FORM 2
THE PATENTS ACT, 1970
(39 of 1970)
&
THE PATENT RULES, 2003
COMPLETE SPECIFICATION (See Section 10 and Rule 13)
Title of invention:
REPRESENTATIONAL LEARNING ASSISTED MATRIX FACTORIZATION FRAMEWORK FOR LOAD DISAGGREGATION
Applicant
Tata Consultancy Services Limited A company Incorporated in India under the Companies Act, 1956
Having address:
Nirmal Building, 9th floor,
Nariman point, Mumbai 400021,
Maharashtra, India
Preamble to the description
The following specification particularly describes the invention and the manner in which it is to be performed.

TECHNICAL FIELD [001] The disclosure herein generally relates to load disaggregation, and, more particularly, to representational learning assisted matrix factorization framework for load disaggregation.
BACKGROUND
[002] Increasing deployment of smart meters in residential and commercial buildings has enabled disaggregation of building power consumption, commonly referred to as Load Disaggregation (LD). LD is the process of obtaining information about the appliances contributing to the aggregate power consumption in a household or a commercial building and such solutions help in understanding the power requirement of buildings in locations of interest. Load Disaggregation is useful to consumers as they can avail information about their electrical power consumption at appliance level and can also receive appliance usage recommendations based on peak pricing scenarios. In turn, this enables consumer segmentation and therefore promotes targeted demand response. Due to the obvious extent of its importance to utilities and consumers, extensive research has been done in the recent past to develop generalized frameworks that can solve this problem for power consumption data from any arbitrary household. Load Disaggregation on high sampled smart meter data (1KHz up to 1MHz) is relatively easy owing to the availability of transients, which are unique for every appliance. Aggregate meter data sampled at 1Hz and above can be disaggregated using event detection algorithms.
[003] For low-sampled data such as one-minute sampled data and above, appropriate modeling and optimisation techniques are required for efficient disaggregation (e.g., refer “Oliver Parson, Siddhartha Ghosh, Mark Weal, and Alex Rogers, “Using hidden markov models for iterative nonintrusive appliance monitoring,” in Neural Information Processing Systems workshop on Machine Learning for Sustainability (17/12/11), December 2011.”, “Nipun Batra, Hongning Wang, Amarjeet Singh, and Kamin Whitehouse, “Matrix factorization for scalable energy breakdown,” in Proceedings of the Thirty-First AAAI Conference on

Artificial Intelligence. 2017, AAAI’17, p. 44674473, AAAI Press. – also referred as Batra et al.”), “J. Z. Kolter, Siddharth Batra, and Andrew Y. Ng, “Energy disaggregation via discriminative sparse coding,” in Advances in Neural Information Processing Systems 23, J. D. Lafferty, C. K. I. Williams, J. Shawe-Taylor, R. S. Zemel, and A. Culotta, Eds., pp. 1153–1161. Curran Associates, Inc., 2010.”,). The performance of disaggregation is measured in terms of load identification accuracy and the energy-consumption estimates of the loads (e.g., refer Batra et al. and “Sagar Verma, Shikha Singh, and Angshul Majumdar, “Multi label restricted boltzmann machine for nonintrusive load monitoring,” in ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2019, pp. 8345–8349. – also referred as Verma et al.”). While all the existing works in this area require large amounts of meta-data and/or training data to train appliance reconstruction but solution for load identification still seems to be sub-optimal. However, only few works are dedicated to the problem of appliance reconstruction. Further, in all existing works, it has been observed that computational costs associated with the training of such models for different geographical locations prevents them for being truly generalizable.
SUMMARY [004] Embodiments of the present disclosure present technological improvements as solutions to one or more of the above-mentioned technical problems recognized by the inventors in conventional systems. For example, in one aspect, there is provided a processor implemented method for representational learning assisted matrix factorization framework for load disaggregation. The method comprises receiving, via one or more hardware processors, an aggregated power consumption data corresponding to a set of appliances, wherein the aggregated power consumption data is time windowed comprising a set of corresponding instances S of a predefined length N; generating, by using a Restricted Boltzmann Machine (RBM) via the one or more hardware processors, a data representation H of the set of corresponding instances S; obtaining, via the one or more hardware processors, an appliance signature matrix L for the set of

appliances based on a set of time windowed historical appliance consumption signatures; and learning, by using a matrix factorization framework via the one or more hardware processors, a mapping matrix M between the data representations H of the set of corresponding instances S and the appliance signature matrix L; obtaining one or more disaggregate estimates for one or more appliances of interest based on the learnt mapping matrix M.
[005] In an embodiment, the data representation H of each instance from the set of corresponding instances is expressed using the equation: p(hJ = l\vi) = (r(bj+Zi(viwij)),
wherein hj is the hidden unit in RBM, p is a probability measure, wherein vt is a visible layer of the RBM, wherein a is sigmoid function, bj is a bias associated with the RBM, and wherein wtj corresponds to one or more weights of one or more hidden units in the RBM.
[006] In an embodiment, the learnt mapping matrix M between the data representations H of the set of corresponding instances S and the appliance signature matrix L is based on the equation: M ← argmin||L - MH\\2F, and
M
wherein F is a Euclidean norm.
[007] In an embodiment, the predefined length N is selected based on a reconstruction efficiency of the RBM for a specific time window.
[008] In an embodiment, the disaggregate estimates are obtained based on the equation:
LestSxN = HSXK * MKXN
wherein K is number of hidden units in the RBM.
[009] In another aspect, there is provided a processor implemented system for representational learning assisted matrix factorization framework for load disaggregation. The system comprises: a memory storing instructions; one or more communication interfaces; and one or more hardware processors coupled to the memory via the one or more communication interfaces, wherein the one or more hardware processors are configured by the instructions to: receive an aggregated power consumption data corresponding to a set of appliances, wherein the

aggregated power consumption data is time windowed comprising a set of corresponding instances S of a predefined length N; generate, by using a Restricted Boltzmann Machine (RBM), a data representation H of the set of corresponding instances 5; obtain an appliance signature matrix L for the set of appliances based on a set of time windowed historical appliance consumption signatures; and learn, by using a matrix factorization framework, a mapping matrix M between the data representations H of the set of corresponding instances S and the appliance signature matrix L; and obtaining disaggregation estimates for one or more appliances of interest based on the learnt mapping matrix M.
[010] In an embodiment, the data representation H of each instance from the set of corresponding instances S is expressed using the equation:
p(hj = l\vi) = (r(bj+Zi(viwij)), wherein hj is the hidden unit in RBM, p is a probability measure, wherein vt is a visible layer of the RBM, wherein σ is sigmoid function, bj is a bias associated with the RBM, and wherein wtJ corresponds to one or more weights of one or more hidden units in the RBM.
[011] In an embodiment, the learnt mapping matrix M between the data representations H of the set of corresponding instances S and the appliance signature matrix L is based on the equation: M ← argmin\\L - MH\\2F, and
M
wherein F is a Euclidean norm.
[012] In an embodiment, the predefined length JV is selected based on a reconstruction efficiency of the RBM for a specific time window.
[013] In an embodiment, the disaggregate estimates are obtained based on the equationLestsxW = HSxK * MKxN, wherein K is number of hidden units in the RBM.
[014] In yet another aspect, there are provided one or more non-transitory machine-readable information storage mediums comprising one or more instructions which when executed by one or more hardware processors cause a method for representational learning assisted matrix factorization framework for load disaggregation. The method comprises receiving, via the one or more hardware

processors, an aggregated power consumption data corresponding to a set of appliances, wherein the aggregated power consumption data is time windowed comprising a set of corresponding instances S of a predefined length N; generating, by using a Restricted Boltzmann Machine (RBM) via the one or more hardware processors, a data representation H of the set of corresponding instances 5; obtaining, via the one or more hardware processors, an appliance signature matrix L for the set of appliances based on a set of time windowed historical appliance consumption signatures; and learning, by using a matrix factorization framework via the one or more hardware processors, a mapping matrix M between the data representations H of the set of corresponding instances S and the appliance signature matrix L; and obtaining disaggregate estimates for one or more appliances of interest from the set of appliances based on the learnt mapping matrix M.
[015] In an embodiment, the data representation H of each instance from the set of corresponding instances S is expressed using the equation:

wherein hj is the hidden unit in RBM, p is a probability measure, wherein vi is a visible layer of the RBM, wherein σ is sigmoid function, bj is a bias associated with the RBM, and wherein wtj corresponds to one or more weights of one or more hidden units in the RBM.
[016] In an embodiment, the learnt mapping matrix M between the data representations H of the set of corresponding instances S and the appliance signature matrix L is based on the equation: M ← argmin\\L - MH\\2F, and
M
wherein F is a Euclidean norm.
[017] In an embodiment, the predefined length JV is selected based on a reconstruction efficiency of the RBM for a specific time window.
[018] In an embodiment, the disaggregate estimates are obtained based on the equation:
LestSxN = HSXK * MKXN
wherein K is number of hidden units in the RBM.

[019] It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention, as claimed.
BRIEF DESCRIPTION OF THE DRAWINGS
[020] The accompanying drawings, which are incorporated in and constitute a part of this disclosure, illustrate exemplary embodiments and, together with the description, serve to explain the disclosed principles:
[021] FIG. 1 depicts an exemplary system for representational learning assisted matrix factorization framework for load disaggregation, in accordance with an embodiment of the present disclosure.
[022] FIG. 2 depicts an exemplary flow chart illustrating a method for representational learning assisted matrix factorization framework for load disaggregation, using the system of FIG. 1, in accordance with an embodiment of the present disclosure.
[023] FIG. 3 depicts a Restricted Boltzmann Machine-Least Squares framework for Load Disaggregation, in accordance with an embodiment of the present disclosure.
[024] FIGS. 4 through 6 depict a graphical representation illustrating appliance reconstruction results for various appliance datasets, in accordance with an example embodiment.
DETAILED DESCRIPTION OF EMBODIMENTS [025] Exemplary embodiments are described with reference to the accompanying drawings. In the figures, the left-most digit(s) of a reference number identifies the figure in which the reference number first appears. Wherever convenient, the same reference numbers are used throughout the drawings to refer to the same or like parts. While examples and features of disclosed principles are described herein, modifications, adaptations, and other implementations are possible without departing from the scope of the disclosed embodiments.

[026] As mentioned above, increasing deployment of smart meters in residential and commercial buildings has enabled disaggregation of building power consumption, commonly referred to as Load Disaggregation (LD). LD is the process of obtaining information about the appliances contributing to the aggregate power consumption in a household or a commercial building and such solutions help in understanding the power requirement of buildings in locations of interest. Load Disaggregation is useful to consumers as they can avail information about their electrical power consumption at appliance level and can also receive appliance usage recommendations based on peak pricing scenarios. In turn, this enables consumer segmentation and therefore promotes targeted demand response. Due to the obvious extent of its importance to utilities and consumers, extensive research has been done in the recent past to develop generalized frameworks that can solve this problem for power consumption data from any arbitrary household. Load Disaggregation on high sampled smart meter data (1KHz up to 1MHz) is relatively easy owing to the availability of transients, which are unique for every appliance. Aggregate meter data sampled at 1Hz and above can be disaggregated using event detection algorithms. In all existing works, it has been observed that computational costs associated with the training of models for different geographical locations prevents them for being truly generalizable.
[027] Business cases require computationally simple generic methods, with little prior information. To this effect, system and method of the present disclosure provide a framework that generates aggregate power data representations using Restricted Boltzmann Machine (RBM) and then use a matrix factorization to express these representations in terms of mapping function and appliance loads of interest. It is observed by the system and method of the present disclosure that the implemented framework does not require the use of iterative procedures to converge the mapping function and the use of RBM in obtaining aggregate data representations promotes such formulation. While other RBM-based LD solutions demonstrated the ability of RBM to act as multi-label classifier (e.g., Verma et al.), the system and method of the present disclosure uses RBM in a regression context

so as to enable accurate appliance reconstructions while simultaneously simplifying Matrix Factorization framework for LD.
[028] It is further observed by the present disclosure that with proper understanding of aggregate and appliance energy consumption patterns, it enables the system and method described herein to provide simplified yet generic approach to load disaggregation and appliance reconstruction. This is possible by exploiting underlying energy structure of time-windowed aggregate power consumption of an arbitrary household using un-supervised learning methods (here, RBM) and by observing individual load consumption patterns of the household of interest. More specifically, the present disclosure and its system and method this formulation using Restricted Boltzmann Machine and Least Squares Solver function and evaluate its performance on UK Dale (e.g., refer “Jack Kelly and William Knottenbelt, “The UK-DALE dataset, domestic appliance-level electricity demand and whole-house demand from five UK homes,” Scientific Data, vol. 2, no. 150007, 2015.” - also referred as Kelly et al.) and AMPds data sets (e.g., “S. Makonin, F. Popowich, L. Bartram, B. Gill, and I. V. Baji, “Ampds: A public dataset for load disaggregation and eco-feedback research,” in 2013 IEEE Electrical Power Energy Conference, 2013, pp. 1-6. - referred as Makonin et al.”). The results in terms of appliance reconstruction, consumption and identification accuracies suggest the applicability of the approach, particularly in real-world scenarios.
[029] Restricted Boltzmann Machines (RBMs) are generative models having a bipartite graphical structure. They are used to learn approximate distributions of an arbitrary input data, so as to generate data similar to the input data. RBMs are classic examples of pattern storage networks and they can be used to learn the implicit patterns in the data.
[030] For a given input data, RBM generates a joint energy function of the visible (observed) units and hidden (feature detectors) units. Denoting the i-th visible unit of the RBM as v, j'-th hidden unit as hj, their biases as ai and bj respectively and the weights between the two units as wij, the energy of the joint configuration (v, h) is given by the following equation:
(1)

The binary state of hidden unit hj is set to 1 using the following equation:
(2) where p(x) represents the logistic sigmoid function and is given by,
(3)
The approximate reconstruction of the input data through the following equation is:
(4)
[031] The system and method henceforth denote the hidden unit values obtained after training the RBM from equation (2) as H, and visible unit values as v for the sake of simplicity.
[032] As RBMs are generative models, they tend to approximate the input data using Kullback-Leibler (KL) divergence method. The approximations, due to KL divergence are not exact replica of the input data but are representative of the structure of the data. The states of the hidden units contain non-linear representations of the data, which are useful for pattern recognition tasks.
[033] Among many applications of RBM, these bipartite graphs are found to be very useful in tasks such as multi-label classification (e.g., refer Verma et. Al) and in collaborative filtering (e.g., “Ruslan Salakhutdinov, Andriy Mnih, and Geoffrey Hinton, “Restricted boltzmann machines for collaborative filtering,” in Proceedings of the 24th international conference on Machine learning, 2007, pp. 791-798.”).
[034] While the use of RBMs as classifiers and generative models is well established, considering RBM features for regression especially in source-separation contexts is not well explored. Time-windowed aggregate data is highly unstructured and here, RBMs are quite efficient in learning the inherent structure in time-windowed chunks of aggregate power data. Data representations thus obtained are evidently regressive in nature and represent the underlying patterns in the data.
[035] In treating Load Disaggregation as a matrix factorization problem, embodiments of the present disclosure present a Matrix Factorization (MF) framework that can express RBM representations of aggregate power consumption in terms of a mapping function M and matrix of appliance signatures L.

[036] While stand-alone RBM based Matrix Factorization solutions exist, they were developed considering the classification ability of RBM. In the present disclosure, system and method chose to ask this question: For a windowed segment of aggregate data representations, what appliances can these representations be mapped onto? In this regard, the system and method obtain the data representations for the aggregate power consumption using RBM(s). It is assumed that the knowledge of the historical consumption behaviors of individual loads of interest is available. Instead of considering their sparse-coded features, the appliance signatures representative of possible consumption patterns directly is used, so as to be able to reduce computational costs. Using a Matrix Factorization approach, factoring of the data representations into the appliance consumption patterns is obtained, using a single-step Least Squares Minimization step.
[037] Referring now to the drawings, and more particularly to FIGS. 1 through 6, where similar reference characters denote corresponding features consistently throughout the figures, there are shown preferred embodiments and these embodiments are described in the context of the following exemplary system and/or method.
[038] FIG. 1 depicts an exemplary system 100 for representational learning assisted matrix factorization framework for load disaggregation, in accordance with an embodiment of the present disclosure. In an embodiment, the system 100 includes one or more hardware processors 104, communication interface device(s) or input/output (I/O) interface(s) 106 (also referred as interface(s)), and one or more data storage devices or memory 102 operatively coupled to the one or more hardware processors 104. The one or more processors 104 may be one or more software processing components and/or hardware processors. In an embodiment, the hardware processors can be implemented as one or more microprocessors, microcomputers, microcontrollers, digital signal processors, central processing units, state machines, logic circuitries, and/or any devices that manipulate signals based on operational instructions. Among other capabilities, the processor(s) is/are configured to fetch and execute computer-readable instructions stored in the memory. In an embodiment, the system 100 can

be implemented in a variety of computing systems, such as laptop computers, notebooks, hand-held devices (e.g., smartphones, tablet phones, mobile communication devices, and the like), workstations, mainframe computers, servers, a network cloud, and the like.
[039] The I/O interface device(s) 106 can include a variety of software and hardware interfaces, for example, a web interface, a graphical user interface, and the like and can facilitate multiple communications within a wide variety of networks N/W and protocol types, including wired networks, for example, LAN, cable, etc., and wireless networks, such as WLAN, cellular, or satellite. In an embodiment, the I/O interface device(s) can include one or more ports for connecting a number of devices to one another or to another server.
[040] The memory 102 may include any computer-readable medium known in the art including, for example, volatile memory, such as static random-access memory (SRAM) and dynamic-random access memory (DRAM), and/or non-volatile memory, such as read only memory (ROM), erasable programmable ROM, flash memories, hard disks, optical disks, and magnetic tapes. In an embodiment, a database 108 is comprised in the memory 102, wherein the database 108 comprises information specific to aggregated power consumption data corresponding to a set of appliances. The database 108 further comprises data representation � of set of instances �, set of time windowed historical appliance consumption signatures for the set of appliances, mapping matrix learnt for obtaining disaggregated data, and the like. The memory 102 further comprises one or more Restricted Boltzmann Machines (RBMs) and one or more matrix factorization framework which when executed perform a method described herein. The memory 102 further comprises (or may further comprise) information pertaining to input(s)/output(s) of each step performed by the systems and methods of the present disclosure. In other words, input(s) fed at each step and output(s) generated at each step are comprised in the memory 102 and can be utilized in further processing and analysis.
[041] FIG. 2, with reference to FIG. 1, depicts an exemplary flow chart illustrating a method for representational learning assisted matrix factorization

framework for load disaggregation, using the system 100 of FIG. 1, in accordance with an embodiment of the present disclosure. In an embodiment, the system(s) 100 comprises one or more data storage devices or the memory 102 operatively coupled to the one or more hardware processors 104 and is configured to store instructions for execution of steps of the method by the one or more processors 104. The steps of the method of the present disclosure will now be explained with reference to components of the system 100 of FIG. 1, and the flow diagram as depicted in FIG. 2. In an embodiment, at step 202 of the present disclosure, the one or more hardware processors 104 receive an aggregated power consumption data corresponding to a set of appliances. In an embodiment, the aggregated power consumption data is time windowed comprising a set of corresponding instances S of a predefined length N. In an embodiment, the set of corresponding instances S may also be referred as the set of instances S or instances S and interchangeably used herein. The predefined length N is selected based on a reconstruction efficiency of a RBM for a specific time window. The length N or the duration of each instance is selected based on the consumption patterns in a particular household. The intent of the method and system described herein is to use such length of time, wherein it is certain that the data possesses a certain pattern of consumption. S such instances contain the overall consumption behaviour of a particular household which can be used for disaggregation. Reconstruction efficiency of the RBM is checked through visual inspection of the RBM reconstructions of the input data. The overall efficiency of the results is computed in terms of the efficiency metrics such as RMS of Percentage of Energy Correctly assigned (PEC) and the efficiency of the framework is computed in terms of r2 score (described below in experimental evaluation and comparison sections). Referring to step 202, it is assumed by the system and method of the present disclosure that there is an aggregated data from a certain household for a month. It is further assumed that historical appliance-level data, for appliances of interest for disaggregation in that specific household is also available. For instance, the appliances of interest are a combination of the most frequently used High Power Appliances such as Microwave Oven, Washer-Dryer, Electric Heater, etc., and always ON appliances such as Refrigerator or Freezer.

Typical duration of high-power operation for Oven spans only over few minutes and in this period, it consumes anywhere to 1K-3K watts depending on the make. The typical power consumption for a Washer-Dryer is around 2-2.5 KW, in the heater mode and around 500-600W in non-heater mode. Its typical duration of operation is 1-1.5 hours at least. Typical power consumption for a Refrigerator or Freezer is rather low around 200-400W and is usually always operational. When any set of arbitrary data is obtained, for say a month or so, the system 100 slices/segments the data into windows of three hours,1 day, etc. during which time, the system 100 observes that a certain structure is obtained across the entire data set. The same time window is naturally applied to historical appliance signatures. A matrix of the historical appliance signature (also referred as appliance matrix and interchangeably used herein) with the same size as the windowed data is then created/formed by the system 100. As mentioned above, it is assumed by the system 100 that there are s instances of data, each of length duration N. The size is then S × N, which applies to both the appliance matrix and data matrix.
[042] Referring to steps of FIG. 2, at step 204 of the present disclosure, the one or more hardware processors 104 generate, by using one or more Restricted Boltzmann Machines (RBMs), a data representation H of the set of corresponding instances S of the aggregated power consumption data. The data matrix is fed to the RBM(s) and the data representations obtained consequently are of size S × K, K being the number of hidden units in the RBM(s). Considering K = 1300 units if N is 1400, and K = 700 if N = 750. The number of hidden units to be considered for each case was obtained using trial and error method and the system 100 selected the above values as the efficiency of the method did not change significantly beyond these values. It is to be understood by a person of ordinary skill in the art or person skilled in the art that such value selection shall not be construed as limiting the scope of the present disclosure.
[043] At step 206 of the present disclosure, the one or more hardware processors 104 obtain an appliance signature matrix for the set of appliances based on a set of time windowed historical appliance consumption signatures. It is also assumed by the system and method of the present disclosure that historical

signatures for freezer, refrigerator, etc. are available. These signatures contain the information of the make and the magnitude of power consumption. Similarly, historical signatures are collected for other mentioned high-power consuming appliances. These historical signatures also reveal the nature of consumption, the usual time of day during which such an appliance is used, and how much power is consumed, therefore. These are the latent features of the consumption data.
[044] Referring to steps of FIG. 2, at step 208 of the present disclosure, the one or more hardware processors learn, by using a matrix factorization framework, a mapping matrix M between the data representations H of the set of corresponding instances S and the appliance signature matrix L, wherein the learnt mapped matrix corresponds to a disaggregated data of one or more appliance of interests from the set of appliances. More specifically, by using the matrix factorization framework, the system 100 equates the appliance matrix with say S = 30 (and N = 1400, 750, 180, etc. for different cases) to the obtain data representations of the size K = 1300, 700, 150, etc. for different cases through a mapping function (refer equation (5) below). This mapping function is obtained using least square solver and is of the size S(=30)*K(= 1300, 700, 150, etc. for different cases.)
[045] At step 210 of the present disclosure, the one or more hardware processors 104 obtain one or more disaggregate estimates for one or more appliances of interest based on the learnt mapping matrix M.
[046] The above steps 202 till 210 are better understood by way of following description. Suppose the aggregate data, is say X, and is time-windowed so that each instance is of length N. Further it is assumed that S such instances are considered to obtain representations using the RBM. The input to RBM is therefore of size S × N. If the number of hidden units are K in number, RBM representations H of size S × K are obtained using equation (2).
[047] For S instances of aggregate data and P appliances of interest for
disaggregation, s­p instances are needed for each appliance of interest that span the appliance’s entire range of consumption behaviors. The appliance consumption

behaviors are time-windowed, so that aggregate data set and the appliance signatures have same length N. The appliance signature matrix, L, would therefore be of size S×N. Using the following equation, the data representation H of each instance from the set of corresponding instances S is expressed using the equation:

wherein hj is the hidden unit in RBM, p is a probability measure, wherein vi is a visible layer of the RBM, wherein σ is sigmoid function, bj is a bias associated with the RBM, and wherein wij corresponds to one or more weights of one or more hidden units in the RBM.
[048] In another words, the data representation H of each instance from the set of corresponding instances S is expressed using the above equation.
[049] In an embodiment, the leant mapping matrix M between the data representations H of the set of corresponding instances S and the appliance signature matrix L is based on the equation: M ← argmin\\L - MH\\2F, and
M
wherein F is a Euclidean norm.
[050] The disaggregated estimates are then obtained using the equation:
Lest S×N = HS×K * MK×N (6)
In other words, the learnt mapping matrix M between the data representations H of the set of corresponding instances S and the appliance signature matrix L is based on the above equation, where K is number of hidden units in the RBM.
[051] It is observed that the formulation as described by the present disclosure is computationally light as it requires learning of only one learnt mapping matrix M as opposed to the need to learn separate models for different appliance signatures (refer conventional research works “Nipun Batra, Hongning Wang, Amarjeet Singh, and Kamin Whitehouse, “Matrix factorization for scalable energy breakdown,” in Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence. 2017, AAAI’17, p. 44674473, AAAI Press.”, “J. Z. Kolter, Siddharth Batra, and Andrew Y. Ng, “Energy disaggregation via discriminative sparse coding,” in Advances in Neural Information Processing Systems 23, J. D. Lafferty,

C. K. I. Williams, J. Shawe-Taylor, R. S. Zemel, and A. Culotta, Eds., pp. 1153– 1161. Curran Associates, Inc., 2010.”, and “Jack Kelly andWilliam Knottenbelt, “Neural nilm: Deep neural networks applied to energy disaggregation,” in Proceedings of the 2nd ACM International Conference on Embedded Systems for Energy-Efficient Built Environments, 2015, pp. 55–64.”). FIG. 3 provides the details of the framework. More specifically, FIG. 3, with reference to FIGS. 1-2, depicts a Restricted Boltzmann Machine-Least Squares framework for Load Disaggregation, in accordance with an embodiment of the present disclosure. The entire framework as depicted in FIG. 3 is therefore a two-step implementation:
1. Representational learning of new data using RBM.
2. Learning the mapping matrix M against these representations and the historical appliance signature matrix.
[052] The different components of the framework depicted in FIG. 3 are obtained as explained below:
1. Aggregate data windows: A matrix of time-windowed aggregate data, the window length selected based on careful inspection of the reconstruction efficiency of the RBM for that specific time-window of consideration.
2. Representational Learning: Output of a single hidden layer RBM, the input to which is the above-mentioned aggregate data matrix.
3. Appliance load matrix: In most cases, historical appliance consumption signatures are available, from which signatures for requisite appliances of interest are drawn and the necessary time-windowed appliance signatures matrix is constructed for the method as described herein. Any further change in consumption patterns of an appliance, new appliance of interest or a new appliance make of interest can be readily incorporated into this matrix.
4. Selecting an objective function: An appropriate objective function that yields right convergence for the mapping function A soft-max layer with a Mean Squared Error (MSE) criterion may not suffice for such supervised regressive learning, hence the system and method adopt Least Squares minimization.

[053] The pseudo code for the method described in FIG. 2 and as implemented by the system 100 and the framework depicted in FIG. 3 is illustrated by way of example below. More specifically, below is an exemplary pseudo code for obtaining/learning mapping matrix M between data representations, H and appliance signatures matrix, L: Step 1: HV←X
Step 2: M ← argmin‖L - MH2‖F.
M
Step 3: Lest S × N=HS×K * MK×N
EXPERIMENTAL EVALUATION:
[054] The implementation procedure of the framework implemented by the system and method of the present disclosure is summarized in above pseudo code. The system of the present disclosure evaluated the method described herein on two publicly available datasets - UK Dale and AMPds. UK Dale consists of aggregate and individual appliance power consumption sampled at 6 seconds interval (e.g., refer Kelly et al.) and AMPds dataset contains whole meter and individual appliance consumption data at one-minute interval (e.g., refer Makonin et al.).
[055] Data was down-sampled to one-minute sampling for UK DALE dataset and hence one-minute sampled data of both houses was used to evaluate the framework implemented by the system and method herein. A combination of high-power consuming and most commonly used appliances was considered for both the datasets. Only few appliance reconstruction results are presented due to space constraints. The list of appliances that were considered for disaggregation for each house is listed in Table 1 (which depicts details of appliances considered for experimental evaluation).
Table 1

Appliances considered in various datasets
AMPds UKDale House 4 UKDale House 1
Washer-Dryer Wash-Oven-Bread Washing Machine

Electric Heater Gas Boiler Refrigerator
Microwave Oven Freezer Microwave Oven
[056] Time windows of observation need to be selected individually for different data sets; this is because each house exhibits a pattern over a certain time window. For instance, the system and method had to choose one day duration for UK Dale House 1, 12 hours of duration for UK Dale House 4 and three hours of duration for AMPds data. The framework described herein was implemented separately on 40 days of data (for both the data sets) as detailed in the above pseudo code. FIGS. 4 through 6 depict a graphical representation illustrating appliance reconstruction results for various appliance datasets, in accordance with an example embodiment. More specifically, FIG. 4, with reference to FIGS. 1 through 3, depicts a graphical representation illustrating refrigeration reconstruction results for UK Dale - House 1 and House 4 and AMPds datasets, in accordance with an example embodiment. FIG. 5, with reference to FIGS. 1 through 4, depicts a graphical representation illustrating washer-dryer reconstruction results for UK Dale - House 1 and House 4 and AMPds datasets, in accordance with an example embodiment. FIG. 6, with reference to FIGS. 1 through 5, depicts a graphical representation illustrating oven reconstruction results for UK Dale - House 1 and House 4 and AMPds datasets, in accordance with an example embodiment. Accurate reconstructions were obtained for all the appliances in all the datasets considered. However, certain false positives were also observed, only in regions where no loads are operational, which reduces the overall load identification accuracy. The effect of false positives on load identification accuracy metric is dealt with in detail below. While such cases have been few, it has been observed by the system and method of the present disclosure that the use of regularizer terms may help in diminishing such errors in such cases. COMPARISON:
[057] The system and method of the present disclosure present the accuracy of disaggregation in terms of the RMS of Percentage of Energy Correctly

assigned (RMS of PEC) metric defined in Batra et. al. The RMS of PEC is given by the following equation:
(7)
where A stands for Appliance under consideration and D for the time-duration considered. Having computed PEC, the system and method of the present disclosure take the RMS of PEC to obtain the error of disaggregation. The system and method of the present disclosure chose the most commonly used appliances -Washer-Dryer, Refrigerator and Microwave for comparing the disaggregation results of the RBM-MF method as described herein. Some of these conventional works used for comparison use a matrix factorization formulation, while the others are representational learning-based methods and hence are appropriate for comparison with the performance of the RBM-MF approach. The results are listed in Table 2. Table 2 below depicts RMS error in percentage of error correctly assigned. Table 3 below depicts performance of RBM-MF in load identification.
Table 2

Appliance RBM-MF MF (prior MLC-RBM MLAE (prior
(method of the
present
disclosure) art) (prior art) art)
Washer-Dryer 0.1 0.49 0.19 0.1
Refrigerator 0.02 3.65 0.164 0.1
Microwave 0.01 0.64 0.1985 0.25
Table 3

Appliance Load identification accuracy
Washer-Dryer 0.99
Refrigerator 0.879
Microwave 0.3
Freezer 0.81
Electric Heater 0.99

[058] Other appliances specific to the houses considered by the present disclosure were Electric Heater and Freezer. The RMS of PEC for these appliances are 0.089 and 0.0026 respectively. These appliances were not available for comparison in the results provided in the conventional works and hence are reported independently.
[059] The system and method of the present disclosure also present the performance of RBM-MF in load identification using Coefficient of Determination regression score function (r2 score) as in the Table 3, as this metric suits the regressive nature of the formulation. Poor r2 scores are observed for Microwave appliance although the RMS of PEC for the appliance is quite low. This is attributed to the existence of false positives in the appliance reconstructions which reduce the r2 scores for the same drastically. Also, while microwave is a high-power consuming appliance, the time-duration for which this consumption occurs is quite small. This aspect also induces reconstruction in non-operational time durations.
[060] It is therefore imperative that the r2 score should be considered together with RMS of PEC metric as the latter accounts for appliance disaggregation error normalized with respect to aggregate data and hence is an accurate representative of the performance of the proposed formulation in such scenarios.
[061] System and method of the present disclosure provide a computationally simple and reliable RBM assisted Matrix Factorization framework for Load Disaggregation. The framework can be adopted to different set of appliances of interest, depending on the geographical context. The use of regularizer terms to reduce false positives is expected to shed more light on the means to make this formulation robust for automated disaggregation results.
[062] The written description describes the subject matter herein to enable any person skilled in the art to make and use the embodiments. The scope of the subject matter embodiments is defined by the claims and may include other modifications that occur to those skilled in the art. Such other modifications are intended to be within the scope of the claims if they have similar elements that do

not differ from the literal language of the claims or if they include equivalent elements with insubstantial differences from the literal language of the claims.
[063] It is to be understood that the scope of the protection is extended to such a program and in addition to a computer-readable means having a message therein; such computer-readable storage means contain program-code means for implementation of one or more steps of the method, when the program runs on a server or mobile device or any suitable programmable device. The hardware device can be any kind of device which can be programmed including e.g., any kind of computer like a server or a personal computer, or the like, or any combination thereof. The device may also include means which could be e.g., hardware means like e.g., an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or a combination of hardware and software means, e.g., an ASIC and an FPGA, or at least one microprocessor and at least one memory with software processing components located therein. Thus, the means can include both hardware means and software means. The method embodiments described herein could be implemented in hardware and software. The device may also include software means. Alternatively, the embodiments may be implemented on different hardware devices, e.g., using a plurality of CPUs.
[064] The embodiments herein can comprise hardware and software elements. The embodiments that are implemented in software include but are not limited to, firmware, resident software, microcode, etc. The functions performed by various components described herein may be implemented in other components or combinations of other components. For the purposes of this description, a computer-usable or computer readable medium can be any apparatus that can comprise, store, communicate, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device.
[065] The illustrated steps are set out to explain the exemplary embodiments shown, and it should be anticipated that ongoing technological development will change the manner in which particular functions are performed. These examples are presented herein for purposes of illustration, and not limitation. Further, the boundaries of the functional building blocks have been arbitrarily

defined herein for the convenience of the description. Alternative boundaries can be defined so long as the specified functions and relationships thereof are appropriately performed. Alternatives (including equivalents, extensions, variations, deviations, etc., of those described herein) will be apparent to persons skilled in the relevant art(s) based on the teachings contained herein. Such alternatives fall within the scope of the disclosed embodiments. Also, the words “comprising,” “having,” “containing,” and “including,” and other similar forms are intended to be equivalent in meaning and be open ended in that an item or items following any one of these words is not meant to be an exhaustive listing of such item or items, or meant to be limited to only the listed item or items. It must also be noted that as used herein and in the appended claims, the singular forms “a,” “an,” and “the” include plural references unless the context clearly dictates otherwise.
[066] Furthermore, one or more computer-readable storage media may be utilized in implementing embodiments consistent with the present disclosure. A computer-readable storage medium refers to any type of physical memory on which information or data readable by a processor may be stored. Thus, a computer-readable storage medium may store instructions for execution by one or more processors, including instructions for causing the processor(s) to perform steps or stages consistent with the embodiments described herein. The term “computer-readable medium” should be understood to include tangible items and exclude carrier waves and transient signals, i.e., be non-transitory. Examples include random access memory (RAM), read-only memory (ROM), volatile memory, nonvolatile memory, hard drives, CD ROMs, DVDs, flash drives, disks, and any other known physical storage media.
[067] It is intended that the disclosure and examples be considered as exemplary only, with a true scope of disclosed embodiments being indicated by the following claims.

We Claim:
1. A processor implemented method, comprising:
receiving, via one or more hardware processors, an aggregated power consumption data corresponding to a set of appliances, wherein the aggregated power consumption data is time windowed comprising a set of corresponding instances S of a predefined length N (202);
generating, by using a Restricted Boltzmann Machine (RBM) via the one or more hardware processors, a data representation H of the set of corresponding instances S associated with the aggregated power consumption data (204);
obtaining, via the one or more hardware processors, an appliance signature matrix L for the set of appliances based on a set of time windowed historical appliance consumption signatures (206);
learning, by using a matrix factorization framework via the one or more hardware processors, a mapping matrix M between the data representations H of the set of corresponding instances S and the appliance signature matrix L (208); and
obtaining one or more disaggregate estimates for one or more appliances of interest based on the learnt mapping matrix M (210).
2. The processor implemented method of claim 1, wherein the data
representation H of each instance from the set of corresponding instances S is
expressed using the equation:
p(hj = 1vi ) = (r(bj+Σi(vi wij )), wherein hj is the hidden unit in RBM, p is a probability measure, wherein vi is a visible layer of the RBM, wherein σ is sigmoid function, bj is a bias associated with the RBM, and wherein wtj corresponds to one or more weights of one or more hidden units in the RBM.
3. The processor implemented method of claim 1, wherein the learnt mapping
matrix M between the data representations H of the set of corresponding instances
S and the appliance signature matrix L is based on the equation:

M ← argmin\\L - MH\\2F, and wherein F is a Euclidean norm.
M
4. The processor implemented method of claim 1, wherein the predefined length N is selected based on a reconstruction efficiency of the RBM for a specific time window.
5. The processor implemented method of claim 1, wherein the one or more disaggregate estimates are obtained based on the equation:
LestS×N = HS×K * MK×N
wherein K is number of hidden units in the RBM.
6. A system (100), comprising:
a memory (102) storing instructions;
one or more communication interfaces (106); and
one or more hardware processors (104) coupled to the memory (102) via the one or more communication interfaces (106), wherein the one or more hardware processors (104) are configured by the instructions to:
receive an aggregated power consumption data corresponding to a set of appliances, wherein the aggregated power consumption data is time windowed comprising a set of corresponding instances S of a predefined length JV;
generate, by using a Restricted Boltzmann Machine (RBM), a data representation H of the set of corresponding instances S associated with the aggregated power consumption data;
obtain an appliance signature matrix L for the set of appliances based on a set of time windowed historical appliance consumption signatures;
learn, by using a matrix factorization framework, a mapping matrix M between the data representations H of the set of corresponding instances S and the appliance signature matrix L; and
obtain one or more disaggregate estimates for one or more appliances of interest based on the learnt mapping matrix M.

7. The system of claim 5, wherein the data representation H of each instance
from the set of corresponding instances S is expressed using the equation:
p(hj = 1\vi) = σ(bj+Σi(viwij)), wherein hj is the hidden unit in RBM, p is a probability measure, wherein vi is a visible layer of the RBM, wherein σ is sigmoid function, bj is a bias associated with the RBM, and wherein wtj corresponds to one or more weights of one or more hidden units in the RBM.
8. The system of claim 6, wherein the learnt mapping matrix M between the
data representations H of the set of corresponding instances S and the appliance
signature matrix L is based on the equation:
M ← argmin\\L - MH\\2F, and wherein F is a Euclidean norm.
M
9. The system of claim 6, wherein the predefined length N is selected based on a reconstruction efficiency of the RBM for a specific time window.
10. The system of claim 6, wherein the disaggregate estimates are obtained based on the equation:
LestS×N = HS×K * MK×N
wherein K is number of hidden units in the RBM.