CN111369120A - Non-invasive load monitoring method based on equipment transfer learning - Google Patents
Non-invasive load monitoring method based on equipment transfer learning Download PDFInfo
- Publication number
- CN111369120A CN111369120A CN202010122993.5A CN202010122993A CN111369120A CN 111369120 A CN111369120 A CN 111369120A CN 202010122993 A CN202010122993 A CN 202010122993A CN 111369120 A CN111369120 A CN 111369120A
- Authority
- CN
- China
- Prior art keywords
- model
- data
- target
- training
- sequence
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 81
- 238000013526 transfer learning Methods 0.000 title claims abstract description 47
- 238000012544 monitoring process Methods 0.000 title claims abstract description 33
- 238000012549 training Methods 0.000 claims abstract description 74
- 230000008569 process Effects 0.000 claims abstract description 50
- 238000012795 verification Methods 0.000 claims abstract description 23
- 238000013508 migration Methods 0.000 claims description 31
- 230000005012 migration Effects 0.000 claims description 31
- 230000006870 function Effects 0.000 claims description 24
- 238000005070 sampling Methods 0.000 claims description 24
- 230000004913 activation Effects 0.000 claims description 20
- 238000013528 artificial neural network Methods 0.000 claims description 17
- 230000014509 gene expression Effects 0.000 claims description 15
- 238000010606 normalization Methods 0.000 claims description 11
- 238000012360 testing method Methods 0.000 claims description 10
- 238000010276 construction Methods 0.000 claims description 9
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 6
- 238000012545 processing Methods 0.000 claims description 3
- 238000009966 trimming Methods 0.000 claims description 3
- 238000013527 convolutional neural network Methods 0.000 description 17
- 238000005265 energy consumption Methods 0.000 description 8
- 230000008901 benefit Effects 0.000 description 4
- 230000005540 biological transmission Effects 0.000 description 3
- 238000013135 deep learning Methods 0.000 description 3
- 239000000126 substance Substances 0.000 description 3
- 230000000694 effects Effects 0.000 description 2
- 208000036829 Device dislocation Diseases 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000002457 bidirectional effect Effects 0.000 description 1
- 125000004122 cyclic group Chemical group 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000002950 deficient Effects 0.000 description 1
- 238000013100 final test Methods 0.000 description 1
- 230000006698 induction Effects 0.000 description 1
- 238000002372 labelling Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
- G06Q10/063—Operations research, analysis or management
- G06Q10/0635—Risk analysis of enterprise or organisation activities
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R21/00—Arrangements for measuring electric power or power factor
- G01R21/133—Arrangements for measuring electric power or power factor by using digital technique
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/06—Energy or water supply
Landscapes
- Engineering & Computer Science (AREA)
- Business, Economics & Management (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Human Resources & Organizations (AREA)
- Economics (AREA)
- Health & Medical Sciences (AREA)
- Strategic Management (AREA)
- Software Systems (AREA)
- Evolutionary Computation (AREA)
- Tourism & Hospitality (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Mathematical Physics (AREA)
- Data Mining & Analysis (AREA)
- General Business, Economics & Management (AREA)
- Marketing (AREA)
- Entrepreneurship & Innovation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Development Economics (AREA)
- Game Theory and Decision Science (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Operations Research (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Biomedical Technology (AREA)
- Educational Administration (AREA)
- Quality & Reliability (AREA)
- Medical Informatics (AREA)
- Power Engineering (AREA)
- Public Health (AREA)
- Water Supply & Treatment (AREA)
- Primary Health Care (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention relates to a non-invasive load monitoring method based on equipment transfer learning, which comprises the steps of collecting load amplitude data of various household electrical appliances and total household load amplitude data, and establishing an original data set; normalizing the data of the original data set; building an initialization sequence to point CNN model to build a device selection model for transfer learning, and selecting initial data and target data for transfer learning; performing a training process and a verification process on the CNN model from the initialization sequence to the point by using initial data to form a transfer learning initial model; fine-tuning a full connection layer of the initial model by adopting target data to form a transfer learning target model; and inputting the normalized household total load amplitude data into a transfer learning target model to obtain the power amplitude of the target equipment, thereby realizing non-invasive load monitoring of the target electrical equipment. The invention reduces the time of model training and the economic cost and time cost of non-invasive load monitoring.
Description
Technical Field
The invention relates to the field of power metering methods, in particular to a non-invasive load monitoring method based on equipment migration learning.
Background
The power consumer power consumption load detail monitoring is mainly used for acquiring information such as the power consumption condition and the power consumption behavior of each load device in a consumer. Traditional invasive power load monitoring all need install induction measurement and data transmission device on every consumer inside the load, and economic input is great, and the management is maintained comparatively complicatedly, is not suitable for popularization on a large scale. Non-Intrusive Load Monitoring (NILM) can refine Monitoring to the inside of the total Load according to the total power consumption of a user without installing a large amount of Monitoring equipment, so that the consumed power and the running state of each electrical appliance are obtained, and the Non-Intrusive Load Monitoring (NILM) has the advantages of strong operability, low implementation cost, high user acceptance, high reliability, easiness in wide popularization and the like. The deep neural network method has the highest precision for realizing NILM, but has the disadvantages that a lot of time is consumed for training a model each time, and the model trained in a data-poor area has a poor effect, so that a proper method is needed for solving the problems.
The existing NILM methods are mainly of two types: one is unsupervised learning without a training process and the other is supervised learning with a priori training process.
The unsupervised learning method is represented by a hidden Markov model, considers the use condition and the working process of equipment, and simplifies the step of labeling data for the electric equipment. However, due to lack of prior supervision, when the number of the electric devices is large, the electric devices are easy to fall into local optimization, and the identification precision is greatly reduced.
The supervised learning algorithm needs to be trained on known data, and the load identification target is achieved by learning the load stamp of the electric equipment. Compared with an unsupervised learning method, the supervised learning method represented by the deep neural network has higher precision and can cope with the complex conditions of similar load marks, occurrence of simultaneous events and the like, but the training of one model usually takes longer time, the models among regions are not universal, and a good solution is not provided for the regions with deficient data.
Disclosure of Invention
In view of this, the present invention provides a non-intrusive load monitoring method based on device migration learning, which employs a sequence-to-point input/output mode and can provide load information to a user more efficiently.
The invention is realized by adopting the following scheme: a non-intrusive load monitoring method based on equipment transfer learning comprises the following steps:
step S1: acquiring load power data of each household appliance and total household load power data, and establishing an original data set; the storage form is as follows:
each household appliance is A, B, … and L, the total household load is U, and the corresponding load data in the data set is
In the formula, 0, 1, …, n represents the sampling point number; a, (n), b (n), …, l (n), u (n) is the amplitude of the power of the specific sampling point; a (n), b (n), …, l (n) represent a set of load sample point magnitudes for the corresponding device; u (n) represents a set of home total load sample point magnitudes;
step S2: normalizing the data of all the stored data sets;
step S3: randomSelecting a section of load sampling point amplitude data of each electrical appliance in the household electrical appliance devices A, B, … and L during respective operation to form a data set Ae(n),Be(n),…,Le(n); set data set Ae(n),Be(n),…,Le(n) the amplitude sets of the total household load sampling points in the corresponding time periods are respectively Ua(n),Ub(n),…,Ul(n), constructing a device selection model for transfer learning by constructing an initialization sequence to point (CNN) model; defining a convolution kernel with a changed weight value in the process of selecting a model by the equipment for constructing the transfer learning as an active convolution kernel, observing the number of the active convolution kernels in the model, and selecting load sampling point amplitude data of the electrical equipment with the largest number of the active convolution kernels as an initial data set O for the transfer learningm(n), the load sampling point amplitude data of the electrical equipment needing to be monitored, namely the target electrical equipment, is used as a target data set P for transfer learningm(n); mixing O withm(n)、Pm(n) two data sets as per 7: 2: 1, respectively randomly dividing the training set, the verification set and the test set; the training set, the verification set and the test set of the formed initial data set are respectively Omt(n)、Omv(n)、Omtt(n), the training set, the verification set and the test set of the target data set are respectively Pmt(n)、Pmv(n)、Pmtt(n);
Step S4: initialization sequence to Point CNN model, set up by step S3, in Um(n)、Omt(n) as a sample, training the model;
step S5: by Um(n)、Omv(n) as a sample, performing a verification process to stop the training process in step S4 to prevent an overfitting phenomenon; at this time, the construction of the transfer learning initial model is finished;
step S6: by Um(n)、Pmt(n) as a sample, carrying out a migration fine tuning process, fine tuning a full connection layer of the migration learning initial model to form a migration learning target model, and finishing the migration learning at the moment; inputting the normalized magnitude data of the total load sampling points of the family into a migration learning target modelAnd obtaining the power amplitude of the target equipment to realize non-invasive load monitoring on the target electrical equipment.
Further, the normalization process uses the following formula:
wherein x ism(i) Representing the power amplitude of each sampling point after normalization processing of the total load of the target equipment or the family; i is 0, 1, …, n denotes the sample point number; x (i) represents the power amplitude of each sample point before normalization,an average value representing the total load of the target device or the home; x after normalizationm(i) Forming a new set of device load data Am(n),Bm(n),…,Lm(n) and a set of Total Home load data Um(n)。
Further, the specific construction process of constructing the device selection model for transfer learning is as follows:
an initialization sequence to point CNN model is built, and the structure is as follows:
an input layer: w samples are input as a window;
the convolutional layer 1: 30 convolution kernels are contained, and each convolution kernel is 10 in size; the stride is 1; the activation function is a linear rectification function ReLU;
and (3) convolutional layer 2: 30 convolution kernels are contained, and each convolution kernel is 8 in size; the stride is 1; the activation function is ReLU;
and (3) convolutional layer: 40 convolution kernels are included, and each convolution kernel is 6 in size; the stride is 1; the activation function is ReLU;
and (4) convolutional layer: 50 convolution kernels are included, and the size of each convolution kernel is 5; the stride is 1; the activation function is ReLU;
and (5) convolutional layer: 50 convolution kernels are included, and the size of each convolution kernel is 5; the stride is 1; the activation function is ReLU;
full connection layer: the dense layer structure comprises 1024 neural network units, and 1 sample is output as a window;
will Ua(n),Ub(n),…,Ul(n) as input sequence, Ae(n),Be(n),…,Le(n) as a target sequence; setting the sequence length of an input window as W, namely the sequence of the input window is u (i) to u (i + W-1); setting the target value under the input window as the midpoint value of the corresponding target sequence, namely ae(τ),be(τ),…,le(τ), where τ is i + W/2;
performing one round of training on the initialization sequence to point CNN model: inputting data of an input sequence into an initialization sequence to point CNN model, and transmitting the data forward through a convolutional layer and a full-link layer to obtain an output value; comparing the output value with the target value to generate an error set epsilon (n); returning epsilon (n) to the network, sequentially obtaining errors of all weights in the full-connection layer and the convolution layer, further updating the weights of the convolution kernel and the full-connection layer neural network unit, and finishing one round of training;
after one round of training is completed, the generated equipment selection model expression of the transfer learning is as follows:
further, the specific content of step S4 is: will Um(n) as input sequence, Omt(n) as a target sequence; setting the sequence length of an input window to W, i.e. the sequence of input windows to um(i) To um(i + W-1); setting the target value under the input window as the midpoint value of the corresponding target sequence, i.e. omt(τ), where τ is i + W/2;
training the initialization sequence to point CNN model: inputting data of an input sequence into an initialization sequence to point CNN model, and transmitting the data forward through a convolutional layer and a full-link layer to obtain an output value; comparing the output value with the target value to generate an error set epsilon1(n); will epsilon1(n) transmitting back to the network, in turn, to find full connectivityErrors of all weights in the layers and the convolutional layers are further updated, and then one round of training is completed; repeating the training process to perform the next round of training; the training process uses an ADAM optimizer to optimize the model.
Further, the specific content of step S5 is:
the expressions defining the training penalty value Lpt and the verification penalty value Lpv for the verification process are as follows:
wherein θ represents a parameter of the network; adopting an early stopping criterion to train the model to stop after the number of rounds meeting the early stopping criterion condition; the early stop criteria are:
Lpυ≥Lpt
at this time, the training process is stopped, the initial model of the transfer learning is constructed, and the expression is as follows:
omt(τ)=f(um(i),um(i+W-1),ε1(n))。
further, the specific content of step S6 is:
will Um(n) as input sequence, Pmt(n) as a target sequence; setting the sequence length of an input window to W, i.e. the sequence of input windows to um(i) To um(i + W-1); setting the target value under the input window as the midpoint value of the corresponding target sequence, i.e. pmt(τ), where τ is i + W/2;
training the migration learning initial model, and simultaneously keeping the weight of the convolutional layer unchanged, wherein the process is as follows:
inputting data of an input sequence into a migration learning initial model, and forward transmitting the data through a convolutional layer and a full-link layer to obtain an output value; comparing the output value with the target value to generate an error setε2(n); will epsilon2(n) returning the weight error of the full connection layer to the network, and further updating the weight of the neural network unit of the full connection layer, wherein one round of training is completed; and repeating the training process to perform the next training round. The training process uses an ADAM optimizer to optimize the model.
The expressions for the training penalty value Lpt1 and the verification penalty value Lpv1 that define the migration trimming process are as follows:
wherein θ represents a parameter of the network; adopting an early stopping criterion to train the model to stop after the number of rounds meeting the early stopping criterion condition; the early stop criteria are:
Lpυ1≥Lpt1
stopping training at the moment, finishing the migration fine tuning process, finishing the construction of the migration learning target model, and obtaining the following expression:
pmt(τ)=f(um(i),um(i+W-1),ε2(n))。
compared with the prior art, the invention has the following beneficial effects:
compared with the existing non-invasive load monitoring method based on the neural network, the method adopts the sequence-to-point input and output mode, can more efficiently provide load information for users, and simplifies the algorithm flow; by means of transfer learning, the universality of the non-invasive load monitoring model is improved, various problems caused by the fact that the sample capacity of the known data set is too small are solved, meanwhile, the time for constructing the target model is reduced, and the economic cost and the time cost are reduced.
Drawings
FIG. 1 is a flow chart of an embodiment of the present invention.
Detailed Description
The invention is further explained below with reference to the drawings and the embodiments.
It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the disclosure. Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.
It is noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of example embodiments according to the present application. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, and it should be understood that when the terms "comprises" and/or "comprising" are used in this specification, they specify the presence of stated features, steps, operations, devices, components, and/or combinations thereof, unless the context clearly indicates otherwise.
As shown in fig. 1, the present embodiment provides a non-intrusive load monitoring method based on device transfer learning, including the following steps:
step S1: acquiring load power data of each household appliance and total household load power data, and establishing an original data set; the storage form is as follows:
each household appliance is A, B, … and L, the total household load is U, and the corresponding load data in the data set is
In the formula, 0, 1, …, n represents the sampling point number; a, (n), b (n), …, l (n), u (n) is the amplitude of the power of the specific sampling point; a (n), b (n), …, l (n) represent a set of load sample point magnitudes for the corresponding device; u (n) represents a set of home total load sample point magnitudes;
step S2: normalizing the data of all the stored data sets;
step S3: randomly selecting a section of load sampling point amplitude data of each electrical appliance in the household electrical appliance devices A, B, … and L during respective operation to form a data set Ae(n),Be(n),…,Le(n);Set data set Ae(n),Be(n),…,Le(n) the amplitude sets of the total household load sampling points in the corresponding time periods are respectively Ua(n),Ub(n),…,Ul(n), constructing a device selection model for transfer learning by constructing an initialization sequence to point (CNN) model; defining a convolution kernel with a changed weight value in the process of selecting a model by the equipment for constructing the transfer learning as an active convolution kernel, observing the number of the active convolution kernels in the model, and selecting load sampling point amplitude data of the electrical equipment with the largest number of the active convolution kernels as an initial data set O for the transfer learningm(n), the load sampling point amplitude data of the electrical equipment needing to be monitored, namely the target electrical equipment, is used as a target data set P for transfer learningm(n); mixing O withm(n)、Pm(n) two data sets as per 7: 2: 1, respectively randomly dividing the training set, the verification set and the test set; the training set, the verification set and the test set of the formed initial data set are respectively Omt(n)、Omv(n)、Omtt(n), the training set, the verification set and the test set of the target data set are respectively Pmt(n)、Pmv(n)、Pmtt(n);
Step S4: initialization sequence to Point CNN model, set up by step S3, in Um(n)、Omt(n) as a sample, training the model;
step S5: by Um(n)、Omv(n) as a sample, performing a verification process to stop the training process in step S4 to prevent an overfitting phenomenon; at this time, the construction of the transfer learning initial model is finished;
step S6: by Um(n)、Pmt(n) as a sample, carrying out a migration fine tuning process, fine tuning a full connection layer of the migration learning initial model to form a migration learning target model, and finishing the migration learning at the moment; and inputting the normalized amplitude data of the sampling points of the total household load into a migration learning target model to obtain the power amplitude of the target equipment so as to realize non-invasive load monitoring on the target electrical equipment.
In this embodiment, the formula adopted by the normalization process is as follows:
wherein x ism(i) Representing the power amplitude of each sampling point after normalization processing of the total load of the target equipment or the family; i is 0, 1, …, n denotes the sample point number; x (i) represents the power amplitude of each sample point before normalization,an average value representing the total load of the target device or the home; treated xm(i) Forming a new set of device load data Am(n),Bm(n),…,Lm(n) and a set of Total Home load data Um(n)。
In this embodiment, the specific construction process of constructing the device selection model for transfer learning is as follows:
an initialization sequence to point CNN model is built, and the structure is as follows:
an input layer: w samples are input as a window;
the convolutional layer 1: 30 convolution kernels are contained, and each convolution kernel is 10 in size; the stride is 1; the activation function is a linear rectification function ReLU;
and (3) convolutional layer 2: 30 convolution kernels are contained, and each convolution kernel is 8 in size; the stride is 1; the activation function is ReLU;
and (3) convolutional layer: 40 convolution kernels are included, and each convolution kernel is 6 in size; the stride is 1; the activation function is ReLU;
and (4) convolutional layer: 50 convolution kernels are included, and the size of each convolution kernel is 5; the stride is 1; the activation function is ReLU;
and (5) convolutional layer: 50 convolution kernels are included, and the size of each convolution kernel is 5; the stride is 1; the activation function is ReLU;
full connection layer: the dense layer structure comprises 1024 neural network units, and 1 sample is output as a window;
will Ua(n),Ub(n),…,Ul(n) as input sequence, Ae(n),Be(n),…,Le(n) as a target sequence; setting the sequence length of an input window as W, namely the sequence of the input window is u (i) to u (i + W-1); setting the target value under the input window as the midpoint value of the corresponding target sequence, namely ae(τ),be(τ),…,le(τ), where τ is i + W/2;
performing one round of training on the initialization sequence to point CNN model: inputting data of an input sequence into an initialization sequence to point CNN model, and transmitting the data forward through a convolutional layer and a full-link layer to obtain an output value; comparing the output value with the target value to generate an error set epsilon (n); returning epsilon (n) to the network, sequentially obtaining errors of all weights in the full-connection layer and the convolution layer, further updating the weights of the convolution kernel and the full-connection layer neural network unit, and finishing one round of training;
after one round of training is completed, the generated equipment selection model expression of the transfer learning is as follows:
in this embodiment, the specific content of step S4 is:
an initialization sequence to point CNN model is built, and the structure is as follows:
an input layer: w samples are input as a window;
the convolutional layer 1: 30 convolution kernels are contained, and each convolution kernel is 10 in size; the stride is 1; the activation function is a linear rectification function ReLU;
and (3) convolutional layer 2: 30 convolution kernels are contained, and each convolution kernel is 8 in size; the stride is 1; the activation function is ReLU;
and (3) convolutional layer: 40 convolution kernels are included, and each convolution kernel is 6 in size; the stride is 1; the activation function is ReLU;
and (4) convolutional layer: 50 convolution kernels are included, and the size of each convolution kernel is 5; the stride is 1; the activation function is ReLU;
and (5) convolutional layer: 50 convolution kernels are included, and the size of each convolution kernel is 5; the stride is 1; the activation function is ReLU;
full connection layer: the dense layer structure comprises 1024 neural network units, and 1 sample is output as a window;
will Um(n) as input sequence, Omt(n) as a target sequence; setting the sequence length of an input window to W, i.e. the sequence of input windows to um(i) To um(i + W-1); setting the target value under the input window as the midpoint value of the corresponding target sequence, i.e. omt(τ), where τ is i + W/2;
training the initialization sequence to point CNN model: inputting data of an input sequence into an initialization sequence to point CNN model, and transmitting the data forward through a convolutional layer and a full-link layer to obtain an output value; comparing the output value with the target value to generate an error set epsilon1(n); will epsilon1(t) returning the error of all weights in the full-connection layer and the convolution layer to the network, and further updating the weights of the model convolution kernel and the full-connection layer neural network unit, wherein one round of training is completed; repeating the training process to perform the next round of training; the training process uses an ADAM optimizer to optimize the model.
In this embodiment, the specific content of step S5 is:
the expressions defining the training penalty value Lpt and the verification penalty value Lpv for the verification process are as follows:
wherein θ represents a parameter of the network; adopting an early stopping criterion to train the model to stop after the number of rounds meeting the early stopping criterion condition; the early stop criteria are:
Lpυ≥Lpt
at this time, the training process is stopped, the initial model of the transfer learning is constructed, and the expression is as follows:
omt(τ)=f(um(i),um(i+W-1),ε1(n))。
in this embodiment, the specific content of step S6 is:
will Um(n) as input sequence, Pmt(n) as a target sequence; setting the sequence length of an input window to W, i.e. the sequence of input windows to um(i) To um(i + W-1); setting the target value under the input window as the midpoint value of the corresponding target sequence, i.e. pmt(τ), where τ is i + W/2;
training the migration learning initial model, and simultaneously keeping the weight of the convolutional layer unchanged, wherein the process is as follows:
inputting data of an input sequence into a migration learning initial model, and forward transmitting the data through a convolutional layer and a full-link layer to obtain an output value; comparing the output value with the target value to generate an error set epsilon2(n); will epsilon2(n) returning the weight error of the full connection layer to the network, and further updating the weight of the neural network unit of the full connection layer, wherein one round of training is completed; and repeating the training process to perform the next training round. The training process uses an ADAM optimizer to optimize the model.
The expressions for the training penalty value Lpt1 and the verification penalty value Lpv1 that define the migration trimming process are as follows:
wherein θ represents a parameter of the network; adopting an early stopping criterion to train the model to stop after the number of rounds meeting the early stopping criterion condition; the early stop criteria are:
Lpυ1≥Lpt1
stopping training at the moment, finishing the migration fine tuning process, finishing the construction of the migration learning target model, and obtaining the following expression:
pmt(τ)=f(um(i),um(i+W-1),ε2(n))。
preferably, in the present embodiment, U is usedm(n)、PmttAnd (n) as a sample, establishing three indexes for testing the precision of the target model.
The first index is the Mean Absolute Error (MAE), which is given by:
wherein the content of the first and second substances,indicates the predicted value, pmtt(τ) represents the actual value, T represents the predicted total time; the index evaluates the absolute error between the predicted value and the actual value;
the second indicator is total energy error (SAE), which is given by the following equation:
wherein the content of the first and second substances,representing the predicted total plant energy consumption, and r representing the actual total plant energy consumption; the index evaluates the relative error of the predicted total energy consumption and the actual total energy consumption;
the third indicator is the daily energy error (EpD), which is given by the formula:
wherein the content of the first and second substances,indicating predicted total energy consumption of the equipment during a day, e indicating actual total energy consumption of the equipment during a day, D indicating testedTotal number of days; the index evaluates the absolute error between the predicted total energy consumption and the actual total energy consumption within one day;
and judging the accuracy of the model according to the average value of the three indexes, wherein the formula is as follows:
wherein α, β and gamma are the weighted numbers of the three indexes;
Preferably, this embodiment
(1) The traditional sequence-to-sequence input and output mode based on the bidirectional cyclic Neural network is improved, and the sequence-to-point input and output mode is formed based on the deep learning of the Convolutional Neural Network (CNN).
(2) The device transfer learning method shown in steps S3-S6 is provided, and a transfer learning initial model is constructed by applying load data training and verification initialization sequence to point CNN model of a certain device aiming at different electrical devices in the same family. And then, fine-tuning the full connection layer of the initial model through the load data of the target equipment, and applying the model to other target equipment.
In the embodiment, the non-intrusive load monitoring is realized by a transfer learning method based on sequence-to-point CNN deep learning. One of the advantages of the monitoring method is that a large number of models do not need to be trained in a time-consuming manner, and only the initial models are widely applied to various target scenes by using transfer learning.
In deep learning, too little training data often reduces the result precision, resulting in poor final test effect. The method provided by the embodiment has the second advantage of effectively overcoming the defect, and the method uses a certain known data set as a sample to construct an initial model, thereby solving the problem of low resolution precision due to too small sample volume of the known data set in the target area.
The conventional sequence-to-sequence mode outputs a target device load sequence by inputting a segment of device total load sequence. In the mode, the monitoring data volume is large, the transmission is complex, the precision is not high enough, and the obtained load information cannot be well included for the user. The sequence-to-point mode provided by this embodiment is to output the median of the target device load sequence by inputting a section of device total load sequence, thereby greatly simplifying the transmission flow, improving the monitoring precision, and providing more practical information for the user.
In summary, compared with the existing non-invasive load monitoring method based on the neural network, the non-invasive load monitoring method based on the equipment transfer learning provided by the invention adopts the sequence-to-point input and output mode, so that the load information can be provided for the user more efficiently, and the algorithm flow is simplified; by means of transfer learning, the universality of the non-invasive load monitoring model is improved, various problems caused by the fact that the sample capacity of the known data set is too small are solved, meanwhile, the time for constructing the target model is reduced, and the economic cost and the time cost are reduced. Therefore, the method provided by the invention has advantages.
The above description is only a preferred embodiment of the present invention, and all equivalent changes and modifications made in accordance with the claims of the present invention should be covered by the present invention.
Claims (6)
1. A non-intrusive load monitoring method based on equipment transfer learning is characterized in that: the method comprises the following steps:
step S1: acquiring load power data of each household appliance and total household load power data, and establishing an original data set; the storage form is as follows:
each household appliance is A, B, … and L, the total household load is U, and the corresponding load data in the data set is
In the formula, 0, 1, …, n represents the sampling point number; a, (n), b (n), …, l (n), u (n) is the amplitude of the power of the specific sampling point; a (n), b (n), …, l (n) represent a set of load sample point magnitudes for the corresponding device; u (n) represents a set of home total load sample point magnitudes;
step S2: normalizing the data of all the stored data sets;
step S3: randomly selecting a section of load sampling point amplitude data of each electrical appliance in the household electrical appliance devices A, B, … and L during respective operation to form a data set Ae(n),Be(n),…,Le(n); set data set Ae(n),Be(n),…,Le(n) the amplitude sets of the total household load sampling points in the corresponding time periods are respectively Ua(n),Ub(n),…,Ul(n), constructing a device selection model for transfer learning by constructing an initialization sequence to point (CNN) model; defining a convolution kernel with a changed weight value in the process of selecting a model by the equipment for constructing the transfer learning as an active convolution kernel, observing the number of the active convolution kernels in the model, and selecting load sampling point amplitude data of the electrical equipment with the largest number of the active convolution kernels as an initial data set O for the transfer learningm(n), the load sampling point amplitude data of the electrical equipment needing to be monitored, namely the target electrical equipment, is used as a target data set P for transfer learningm(n); mixing O withm(n)、Pm(n) two data sets as per 7: 2: 1, respectively randomly dividing the training set, the verification set and the test set; the training set, the verification set and the test set of the formed initial data set are respectively Omt(n)、Omv(n)、Omtt(n), the training set, the verification set and the test set of the target data set are respectively Pmt(n)、Pmv(n)、Pmtt(n);
Step S4: initialization sequence to Point CNN model, set up by step S3, in Um(n)、Omt(n) as a sample, training the model;
step S5: by Um(n)、Omv(n) as a sample, performing a verification process to stop the training process in step S4 to prevent an overfitting phenomenon; at this time, the construction of the transfer learning initial model is finished;
step S6: by Um(n)、Pmt(n) as a sample, carrying out a migration fine tuning process, fine tuning a full connection layer of the migration learning initial model to form a migration learning target model, and finishing the migration learning at the moment; and inputting the normalized amplitude data of the sampling points of the total household load into a migration learning target model to obtain the power amplitude of the target equipment so as to realize non-invasive load monitoring on the target electrical equipment.
2. The non-intrusive load monitoring method based on equipment transfer learning according to claim 1, characterized in that: the normalization process uses the following formula:
wherein x ism(i) Representing the power amplitude of each sampling point after normalization processing of the total load of the target equipment or the family; i is 0, 1, …, n denotes the sample point number; x (i) represents the power amplitude of each sample point before normalization,an average value representing the total load of the target device or the home; x after normalizationm(i) Forming a new set of device load data Am(n),Bm(n),…,Lm(n) and a set of Total Home load data Um(n)。
3. The non-intrusive load monitoring method based on equipment transfer learning according to claim 1, characterized in that: the specific construction process of the equipment selection model for constructing the transfer learning is as follows:
an initialization sequence to point CNN model is built, and the structure is as follows:
an input layer: w samples are input as a window;
the convolutional layer 1: 30 convolution kernels are contained, and each convolution kernel is 10 in size; the stride is 1; the activation function is a linear rectification function ReLU;
and (3) convolutional layer 2: 30 convolution kernels are contained, and each convolution kernel is 8 in size; the stride is 1; the activation function is ReLU;
and (3) convolutional layer: 40 convolution kernels are included, and each convolution kernel is 6 in size; the stride is 1; the activation function is ReLU;
and (4) convolutional layer: 50 convolution kernels are included, and the size of each convolution kernel is 5; the stride is 1; the activation function is ReLU;
and (5) convolutional layer: 50 convolution kernels are included, and the size of each convolution kernel is 5; the stride is 1; the activation function is ReLU;
full connection layer: the dense layer structure comprises 1024 neural network units, and 1 sample is output as a window;
will Ua(n),Ub(n),…,Ul(n) as input sequence, Ae(n),Be(n),…,Le(n) as a target sequence; setting the sequence length of an input window as W, namely the sequence of the input window is u (i) to u (i + W-1); setting the target value under the input window as the midpoint value of the corresponding target sequence, namely ae(τ),be(τ),…,le(τ), where τ is i + W/2;
performing one round of training on the initialization sequence to point CNN model: inputting data of an input sequence into an initialization sequence to point CNN model, and transmitting the data forward through a convolutional layer and a full-link layer to obtain an output value; comparing the output value with the target value to generate an error set epsilon (n); returning epsilon (n) to the network, sequentially obtaining errors of all weights in the full-connection layer and the convolution layer, further updating the weights of the convolution kernel and the full-connection layer neural network unit, and finishing one round of training;
after one round of training is completed, the generated equipment selection model expression of the transfer learning is as follows:
4. the non-intrusive load monitoring method based on equipment transfer learning according to claim 1, characterized in that: the specific content of step S4 is:
will Um(n) as input sequence, Omt(n) as a target sequence; setting the sequence length of an input window to W, i.e. the sequence of input windows to um(i) To um(i + W-1); setting the target value under the input window as the midpoint value of the corresponding target sequence, i.e. omt(τ), where τ is i + W/2;
training the initialization sequence to point CNN model: inputting data of an input sequence into an initialization sequence to point CNN model, and transmitting the data forward through a convolutional layer and a full-link layer to obtain an output value; comparing the output value with the target value to generate an error set epsilon1(n); will epsilon1(n) returning the error of all weights in the full-connection layer and the convolution layer to the network, and further updating the weights of the model convolution kernel and the full-connection layer neural network unit, wherein one round of training is completed; repeating the training process to perform the next round of training; the training process uses an ADAM optimizer to optimize the model.
5. The non-intrusive load monitoring method based on equipment transfer learning according to claim 1, characterized in that: the specific content of step S5 is:
the expressions defining the training penalty value Lpt and the verification penalty value Lpv for the verification process are as follows:
wherein θ represents a parameter of the network; adopting an early stopping criterion to train the model to stop after the number of rounds meeting the early stopping criterion condition; the early stop criteria are:
Lpv≥Lpt
at this time, the training process is stopped, the initial model of the transfer learning is constructed, and the expression is as follows:
omt(τ)=f(um(i),um(i+W-1),ε1(n))。
6. the non-intrusive load monitoring method based on equipment transfer learning according to claim 1, characterized in that: the specific content of step S6 is:
will Um(n) as input sequence, Pmt(n) as a target sequence; setting the sequence length of an input window to W, i.e. the sequence of input windows to um(i) To um(i + W-1); setting the target value under the input window as the midpoint value of the corresponding target sequence, i.e. pmt(τ), where τ is i + W/2;
training the migration learning initial model, and simultaneously keeping the weight of the convolutional layer unchanged, wherein the process is as follows:
inputting data of an input sequence into a migration learning initial model, and forward transmitting the data through a convolutional layer and a full-link layer to obtain an output value; comparing the output value with the target value to generate an error set epsilon2(n); will epsilon2(n) returning the weight error of the full connection layer to the network, and further updating the weight of the neural network unit of the full connection layer, wherein one round of training is completed; and repeating the training process to perform the next training round. The training process uses an ADAM optimizer to optimize the model.
The expressions for the training penalty value Lpt1 and the verification penalty value Lpv1 that define the migration trimming process are as follows:
wherein θ represents a parameter of the network; adopting an early stopping criterion to train the model to stop after the number of rounds meeting the early stopping criterion condition; the early stop criteria are:
Lpv1≥Lpt1
stopping training at the moment, finishing the migration fine tuning process, finishing the construction of the migration learning target model, and obtaining the following expression:
pmt(τ)=f(um(i),um(i+W-1),ε2(n))。
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010122993.5A CN111369120A (en) | 2020-02-27 | 2020-02-27 | Non-invasive load monitoring method based on equipment transfer learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010122993.5A CN111369120A (en) | 2020-02-27 | 2020-02-27 | Non-invasive load monitoring method based on equipment transfer learning |
Publications (1)
Publication Number | Publication Date |
---|---|
CN111369120A true CN111369120A (en) | 2020-07-03 |
Family
ID=71211576
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202010122993.5A Pending CN111369120A (en) | 2020-02-27 | 2020-02-27 | Non-invasive load monitoring method based on equipment transfer learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN111369120A (en) |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111753968A (en) * | 2020-07-07 | 2020-10-09 | 苏州联电能源发展有限公司 | Non-invasive load monitoring intelligent ammeter and electricity decomposition method |
CN112736907A (en) * | 2020-12-28 | 2021-04-30 | 国网电子商务有限公司 | Non-invasive household power load decomposition characteristic selection method and device |
CN113205177A (en) * | 2021-04-25 | 2021-08-03 | 广西大学 | Electric power terminal identification method based on incremental collaborative attention mobile convolution |
CN113970667A (en) * | 2021-10-10 | 2022-01-25 | 上海梦象智能科技有限公司 | Non-invasive load monitoring method based on midpoint value of prediction window |
CN114925115A (en) * | 2022-05-27 | 2022-08-19 | 天津大学 | Light generalized load decomposition method based on depth separable convolution |
CN115204529A (en) * | 2022-09-15 | 2022-10-18 | 之江实验室 | Non-invasive load monitoring method and device based on time attention mechanism |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108899892A (en) * | 2018-06-08 | 2018-11-27 | 上海电力学院 | One kind being based on CNN non-intrusion type power load decomposition method |
CN110264041A (en) * | 2019-05-24 | 2019-09-20 | 澳门大学 | A kind of non-intrusion type load decomposition and monitoring method based on difference |
CN110376457A (en) * | 2019-06-28 | 2019-10-25 | 同济大学 | Non-intrusion type load monitoring method and device based on semi-supervised learning algorithm |
-
2020
- 2020-02-27 CN CN202010122993.5A patent/CN111369120A/en active Pending
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108899892A (en) * | 2018-06-08 | 2018-11-27 | 上海电力学院 | One kind being based on CNN non-intrusion type power load decomposition method |
CN110264041A (en) * | 2019-05-24 | 2019-09-20 | 澳门大学 | A kind of non-intrusion type load decomposition and monitoring method based on difference |
CN110376457A (en) * | 2019-06-28 | 2019-10-25 | 同济大学 | Non-intrusion type load monitoring method and device based on semi-supervised learning algorithm |
Non-Patent Citations (3)
Title |
---|
CHAOYUN ZHANG ET AL.: "sequence-to-point learning with neural networks for non-intrusive load monitoring", 《ARXIV》 * |
MICHELE D’INCECCO ET AL.: "Transfer Learning for Non-Intrusive Load Monitoring", 《 IEEE TRANSACTIONS ON SMART GRID》 * |
陈小闵: "非侵入式负荷分解技术及应用研究", 《中国优秀硕士学位论文全文数据库 工程科技Ⅱ辑》 * |
Cited By (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111753968A (en) * | 2020-07-07 | 2020-10-09 | 苏州联电能源发展有限公司 | Non-invasive load monitoring intelligent ammeter and electricity decomposition method |
CN111753968B (en) * | 2020-07-07 | 2023-08-08 | 苏州联电能源发展有限公司 | Non-invasive load monitoring intelligent ammeter and electric quantity decomposition method |
CN112736907A (en) * | 2020-12-28 | 2021-04-30 | 国网电子商务有限公司 | Non-invasive household power load decomposition characteristic selection method and device |
CN113205177A (en) * | 2021-04-25 | 2021-08-03 | 广西大学 | Electric power terminal identification method based on incremental collaborative attention mobile convolution |
CN113205177B (en) * | 2021-04-25 | 2022-03-25 | 广西大学 | Electric power terminal identification method based on incremental collaborative attention mobile convolution |
CN113970667A (en) * | 2021-10-10 | 2022-01-25 | 上海梦象智能科技有限公司 | Non-invasive load monitoring method based on midpoint value of prediction window |
CN113970667B (en) * | 2021-10-10 | 2024-04-05 | 上海梦象智能科技有限公司 | Non-invasive load monitoring method based on predicted window midpoint value |
CN114925115A (en) * | 2022-05-27 | 2022-08-19 | 天津大学 | Light generalized load decomposition method based on depth separable convolution |
CN115204529A (en) * | 2022-09-15 | 2022-10-18 | 之江实验室 | Non-invasive load monitoring method and device based on time attention mechanism |
CN115204529B (en) * | 2022-09-15 | 2022-12-20 | 之江实验室 | Non-invasive load monitoring method and device based on time attention mechanism |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111369120A (en) | Non-invasive load monitoring method based on equipment transfer learning | |
CN110705743B (en) | New energy consumption electric quantity prediction method based on long-term and short-term memory neural network | |
CN109145949A (en) | Non-intrusive electrical load monitoring and decomposition method and system based on integrated study | |
CN109407654B (en) | Industrial data nonlinear causal analysis method based on sparse deep neural network | |
Fan et al. | A novel machine learning method based approach for Li-ion battery prognostic and health management | |
CN110807554B (en) | Generation method and system based on wind power/photovoltaic classical scene set | |
CN111860982A (en) | Wind power plant short-term wind power prediction method based on VMD-FCM-GRU | |
CN108197648A (en) | A kind of Fault Diagnosis Method of Hydro-generating Unit and system based on LSTM deep learning models | |
CN108171259A (en) | A kind of wisdom tip water quality monitoring system and method based on NB-IoT | |
CN110889085A (en) | Intelligent wastewater monitoring method and system based on complex network multiple online regression | |
CN110824914B (en) | Intelligent wastewater treatment monitoring method based on PCA-LSTM network | |
CN104636985A (en) | Method for predicting radio disturbance of electric transmission line by using improved BP (back propagation) neural network | |
CN112130086B (en) | Method and system for predicting remaining life of power battery | |
CN108091166A (en) | Forecasting Methodology, device, equipment and the storage medium of available parking places number of variations | |
CN111160659B (en) | Power load prediction method considering temperature fuzzification | |
CN111612262A (en) | Wind power probability prediction method based on quantile regression | |
CN115758290A (en) | Fan gearbox high-speed shaft temperature trend early warning method based on LSTM | |
CN116707331B (en) | Inverter output voltage high-precision adjusting method and system based on model prediction | |
CN113988441A (en) | Power wireless network link quality prediction and model training method and device | |
CN113627091A (en) | Device and method for predicting energy load | |
CN115561416A (en) | Method and device for detecting inlet water quality of sewage treatment plant in real time | |
CN113902183B (en) | BERT-based non-invasive transformer area charging pile state monitoring and electricity price adjusting method | |
CN106295877B (en) | Method for predicting electric energy consumption of smart power grid | |
CN116930609A (en) | Electric energy metering error analysis method based on ResNet-LSTM model | |
CN109408896B (en) | Multi-element intelligent real-time monitoring method for anaerobic sewage treatment gas production |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20200703 |