The system 100 shows a flow for training and testing of a plurality of joint ML models corresponding to a plurality of devices of an industrial facility.

[0082] The system 100 comprises a data annotation module 101, which may be similar to the module 61 (described with reference to FIG. 6) and the system 70 (described with reference to FIG. 7). Data annotation may be performed at the data annotation module 101 for the plurality of devices, as described with reference to FIG. 7. The system 100 further comprises a training module 103 (similar to the training module 62 and system 80) and a testing module 104 (similar to the testing module 63 and system 90). The training module 103 may receive as inputs: energy data corresponding to the plurality of devices of the industrial facility (which may be provided from a database 102); and annotated training data from the data annotation module 101. The training module 103 may train respective joint ML models 103-1, 103-2,...103-N for each of the plurality of devices (appliance 1, appliance 2,... appliance N respectively) of the industrial facility. Each trained model 103-1, 103-2, ... 103-N may then be provided for testing to a testing module 104 which individually tests models 104-1, 104-2,... 104-N for the plurality of devices (appliance 1, appliance 2,... appliance N, respectively) of the industrial facility. The testing module 104 also receives, as inputs, annotated test data from the data annotation module 101. The testing module then provides evaluation scores as outputs, such as evaluation score 105-1, 105-2,...105-N for the plurality of devices (appliance 1, appliance 2,... appliance N, respectively).

[0083] The trained and tested model may then be used at an inference engine 106 (e.g. joint ML model 21), which may perform load identification and load disaggregation based on input energy data. The inference engine 106 may provide output 107 comprising load identification data and final load disaggregation data, as described by the algorithm 30 with reference to FIG. 3.

[0084] It will be appreciated that industrial load disaggregation, as described by the above example embodiments, may help to monitor individual appliances' energy consumption at regular intervals. Such load disaggregation may identify the appliances that are consuming more power / energy, or the appliances that are mostly in use. This may provide insight for production planning and / or energy savings. The model-based appliance load prediction approach (non-intrusive load monitoring) replaces a device-based appliance energy collection process (intrusive load monitoring relying on sensors, power meters, or the like) and reduces the installation and maintenance costs of devices to collect appliance energy. The model-based approach may also avoid human errors. The data annotation process described above may be performed based on data analysis and domain expert knowledge to discretize continuous appliance energies into appliance operational modes in industrial domain, such that appliance operational modes may be learnt using model training.

[0085] The above example embodiments provide a joint ML model learning method that may have the capability to learn appliance load identification and disaggregation though joint loss optimization from the main industrial energy. The joint ML model may be used for assigning load disaggregation values to various appliance operational modes. In inference, the joint ML model may firstly determine if an appliance is in off mode, and secondly may determine that the predicted initial load disaggregation data is a false positive if the appliance is determined to be off. If a false positive is detected, the predicted values may be excluded from determining the final load disaggregation data so as to minimize the false positives. If an appliance is not in an OFF state, then the predicted values are considered to be valid. The example embodiments described also provide the ability to identify faulty appliances when appliance energy predictions are deviating from the normal appliance energy. Also, load identification or load disaggregation associating to the appliance mode of operations may assist in identifying a fault condition in one or more operational modes of an appliance. Load disaggregation may help to reduce the cost associated for industrial energy auditing process. The above embodiments described for load disaggregation may help in improving accuracy in auditing by providing all appliance energy predictions.

[0086] For completeness, FIG. 11 is a schematic diagram of components of one or more of the example embodiments described previously, which hereafter are referred to generically as processing systems 300. A processing system 300 may have a processor 302, a memory 304 closely coupled to the processor and comprised of a RAM 314 and ROM 312, and, optionally, user input 310 and a display 318. The processing system 300 may comprise one or more network / apparatus interfaces 308 for connection to a network / apparatus, e.g., a modem which may be wired or wireless. Interface 308 may also operate as a connection to at least one other apparatus such as at least one device / apparatus which is not network side apparatus. Thus, direct connection between devices / apparatus without network participation is possible. The processor 302 is connected to each of the other components in order to control operation thereof.

[0087] The memory 304 may comprise a non-volatile memory, such as a hard disk drive (HDD) or a solid-state drive (SSD). The ROM 312 of the memory 304 stores, amongst other things, an operating system 315 and may store software applications 316. The RAM 314 of the memory 304 is used by the processor 302 for the temporary storage of data. The operating system 315 may contain computer program code which, when executed by the processor implements aspects of the algorithms 30, 50 and 70 described above. The operating system 315 may further contain computer program code which, when executed by the processor implements aspects of the joint ML model 21, 42-45, modules 61, 62, 63, 82, 93, 101, 103, and / or 104 as described above. Note that in the case of small device / apparatus the memory can be most suitable for small size usage (i.e., not always hard disk drive (HDD) or solid-state drive (SSD) is used).

[0088] The processor 302 may take any suitable form. For instance, it may be a microcontroller, a plurality of microcontrollers, a processor, or a plurality of processors.

[0089] The processing system 300 may be a standalone computer, a server, a console, or a network thereof. The processing system 300 and needed structural parts may be all inside device / apparatus such as loT device / apparatus (i.e., embedded to very small size).

[0090] In some example embodiments, the processing system 300 may also be associated with external software applications. These may be applications stored on a remote server device / apparatus and may run partly or exclusively on the remote server device / apparatus. These applications may be termed cloud-hosted applications. The processing system 300 may be in communication with the remote server device / apparatus in order to utilize the software application stored there.

[0091] FIG. 12 shows tangible media, specifically a removable memory unit 365, storing computer-readable code which when run by a computer may perform methods according to example embodiments described above. The removable memory unit 365 may be a memory stick, e.g. a USB memory stick, having internal memory 366 for storing the computer-readable code. The internal memory 366 may be accessed by a computer system via a connector 367. Other forms of tangible storage media may be used. Tangible media can be any device / apparatus capable of storing data / information which data / information can be exchanged between devices / apparatus / network.

[0092] Embodiments of the present invention may be implemented in software, hardware, application logic or a combination of software, hardware and application logic. The software, application logic and / or hardware may reside on memory, or any computer media. In example embodiments, the application logic, software or an instruction set is maintained on any one of various conventional computer-readable media. In the context of this document, a "memory" or "computer-readable medium" may be any non-transitory media or means that can contain, store, communicate, propagate or transport the instructions for use by or in connection with an instruction execution system, apparatus, or device, such as a computer.

[0093] Reference to, where relevant, "computer-readable storage medium", "computer program product", "tangibly embodied computer program" etc., or a "processor" or "processing circuitry" etc. should be understood to encompass not only computers having differing architectures such as single / multi-processor architectures and sequencers / parallel architectures, but also specialised circuits such as field programmable gate arrays FPGA, application specify circuits ASIC, signal processing devices / apparatus and other devices / apparatus. References to computer program, instructions, code etc. should be understood to express software for a programmable processor firmware such as the programmable content of a hardware device / apparatus as instructions for a processor or configured or configuration settings for a fixed function device / apparatus, gate array, programmable logic device / apparatus, etc.

[0094] As used in this application, the term "circuitry" refers to all of the following: (a) hardware-only circuit implementations (such as implementations in only analogue and / or digital circuitry) and (b) to combinations of circuits and software (and / or firmware), such as (as applicable): (i) to a combination of processor(s) or (ii) to portions of processor(s) / software (including digital signal processor(s)), software, and memory(ies) that work together to cause an apparatus, such as a server, to perform various functions) and (c) to circuits, such as a microprocessor(s) or a portion of a microprocessor(s), that require software or firmware for operation, even if the software or firmware is not physically present. If desired, the different functions discussed herein may be performed in a different order and / or concurrently with each other. Furthermore, if desired, one or more of the above-described functions may be optional or may be combined. Similarly, it will also be appreciated that the flow charts of Figures 3, 5 and 7 are examples only and that various operations depicted therein may be omitted, reordered and / or combined.

[0095] It will be appreciated that the above-described example embodiments are purely illustrative and do not limit the scope of the invention. Other variations and modifications will be apparent to persons skilled in the art upon reading the present specification.

[0096] Moreover, the disclosure of the present application should be understood to include any novel features or any novel combination of features either explicitly or implicitly disclosed herein or any generalization thereof and during the prosecution of the present application or of any application derived therefrom, new claims may be formulated to cover any such features and / or combination of such features.

Claims

- 24 -Claims1. A method for load monitoring for use in an industrial facility comprising a plurality of devices, the method comprising: obtaining energy data associated with the industrial facility, wherein the energy data relates to an aggregate energy consumption of the plurality of devices within the industrial facility over a time period; determining load identification data and initial load disaggregation data for each of the plurality of devices using a respective joint machine learning model; and providing an output comprising a final load disaggregation data for the plurality of devices based, at least in part, on the load identification data and the initial load disaggregation data.

2. A method as claimed in claim 1, further comprising: if the load identification data indicates an OFF status for one or more of the plurality of devices, determining the initial load disaggregation data for the respective one or more devices to be a false positive; and modifying the initial load disaggregation data based at least in part, on the determined false positives, wherein the final load disaggregation data is further based on the modified initial load disaggregation data.

3. A method as claimed in claim 2, wherein modifying the initial load disaggregation data comprises discarding load aggregation data corresponding to the determined false positives.

4. A method as claimed in any one of the preceding claims, wherein the load identification data indicates operational modes for one or more of the plurality of devices having an ON status.

5. A method as claimed in claim 4, further comprising mapping the final load disaggregation data for one or more of the plurality of devices on one or more discrete labels corresponding to an operational mode indicated for the respective one or more of the plurality of devices.

6. A method as claimed in any one of claims 2 to 5, further comprising determining a fault condition in one or more of the plurality of devices if number of false positives determined for the respective one or more of the plurality of devices exceeds a threshold number.

7. A method as claimed in any one of the preceding claims, wherein each joint machine learning model comprises: a classification model for determining the load identification data corresponding to a respective device of the plurality of devices, and a prediction model for determining the load disaggregation data corresponding to the respective device of the plurality of devices.

8. A method as claimed in any one of the preceding claims, wherein the industrial facility comprises a plurality of devices having variable energy consumption modes.

9. A method for training a joint machine learning model for one or more of a plurality of devices in an industrial facility comprising a plurality of devices, the method comprising: obtaining annotated load identification data and load disaggregation data, wherein the annotated load identification data comprises information of one or more operational modes of one or more devices in the industrial facility; training a classification model within the joint machine learning model with the annotated load identification data, and corresponding energy data associated with the industrial facility, and training a prediction model within the joint machine learning model with the load disaggregation data and corresponding energy data associated with the industrial facility, wherein the energy data relates to an aggregate energy consumption of the plurality of devices within the industrial facility.

10. A method as claimed in claim 9, further comprising performing data annotation for obtaining the annotated load identification data, wherein performing data annotation comprises: analysing energy usage data corresponding to one or more of the plurality of devices in the industrial facility in one or more operational modes at an ON status of the respective one or more of the plurality of devices; and annotating the energy usage data with discrete labels based on the one or more operational modes of the respective one or more of the plurality of devices.

11. A method as claimed in any one of claims 9 or 10, wherein the data annotation is performed based on information obtained from sensors corresponding to one or more of the plurality of devices.

12. A method as claimed in any one of claims 7 or 8, further comprising testing the joint machine learning model, comprising: determining a first evaluation score for the classification model by comparing predicted load identification data with ground truth load identification data based on a first one or more metrics, wherein the predicted load identification data is an output from the classification model, and the ground truth load identification data is based, at least in part, on the annotated load identification data; and / or determining a second evaluation score for the prediction model by comparing the predicted load disaggregation data with ground truth load disaggregation data based on a second one or more metrics, wherein the predicted load disaggregation data is an output from the predicted model, and the ground truth load disaggregation model is based, at least in part, on the load disaggregation data.

13. A method as claimed in claim 12, wherein the first one or more metrics comprise one or more of a precision metric and / or a recall metric.

14. A method as claimed in any one of claims 12 or 13, wherein the second one or more metrics comprise a mean absolute error metric, a mean absolute percentage error metric, and / or a normalized error in assigned energy metric.

15. A system comprising: a first apparatus configured to train a joint machine learning model for one or more of a plurality of devices in an industrial facility comprising a plurality of devices, wherein training the joint machine learning model comprises: obtaining annotated load identification data and load disaggregation data, wherein the annotated load identification data comprises information of one or more operational modes of one or more devices in the industrial facility; training a classification model within the joint machine learning model with the annotated load identification data and corresponding energy data associated with the industrial facility, and training a prediction model within the joint machine learning model with the load disaggregation data and corresponding energy data associated with the industrial facility, wherein the energy data relates to an aggregate energy consumption of the plurality of devices within the industrial facility;- 27 - a second apparatus configured to perform load monitoring for use in the industrial facility comprising a plurality of devices, wherein load monitoring comprises: obtaining energy data associated with the industrial facility, wherein the energy data relates to an aggregate energy consumption of the plurality of devices within the industrial facility over a time period; determining load identification data and initial load disaggregation data for each of the plurality of devices using a respective joint machine learning model, wherein the respective joint machine learning model is an output of the first apparatus; and providing an output comprising a final load disaggregation data for the plurality of devices based, at least in part, on the load identification data and the initial load disaggregation data.

Images