CN114511058B - Load element construction method and device for electric power user portrait - Google Patents
Load element construction method and device for electric power user portrait Download PDFInfo
- Publication number
- CN114511058B CN114511058B CN202210101090.8A CN202210101090A CN114511058B CN 114511058 B CN114511058 B CN 114511058B CN 202210101090 A CN202210101090 A CN 202210101090A CN 114511058 B CN114511058 B CN 114511058B
- Authority
- CN
- China
- Prior art keywords
- user
- active power
- feature vector
- load element
- layer
- 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.)
- Active
Links
- 238000010276 construction Methods 0.000 title claims abstract description 15
- 239000013598 vector Substances 0.000 claims abstract description 217
- 238000012549 training Methods 0.000 claims abstract description 30
- 238000012545 processing Methods 0.000 claims abstract description 24
- 238000000034 method Methods 0.000 claims abstract description 18
- 230000004913 activation Effects 0.000 claims abstract description 13
- 238000007477 logistic regression Methods 0.000 claims abstract description 11
- 238000004364 calculation method Methods 0.000 claims abstract description 8
- 239000011159 matrix material Substances 0.000 claims description 14
- 229910002056 binary alloy Inorganic materials 0.000 claims description 10
- 230000005611 electricity Effects 0.000 claims description 5
- 238000009499 grossing Methods 0.000 claims description 4
- 230000006870 function Effects 0.000 description 17
- 238000005516 engineering process Methods 0.000 description 6
- 238000012986 modification Methods 0.000 description 4
- 230000004048 modification Effects 0.000 description 4
- 238000007476 Maximum Likelihood Methods 0.000 description 2
- 230000004075 alteration Effects 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 2
- 238000007405 data analysis Methods 0.000 description 2
- 230000003213 activating effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000015572 biosynthetic process Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000011478 gradient descent method Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/31—Indexing; Data structures therefor; Storage structures
- G06F16/316—Indexing structures
- G06F16/325—Hash tables
-
- 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
-
- 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
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Software Systems (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Biophysics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Databases & Information Systems (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Evolutionary Biology (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a load element construction method and a device for an electric power user portrait, wherein the method comprises the following steps: collecting total power of a household user and active power of a single load element as a training set; carrying out thermal single code coding on the total power of the home subscriber and the active power of a single load element; carrying out hash processing on the user characteristic vector and the active power characteristic vector; respectively inputting the low-dimensional user feature vector and the low-dimensional active power feature vector into a first multi-layer perceptron network and a second multi-layer perceptron network for training, and respectively outputting each user embedded feature vector and each load element embedded feature vector; calculating the relevance of each user embedded feature vector and the load element embedded feature vector based on cosine similarity, and converting a relevance calculation result into posterior probability by utilizing a logistic regression activation function to obtain a load element with highest relevance to the user; the method can improve the efficiency of establishing the user portrait model and has strong reliability.
Description
Technical Field
The invention relates to the technical field of power, in particular to a method and a device for constructing a load element for a power consumer portrait.
Background
With the continuous development of electronic technology and big data technology, the formation of household electricity information is gradually converted into a data form, so that willingness information can be intuitively and conveniently obtained, the time consumption is further reduced, and the efficiency is improved. In order to find the target data more quickly and accurately in the network, various attribute information of household user electricity consumption needs to be divided by using a related classification technology. Supply and demand information of the home user can be further known through portrait construction, data such as behavior habits and the like are accurately positioned, and the overall view of the user information is conveniently known.
In the technical field of portrayal, embedded features are generated by utilizing artificially constructed features through a deep learning method and are input into a load element which is predicted to be opened by a user in a neural network, and finally, the habit of electricity utilization behavior of a household user is displayed. Load element opening prediction is the most important consideration of model effect, and a user is crucial to the historical opening behavior of the load element. However, most current power consumer representation models utilize some artificial features as raw features to input into the multi-layer neural network, which are derived from the consumer representation, the consumer's historical behavior, the load element properties, etc.
For example, patent document CN112417308A discloses a user portrait tag generating method based on electric power big data, which generates a user characteristic tag by using big data processing technology, wherein a basic database configuration of the big data processing technology is constructed around customer appeal, opinion and consultation data streams of channels such as electric power official telephones, electric power intranet and extranet, mobile phone APP, weChat public number lines, business hall opinion books are imported into the basic database as tag data original sources, and clients are marked in a tag form through data analysis, so that a user portrait is established.
The method can integrate various source data, build a multi-dimensional and three-dimensional customer portrait by means of big data analysis technology, and describe deep behavior characteristics of users through labels. However, the method is biased to adopt consultation opinions provided by users to describe the characteristics of the users, lacks acquisition of direct electricity information, does not well integrate structural features and node features of complex recommended network scenes, is poor in model reliability, is low in operation speed, and is low in user portrait model establishment efficiency.
Disclosure of Invention
The invention provides a method and a device for constructing a load element for an electric power user portrait, which are used for carrying out model training based on the total power of a household user and the active power of a single load element, and are used for integrating the structural characteristics and the node characteristics of a complex recommended network scene, so that the efficiency of establishing a user portrait model is improved, and the reliability is high.
A load element construction method for an electric power user portrait includes:
collecting total power of a household user and active power of a single load element as a training set;
performing thermal single code coding on the total power of the home user and the active power of a single load element to respectively obtain a user characteristic vector and an active power characteristic vector;
respectively carrying out hash processing on the user characteristic vector and the active power characteristic vector to respectively obtain a low-dimensional user characteristic vector and a low-dimensional active power characteristic vector;
the low-dimensional user feature vector and the low-dimensional active power feature vector are respectively input into a first multi-layer perceptron network and a second multi-layer perceptron network for training, and each user embedded feature vector and each load element embedded feature vector are respectively output;
and calculating the correlation between each user embedded feature vector and the load element embedded feature vector based on the cosine similarity, and converting the correlation calculation result into posterior probability by utilizing a logistic regression activation function to obtain the load element with the highest degree of correlation with the user.
Further, performing hot unique code encoding on the total power of the home user, including:
dividing the same family user total power into user similar characteristics;
the similar characteristics of each user are expressed by binary system, and the user characteristic vector is obtained;
performing thermal single code encoding on the active power of the single load element to obtain an active power characteristic vector of the single load element, including:
dividing the active power of the same single load element into active power similar characteristics;
and (5) representing each active power similar characteristic by using a binary system to obtain an active power characteristic vector.
Further, hash processing is performed on the user feature vector to obtain a low-dimensional user feature vector, including:
respectively adding a start mark and an end mark before and after the user feature vector;
setting step length and sliding window based on N-GRAM algorithm, recording user characteristic vector slices of each scratch, recording all user characteristic vector slices as vector representations composed of a plurality of numbers, and obtaining low-dimension user characteristic vectors;
performing hash processing on the active power feature vector to obtain a low-dimension active power feature vector, including:
respectively adding a start mark and an end mark before and after the active power characteristic vector;
based on an N-GRAM algorithm, setting a step length and a sliding window, recording active power characteristic vector slices which are marked each time, recording all the active power characteristic vector slices as vector representations consisting of a plurality of numbers, and obtaining a low-dimensional active power characteristic vector.
Further, the low-dimensional user feature vector and the low-dimensional active power feature vector are respectively input into a first multi-layer perceptron network and a second multi-layer perceptron network for training, and parameters of the first multi-layer perceptron network and the second multi-layer perceptron network are respectively obtained according to a minimized loss function.
Further, the parameters of the first multi-layer perceptron network and the second multi-layer perceptron network comprise weight matrixes and bias items of all hidden layers.
Further, the load element embedding feature vector is calculated by the following formula:
l i ＝f(W i l i-1 +b i )i＝2,…,N-1；
y L ＝f(W N l N-1 +b N )；
wherein ,yL Embedding feature vectors for load elements, l i Hidden layer of the ith layer of the first multi-layer perceptron network, W i B is a weight matrix of an ith hidden layer of the first multi-layer perceptron network i Bias item representing ith hidden layer of first multi-layer perceptron network, N is hidden layer number of first multi-layer perceptron network, b N Bias items for the output layer of the first multi-layer perceptron network;
the user-embedded feature vector is calculated by the following formula:
l m ＝f(W m l m-1 +b m )m＝2,…,M-1；
y U ＝f(W M l M-1 +b M )；
wherein ,yU Embedding feature vectors for users, l m Hidden layer of the m-th layer of the second multi-layer perceptron network, W m B is a weight matrix of an mth hidden layer of the second multi-layer perceptron network m Bias term representing mth hidden layer of second multi-layer perceptron network, M is hidden layer number of second multi-layer perceptron network, b M And the bias item of the output layer of the second multi-layer perceptron network.
Further, the correlation of the user embedded feature vector and the load element embedded feature vector is calculated by the following formula:
wherein ,yU Embedding feature vectors for users, y L The feature vector is embedded for the load element,
representing a transpose of the user embedded feature vector, R (U, L) representing a correlation of the user embedded feature vector with the load element embedded feature vector.
Further, the posterior probability is calculated by the following formula:
wherein ,P(Lj U) is a posterior probability, γ is a smoothing factor of a logistic regression activation function, L represents a set of load elements, U represents a set of users, L' represents a sum of all load elements in the set of load elements, j is "+" or "-", L + Representing a positive sample of the load cell, L - Representing a load cell negative sample.
Further, the minimized loss function is expressed by the following formula:
wherein Λ is a parameter of the multi-layer perceptron network, P (L + U) represents the posterior probability, U represents the set of users, L + Representing a positive sample of the load cell.
A load cell construction apparatus for a representation of an electric power consumer, comprising:
the acquisition module is used for acquiring the total power of the household users and the active power of the single load element as a training set;
the coding module is used for carrying out hot unique code coding on the total power of the home user and the active power of the single load element to respectively obtain a user characteristic vector and an active power characteristic vector;
the processing module is used for respectively carrying out hash processing on the user characteristic vector and the active power characteristic vector to respectively obtain a low-dimensional user characteristic vector and a low-dimensional active power characteristic vector;
the training module is used for respectively inputting the low-dimensional user characteristic vector and the low-dimensional active power characteristic vector into a first multi-layer perceptron network and a second multi-layer perceptron network for training, and respectively outputting each user embedded characteristic vector and each load element embedded characteristic vector;
and the calculating module is used for calculating the correlation between each user embedded feature vector and the load element embedded feature vector based on the cosine similarity, converting the correlation calculation result into posterior probability by utilizing a logistic regression activation function, and obtaining the load element with the highest degree of correlation with the user.
The invention provides a load element construction method and a load element construction device for an electric power user portrait, which at least comprise the following beneficial effects:
the method comprises the steps of carrying out model training based on total power of household users and active power of load elements, carrying out hash processing on characteristics obtained through thermal single code coding, carrying out multi-layer perceptron network training, generating similarity between users and loads by means of a double-tower model, and finally obtaining the load element with highest degree of association with the users, wherein structural characteristics and node characteristics of complex recommended network scenes are fused, manual priori deviation is avoided to a certain extent, operation speed during on-line prediction is effectively improved, user portrait establishment efficiency is high, and reliability is high.
Drawings
FIG. 1 is a flow chart of one embodiment of a method for constructing a load cell for an electrical consumer representation provided by the present invention.
FIG. 2 is a schematic diagram of an embodiment of a load cell construction apparatus for consumer representation of electrical power provided by the present invention.
Description of the drawings: the system comprises a 1-acquisition module, a 2-coding module, a 3-processing module, a 4-training module and a 5-calculating module.
Detailed Description
In order to better understand the above technical solutions, the following detailed description will be given with reference to the accompanying drawings and specific embodiments.
Referring to FIG. 1, in some embodiments, a load cell construction method for a representation of an electrical consumer is provided, comprising:
s1, collecting total power of a household user and active power of a single load element as a training set;
s2, performing hot unique code coding on the total power of the home user and the active power of a single load element to respectively obtain a user characteristic vector and an active power characteristic vector;
s3, carrying out hash processing on the user characteristic vector and the active power characteristic vector respectively to obtain a low-dimensional user characteristic vector and a low-dimensional active power characteristic vector respectively;
s4, respectively inputting the low-dimensional user feature vector and the low-dimensional active power feature vector into a first multi-layer perceptron network and a second multi-layer perceptron network for training, and respectively outputting each user embedded feature vector and each load element embedded feature vector;
s5, calculating the correlation between each user embedded feature vector and the load element embedded feature vector based on cosine similarity, and converting a correlation calculation result into posterior probability by using a logistic regression activation function to obtain the load element with the highest degree of correlation with the user.
Specifically, in step S2, the hot-unique code encoding is performed on the total power of the home subscriber, including:
s211, dividing the total power of the same household user into user similar characteristics;
s212, representing the similar characteristics of each user by using binary system to obtain the user characteristic vector;
performing thermal single code encoding on the active power of the single load element to obtain an active power characteristic vector of the single load element, including:
s221, dividing the active power of the same single load element into active power similar characteristics;
s222, representing each active power similar characteristic by binary system to obtain an active power characteristic vector.
In a specific embodiment, the total power of the home subscriber and the active power of each load element are processed separately, the first multi-layer perceptron network is used for processing the total power of the home subscriber, the second multi-layer perceptron network is used for processing the active power of each load element, and the first multi-layer perceptron network and the second multi-layer perceptron network form a double-tower model.
Specifically, the present active power data is first used as a dictionary, the number of word banks is the dimension of the thermal single code coding, the dictionary is a zero vector of the word bank number x 1 dimension, and the zero vector is 1 only where the current active power value is represented, for example, the power dictionary of a refrigerator is [0,1,2,3,4 … … ], and if the current power is 0, the thermal single code coding result is [1,0 … ].
When the active power of the same single load element is divided into the same kind of characteristics, only one bit of each characteristic in each sample needs to be enabled, and the rest bits are disabled.
In step S3, hash processing is performed on the user feature vector to obtain a low-dimensional user feature vector, including:
s311, respectively adding a start mark and an end mark before and after the user feature vector;
s312, setting step length and sliding window based on an N-GRAM algorithm, recording user feature vector slices scratched each time, recording all user feature vector slices as vector representations consisting of a plurality of numbers, and obtaining a low-dimensional user feature vector;
performing hash processing on the active power feature vector to obtain a low-dimension active power feature vector, including:
s321, respectively adding a start mark and an end mark before and after the active power characteristic vector;
s322, setting step length and sliding window based on an N-GRAM algorithm, recording active power characteristic vector slices of each pass, recording all active power characteristic vector slices as vector representations composed of a plurality of numbers, and obtaining a low-dimensional active power characteristic vector. The algorithm may be represented by the following formula:
l 1 ＝W 1 x
wherein x represents an input feature vector of the thermal single code coding, l 1 Represents a first hidden layer, W 1 Representing the weight matrix of the first layer.
In a specific embodiment, taking a refrigerator as an example of a load element, when the active power of the refrigerator at a certain moment is 258W, a sign #258# is added to the beginning and the end of a user feature vector, then the active power is divided into n numbers, and when n=3 is taken, the numbers are #25, 258 and 58#, so that the dictionary obtained by the hot unique code coding can be represented by a vector formed by a plurality of 3 numbers, and the dimension of the dictionary is greatly reduced. For example, when the active power is 259W and the hot-unique code dictionary is [ … #25, … 258, 259, 260, … 59# … ], the hash-processed code is [ … 1, … 0,1,0 … 1 … ].
In step S4, the low-dimensional user feature vector and the low-dimensional active power feature vector are respectively input into a first multi-layer perceptron network and a second multi-layer perceptron network for training, and parameters of the first multi-layer perceptron network and the second multi-layer perceptron network are respectively obtained according to a minimized loss function.
The parameters of the first multi-layer perceptron network and the second multi-layer perceptron network comprise weight matrixes and bias items of all hidden layers.
The load element embedding feature vector is calculated by the following formula:
l i ＝f(W i l i-1 +b i )i＝2,…,N-1；
y L ＝f(W N l N-1 +b N )；
wherein ,yL Embedding feature vectors for load elements, l i Hidden layer of the ith layer of the first multi-layer perceptron network, W i B is a weight matrix of an ith hidden layer of the first multi-layer perceptron network i Bias item representing ith hidden layer of first multi-layer perceptron network, N is hidden layer number of first multi-layer perceptron network, b N Bias items for the output layer of the first multi-layer perceptron network;
the user-embedded feature vector is calculated by the following formula:
l m ＝f(W m l m-1 +b m )m＝2,…,M-1；
y U ＝f(W M l M-1 +b M )；
wherein ,yU Embedding feature vectors for users, l m Hidden layer of the m-th layer of the second multi-layer perceptron network, W m B is a weight matrix of an mth hidden layer of the second multi-layer perceptron network m Bias term representing mth hidden layer of second multi-layer perceptron network, M is hidden layer number of second multi-layer perceptron network, b M And the bias item of the output layer of the second multi-layer perceptron network.
Wherein, the activation functions of the hidden layer and the output layer adopt a tanh function, which is expressed by the following formula:
in a specific embodiment, the first multi-layer perceptron network and the second multi-layer perceptron network each comprise an input layer, a hidden layer and an output layer, after the low-dimensional user feature vector and the low-dimensional active power feature vector are input to the input layer, activating and embedding feature vector calculation is performed in the hidden layer and the output layer, and then the finally obtained 128-dimensional vector is output as an embedding feature vector.
In step S5, the correlation between the user embedded feature vector and the load element embedded feature vector is calculated by the following formula:
wherein ,yU For embedding feature vectors for users, y L The feature vector is embedded for the load element,
representing a transpose of the user embedded feature vector, R (U, L) representing a correlation of the user embedded feature vector with the load element embedded feature vector.
In order to obtain parameters of the multi-layer perceptron network and the embedded feature vector output network, the parameters comprise weight matrixes and bias items, and the parameters of the multi-layer perceptron network and the embedded feature vector output network are obtained based on a maximum likelihood estimation method according to the condition probability that all loads are started on the premise of maximizing the total power data of power consumption of a given household user.
The posterior probability is calculated by the following formula:
wherein ,P(Lj U) is a posterior probability, γ is a smoothing factor of a logistic regression activation function, L represents a set of load elements, U represents a set of users, L' represents a sum of all load elements in the set of load elements, j is "+" or "-", L + Representing a positive sample of the load cell, L - Representing a load cell negative sample.
In one embodiment, the activation function may employ a softmax function to convert cosine similarity of the user and the load element to posterior probability, with gamma values typically set based on human experience.
Ideally, L should contain all load elements that may participate in the matching. For a scenario where one query of home users and turn on load element behavior occurs, use L + Representing the load that was turned on, i.e., positive samples, using negative sampling to randomly select 4 different loads as negative samples for which no turn-on behavior occurred, specifically expressed as
In training, model parameter estimation aims at maximizing the probability of load element start-up behavior to occur given the entire training of querying home users. Therefore, the minimization loss function employs maximum likelihood estimation, which is expressed by the following formula:
wherein Λ is a parameter of the multi-layer perceptron network, P (L + U) represents the posterior probability, U represents the set of users, L + Representing a positive sample of the load cell.
The posterior probability accuracy obtained according to the initial parameters of the double-tower model is low, so that the parameters of the model can be optimized through a random gradient descent method, the posterior probability is maximized, and the parameters of each layer of the double-tower model are updated.
Finally, the load element with the highest degree of association with the user is obtained, the obtained embedded feature vector can be used for comparing the similarity between the user and the load element, so that the load possibly started at the next moment can be found, the similarity between different family users can be compared, the family users with the same type can be found, and the information features of the family users can be collected.
Referring to FIG. 2, in some embodiments, there is provided a load cell construction apparatus for a representation of an electric utility, comprising:
the acquisition module 1 is used for acquiring the total power of the household users and the active power of a single load element as a training set;
the encoding module 2 is used for performing hot unique code encoding on the total power of the home user and the active power of the single load element to respectively obtain a user characteristic vector and an active power characteristic vector;
the processing module 3 is used for respectively carrying out hash processing on the user characteristic vector and the active power characteristic vector to respectively obtain a low-dimensional user characteristic vector and a low-dimensional active power characteristic vector;
the training module 4 is used for respectively inputting the low-dimensional user feature vector and the low-dimensional active power feature vector into a first multi-layer perceptron network and a second multi-layer perceptron network for training, and respectively outputting each user embedded feature vector and each load element embedded feature vector;
and the calculating module 5 is used for calculating the correlation between each user embedded feature vector and the load element embedded feature vector based on cosine similarity, and converting the correlation calculation result into posterior probability by utilizing a logistic regression activation function to obtain the load element with the highest degree of correlation with the user.
Specifically, the encoding module 2 is further configured to perform hot-independent encoding on the total power of the home subscriber, and includes:
dividing the same family user total power into user similar characteristics;
the similar characteristics of each user are expressed by binary system, and the user characteristic vector is obtained;
performing thermal single code encoding on the active power of the single load element to obtain an active power characteristic vector of the single load element, including:
dividing the active power of the same single load element into active power similar characteristics;
and (5) representing each active power similar characteristic by using a binary system to obtain an active power characteristic vector.
The processing module 3 is further configured to hash the user feature vector to obtain a low-dimensional user feature vector, and includes:
respectively adding a start mark and an end mark before and after the user feature vector;
setting step length and sliding window based on N-GRAM algorithm, recording user characteristic vector slices of each scratch, recording all user characteristic vector slices as vector representations composed of a plurality of numbers, and obtaining low-dimension user characteristic vectors;
performing hash processing on the active power feature vector to obtain a low-dimension active power feature vector, including:
respectively adding a start mark and an end mark before and after the active power characteristic vector;
based on an N-GRAM algorithm, setting a step length and a sliding window, recording active power characteristic vector slices which are marked each time, recording all the active power characteristic vector slices as vector representations consisting of a plurality of numbers, and obtaining a low-dimensional active power characteristic vector.
The training module 4 is further configured to input the low-dimensional user feature vector and the low-dimensional active power feature vector to the first multi-layer perceptron network and the second multi-layer perceptron network, respectively, to perform training, and obtain parameters of the first multi-layer perceptron network and the second multi-layer perceptron network, respectively, according to a minimized loss function.
The parameters of the first multi-layer perceptron network and the second multi-layer perceptron network comprise weight matrixes and bias items of all hidden layers.
The load element embedding feature vector is calculated by the following formula:
l i ＝f(W i l i-1 +b i )i＝2,…,N-1；
y L ＝f(W N l N-1 +b N )；
wherein ,yL Embedding feature vectors for load elements, l i Hidden layer of the ith layer of the first multi-layer perceptron network, W i B is a weight matrix of an ith hidden layer of the first multi-layer perceptron network i The bias item representing the ith hidden layer of the first multi-layer perceptron network, wherein N is the number of hidden layers of the first multi-layer perceptron network;
the user-embedded feature vector is calculated by the following formula:
l m ＝f(W m l m-1 +b m )m＝2,…,M-1；
y U ＝f(W M l M-1 +b M )；
wherein ,yU Embedding feature vectors for users, l m Hidden layer of the m-th layer of the second multi-layer perceptron network, W m B is a weight matrix of an mth hidden layer of the second multi-layer perceptron network m And the bias item representing the M-th hidden layer of the second multi-layer perceptron network, wherein M is the number of hidden layers of the second multi-layer perceptron network.
The calculating module 5 is further configured to calculate a correlation between the user embedded feature vector and the load element embedded feature vector, where the calculation is performed by the following formula:
wherein ,yU for embedding feature vectors for users, y L The feature vector is embedded for the load element,
representing a transpose of the user embedded feature vector, R (U, L) representing a correlation of the user embedded feature vector with the load element embedded feature vector.
The posterior probability is calculated by the following formula:
wherein ,P(Lj U) is a posterior probability, γ is a smoothing factor of a logistic regression activation function, L represents a set of load elements, U represents a set of users, L' represents a sum of all load elements in the set of load elements, j is "+" or "-", L + Representing a positive sample of the load cell, L - Representing a load cell negative sample.
By setting L', traversing the load elements within the set L can be achieved.
The minimized loss function is represented by the following formula:
wherein Λ is a parameter of the multi-layer perceptron network, P (L + U) represents the posterior probability, U represents the set of users, L + Representing a positive sample of the load cell.
According to the load element construction method and device for the electric power user portrait, model training is carried out based on total power of a household user and active power of the load element, hash processing is carried out on the characteristics obtained through thermal unique code coding, multi-layer perceptron network training is carried out, similarity between the user and each load is generated by means of a double-tower model, and finally the load element with the highest degree of association with the user is obtained, wherein structural characteristics and node characteristics of a complex recommended network scene are fused, manual priori deviation is avoided to a certain extent, operation speed during online prediction is effectively improved, user portrait model construction efficiency is high, and reliability is strong.
While preferred embodiments of the present invention have been described, additional variations and modifications in those embodiments may occur to those skilled in the art once they learn of the basic inventive concepts. It is therefore intended that the following claims be interpreted as including the preferred embodiments and all such alterations and modifications as fall within the scope of the invention. It will be apparent to those skilled in the art that various modifications and variations can be made to the present invention without departing from the spirit or scope of the invention. Thus, it is intended that the present invention also include such modifications and alterations insofar as they come within the scope of the appended claims or the equivalents thereof.
Claims (9)
1. A load cell construction method for an electricity consumer representation, comprising:
collecting total power of a household user and active power of a single load element as a training set;
performing thermal single code coding on the total power of the home user and the active power of a single load element to respectively obtain a user characteristic vector and an active power characteristic vector;
respectively carrying out hash processing on the user characteristic vector and the active power characteristic vector to respectively obtain a low-dimensional user characteristic vector and a low-dimensional active power characteristic vector;
the low-dimensional user feature vector and the low-dimensional active power feature vector are respectively input into a first multi-layer perceptron network and a second multi-layer perceptron network for training, and each user embedded feature vector and each load element embedded feature vector are respectively output;
calculating the relevance of each user embedded feature vector and the load element embedded feature vector based on cosine similarity, and converting a relevance calculation result into posterior probability by utilizing a logistic regression activation function to obtain a load element with highest relevance to the user;
performing hot unique code encoding on the total power of the home user, including:
dividing the same family user total power into user similar characteristics;
the similar characteristics of each user are expressed by binary system, and the user characteristic vector is obtained;
performing thermal single code encoding on the active power of the single load element to obtain an active power characteristic vector of the single load element, including:
dividing the active power of the same single load element into active power similar characteristics;
and (5) representing each active power similar characteristic by using a binary system to obtain an active power characteristic vector.
2. The method of claim 1, wherein hashing the user feature vector to obtain a low-dimensional user feature vector comprises:
respectively adding a start mark and an end mark before and after the user feature vector;
setting step length and sliding window based on N-GRAM algorithm, recording user characteristic vector slices of each scratch, recording all user characteristic vector slices as vector representations composed of a plurality of numbers, and obtaining low-dimension user characteristic vectors;
performing hash processing on the active power feature vector to obtain a low-dimension active power feature vector, including:
respectively adding a start mark and an end mark before and after the active power characteristic vector;
based on an N-GRAM algorithm, setting a step length and a sliding window, recording active power characteristic vector slices which are marked each time, recording all the active power characteristic vector slices as vector representations consisting of a plurality of numbers, and obtaining a low-dimensional active power characteristic vector.
3. The method of claim 1, wherein the low-dimensional user feature vector and the low-dimensional active power feature vector are input to a first multi-layer perceptron network and a second multi-layer perceptron network, respectively, for training, and parameters of the first multi-layer perceptron network and the second multi-layer perceptron network are obtained, respectively, according to a minimization loss function.
4. The method of claim 3, wherein the parameters of the first and second multi-layer perceptron networks comprise a weight matrix and bias terms for each hidden layer.
5. The method of claim 4, wherein the load element embedded feature vector is calculated by the formula:
wherein ,
embedding feature vectors for load elements, < >>
Conceal layer for the i-th layer of the first multi-layer perceptron network->
Weight matrix for the i-th hidden layer of the first multi-layer perceptron network,/for the i-th hidden layer>
Bias item representing ith hidden layer of first multi-layer perceptron network, number of hidden layers of first multi-layer perceptron network is N-2, < ->
Bias items for the output layer of the first multi-layer perceptron network; />
Input feature vector representing a thermal unique code, < ->
Representing the input layer->
Weight matrix representing input layer, W N A weight matrix of the output layer of the first multi-layer perceptron network;
the user-embedded feature vector is calculated by the following formula:
wherein ,
embedding feature vectors for a user->
Conceal layer for the m-th layer of the second multi-layer perceptron network->
Weight matrix for the m-th hidden layer of the second multi-layer perceptron network,/for the m-th hidden layer>
Bias item representing mth hidden layer of second multi-layer perceptron network, number of hidden layers of second multi-layer perceptron network is M-2,/or->
Bias item W for output layer of second multi-layer perceptron network M And the weight matrix is the weight matrix of the second multi-layer perceptron network output layer.
6. The method of claim 5, wherein the correlation of the user embedded feature vector with the load element embedded feature vector is calculated by the following formula:
wherein ,
embedding feature vectors for a user->
Embedding feature vectors for load elements, < >>
Representing the transpose of the user's embedded feature vector, +.>
The correlation of the user embedded feature vector with the load element embedded feature vector is represented.
7. The method of claim 6, wherein the posterior probability is calculated by the following formula:
wherein ,
for posterior probability>
For the smoothing factor of the logistic regression activation function, L represents the set of load elements, U represents the set of users, L' represents the sum of all load elements in the set of load elements, j is "+" or "-", "and>
representing a positive sample of the load cell, < >>
Representing a load cell negative sample.
8. The method of claim 7, wherein the minimization of the loss function is represented by the following formula:
wherein ,
is a parameter of a multi-layer perceptron network, +.>
Representing posterior probability, U representing the set of users, +.>
Representing a positive sample of the load cell.
9. A load cell construction device for an electric power consumer figure, comprising:
the acquisition module is used for acquiring the total power of the household users and the active power of the single load element as a training set;
the coding module is used for carrying out hot unique code coding on the total power of the home user and the active power of the single load element to respectively obtain a user characteristic vector and an active power characteristic vector;
the processing module is used for respectively carrying out hash processing on the user characteristic vector and the active power characteristic vector to respectively obtain a low-dimensional user characteristic vector and a low-dimensional active power characteristic vector;
the training module is used for respectively inputting the low-dimensional user characteristic vector and the low-dimensional active power characteristic vector into a first multi-layer perceptron network and a second multi-layer perceptron network for training, and respectively outputting each user embedded characteristic vector and each load element embedded characteristic vector;
the computing module is used for computing the relevance of each user embedded feature vector and the load element embedded feature vector based on cosine similarity, converting a relevance computing result into posterior probability by utilizing a logistic regression activation function, and obtaining the load element with the highest relevance to the user;
the encoding module is also used for carrying out hot unique code encoding on the total power of the home user, and comprises the following steps:
dividing the same family user total power into user similar characteristics;
the similar characteristics of each user are expressed by binary system, and the user characteristic vector is obtained;
performing thermal single code encoding on the active power of the single load element to obtain an active power characteristic vector of the single load element, including:
dividing the active power of the same single load element into active power similar characteristics;
and (5) representing each active power similar characteristic by using a binary system to obtain an active power characteristic vector.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210101090.8A CN114511058B (en) | 2022-01-27 | 2022-01-27 | Load element construction method and device for electric power user portrait |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210101090.8A CN114511058B (en) | 2022-01-27 | 2022-01-27 | Load element construction method and device for electric power user portrait |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114511058A CN114511058A (en) | 2022-05-17 |
| CN114511058B true CN114511058B (en) | 2023-06-02 |
Family
ID=81549435
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210101090.8A Active CN114511058B (en) | 2022-01-27 | 2022-01-27 | Load element construction method and device for electric power user portrait |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114511058B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115017333A (en) * | 2022-06-13 | 2022-09-06 | 四川大学 | Method for converting modeless data of material genetic engineering into knowledge map |
| CN114881538B (en) * | 2022-06-24 | 2022-12-13 | 东南大学溧阳研究院 | Demand response user selection method based on perceptron |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109598446A (en) * | 2018-12-09 | 2019-04-09 | 国网江苏省电力有限公司扬州供电分公司 | A kind of tariff recovery Warning System based on machine learning algorithm |
| WO2019141040A1 (en) * | 2018-01-22 | 2019-07-25 | 佛山科学技术学院 | Short term electrical load predication method |
| WO2019238096A1 (en) * | 2018-06-13 | 2019-12-19 | 国网江苏省电力有限公司 | Method and apparatus for estimating weather-sensitive load power |
| CN110928990A (en) * | 2019-10-31 | 2020-03-27 | 南方电网调峰调频发电有限公司 | Method special for recommending standing book data of power equipment based on user portrait |
| CN111428355A (en) * | 2020-03-18 | 2020-07-17 | 东南大学 | Modeling method for power load digital statistics intelligent synthesis |
| CN111651668A (en) * | 2020-05-06 | 2020-09-11 | 上海晶赞融宣科技有限公司 | User portrait label generation method and device, storage medium and terminal |
| CN112598303A (en) * | 2020-12-28 | 2021-04-02 | 宁波迦南智能电气股份有限公司 | Non-invasive load decomposition method based on combination of 1D convolutional neural network and LSTM |
| CN113554241A (en) * | 2021-09-02 | 2021-10-26 | 国网山东省电力公司泰安供电公司 | User layering method and prediction method based on user electricity complaint behaviors |
Family Cites Families (32)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101297318B (en) * | 2005-08-23 | 2013-01-23 | 株式会社理光 | Data organization and access for mixed media document system |
| JP5690296B2 (en) * | 2012-03-01 | 2015-03-25 | 日本電信電話株式会社 | Load balancing program and load balancing apparatus |
| CN104777383B (en) * | 2015-04-16 | 2018-01-05 | 武汉阿帕科技有限公司 | A kind of non-intrusion type electrical load monitoring and load decomposition device |
| CN106936129B (en) * | 2017-03-23 | 2020-04-24 | 东北大学 | Power load identification method and system based on multi-feature fusion |
| JP6832783B2 (en) * | 2017-04-20 | 2021-02-24 | 株式会社日立製作所 | Data analyzers, data analysis methods, and data analysis programs |
| CN107121210B (en) * | 2017-05-19 | 2019-12-03 | 四川成瑞科技有限公司 | Temp monitoring alarm, method and system |
| CN107578288B (en) * | 2017-09-08 | 2020-09-18 | 东南大学 | Non-invasive load decomposition method considering user power consumption mode difference |
| US20190325293A1 (en) * | 2018-04-19 | 2019-10-24 | National University Of Singapore | Tree enhanced embedding model predictive analysis methods and systems |
| CN109101620B (en) * | 2018-08-08 | 2022-07-05 | 阿里巴巴（中国）有限公司 | Similarity calculation method, clustering method, device, storage medium and electronic equipment |
| CN109523057A (en) * | 2018-10-18 | 2019-03-26 | 国网山东省电力公司经济技术研究院 | A kind of regional power grid Methods of electric load forecasting considering economic transition background |
| CN110212542A (en) * | 2019-05-28 | 2019-09-06 | 国网江苏省电力有限公司 | A kind of accurate cutting load method and device |
| CN110472041B (en) * | 2019-07-01 | 2021-08-03 | 浙江工业大学 | Text classification method for customer service online quality inspection |
| JP6691280B1 (en) * | 2019-07-30 | 2020-04-28 | 特許庁長官 | Management system and management method |
| CN110635833B (en) * | 2019-09-25 | 2020-12-15 | 北京邮电大学 | Power distribution method and device based on deep learning |
| CN110866471A (en) * | 2019-10-31 | 2020-03-06 | Oppo广东移动通信有限公司 | Face image quality evaluation method and device, computer readable medium and communication terminal |
| CN110991263B (en) * | 2019-11-12 | 2022-03-18 | 华中科技大学 | Non-invasive load identification method and system for resisting background load interference |
| CN113128494A (en) * | 2019-12-30 | 2021-07-16 | 华为技术有限公司 | Method, device and system for recognizing text in image |
| CN111489036B (en) * | 2020-04-14 | 2023-06-09 | 天津相和电气科技有限公司 | Resident load prediction method and device based on electrical load characteristics and deep learning |
| CN111428816B (en) * | 2020-04-17 | 2023-01-20 | 贵州电网有限责任公司 | Non-invasive load decomposition method |
| CN111612319A (en) * | 2020-05-11 | 2020-09-01 | 上海电力大学 | Load curve depth embedding clustering method based on one-dimensional convolution self-encoder |
| CN111799782A (en) * | 2020-06-29 | 2020-10-20 | 中国电力科学研究院有限公司 | Power equipment power failure window period correction method and system based on machine learning |
| CN111860977B (en) * | 2020-06-30 | 2022-07-01 | 清华大学 | Probability prediction method and probability prediction device for short-term load |
| CN111753058B (en) * | 2020-06-30 | 2023-06-02 | 北京信息科技大学 | Text viewpoint mining method and system |
| CN111949707B (en) * | 2020-08-06 | 2022-03-22 | 杭州电子科技大学 | Shadow field-based hidden Markov model non-invasive load decomposition method |
| CN112379159B (en) * | 2020-11-09 | 2021-10-26 | 北华航天工业学院 | Non-invasive household load decomposition method based on electric appliance operation mode |
| CN112364203B (en) * | 2020-11-30 | 2023-04-28 | 未来电视有限公司 | Television video recommendation method, device, server and storage medium |
| AU2020104000A4 (en) * | 2020-12-10 | 2021-02-18 | Guangxi University | Short-term Load Forecasting Method Based on TCN and IPSO-LSSVM Combined Model |
| CN113094198A (en) * | 2021-04-13 | 2021-07-09 | 中国工商银行股份有限公司 | Service fault positioning method and device based on machine learning and text classification |
| CN113761250A (en) * | 2021-04-25 | 2021-12-07 | 腾讯科技（深圳）有限公司 | Model training method, merchant classification method and device |
| CN113687182A (en) * | 2021-08-12 | 2021-11-23 | 湖南工学院 | Household load identification method, program and system based on noise reduction automatic encoder |
| CN113935401A (en) * | 2021-09-18 | 2022-01-14 | 北京三快在线科技有限公司 | Article information processing method, article information processing device, article information processing server and storage medium |
| CN113887833A (en) * | 2021-10-28 | 2022-01-04 | 西安热工研究院有限公司 | Distributed energy user side time-by-time load prediction method and system |
-
2022
- 2022-01-27 CN CN202210101090.8A patent/CN114511058B/en active Active
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2019141040A1 (en) * | 2018-01-22 | 2019-07-25 | 佛山科学技术学院 | Short term electrical load predication method |
| WO2019238096A1 (en) * | 2018-06-13 | 2019-12-19 | 国网江苏省电力有限公司 | Method and apparatus for estimating weather-sensitive load power |
| CN109598446A (en) * | 2018-12-09 | 2019-04-09 | 国网江苏省电力有限公司扬州供电分公司 | A kind of tariff recovery Warning System based on machine learning algorithm |
| CN110928990A (en) * | 2019-10-31 | 2020-03-27 | 南方电网调峰调频发电有限公司 | Method special for recommending standing book data of power equipment based on user portrait |
| CN111428355A (en) * | 2020-03-18 | 2020-07-17 | 东南大学 | Modeling method for power load digital statistics intelligent synthesis |
| CN111651668A (en) * | 2020-05-06 | 2020-09-11 | 上海晶赞融宣科技有限公司 | User portrait label generation method and device, storage medium and terminal |
| CN112598303A (en) * | 2020-12-28 | 2021-04-02 | 宁波迦南智能电气股份有限公司 | Non-invasive load decomposition method based on combination of 1D convolutional neural network and LSTM |
| CN113554241A (en) * | 2021-09-02 | 2021-10-26 | 国网山东省电力公司泰安供电公司 | User layering method and prediction method based on user electricity complaint behaviors |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114511058A (en) | 2022-05-17 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Zhang et al. | Improved deep hashing with soft pairwise similarity for multi-label image retrieval | |
| CN112214685B (en) | Knowledge graph-based personalized recommendation method | |
| CN111291836B (en) | Method for generating student network model | |
| CN114511058B (en) | Load element construction method and device for electric power user portrait | |
| CN111414461B (en) | Intelligent question-answering method and system fusing knowledge base and user modeling | |
| CN111897913A (en) | Semantic tree enhancement based cross-modal retrieval method for searching video from complex text | |
| CN101346718A (en) | Method for providing user of chosen content item | |
| CN109033107A (en) | Image search method and device, computer equipment and storage medium | |
| CN113177141B (en) | Multi-label video hash retrieval method and device based on semantic embedded soft similarity | |
| CN113343125B (en) | Academic accurate recommendation-oriented heterogeneous scientific research information integration method and system | |
| CN114693397B (en) | Attention neural network-based multi-view multi-mode commodity recommendation method | |
| CN109272332B (en) | Client loss prediction method based on recurrent neural network | |
| CN111737578A (en) | Recommendation method and system | |
| CN111159485A (en) | Tail entity linking method, device, server and storage medium | |
| CN112800344B (en) | Deep neural network-based movie recommendation method | |
| CN112860930B (en) | Text-to-commodity image retrieval method based on hierarchical similarity learning | |
| CN113569001A (en) | Text processing method and device, computer equipment and computer readable storage medium | |
| CN114462420A (en) | False news detection method based on feature fusion model | |
| CN115408603A (en) | Online question-answer community expert recommendation method based on multi-head self-attention mechanism | |
| CN114781503A (en) | Click rate estimation method based on depth feature fusion | |
| CN114004220A (en) | Text emotion reason identification method based on CPC-ANN | |
| CN116244484B (en) | Federal cross-modal retrieval method and system for unbalanced data | |
| CN116108836B (en) | Text emotion recognition method and device, computer equipment and readable storage medium | |
| CN111859979A (en) | Ironic text collaborative recognition method, ironic text collaborative recognition device, ironic text collaborative recognition equipment and computer readable medium | |
| CN115952360A (en) | Domain-adaptive cross-domain recommendation method and system based on user and article commonality modeling |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |