CN114545066A - Non-invasive load monitoring model polymerization method and system - Google Patents

Non-invasive load monitoring model polymerization method and system Download PDF

Info

Publication number
CN114545066A
CN114545066A CN202210025465.7A CN202210025465A CN114545066A CN 114545066 A CN114545066 A CN 114545066A CN 202210025465 A CN202210025465 A CN 202210025465A CN 114545066 A CN114545066 A CN 114545066A
Authority
CN
China
Prior art keywords
model
load
active power
data
aggregation
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
Application number
CN202210025465.7A
Other languages
Chinese (zh)
Inventor
缪巍巍
曾锃
滕昌志
毕思博
张瑞
李世豪
张明轩
张震
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Information and Telecommunication Branch of State Grid Jiangsu Electric Power Co Ltd
Original Assignee
Information and Telecommunication Branch of State Grid Jiangsu Electric Power Co Ltd
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Information and Telecommunication Branch of State Grid Jiangsu Electric Power Co Ltd filed Critical Information and Telecommunication Branch of State Grid Jiangsu Electric Power Co Ltd
Priority to CN202210025465.7A priority Critical patent/CN114545066A/en
Publication of CN114545066A publication Critical patent/CN114545066A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
    • G01R21/00—Arrangements for measuring electric power or power factor
    • G01R21/001—Measuring real or reactive component; Measuring apparent energy
    • 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
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • 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
    • 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
    • G06N3/048—Activation functions
    • 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
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
    • Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
    • Y04S20/00—Management or operation of end-user stationary applications or the last stages of power distribution; Controlling, monitoring or operating thereof
    • Y04S20/20—End-user application control systems
    • Y04S20/222—Demand response systems, e.g. load shedding, peak shaving

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Power Engineering (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses a non-invasive load monitoring model polymerization method and a system, wherein the method comprises the following steps: acquiring active power of a user bus and each load; pairing the active power of the user bus and the active power of each load to form a data set and preprocessing the data set; the method comprises the steps that three preset base models are trained for the first time, and the three base models after the first training are used for decomposing active power of each load to obtain a decomposition result; aggregating the decomposition results by using a preset perceptron aggregation model and optimizing the preset perceptron aggregation model to obtain preliminary aggregation data of the active power of each load; and optimizing the pre-established DAE model based on the decomposition result and the preliminary aggregation data, wherein the optimized DAE model is an aggregation model of the non-intrusive load monitoring model and is used for outputting data of active power of each load. The invention can aggregate the results of different load decompositions to obtain a decomposition result superior to that of a single model.

Description

Non-invasive load monitoring model polymerization method and system
Technical Field
The invention relates to a non-invasive load monitoring model polymerization method and a system, belonging to the technical field of non-invasive electric load monitoring.
Background
Originally when measuring the inside electrical apparatus of family and obtaining data, a lot of all are invasive monitoring, to installing the sensor that has advanced communication function between every electrical apparatus and the block terminal promptly, although this kind of mode measurement accuracy is high, but hardware installation or transformation can bring inconvenience for resident's life, and the structure is often complicated, with high costs, and later maintenance is also comparatively loaded down with trivial details. Therefore, non-intrusive load monitoring structures have been proposed to fill this deficiency, which corresponds to a one-to-many relationship. For a single family, if an intelligent electric meter is installed at a home, the electricity utilization information of the user can be collected, and the electricity utilization condition of each device can be obtained, so that the hardware cost is greatly reduced, the installation is convenient and fast, and the intelligent electric meter is suitable for family load monitoring.
As for the non-intrusive load identification algorithm, the current non-intrusive power load intelligent identification system generally performs aggregation based on a single model, a common model deep learning model, and the like. The single model is usually good in identification effect only for certain types of electric appliance loads, and universality are required to be improved. Many deep learning algorithms with good load monitoring effect are applied to non-intrusive load monitoring, for example, a Seq2Point model, a DAE model, a residual network model based on Attention, and the like. The accuracy of the corresponding decomposition power of each model for different load devices is different, and a method for aggregating the decomposition results of each model is lacked at present. The existing model is used for load decomposition, so that a certain electric appliance is better identified, but other electric appliances are poorer in identification effect.
Disclosure of Invention
The invention aims to overcome the defects in the prior art and provides a non-invasive load monitoring model polymerization method and system, which can polymerize the decomposition results of different loads and obtain the decomposition result superior to that of a single model.
In order to achieve the purpose, the invention is realized by adopting the following technical scheme:
in a first aspect, the present invention provides a method for aggregating non-intrusive load monitoring models, including:
acquiring active power of a user bus and each load;
pairing the active power of the user bus and the active power of each load to form a data set, and processing the data set;
performing primary training on three preset base models based on the processed data set, and decomposing the active power of each load by adopting the three base models after the primary training to obtain a decomposition result;
aggregating the decomposition results by using a preset perceptron aggregation model, optimizing parameters of the preset perceptron aggregation model, and obtaining preliminary aggregation data of active power of each load based on the trained perceptron aggregation model;
and optimizing the pre-established DAE model based on the decomposition result and the preliminary aggregation data, wherein the optimized DAE model is the aggregation model of the non-invasive load monitoring model and is used for outputting the data of the active power of each load.
With reference to the first aspect, further, the obtaining the active power of the user bus and each load includes: and extracting active power data of each load and each load cluster of the user according to the frequency of times/minute by using active power monitoring equipment, and arranging the active power data in a time sequence to obtain the active power of a user bus and the active power of each load.
With reference to the first aspect, further, the composing the data set includes:
the sliding window obtains active power data with a specified length, and the method comprises the following steps: setting the length of the time sequence data intercepted by the sliding window as T, gradually carrying out sliding interception on the acquired active power of the user bus and the active power of each load according to time sequence by taking 1 as a step length, setting the total length of the data as N,
then, the data of the active power of the user bus is obtained as follows:
Figure BDA0003464365700000031
the data of the active power of the ith load are as follows:
Figure BDA0003464365700000032
the active power of the user bus and the active power of each load are paired to form a data set, and the method comprises the following steps:
data for making user bus active power
Figure BDA0003464365700000033
Wherein xnT is a vector, N is 1,2, …, N-T +1,
data of active power of ith load
Figure BDA0003464365700000034
Wherein
Figure BDA0003464365700000035
T is vector, N is 1,2, …, N-T + 1;
then
Figure BDA0003464365700000036
And the data sets are paired with the active power of the user bus and the active power of each load.
With reference to the first aspect, further, the processing the data set includes:
carrying out normalization processing on the data, wherein the normalization processing comprises the following steps: and taking out the maximum value of the active power of the user bus, and dividing the active power of the user bus and the active power of each load by the maximum value of the active power of the user bus respectively.
With reference to the first aspect, further, the performing primary training on the preset three base models includes:
training a preset Seq2Point model by using the data set subjected to the ith load processing, and updating model parameters by adopting a reverse optimization algorithm to obtain a trained Seq2Point model;
training a preset DAE model by using the data set subjected to the ith load processing, and updating model parameters by adopting a random gradient descent algorithm to obtain a trained DAE model;
training a preset Attention and a residual network model by using the ith load-processed data set, updating an Attention parameter by adopting a BP (back propagation) algorithm to obtain a trained Attention, and updating a model parameter by adopting a random gradient descent algorithm based on the trained Attention to obtain a trained residual network model;
processing x in the obtained data set by adopting three base model pairs finished by primary trainingnAnd performing active power decomposition on each load, wherein the decomposition results of the ith load decomposed by the three trained basic models are respectively as follows:
the decomposition result of the trained Seq2Point model:
Figure BDA0003464365700000041
decomposition results of the trained DAE model:
Figure BDA0003464365700000042
and (3) decomposing the trained residual error network model:
Figure BDA0003464365700000043
with reference to the first aspect, optionally, the preset Seq2Point model includes:
an input layer: will be provided with
Figure BDA0003464365700000044
The T sampling points are input into a preset Seq2Point model, and the tensor of the input data is (T, 1);
forward GRU layer: mining a change rule in the data of the input layer, wherein the size of a hidden layer is 2T, the tensor of input data of the second forward GRU layer is (T,2T), and the tensor of output data is (4T, 1);
full connection layer: mapping the high-dimensional data to the low-dimensional data, wherein the input data tensor is (4T,1), and the output data tensor is (2T, 1);
an output layer: the normalized power value of the ith load is output, the input data tensor is (2T,1), and the output data tensor is (1, 1).
With reference to the first aspect, optionally, the training a preset Seq2Point model includes:
the activation function employed in each cell is the Relu function, expressed as:
Figure BDA0003464365700000051
the calculation formula of the GRU layer is expressed as:
Figure BDA0003464365700000052
in the formula (2), rtIndicating gating of reset, WrRepresents the weight of the reset gate, sigma represents the sigmoid function, ht-1Represents the state of the last neuron descending, ztIndicates the state of the last neuron, WzIndicating that the weight of the door is updated,
Figure BDA0003464365700000053
represents the current state of the neuron or neurons,
Figure BDA0003464365700000054
represents a weight value;
using the mean square error as a loss function, the expression is:
Figure BDA0003464365700000055
in equation (3), MSE represents a loss function, m represents the number of samples of the data being trained,
Figure BDA0003464365700000056
representing the active power predicted by a preset Seq2Point model;
and updating the model parameters by adopting a reverse optimization algorithm, so that the model is converged to obtain the trained Seq2Point model.
With reference to the first aspect, optionally, the preset DAE model includes:
an encoder: training data and generating a set of feature maps, which sequentially comprise a convolutional layer, a linear activation function, a maximum pooling layer, an additional convolutional and pooling layer, a full link layer and a ReLU activation function shutdown encoder network;
a decoder: the system sequentially comprises a ReLU activation function, a full connection layer, an additional convolution and pooling layer, an up-sampling layer, a linear activation function and a convolution layer.
With reference to the first aspect, optionally, the preset Attention and residual network model includes:
an input layer: for inputting the feature x of each samplen；
Embedding layer: the method is used for generating low-dimensional dense vectors, and is convenient for feature extraction, and the transformed low-dimensional dense vectors are as follows:
h＝WX+b (4)
in the formula (4), h represents the output of a preset Attention and a hidden layer of a residual network model, W represents a transformation matrix, b represents a coefficient matrix, and the parameters of W and b are updated based on a BP algorithm in the model training process;
an encoder and a decoder: the vector used for splicing forward GRU and reverse GRU outputs;
an output layer: for outputting the result vector.
With reference to the first aspect, optionally, the training of the preset Attention includes:
according to the attention mechanism, a hidden state sequence [ h ] is utilized0,h1,...,hn-1]Converting the fixed semantic vector c into a dynamic variable semantic vector ctA parameter vector Q, K, V is defined, represented as follows:
Figure BDA0003464365700000071
in equation (5), the parameter vector Q, K, V is updated based on the BP algorithm during the model training process, and then:
Q＝K＝V＝h＝[h0,h1,...,hn-1] (6)
let the function between the residual connections of the residual network be f (X), then the output of the residual network be g (X) ═ X + f (X);
and (4) updating the model parameters by using the cross entropy as a loss function and adopting a random gradient descent algorithm to obtain a trained residual error network model.
With reference to the first aspect, further, the obtaining preliminary aggregation data of the active power of each load includes:
grouping the decomposition results into decomposition datasets
Figure BDA0003464365700000072
N-1, 2, …, N-T +1, wherein
Figure BDA0003464365700000073
The characteristics are represented by a plurality of symbols,
Figure BDA0003464365700000074
represents a target value;
aggregating features using pre-established perceptron aggregation models
Figure BDA0003464365700000075
Obtaining:
Figure BDA0003464365700000076
training a pre-established perceptron aggregate model by using a decomposition data set, selecting a mean square error as a loss function, and updating parameters of the pre-established perceptron aggregate model by adopting a random gradient descent algorithm to obtain the trained perceptron aggregate model;
obtaining preliminary aggregation data of active power of each load based on a trained perceptron aggregation model
Figure BDA0003464365700000077
n＝1,2,…,N-T+1。
With reference to the first aspect, further, the optimizing the pre-established DAE model includes:
decomposing the result
Figure BDA0003464365700000081
And preliminary aggregated data
Figure BDA0003464365700000082
Forming a training data set;
and training the pre-established DAE model by using a training data set, selecting a mean square error as a loss function, and updating parameters of the pre-established DAE model by adopting a random gradient descent algorithm to obtain the trained DAE model.
With reference to the first aspect, optionally, the model structure of the pre-established DAE model has 7 layers, which sequentially include:
the length of the layer 1 input layer is determined by the length of the input sequence;
processing input signals by the 2 nd layer convolution layer, extracting features to form a feature map, wherein 8 convolution kernels with the length of 4 are used, and a linear activation function is used;
expanding the feature diagram through a layer 3, then compressing the size through two full connection layers of a layer 4 and a layer 5, reducing the dimension of the sequence to obtain a compact form of the input sequence, and keeping effective information and removing noise information in the effective information;
taking the 5 th layer as a middle point, the 6 th and 7 th layers are symmetrical to the 2 nd and 4 th layers, the 7 th layer is a convolution layer, and the 6 th and 7 th layers are used for decoding data in a compact form output by the first 5 th layers to obtain an output sequence in the same form as the input sequence.
In a second aspect, the present invention provides a non-intrusive electrical load monitoring multi-model aggregation system, including:
an acquisition module: the system comprises a power supply, a control unit, a power supply and a power supply, wherein the power supply is used for acquiring active power of a user bus and each load;
a data preprocessing module: the device is used for pairing the active power of the user bus and the active power of each load to form a data set and processing the data set;
a decomposition module: the system comprises a data set, a load real-time analysis unit and a load real-time analysis unit, wherein the data set is used for carrying out primary training on three preset base models based on the processed data set, and decomposing the active power of each load by adopting the three base models after the primary training to obtain a decomposition result;
a preliminary polymerization module: the system comprises a decomposition module, a sensor aggregation module, a parameter setting module and a parameter setting module, wherein the decomposition module is used for aggregating decomposition results by using a pre-established sensor aggregation module, optimizing parameters of the pre-established sensor aggregation module and obtaining preliminary aggregation data of active power of each load based on the trained sensor aggregation module;
an output module: and the optimization module is used for optimizing the pre-established DAE model based on the decomposition result and the preliminary aggregation data, and the optimized DAE model is the aggregation model of the non-invasive load monitoring model and is used for outputting the data of the active power of each load.
In a third aspect, the present invention provides a computing device comprising a processor and a storage medium;
the storage medium is used for storing instructions;
the processor is configured to operate in accordance with the instructions to perform the steps of the method of the first aspect.
In a fourth aspect, the invention provides a computer-readable storage medium having stored thereon a computer program which, when executed by a processor, performs the steps of the method of the first aspect.
Compared with the prior art, the non-intrusive load monitoring model polymerization method and system provided by the embodiment of the invention have the beneficial effects that:
the invention obtains the active power of the user bus and each load; pairing the active power of the user bus and the active power of each load to form a data set, and processing the data set; and performing primary training on the three preset base models based on the processed data set, and decomposing the active power of each load by adopting the three base models after the primary training to obtain a decomposition result. According to the invention, through the primary training of the three base models, the advantages and disadvantages of each model for load identification of different electrical appliances are obtained through transverse comparison, and the optimal model for load identification of each electrical appliance can be obtained;
according to the method, a pre-established perceptron aggregation model is used for aggregating decomposition results, parameters of the pre-established perceptron aggregation model are optimized, and preliminary aggregation data of active power of each load are obtained based on the trained perceptron aggregation model. According to the invention, the trained perceptron aggregate model is used for preliminarily fusing the results of different models for identifying the same electric appliance load, so that the model with good electric appliance identification effect can obtain higher aggregate proportion in preliminary aggregation, and the identification precision is increased;
the DAE model established in advance is optimized based on the decomposition result and the preliminary aggregation data, and the optimized DAE model is the aggregation model of the non-invasive load monitoring model and is used for outputting the data of the active power of each load. The method can be combined with the three models for identification, has higher universality and universality, can obtain more accurate identification results, and can aggregate the results of different load decompositions to obtain the decomposition result superior to that of a single model.
Drawings
Fig. 1 is a block flow diagram of a non-intrusive load monitoring model aggregation method according to an embodiment of the present invention;
fig. 2 is a frame of a preset Seq2Point model in a non-intrusive load monitoring model aggregation method according to an embodiment of the present invention;
fig. 3 is a framework of a preset DAE model in a non-invasive load monitoring model aggregation method according to an embodiment of the present invention;
fig. 4 is a framework of the Attention and residual network model preset in the non-intrusive load monitoring model aggregation method according to an embodiment of the present invention.
Detailed Description
The invention is further described below with reference to the accompanying drawings. The following examples are only for illustrating the technical solutions of the present invention more clearly, and the protection scope of the present invention is not limited thereby.
The first embodiment is as follows:
as shown in fig. 1, an embodiment of the present invention provides a non-intrusive load monitoring model aggregation method, including:
acquiring active power of a user bus and each load;
pairing the active power of the user bus and the active power of each load to form a data set, and processing the data set;
performing primary training on three preset base models based on the processed data set, and decomposing the active power of each load by adopting the three base models after the primary training to obtain a decomposition result;
aggregating the decomposition results by using a preset perceptron aggregation model, optimizing parameters of the preset perceptron aggregation model, and obtaining preliminary aggregation data of active power of each load based on the trained perceptron aggregation model;
and optimizing the pre-established DAE model based on the decomposition result and the preliminary aggregation data, wherein the optimized DAE model is the aggregation model of the non-invasive load monitoring model and is used for outputting the data of the active power of each load.
The method comprises the following specific steps:
step 1: and acquiring the active power of the user bus and each load.
And extracting active power data of each load and each load cluster of the user according to the frequency of times/minute by using active power monitoring equipment, and arranging the active power data in a time sequence to obtain the active power of a user bus and the active power of each load.
Step 2: and pairing the active power of the user bus and the active power of each load to form a data set, and processing the data set.
Step 2.1: and pairing the active power of the user bus and the active power of each load to form a data set.
The sliding window obtains active power data with a specified length, and the method comprises the following steps: setting the length of the time sequence data intercepted by the sliding window as T, gradually carrying out sliding interception on the acquired active power of the user bus and the active power of each load according to time sequence by taking 1 as a step length, setting the total length of the data as N,
then, the data of the active power of the user bus is obtained as follows:
Figure BDA0003464365700000121
the data of the active power of the ith load are as follows:
Figure BDA0003464365700000122
the active power of the user bus and the active power of each load are paired to form a data set, and the method comprises the following steps:
data for making user bus active power
Figure BDA0003464365700000123
Wherein xnT is a vector, N is 1,2, …, N-T +1,
data of active power of ith load
Figure BDA0003464365700000124
Wherein
Figure BDA0003464365700000125
T is vector, N is 1,2, …, N-T + 1;
then
Figure BDA0003464365700000126
For a subscriber busAnd the active power of each load are paired.
Step 2.2: the data set is processed.
Carrying out normalization processing on the data, wherein the normalization processing comprises the following steps: and taking out the maximum value of the active power of the user bus, and dividing the active power of the user bus and the active power of each load by the maximum value of the active power of the user bus respectively.
And step 3: and performing primary training on the three preset base models based on the processed data set, and decomposing the active power of each load by adopting the three base models after the primary training to obtain a decomposition result.
Training a preset Seq2Point model by using the data set subjected to the ith load processing, and updating model parameters by adopting a reverse optimization algorithm to obtain a trained Seq2Point model;
training a preset DAE model by using the data set subjected to the ith load processing, and updating model parameters by adopting a random gradient descent algorithm to obtain a trained DAE model;
and training a preset Attention and a residual error network model by using the data set after the ith load processing, updating the Attention parameter by adopting a BP algorithm to obtain the Attention after the training is finished, and updating the model parameter by adopting a random gradient descent algorithm based on the Attention after the training is finished to obtain the residual error network model after the training is finished.
Specifically, as shown in fig. 2, the preset Seq2Point model includes:
an input layer: will be provided with
Figure BDA0003464365700000131
The T sampling points are input into a preset Seq2Point model, and the tensor of the input data is (T, 1);
forward GRU layer: mining a change rule in the data of the input layer, wherein the size of a hidden layer is 2T, the tensor of input data of the second forward GRU layer is (T,2T), and the tensor of output data is (4T, 1);
full connection layer: mapping the high-dimensional data to the low-dimensional data, wherein the input data tensor is (4T,1), and the output data tensor is (2T, 1);
an output layer: the normalized power value of the ith load is output, the input data tensor is (2T,1), and the output data tensor is (1, 1).
Training a preset Seq2Point model, comprising:
the activation function employed in each cell is the Relu function, expressed as:
Figure BDA0003464365700000132
the calculation formula of the GRU layer is expressed as follows:
Figure BDA0003464365700000133
in the formula (2), rtIndicating gating of reset, WrRepresents the weight of the reset gate, sigma represents the sigmoid function, ht-1Representing the state of the last neuron descended, ztIndicates the state of the last neuron, WzIndicating that the weight of the door is updated,
Figure BDA0003464365700000141
is indicative of the current state of the neuron,
Figure BDA0003464365700000142
represents a weight value;
using the mean square error as a loss function, the expression is:
Figure BDA0003464365700000143
in equation (3), MSE represents a loss function, m represents the number of samples of the data being trained,
Figure BDA0003464365700000144
representing the active power predicted by a preset Seq2Point model;
and updating the model parameters by adopting a reverse optimization algorithm, so that the model is converged to obtain the trained Seq2Point model.
Specifically, as shown in fig. 3, the preset DAE model includes:
an encoder: training data and generating a set of feature maps, which sequentially comprise a convolutional layer, a linear activation function, a maximum pooling layer, an additional convolutional and pooling layer, a full link layer and a ReLU activation function shutdown encoder network;
a decoder: the system sequentially comprises a ReLU activation function, a full connection layer, an additional convolution and pooling layer, an up-sampling layer, a linear activation function and a convolution layer.
Specifically, as shown in fig. 4, the preset Attention and residual network model includes:
an input layer: for inputting the feature x of each samplen；
Embedding layer: the method is used for generating low-dimensional dense vectors, and is convenient for feature extraction, and the transformed low-dimensional dense vectors are as follows:
h＝WX+b (4)
in the formula (4), h represents the output of a preset Attention and a hidden layer of a residual error network model, W represents a transformation matrix, b represents a coefficient matrix, and the parameters of W and b are updated based on a BP algorithm in the model training process;
an encoder and a decoder: the vector used for splicing forward GRU and reverse GRU outputs;
an output layer: for outputting the result vector.
Training a preset Attention, comprising:
according to the attention mechanism, a hidden state sequence [ h ] is utilized0,h1,...,hn-1]Converting the fixed semantic vector c into a dynamic variable semantic vector ctA parameter vector Q, K, V is defined, represented as follows:
Figure BDA0003464365700000151
in equation (5), the parameter vector Q, K, V is updated based on the BP algorithm during the model training process, and then:
Q＝K＝V＝h＝[h0,h1,...,hn-1] (6)
let the function between the residual connections of the residual network be f (X), then the output of the residual network be g (X) ═ X + f (X);
and (3) updating the model parameters by using the cross entropy as a loss function and adopting a random gradient descent algorithm to obtain a trained residual error network model.
Processing x in the obtained data set by adopting three base model pairs after primary trainingnAnd performing active power decomposition on each load, wherein the decomposition results of the ith load decomposed by the three trained basic models are respectively as follows:
the decomposition result of the trained Seq2Point model:
Figure BDA0003464365700000152
decomposition results of the trained DAE model:
Figure BDA0003464365700000153
decomposing the trained residual error network model:
Figure BDA0003464365700000161
according to the invention, through the primary training of the three base models, the advantages and disadvantages of each model for load identification of different electrical appliances are obtained through transverse comparison, the optimal model for load identification of each electrical appliance can be obtained, and the optimal decomposition result for load identification of each electrical appliance can be obtained.
And 5: and aggregating the decomposition results by using a preset perceptron aggregation model, optimizing parameters of the preset perceptron aggregation model, and obtaining preliminary aggregation data of the active power of each load based on the trained perceptron aggregation model.
Composing the decomposition results into a decomposition dataset
Figure BDA0003464365700000162
N-1, 2, …, N-T +1, wherein
Figure BDA0003464365700000163
The characteristics are represented by a set of data,
Figure BDA0003464365700000164
represents a target value;
aggregating features using pre-established perceptron aggregation models
Figure BDA0003464365700000165
Obtaining:
Figure BDA0003464365700000166
training a pre-established perceptron aggregate model by using a decomposition data set, selecting a mean square error as a loss function, and updating parameters of the pre-established perceptron aggregate model by using a random gradient descent algorithm to obtain the trained perceptron aggregate model;
obtaining preliminary aggregation data of active power of each load based on a trained perceptron aggregation model
Figure BDA0003464365700000167
n＝1,2,…,N-T+1。
According to the invention, the trained perceptron aggregate model is used for preliminarily fusing the results of different models for identifying the same electric appliance load, and the model with good electric appliance identification effect can obtain higher aggregate proportion in preliminary aggregation, so that the identification precision is increased.
And 6: and optimizing the pre-established DAE model according to the decomposition result and the preliminary aggregation data, wherein the optimized DAE model is the aggregation model of the non-invasive load monitoring model and is used for outputting the data of the active power of each load.
The model structure of the pre-established DAE model has 7 layers, which are sequentially as follows:
the length of the layer 1 input layer is determined by the length of the input sequence;
processing input signals by the 2 nd layer convolution layer, extracting features to form a feature map, wherein 8 convolution kernels with the length of 4 are used, and a linear activation function is used;
expanding the characteristic diagram through a 3 rd layer, then compressing the size through two full connection layers of a 4 th layer and a 5 th layer, reducing the dimension of the sequence to obtain a compact form of the input sequence, and keeping effective information and removing noise information in the effective information;
taking the 5 th layer as a middle point, the 6 th and 7 th layers are symmetrical to the 2 nd and 4 th layers, the 7 th layer is a convolution layer, and the 6 th and 7 th layers are used for decoding data in a compact form output by the first 5 th layers to obtain an output sequence in the same form as the input sequence.
Optimizing a pre-established DAE model, comprising:
decomposing the result
Figure BDA0003464365700000171
And preliminarily aggregate the data
Figure BDA0003464365700000172
Forming a training data set;
and training the pre-established DAE model by using a training data set, selecting the mean square error as a loss function, and updating the parameters of the pre-established DAE model by adopting a random gradient descent algorithm to obtain the trained DAE model.
The method can be combined with three models for identification, has higher universality and universality, can obtain more accurate identification results, and can aggregate the results of different load decompositions to obtain a decomposition result superior to that of a single model.
Example two:
the embodiment of the invention provides a non-invasive electrical load monitoring multi-model polymerization system, which comprises:
an acquisition module: the system comprises a power supply, a control unit, a power supply and a power supply, wherein the power supply is used for acquiring active power of a user bus and each load;
a data preprocessing module: the device is used for pairing the active power of the user bus and the active power of each load to form a data set and processing the data set;
a decomposition module: the system comprises a data set, a load real-time analysis unit and a load real-time analysis unit, wherein the data set is used for carrying out primary training on three preset base models based on the processed data set, and decomposing the active power of each load by adopting the three base models after the primary training to obtain a decomposition result;
a preliminary polymerization module: the system comprises a decomposition module, a sensor aggregation module, a parameter setting module and a parameter setting module, wherein the decomposition module is used for aggregating decomposition results by using a pre-established sensor aggregation module, optimizing parameters of the pre-established sensor aggregation module and obtaining preliminary aggregation data of active power of each load based on the trained sensor aggregation module;
an output module: and the optimization module is used for optimizing the pre-established DAE model based on the decomposition result and the preliminary aggregation data, and the optimized DAE model is the aggregation model of the non-invasive load monitoring model and is used for outputting the data of the active power of each load.
Example three:
the embodiment of the invention provides a computing device, which comprises a processor and a storage medium;
the storage medium is to store instructions;
the processor is configured to operate in accordance with the instructions to perform the steps of the method of embodiment one.
Example four:
embodiments of the present invention further provide a computer-readable storage medium, on which a computer program is stored, where the computer program, when executed by a processor, implements the steps of the method of an embodiment.
As will be appreciated by one skilled in the art, embodiments of the present application may be provided as a method, system, or computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.
The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the application. It will be understood that each flow and/or block of the flow diagrams and/or block diagrams, and combinations of flows and/or blocks in the flow diagrams and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
The above description is only a preferred embodiment of the present invention, and it should be noted that, for those skilled in the art, it is possible to make various improvements and modifications without departing from the technical principle of the present invention, and those improvements and modifications should be considered as the protection scope of the present invention.

Claims (10)

1. A method for non-intrusive load monitoring model aggregation, comprising:
acquiring active power of a user bus and each load;
pairing the active power of the user bus and the active power of each load to form a data set, and processing the data set;
performing primary training on three preset base models based on the processed data set, and decomposing active power of each load by adopting the three base models after primary training to obtain a decomposition result;
aggregating the decomposition results by using a preset perceptron aggregation model, optimizing parameters of the preset perceptron aggregation model, and obtaining preliminary aggregation data of active power of each load based on the trained perceptron aggregation model;
and optimizing the pre-established DAE model based on the decomposition result and the preliminary aggregation data, wherein the optimized DAE model is the aggregation model of the non-invasive load monitoring model and is used for outputting the data of the active power of each load.
2. The method for aggregating the non-intrusive load monitoring models as defined in claim 1, wherein the obtaining the active power of the user bus and the loads comprises: and extracting active power data of each load and each load cluster of the user according to the frequency of times/minute by using active power monitoring equipment, and arranging the active power data in a time sequence to obtain the active power of a user bus and the active power of each load.
3. The method of non-intrusive load monitoring model aggregation according to claim 1, wherein the composition data set comprises:
the sliding window obtains active power data with a specified length, and the method comprises the following steps: setting the length of the time sequence data intercepted by the sliding window as T, gradually carrying out sliding interception on the acquired active power of the user bus and the active power of each load according to time sequence by taking 1 as a step length, setting the total length of the data as N,
then, the data of the active power of the user bus is obtained as follows:
Figure FDA0003464365690000021
the data of the active power of the ith load are as follows:
Figure FDA0003464365690000022
the active power of the user bus and the active power of each load are paired to form a data set, and the method comprises the following steps:
data for making user bus active power
Figure FDA0003464365690000023
Wherein x isnT is a vector, N is 1,2, …, N-T +1,
data of active power of ith load
Figure FDA0003464365690000024
Wherein
Figure FDA0003464365690000025
T is vector, N is 1,2, …, N-T + 1;
then the
Figure FDA0003464365690000026
And the data sets are paired with the active power of the user bus and the active power of each load.
4. The method for non-intrusive load monitoring model aggregation as defined in claim 1, wherein the processing the data set comprises:
carrying out normalization processing on the data, wherein the normalization processing comprises the following steps: and taking out the maximum value of the active power of the user bus, and dividing the active power of the user bus and the active power of each load by the maximum value of the active power of the user bus respectively.
5. The method for aggregating non-intrusive load monitoring models as defined in claim 1, wherein the initial training of the predetermined three base models comprises:
training a preset Seq2Point model by using the data set subjected to the ith load processing, and updating model parameters by adopting a reverse optimization algorithm to obtain a trained Seq2Point model;
training a preset DAE model by using the data set subjected to the ith load processing, and updating model parameters by adopting a random gradient descent algorithm to obtain a trained DAE model;
training a preset Attention and a residual network model by using the data set subjected to the ith load processing, updating an Attention parameter by adopting a BP algorithm to obtain a trained Attention, and updating a model parameter by adopting a random gradient descent algorithm based on the trained Attention to obtain a trained residual network model;
processing x in the obtained data set by adopting three base model pairs finished by primary trainingnAnd performing active power decomposition on each load, wherein the decomposition results of the ith load decomposed by the three trained basic models are respectively as follows:
the decomposition result of the trained Seq2Point model:
Figure FDA0003464365690000031
decomposition results of the trained DAE model:
Figure FDA0003464365690000032
and (3) decomposing the trained residual error network model:
Figure FDA0003464365690000033
6. the method for polymerizing the non-intrusive load monitoring model according to claim 5, wherein the obtaining of the preliminary polymerization data of the active power of each load comprises:
composing the decomposition results into a decomposition dataset
Figure FDA0003464365690000034
Wherein
Figure FDA0003464365690000035
The characteristics are represented by a plurality of symbols,
Figure FDA0003464365690000036
represents a target value;
aggregating features using pre-established perceptron aggregation models
Figure FDA0003464365690000037
Obtaining:
Figure FDA0003464365690000041
training a pre-established perceptron aggregate model by using a decomposition data set, selecting a mean square error as a loss function, and updating parameters of the pre-established perceptron aggregate model by using a random gradient descent algorithm to obtain the trained perceptron aggregate model;
obtaining preliminary aggregation data of active power of each load based on a trained perceptron aggregation model
Figure FDA0003464365690000042
7. The method for non-intrusive load monitoring model aggregation as defined in claim 6, wherein the optimizing the pre-established DAE model comprises:
decomposing the result
Figure FDA0003464365690000043
And preliminary aggregated data
Figure FDA0003464365690000044
Forming a training data set;
and training the pre-established DAE model by using a training data set, selecting the mean square error as a loss function, and updating the parameters of the pre-established DAE model by adopting a random gradient descent algorithm to obtain the trained DAE model.
8. A non-intrusive electrical load monitoring multimodal aggregation system, comprising:
an acquisition module: the system comprises a power supply, a control unit, a power supply and a power supply, wherein the power supply is used for acquiring active power of a user bus and each load;
a data preprocessing module: the device is used for pairing the active power of the user bus and the active power of each load to form a data set and processing the data set;
a decomposition module: the system comprises a data set, a load real-time analysis unit and a load real-time analysis unit, wherein the data set is used for carrying out primary training on three preset base models based on the processed data set, and decomposing the active power of each load by adopting the three base models after the primary training to obtain a decomposition result;
a preliminary polymerization module: the system comprises a decomposition module, a sensor aggregation module, a parameter setting module and a parameter setting module, wherein the decomposition module is used for aggregating decomposition results by using a pre-established sensor aggregation module, optimizing parameters of the pre-established sensor aggregation module and obtaining preliminary aggregation data of active power of each load based on the trained sensor aggregation module;
an output module: and the optimization module is used for optimizing a pre-established DAE model based on the decomposition result and the preliminary aggregation data, and the optimized DAE model is an aggregation model of the non-intrusive load monitoring model and is used for outputting data of active power of each load.
9. A computing device comprising a processor and a storage medium;
the storage medium is used for storing instructions;
the processor is configured to operate in accordance with the instructions to perform the steps of the method of any of claims 1 to 7.
10. Computer-readable storage medium, on which a computer program is stored which, when being executed by a processor, carries out the steps of the method according to any one of claims 1 to 7.
CN202210025465.7A 2022-01-11 2022-01-11 Non-invasive load monitoring model polymerization method and system Pending CN114545066A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210025465.7A CN114545066A (en) 2022-01-11 2022-01-11 Non-invasive load monitoring model polymerization method and system

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210025465.7A CN114545066A (en) 2022-01-11 2022-01-11 Non-invasive load monitoring model polymerization method and system

Publications (1)

Publication Number Publication Date
CN114545066A true CN114545066A (en) 2022-05-27

Family

ID=81670339

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210025465.7A Pending CN114545066A (en) 2022-01-11 2022-01-11 Non-invasive load monitoring model polymerization method and system

Country Status (1)

Country Link
CN (1) CN114545066A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114977176A (en) * 2022-07-19 2022-08-30 深圳江行联加智能科技有限公司 Power load decomposition method, device, equipment and storage medium

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114977176A (en) * 2022-07-19 2022-08-30 深圳江行联加智能科技有限公司 Power load decomposition method, device, equipment and storage medium

Similar Documents

Publication Publication Date Title
CN111428201B (en) Prediction method for time series data based on empirical mode decomposition and feedforward neural network
CN109255728B (en) Photovoltaic power generation power neural network prediction method based on chaotic phase space optimization reconstruction
CN111008726B (en) Class picture conversion method in power load prediction
CN106067034B (en) Power distribution network load curve clustering method based on high-dimensional matrix characteristic root
CN111861023A (en) Statistical-based hybrid wind power prediction method and device
CN114969060B (en) Industrial equipment time sequence data compression storage method and device
CN116702831A (en) Hybrid short-term wind power prediction method considering massive loss of data
CN114545066A (en) Non-invasive load monitoring model polymerization method and system
CN116599050A (en) Photovoltaic prediction method and related device based on self-attention mechanism
CN117688362A (en) Photovoltaic power interval prediction method and device based on multivariate data feature enhancement
CN116384574A (en) W-LSTM-based non-parameter load interval prediction method
CN108459585B (en) Power station fan fault diagnosis method based on sparse local embedded deep convolutional network
CN116679231A (en) Lithium battery SoH estimation method based on gram angle field and VGG16 model
CN116338322A (en) Power grid line impedance prediction method and system
CN112561050A (en) Neural network model training method and device
CN113111588B (en) NO of gas turbine X Emission concentration prediction method and device
CN113947237A (en) AMS-TCN-based wind power error correction method
CN114118568A (en) Multi-attention-based power consumption prediction method for convolutional neural network
CN114792026A (en) Method and system for predicting residual life of aircraft engine equipment
CN116698410B (en) Rolling bearing multi-sensor data monitoring method based on convolutional neural network
CN117497194B (en) Biological information processing method and device, electronic equipment and storage medium
CN117522049A (en) Method, device, equipment and storage medium for predicting energy load
CN106656196A (en) Algorithm integrated with forward backward pursuit and based on compression perception theory
CN114352485A (en) Wind speed prediction method and device for wind driven generator and computer equipment
CN116953379A (en) Non-invasive load monitoring method and terminal based on compressive sensing blind analysis

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