Method and system for low sampling rate electrical load disaggregation
This disclosure relates generally to method and system for low sampling rate electrical load disaggregation. At low sampling rates, disaggregation of energy load is challenging due to unavailability of events and signatures of the constituent loads. The disclosed energy disaggregation technique receives aggregated load data from a utility meter and sequentially obtains training data for determining disaggregated energy load at low sampling rate. Dictionaries are used to characterize the different loads in terms of power values and time of operation. The obtained dictionary coefficients are treated as graph signals and graph smoothness is used for propagating the coefficients from the training phase to the test phase by formulating an optimization model. The derivation of the optimization model identifies the load of interest and estimate their power consumption based on optimization model constraints. This method achieves accuracy greater than 70% for the loads of interest at low sampling rates.
Latest Tata Consultancy Services Limited Patents:
 Systems and methods for assessing quality of input text using recurrent neural networks
 DISCRIMINATING FEATURES BASED SEPSIS PREDICTION
 METHOD AND SYSTEM FOR HIERARCHICAL TIMESERIES CLUSTERING WITH AUTO ENCODED COMPACT SEQUENCE (AECS)
 Systems and methods for generating control system solutions for robotics environments
 Method and system for augmented reality based liquid measurement and mapping
The U.S. patent application claims priority under 35 U.S.C § 119 to Indian patent application no. (201921022305), filed on Jun. 5, 2019.
TECHNICAL FIELDThe disclosure herein generally relates to utility consumption, and, more particularly, to method and system for low sampling rate electrical load disaggregation.
BACKGROUNDIn modern power systems, smart meter data analytics is crucial for various entities involved in energy management. Present power sector is witnessing demandside deregulations across the world, coupled with the ongoing deployment of smart meters which provides energy consumption data at regular intervals. Electrical load disaggregation is one of the smart meter analytics techniques which provide information about the usage and consumption patterns of the individual loads using the aggregate power consumption measurements. It can be leveraged for a variety of functionalities and value additions in the emerging smart grid scenario. Many existing and emerging smart meters provide low granular aggregate power measurements at 15 or 30minutes intervals. However, for these low sampling rates, disaggregation of loads is tedious due to unavailability of events and signatures of the constituent loads. Hence, there is a necessity for a technique to analyze this low sampled aggregated power data to provide valuable inputs and insights for different entities in the ecosystem like, consumers, aggregators, retailers and utilities.
While disaggregation of loads using high power sampled data is studied extensively in literature, limited results are available for the low sampled case. Most of the recent methods for low sampled power data utilize the emerging concepts of Graph Signal Processing (GSP) for load disaggregation. One method is based on greedily solving the graph smoothness term for load identification from 1minute sampled aggregate power data. Although the detection accuracy is good, the consumption estimates are poor as they are merely the average consumption of the load. Another method uses graph signal smoothnessbased label propagation for load identification and actual power consumption estimation for 15 minutes aggregate power data. Both these methods use the assumption that the load signals are smooth with respect to the graph created using aggregate power measurements to carry out load disaggregation. Another method on the similar lines, utilizes matrix factorization and graph shift quadratic form constraint to carry out load disaggregation. However, this method estimates the individual load signals as piecewise constant values instead of real values. Thus, conventional techniques lack determining disaggregated energy loads at low sampling frequency.
SUMMARYEmbodiments of the present disclosure present technological improvements as solutions to one or more of the abovementioned technical problems recognized by the inventors in conventional systems. For example, in one embodiment, a system for is provided. The system includes a processor, an Input/output (I/O) interface and a memory coupled to the processor is capable of executing programmed instructions stored in the processor in the memory to preprocess to receive a aggregated load data (
In another aspect, provides a method that includes a processor, an Input/output (I/O) interface and a memory coupled to the processor is capable of executing programmed instructions stored in the processor in the memory to preprocess receive a aggregated load data (
In yet another aspect, provides one or more nontransitory machinereadable information storage mediums comprising one or more instructions, which when executed by one or more hardware processors perform actions includes preprocessing receive a aggregated load data (
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:
Exemplary embodiments are described with reference to the accompanying drawings. In the figures, the leftmost digit(s) of a reference number identifies the
The embodiments herein provide a method and a system, interchangeably referred as energy load disaggregation system for determining the disaggregated energy of the electrical load(s) or appliance (s) from the power data at low sampling rate. The electrical load may also be referred as aggregated power data. The aggregate power data is the total amount of energy load consumed by all the electrical appliances configured to the corresponding utility meter among a plurality of utility meters. The method disclosed determines energy consumed by all the individual appliance(s) of a house or a building from the aggregate power data using an energy disaggregation technique. The term ‘disaggregated energy load’ may refer to an amount of energy that may be consumed by a single or individual appliance configured to the corresponding utility meters available in the electrical network. Further, as used in the present disclosure, the term “load or appliance” may refer to any device that is capable of consuming electricity. For example, it may be a dish washer, refrigerator, dryer or any other device which consumes energy.
In an embodiment to effectively address 15 or 30minutes energy load disaggregation, and to obtain a good estimate of load consumption, the present disclosure presents an analytical formulation blending the sparse dictionary representation and graph signal smoothness based on label propagation for energy load disaggregation. The dictionaries learn individual load characteristics in terms of magnitude and time of operation. The coefficients select the right dictionary columns (atoms) and scale them appropriately to facilitate the discrimination of loads. The graphsignal based label propagation identifies the timings of load operation in a more accurate way though having small amount of training data and further refines the individual load estimation. This technique achieved an accuracy greater than or equal to 70% for the loads of interest. The disclosed technique performs better accuracy than the standard GSP and dictionarybased methods which are known in the art. This technique receives aggregate power data at predefined regular intervals for the duration of N days along with individual power consumed by the M appliances for n days at a low rate (T samples in a day).
Referring now to the drawings, and more particularly to
The I/O interface(s) 206 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 for verifying software code.
The memory 208 may include any computerreadable 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 nonvolatile memory, such as read only memory (ROM), erasable programmable ROM, flash memories, hard disks, optical disks, and magnetic tapes. The memory 208 further may include modules 210. In an embodiment, the modules 210 include an energy load disaggregation module 212, for implementing functions of the system 102. In an embodiment, the modules 210 can be an Integrated Circuit (IC) (not shown), external to the memory 208, implemented using a FieldProgrammable Gate Array (FPGA) or an ApplicationSpecific Integrated Circuit (ASIC). The names (or expressions or terms) of the modules of functional block within the modules 210 referred herein, are used for explanation and are not construed to be limitation(s). Further, the memory 208 can also include the repository 214. The repository 214 may store the training data. The memory 208 may further comprise information pertaining to input(s)/output(s) of each step performed by the system 10 and methods of the present disclosure. The system 102 can be configured to process the aggregated load data (
The steps of the method 300 of the flow diagram will now be explained with reference to the components or blocks of the system 100 in conjunction with the example architecture of the system as depicted in
At step 302 of the method 300, the processor 204 is configured to receive a aggregated load data (
p(i)=Σ_{m=1}^{M}Pm(i)+n(i) (1)
where,
p(i) is the aggregate power measurement sampled at time instants i=1, 2, 3 . . . , N.
Pm(i) are the individual loads for m∈M contributed to that measurement and
n(i) is the measurement noise which includes the base loads (combined low power loads).
In an embodiment, the system 102 is configured to receive the aggregated load data (
At step 304 of the method 300, the processor 204 is configured to represent, the aggregate load data (
where, B includes the base loads and other loads not considered in M. For each of the individual loads X_{m }a sparse representation using the corresponding overcomplete dictionary D_{m }is learnt. Since we are considering the power consumption, nonnegative sparse coding version of KSVD is used to learn sparse representation of the loads which leads to the following formulation as represented in equation 3,
where, D_{m}∈R^{T*K}, is the dictionary with K atoms and Z_{m}^{train}∈R^{K*n }are the dictionary coefficients or activations. The second term, ∥Z_{m}^{train}∥_{0 }is the l_{o }sparsity measure which ensures a sparse representation of data is learnt using at most τ atoms (columns) of the dictionary. In the actual implementation of the optimization step, the convex surrogate, l_{1 }norm of the coefficients is used instead of the l_{0 }norm.
At step 306 of the method 300, the processor 204 is configured to determine the disaggregated load energy at low sampling rate for the aggregated load data (
Total load consumption (
For the given aggregated load data (
where,
Further, the method constructs a Laplacian graph (L_{m}) for each electrical load (m) using the total load consumption (
Further, the method formulates, an optimization model for estimating the dictionary coefficient (Z_{m}^{test}) of each electrical loads (m). The first term of the optimization model as specified in equation 5, estimates the coefficients using the learnt dictionaries such that the error between
Further, the method, constructs a graph using the total load consumption (
Since the dictionaries for the electrical loads are computed only once at the time of training and are considered fixed, the graph smoothness term helps in accurately picking up the timing instants of the electrical load operation in the test phase. This addresses the problem of capturing the random switching of loads mentioned earlier. A graph of N nodes is constructed using the aggregate power measurements
W(i,j)=exp(−dist(i,j)^{2}/σ^{2}) (6)
where, σ is a scaling factor and (dist(i,j) is the distance measure between nodes i and j which can be chosen based on the application. Since a vector resides on each node, here, Euclidean distance is considered to compute the weights of the graph. The graph Laplacian L is obtained as: L=D^{G}−W, where, D^{G }is the degree matrix and W is the weight matrix obtained from equation 5. The rationale behind using the graph is that, similar load operations will have similar activations or coefficients and the same can be propagated from the training to the test phase. This enables the timing instants of the load operation to be picked up more accurately. Since the coefficients Z_{m}^{train }are known for the training data, the optimization model represented in equation 5 and is solved for Z_{m}^{test }in a greedy manner by estimating one load at a time. The optimization model problem is reformulated as mentioned in equation 7,
Constraints of Optimization Model,

 1. estimating, the dictionary coefficient (Z_{m}^{test}) using the dictionary (D_{m}) such that error between (
X _{test}) and the product of the dictionary (D_{m}) and the dictionary coefficient (Z_{m}^{test}) is minimized and  2. determining, the dictionary coefficient (Z_{m}^{test}) for load m such that the total variation over the graph topology (L_{m}) is minimized, in other words, the coefficients are smooth with respect to the graph.
The graph smoothness in equation 8 is expressed as:
Z_{m}L_{m}Z_{m}^{T}=Z_{m}(:,1:n)L_{m}(1:n,1:n)Z_{m}(:,1:n)^{T}+Z_{m}(:,1:n)L_{m}(1:n,1:n+1:N)Z_{m}(:,n+1:N)^{T}+Z_{m}(:,n+1:N)L(n+1:N,1:n)Z_{m}(:,n+1:n)^{T}+Z_{m}(:,n+1:N)L_{m}(n+1:N,n+1:N)Z_{m}(:,n+1:N)^{T} (8)
Where, Z_{m}(:,1:n)=Z_{m}^{train }and Z_{m}(:,n+1:N)=Z_{m}^{test}. This is an extension of the graph smoothness term expansion to accommodate vector valued graph signals. Since, Z_{m}^{train }is known L_{m }is symmetric, minimizing the graph smoothness term results in represented in equation 9,
 1. estimating, the dictionary coefficient (Z_{m}^{test}) using the dictionary (D_{m}) such that error between (
Solving for Z_{m}^{test }by taking a derivative of equation 7 with respect to Z_{m}^{test }and equating it to zero results in Sylvester equation (5) of the form as represented in equation 10,
D_{m}^{T}D_{m}Z_{m}^{test}+βZ_{m}^{test}L_{m}(n+1:N,n+1:N)=D_{m}^{T}
Since the graph construction at a given iteration is more influenced by the highest power consuming load, the latter is disaggregated first treating the contribution of other loads as interference.
Further, the method estimates load consumption () for each electrical load (m) among the plurality of electrical loads (M) using the estimated energy coefficient (Z_{m}^{test}) with the corresponding dictionary (D_{m}) as represented in equation 11,
D_{m}Z_{m}^{test} (11)
Once the aggregated load data is disaggregated its contribution is removed from the aggregate power measurement for
The
and the procedure repeats for the next load till all the loads of interest are disaggregated.
In an embodiment, the potential of the disclosed methodology for load disaggregation is demonstrated using two datasets, the simulated one using Home Energy Simulator (HES) and the public UMass Smart Home Dataset, with power data sampled at 15 and 30 minutes respectively. Here, dictionaries for the different loads are learnt using the KSVD toolbox in Matlab with K=100 atoms and sparsity threshold T=20. The value of β is chosen to provide more importance to the graph smoothness term while estimating the load activations. The disaggregation results are evaluated for both load identification and consumption estimation. The wellknown F1 score, which is the harmonic mean of precision and recall is used for load identification. For load consumption estimation, normalized Root Mean Squared Error (NRMSE) and normalized Mean Absolute Error (NMAE) metrics are considered to evaluate the waveform reconstruction capability of the respective loads.
These performance metrics are defined as represented in equation 14,
Where, Precsision=TP/(TP+FP) and Recall=TP/(TP+FN)
TP=true positives, FP=false positives and FN=false negatives
The normalized Root Mean Squared Error (NRMSE) is represented in equation 15,
The normalized Mean Absolute Error (NMAE) metrics is represented in equation 16,
where X_{act }is the actual and X_{pred }is the estimated power consumption of the load. The normalized values are obtained by dividing these metrics by the average power consumption of the respective loads. electrical loads (m) with distinct power levels are only considered for disaggregation. Base load (combined low power) removal is not considered here as they may remove the loads of interest and adversely affect the accuracy of both load detection and estimation. In addition, the average percentage accuracy of the load consumption estimation is provided for comparative analysis. It is known, in the low sampling regime it is difficult to disaggregate all possible loads.
Reasonable detection and consumption accuracy of major power consuming loads can provide valuable insights to both consumers and utilities. Hence, major power consuming loads with distinct power levels are considered for disaggregation. Since the suggested methodology systematically combine disaggregation vis discriminative sparse coding (DDSC) and graph signal processing (GSP) in the directions of providing comparative results for the available prior arts are evaluated for both the data sets.
In an embodiment, the disclosed method is able to disaggregate only three loads namely, Dryer, Dishwasher and Refrigerator with good accuracy. The identification in terms of F1 score and consumption estimates in terms of average percentage accuracy (% Acc) obtained by all the three methods for identification and consumption estimates obtained as represented as mentioned below in Table 1,
It can be seen, both the GSP technique and the disclosed method have similar detection accuracy (F1 score), however, the consumption accuracy of latter is higher. This is because the dictionary representation helps in better load waveform reconstruction. On the other hand, the DDSC method on its own has low detection accuracy since it requires enormous historical data to capture different possible operational instants of loads to perform effectively. Even with the reasonable amount of training data considered for experimentation, the technique resulted in poor accuracy. The combination of Dictionary representation and the GSP technique resulted in improved accuracy metrics.
For this case also, the disclosed technique performed better than the GSP method both in terms of identification and consumption estimation. The performance of DDSC method is very poor and is not considered for comparison.
In one embodiment, the energy loads are disaggregated in an iterative manner starting from the highpower consuming load, false negatives leads to false positives for the next load being considered and viceversa. To reduce this error propagation, appropriate post processing is done on the load estimation. Here, a threshold is applied on the load estimations, where the threshold is half of the maximum magnitude of the load power waveform. The load values above the threshold is considered as valid estimations. This simple step results in a decent identification of loads with distinct power levels. For the case of continuously ON loads, although the contributions of other similar power loads are subsumed, still the algorithm can detect them due to their regularity of operation. It is worth noting that, the disclosed optimization formulation exploits the energy consumption behavior of the consumer. It estimates the load activations based on the similar power consumption profile seen in the training phase. Hence, it is able to accurately pick the timing instants of load operation even when these instances are not seen in the training. This leads to an advantage in terms of working with less amount of training data. The accuracy of load consumption estimation in terms of NRMSE and NMAE obtained with the disclosed method for both the datasets are summarized in table 3.
As is evident from the results (Tables 13), the accuracy of load identification and consumption estimation goes down as the sampling intervals increases from 15 minutes to 30 minutes. This is because at those sampling instants, some of the loads may exhibit similar power values leading to more complex superpositions. It is useful to note that the deterioration of the disclosed method is less compared to GSP technique, when we move from 15 minutes to 30 minutes data. This justifies the importance of having a suitable representation together with GSP to effectively handle disaggregation of lowsampled data. Further, the good estimation and consumption accuracies achieved through this method can facilitate load profiling of consumers and their segmentation in a more effective manner.
The embodiments of present disclosure herein addresses unresolved problem of low sampling rate electrical load disaggregation. The embodiment thus provides the necessity of systematically combining signal processing and machine learning based on mathematical formulations to arrive at algorithms for load disaggregation from low sampled power data. Moreover, the embodiments herein further provide an appropriate data/signal representation coupled with graph signal processing for low sampled load disaggregation. Here, the concepts of dictionary learning and graph signal processing are combined to arrive at a novel methodology for disaggregating major power consuming loads in residential buildings. Dictionaries are used to characterize the different loads in terms of power values and time of operation. The coefficients thus obtained are treated as graph signals and graph smoothness is used to propagate the coefficients from the known training phase to the test phase. The present disclosure provides the optimization formulation, the derivation of the requisite solution steps to identify the loads of interest and estimate their power consumption. Towards demonstrating the usefulness of the technique, typical results for both simulated and real data are provided.
It is to be understood that the scope of the protection is extended to such a program and in addition to a computerreadable means having a message therein; such computerreadable storage means contain programcode 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 applicationspecific integrated circuit (ASIC), a fieldprogrammable 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.
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 computerusable 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.
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.
Furthermore, one or more computerreadable storage media may be utilized in implementing embodiments consistent with the present disclosure. A computerreadable storage medium refers to any type of physical memory on which information or data readable by a processor may be stored. Thus, a computerreadable 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 “computerreadable medium” should be understood to include tangible items and exclude carrier waves and transient signals, i.e., be nontransitory. Examples include random access memory (RAM), readonly memory (ROM), volatile memory, nonvolatile memory, hard drives, CD ROMs, DVDs, flash drives, disks, and any other known physical storage media.
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.
Claims
1. A processor implemented (102) method for electrical load disaggregation, wherein the method comprises:
 receiving, by a processor (204), an aggregated load data (Xtest), obtained from a utility meter among a plurality of utility meters at a plurality of regular time intervals, wherein the aggregated load data comprises total energy consumed by a plurality of electrical loads (m) connected to a utility meter and base loads which are low power loads; and a training data (Xtrain), wherein the training data is generated from the plurality of utility meters;
 representing, by the processor (204), the aggregate load data (Xtest) and the training data (Xtrain) as a matrix, where a column vector represents the plurality of regular time intervals covered in a predefined time period and a row vector represents a total number of predefined time periods considered; and
 determining, by the processor (204), from the aggregated load data (Xtest) of the utility meter, a disaggregated energy load () at low sampling rate, using an energy disaggregation technique comprising: obtaining, a total load consumption (X) using the aggregated load data (Xtest) and the training data (Xtrain); constructing, a laplacian graph (Lm) for each of the plurality of electrical load (m) using the total load consumption (X); formulating, an optimization model for estimating the dictionary coefficient (Zmtest) of each of the plurality of electrical loads (m); estimating, the dictionary coefficient (Zmtest) of each electrical loads (m), using the laplacian graph (Lm), a dictionary (Dm) and a dictionary coefficient (Zmtrain), wherein, the dictionary (Dm) and the dictionary coefficient (Zmtrain) are obtained for each of the electrical load (m) using a pretrained dictionary, and wherein, the dictionary is pretrained using the dictionary learning for each electrical load (m) load using a load specific consumption (Xm) corresponding to the electrical load (m); estimating, the load consumption () for each electrical load (m) among the plurality of electrical loads (M) using the estimated energy coefficient (Zmtest) with the corresponding dictionary (Dm); and iteratively, performing for each electrical load (m) among the plurality of electrical loads (M) steps of: recomputing, a new aggregated load data (Xtest) by subtracting the estimated load energy consumption () from the previous load consumption (Xtest) value; and recomputing, the new total load consumption (X) by subtracting the estimated load energy consumption () and the load specific consumption (Xm) from the previous total load consumption (X).
2. The method as claimed in claim 1, wherein the optimization model for estimating the dictionary coefficient (Zmtest) of each electrical loads (m) by,
 construct, a graph using the total load consumption (X), considering the dictionary coefficient (Zm) as the graph signal.
3. The method as claimed in claim 2, wherein the optimization model constraints includes,
 estimating, the dictionary coefficient (Zmtest)) using the dictionary (Dm) such that error between (Xtest) and the product of the dictionary (Dm) and the dictionary coefficient (Zmtest)) is minimized; and
 determining, the dictionary coefficient (Zmtest) for each electrical load (m) such that the total variation over the graph topology (Lm) is minimized.
4. The method as claimed in claim 2, wherein the solution to the optimization model is determined by taking derivative of the objective function with respect to the dictionary coefficient (Zmtest)) and equating it to zero.
5. The method as claimed in claim 1, wherein the load consumption () is estimated based on the product of estimated energy coefficient (Zmtest) with the corresponding dictionary (Dm).
6. A system (102) for electrical load disaggregation, the system (102) comprising:
 a memory (208) storing instructions;
 one or more Input/output (I/O) interfaces (206);
 and one or more hardware processors (204) coupled to the memory (208) via the one or more 1/O interfaces (206), wherein the one or more hardware processors (204) are configured by the instructions to: receive from the energy load disaggregation module (212), an aggregated load data (Xtest), obtained from a utility meter among a plurality of utility meters at a plurality of regular time intervals, wherein the aggregated load data comprises total energy consumed by a plurality of electrical loads connected to the corresponding utility meter and base loads which are low power loads; and a training data (Xtrain), wherein the training data is generated from the plurality of utility meters; represent, the aggregate load data (Xtest) and the training data (Xtrain) as a matrix, where a column vector represents the plurality of regular time intervals covered in a predefined time period and a row vector represents a total number of predefined time periods considered; and determine, a disaggregated energy load at low sampling rate for the aggregated load data (Xtest) of each utility meter, using an energy disaggregation technique comprising: obtain, a total load consumption (X) using the aggregated load data (Xtest) and the training data (Xtrain); construct, a laplacian graph (Lm) for each electrical load (m) using the total load consumption (X); formulate, an optimization model for estimating the dictionary coefficient (Zmtest) of each electrical loads (m); estimate, the dictionary coefficient (Zmtest)) of each electrical loads (m), using the laplacian graph (Lm), a dictionary (Dm) and a dictionary coefficient (Zmtest), wherein, the dictionary (Dm) and the dictionary coefficient (Zmtrain) are obtained corresponding to the electrical load (m) using a pretrained dictionary, wherein, the dictionary is pretrained using the dictionary learning for every load specific consumption (Xm) corresponding to the electrical load (m); estimate, the load consumption () for each electrical load (m) among the plurality of electrical loads (M) using the estimated energy coefficient (Zmtest) with the corresponding dictionary (Dm); and iteratively, perform for each electrical load (m+1) among the plurality of electrical loads (M) to, recompute, a new aggregated load data (Xtest) by subtracting the estimated load energy consumption () from the previous load consumption (Xtest) value; and recompute, the new total load consumption (X) by subtracting the estimated load energy consumption () and the load specific consumption (Xm) from the previous total load consumption (X).
7. The system (102) as claimed in claim 6, wherein the one or more hardware processors (204) are configured for the optimization model to estimate the dictionary coefficient (Zmtest) of each electrical loads (m) by constructing, a graph using the total load consumption (X), considering the dictionary coefficient (Zm) as graph signal.
8. The system (102) as claimed in claim 7, wherein the one or more hardware processors (204) are configured wherein the optimization model constraints includes,
 estimating, the dictionary coefficient (Zmtest) using the dictionary (Dm) such that error between (Xtest) and the product of the dictionary (Dm) and the dictionary coefficient (Zmtest)) is minimized; and
 determining, the dictionary coefficient (Zmtest)) for load m such that the total variation over the graph topology (Lm) is minimized, in other words, the coefficients are smooth with respect to the graph.
9. The system (102) as claimed in claim 7, wherein the one or more hardware processors (204) is configured wherein the solution to the optimization model is determined by taking derivative of the objective function with respect to the dictionary coefficient (Zmtest) and equating it to zero.
10. The system (102) as claimed in claim 6, wherein the one or more hardware processors (204) is configured wherein the load consumption () is estimated based on the product of estimated energy coefficient (Zmtest) with the corresponding dictionary (Dm).
11. One or more nontransitory machinereadable information storage mediums comprising one or more instructions which when executed by one or more hardware processors perform actions comprising:
 receiving, by a processor (204), an aggregated load data (Xtest), obtained from a utility meter among a plurality of utility meters at a plurality of regular time intervals, wherein the aggregated load data comprises total energy consumed by a plurality of electrical loads (m) connected to a utility meter and base loads which are low power loads; and a training data (Xtrain), wherein the training data is generated from the plurality of utility meters;
 representing, by the processor (204), the aggregate load data (Xtest) and the training data (Xtrain) as a matrix, where a column vector represents the plurality of regular time intervals covered in a predefined time period and a row vector represents a total number of predefined time periods considered; and
 determining, by the processor (204), from the aggregated load data (Xtest) of the utility meter, a disaggregated energy load () at low sampling rate, using an energy disaggregation technique comprising: obtaining, a total load consumption (X) using the aggregated load data (Xtest) and the training data (Xtrain); constructing, a laplacian graph (Lm) for each of the plurality of electrical load (m) using the total load consumption (X); formulating, an optimization model for estimating the dictionary coefficient (Zmtest) of each of the plurality of electrical loads (m); estimating, the dictionary coefficient (Zmtest) of each electrical loads (m), using the laplacian graph (Lm), a dictionary (Dm) and a dictionary coefficient (Zmtrain), wherein, the dictionary (Dm) and the dictionary coefficient (Zmtrain) are obtained for each of the electrical load (m) using a pretrained dictionary, and wherein, the dictionary is pretrained using the dictionary learning for each electrical load (m) load using a load specific consumption (Xm) corresponding to the electrical load (m); estimating, the load consumption () for each electrical load (m) among the plurality of electrical loads (M) using the estimated energy coefficient (Zmtest) with the corresponding dictionary (Dm); and iteratively, performing for each electrical load (m) among the plurality of electrical loads (M) steps of: recomputing, a new aggregated load data (Xtest) by subtracting the estimated load energy consumption () from the previous load consumption (Xtest) value; and recomputing, the new total load consumption (X) by subtracting the estimated load energy consumption () and the load specific consumption (Xm) from the previous total load consumption (X).
12. The one or more nontransitory machinereadable information storage mediums of claim 11, wherein the optimization model for estimating the dictionary coefficient (Zmtest) of each electrical loads (m) by,
 construct, a graph using the total load consumption (X), considering the dictionary coefficient (Zm) as the graph signal.
13. The one or more nontransitory machinereadable information storage mediums of claim 12, wherein the optimization model constraints includes,
 estimating, the dictionary coefficient (Zmtest)) using the dictionary (Dm) such that error between (Xtest) and the product of the dictionary (Dm) and the dictionary coefficient (Zmtest)) is minimized; and
 determining, the dictionary coefficient (Zmtest) for each electrical load (m) such that the total variation over the graph topology (Lm) is minimized.
14. The one or more nontransitory machinereadable information storage mediums of claim 12, wherein the solution to the optimization model is determined by taking derivative of the objective function with respect to the dictionary coefficient (Zmtest) and equating it to zero.
15. The one or more nontransitory machinereadable information storage mediums of claim 11, wherein the load consumption () is estimated based on the product of estimated energy coefficient (Zmtest) with the corresponding dictionary (Dm).
9256908  February 9, 2016  Dai et al. 
20120278272  November 1, 2012  Kim 
20120290230  November 15, 2012  Berges Gonzalez 
20130103215  April 25, 2013  Dai 
20130238266  September 12, 2013  Savvides 
20170030949  February 2, 2017  Yu 
 Kumar, K. et al. “An Intuitive Explanation of Graph Signal ProcessingBased Electrical Load Disaggregation,” 2017 IEEE 13^{th }International Colloquium on Signal Processing & Its Applications (CSPA), Mar. 1012, 2017, Penang, Malaysia; 7 pages.
 Gaur, M. et al. (2018). “Disaggregating Transform Learning for NonIntrusive Load Monitoring,” IEEE Access, vol. 6; pp. 4625646265.
Type: Grant
Filed: Mar 10, 2020
Date of Patent: Sep 14, 2021
Patent Publication Number: 20210011062
Assignee: Tata Consultancy Services Limited (Mumbai)
Inventors: Kriti Kumar (Bangalore), Mariswamy Girish Chandra (Bangalore), Achanna Anil Kumar (Bangalore), Naveen Kumar Thokala (Bangalore)
Primary Examiner: Raymond L Nimox
Application Number: 16/813,794
International Classification: G01R 21/133 (20060101); G06N 20/00 (20190101); G06N 7/00 (20060101);