CN113222722B - Power budget assessment method based on gate structure depth crossing network - Google Patents

Power budget assessment method based on gate structure depth crossing network Download PDF

Info

Publication number
CN113222722B
CN113222722B CN202110514430.5A CN202110514430A CN113222722B CN 113222722 B CN113222722 B CN 113222722B CN 202110514430 A CN202110514430 A CN 202110514430A CN 113222722 B CN113222722 B CN 113222722B
Authority
CN
China
Prior art keywords
network
power budget
gate structure
vectors
features
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
Application number
CN202110514430.5A
Other languages
Chinese (zh)
Other versions
CN113222722A (en
Inventor
许斌锋
仲田
王青国
胡扬波
陆野
徐进
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.)
Jiangsu Electric Power Information Technology Co Ltd
Original Assignee
Jiangsu Electric Power Information Technology 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 Jiangsu Electric Power Information Technology Co Ltd filed Critical Jiangsu Electric Power Information Technology Co Ltd
Priority to CN202110514430.5A priority Critical patent/CN113222722B/en
Publication of CN113222722A publication Critical patent/CN113222722A/en
Application granted granted Critical
Publication of CN113222722B publication Critical patent/CN113222722B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q40/00—Finance; Insurance; Tax strategies; Processing of corporate or income taxes
    • G06Q40/12—Accounting
    • G06Q40/125—Finance or payroll
    • 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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
    • G06Q10/063—Operations research, analysis or management
    • G06Q10/0631—Resource planning, allocation, distributing or scheduling for enterprises or organisations
    • G06Q10/06313—Resource planning in a project environment
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
    • G06Q50/06—Energy or water supply
    • 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
    • Y02E—REDUCTION OF GREENHOUSE GAS [GHG] EMISSIONS, RELATED TO ENERGY GENERATION, TRANSMISSION OR DISTRIBUTION
    • Y02E40/00—Technologies for an efficient electrical power generation, transmission or distribution
    • Y02E40/70—Smart grids as climate change mitigation technology in the energy generation sector
    • 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
    • Y04S10/00—Systems supporting electrical power generation, transmission or distribution
    • Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications

Landscapes

  • Engineering & Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Human Resources & Organizations (AREA)
  • Economics (AREA)
  • General Physics & Mathematics (AREA)
  • Strategic Management (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Marketing (AREA)
  • General Health & Medical Sciences (AREA)
  • General Business, Economics & Management (AREA)
  • Tourism & Hospitality (AREA)
  • Finance (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Data Mining & Analysis (AREA)
  • Computational Linguistics (AREA)
  • Software Systems (AREA)
  • Development Economics (AREA)
  • Biophysics (AREA)
  • Mathematical Physics (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • Accounting & Taxation (AREA)
  • Molecular Biology (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • Game Theory and Decision Science (AREA)
  • Quality & Reliability (AREA)
  • Operations Research (AREA)
  • Biodiversity & Conservation Biology (AREA)
  • Primary Health Care (AREA)
  • Water Supply & Treatment (AREA)
  • Public Health (AREA)
  • Educational Administration (AREA)
  • Technology Law (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses a power budget assessment method based on a gate structure depth crossing network, which extracts representative characteristics affecting power budget from 2 dimensions of a primary profit center and a service domain; converting the features from discrete and continuous network layers into low-dimensional dense characterization vectors respectively; a multi-head attention network in an open-source transducer model is adopted to capture interaction among different features and learn ambiguity brought by diversified feature interaction, so that a high-quality feature expression vector is output. The characterization vectors are respectively transmitted into a deep network and a cross network for fusion; based on the trained model, a future year power budget estimate for the samples in the test set is calculated. The invention solves the difficult problem of lack of decision guidance in the annual budget and allocation process of the power system units and professional departments, not only maintains the advantage of the depth network, but also can utilize the cross network to carry out explicit cross calculation on the characteristics.

Description

Power budget assessment method based on gate structure depth crossing network
Technical Field
The invention belongs to the field of electric power, and particularly relates to an electric power budget assessment method based on a gate structure depth crossing network.
Background
With the rapid development of national economy, in order to meet the actual development demands of society, electric power companies need to be intelligently innovated and reformed from the aspects of internal systems and development strategies. The electric power financial budget assessment is an important component in the management work of the electric power enterprises, and in the actual development and management process, the electric power enterprises only pay full attention to the electric power financial budget management work and can effectively solve the problems in the electric power financial budget management by combining with the actual development condition, so that the electric power guarantee is fundamentally provided for the development of national economy in China. When the existing electric power company performs budget allocation on a first-level profit center and a business domain, the problem of experience is presented, a complete scientific system structure and model are lacked to ensure the reasonability and feasibility of budget allocation, and related technical support is lacked.
At present, the existing research work aiming at power budget prediction mainly focuses on extracting relevant characteristic basic knowledge, adopts a machine learning algorithm as a main part, and combines social economic data such as power budget, urban GDP, urban population and the like in the past year to build a prediction model so as to improve scientificity of budget allocation, such as a linear regression model, a logistic regression model, a decision tree model, a depth network, a cross network and the like. The above models often ignore the synergistic effect between features, nor do they disambiguate semantics between features. In addition, the difference in importance of the vectors of these class 2 network learning is also ignored in terms of vector fusion of deep network and cross network learning.
Disclosure of Invention
The invention aims to provide a power budget assessment method based on a gate structure depth cross network, which solves the problems of scientificity and irrational distribution in the prior power budget distribution, thereby developing a new thought for planning and cultivating new financial management methods and new modes for professional resources and being applicable to different power financial prediction scenes.
In order to solve the problems, the invention adopts the following technical scheme:
the power budget assessment method based on the gate structure depth cross network comprises the steps of firstly, preprocessing related operations on primary profit center dimension data and business domain dimension data of an enterprise; secondly, starting from the 2 dimensions of a first-level profit center and a service domain, respectively extracting 13 representative features affecting power budget from the 2 types of data; thirdly, converting the characteristics from discrete and continuous characteristics into low-dimensional dense characterization vectors by adopting an Embedding network layer; and thirdly, capturing interaction among different features and learning ambiguity brought by diversified feature interaction by adopting a multi-head attention network of a Google transform model, so as to output a high-quality feature expression vector. Then, the characterization vectors are respectively transmitted into a depth network and a cross network, the depth network is utilized to implicitly generate interaction between the features and the cross network to carry out explicit cross calculation on the features, the workload of manually carrying out feature cross is reduced, and the vectors learned by the 2 types of networks are fused through a gate structure network; finally, based on the trained model, a future year power budget estimate for the samples in the test set is calculated.
The method comprises the following steps:
1) From the first level profit center and 2 dimensions of the business domain, the power budget allocation data over the years is analyzed. And mining influence characteristics affecting the power budget allocation data of the past year, wherein the influence characteristics comprise a first-level profit center dimension data characteristic and a service domain dimension data characteristic 2 part of an enterprise, preprocessing related operations are carried out on the characteristics, and a training set and a testing set are divided.
2) Establishing a power budget evaluation model based on a gate structure depth crossing network, implicitly generating interactions among features and a crossing network by using the depth network to perform explicit crossing calculation on the features, and fusing vectors learned by the 2 types of networks through the gate structure network. Inputting all the characteristic values in the training set obtained in the step 1) into the method for training and learning, obtaining a high-order characteristic expression vector, and establishing a nonlinear relation between the annual power budget allocation value and the characteristic value.
3) And (3) optimizing the power budget evaluation model designed in the step (2) based on the gate structure depth crossing network in a training stage by means of the Adam optimizer to minimize a loss function so as to obtain optimal parameters.
4) And 3) according to the optimized model obtained in the step 3), finally, automatically generating an electric power budget proposal scheme by inputting relevant characteristic values in a test stage.
The specific method for analyzing the annual power budget allocation data and the influencing factors thereof in the step 1) is as follows: and analyzing the primary profit center dimension data of the power enterprise, and finding out parameters with higher relativity with annual power budget allocation data, such as the number of staff of each power supply company, the number of equipment, the number of substations of each voltage class, the number of transmission lines, the sales amount, the transmission amount, the original value of assets, standard operation cost (unit price) and the like. And analyzing the business domain dimension data to find out parameters with higher relativity with annual power budget allocation, such as city GDP, population, other economic indexes and the like. From the selected features, the features with data in calendar year and reasonable are selected, the data value or accumulated value of the calendar year is analyzed, the proper value is selected for each feature, and the features are divided into discrete type and continuous type and are input of a deep learning model. The data sets are partitioned according to the ratio of training set to test set 9:1.
The specific operation of training and learning by using the power budget assessment model based on the gate structure depth crossing network in the step 2) is as follows:
first, the class 2 feature sets are grouped into vectorsAnd->Two Embedding network layers are input, which are aimed at converting these features from discrete and continuous to low-dimensional dense characterization vectors, respectively, denoted asThe finally learned class 2 vectors are spliced by means of a Concat function in the neural network and recorded as。
Secondly, the first step of the method comprises the steps of,and considering the influence of the cooperative influence among the features and the influence of the feature interaction of the conflict semantics on the prediction result. The invention adopts the multi-head attention network of the Google transform model to capture the interaction among different features and learn the ambiguity brought by diversified feature interaction, and meanwhile, the model has strong parallel computing performance and can efficiently output high-quality feature expression vectors. Input vector given a transducer modelTransformer>Potential expression vector of individual head->The product attention (Dot-Product Attention) can be obtained by scaling the Dot product:
wherein Q, K, V respectively represents three vectors of Query, key and Value in a transducer model,and->Is used for learning the transducer>Weight parameters of individual head->Is the dimension of vector K. Hidden feature->Form an enhanced token vector +.>Information inherent to each feature and ambiguous information are stored. The invention combines a feed forward data network with an excitation function to learn a nonlinear combination of information:
wherein,is a trainable weight, +.>Is the number of attention heads; representative vector concatenation.
Again, the characterization vector of the featureThe method is respectively input into a depth network and a crossover network, and aims to implicitly generate interaction between the features by using the depth network, and simultaneously, perform explicit crossover calculation on the features by using the crossover network, so that the workload of manually performing feature crossover is reduced. Specifically, the Cross Network (Cross Network) is composed of a plurality of Cross layers, and the first layer is calculated by:>each of the intersecting layers after the second layer is calculated as follows:
here the number of the elements is the number,and->Respectively represent->Layer and->Input vector of layer crossing network, +.>And->Are respectively->The weights and offsets of the layer crossing network.
The Deep Network (Deep Network) is a fully connected feed-forward neural Network, and the first layer is calculated by the following method:the calculation mode of each depth network behind the second layer is as follows:
wherein,and->Are respectively->Layer and->Deep network of layersOutput vector of>And->Respectively the firstWeights and biases for the depth of layer network. The total layer number of the depth network and the cross network isnThe expression vectors of the finally obtained features are denoted +.>And->。
Finally, the class 2 network learning vectors are measured by a gate structure networkAnd->And fusing the importance of the two, and the calculation mode is as follows:
wherein,、/>and->Is a parameter in the door structure fusion network, +.>The excitation function selected for the present invention. Finally, pass the gate structure pair->And->Fusion can be carried out to obtain the final characterization vector +.>The calculation method is as follows:
finally, the step of obtaining the product,hadamard product (Hadamard product) representing two vectors. Use->As an excitation function, will->Converting into predicted power budget values:
the sample loss function in the step 3) is a square loss function, and a training set is given，tFor the collection->Of the label (real electricity)Force budget value) is +.>The final squaring loss function is defined as follows:
(/>
the invention minimizes the above loss function by means of Adam optimizer, thereby tuning the parameters in the predictive model to optimal configuration.
In the step 4), the power budget evaluation model based on the gate structure depth crossing network trained in the step 3) is used for realizing the automatic generation of a power budget proposal by inputting relevant characteristic values in a test set and finally outputting the power budget value.
According to the invention, interaction among different features and ambiguity brought by learning diversified feature interaction are captured by adopting the multi-head attention network of the Google transform model, on the basis, a gate structure depth crossing network is designed, the interaction among the features and the crossing network are implicitly generated by using the depth network to carry out explicit crossing calculation on the features, the workload of manually carrying out feature crossing is reduced, and vectors learned by the 2 types of networks are fused through the gate structure network, so that high-precision electric power financial budget assessment is finally realized.
The beneficial effects of the invention are as follows: from the two dimensions of the first-level profit center and the business domain, the power budget allocation data of the past year and the corresponding relevant characteristics thereof are analyzed, and a model based on a gate structure depth crossing network is designed to predict the power budget allocation data of the future year.
In particular, the invention has the following advantages:
1. the model based on the gate structure depth crossing network can capture interaction among different features and learn ambiguity brought by diversified feature interaction, on the basis, the depth network is utilized to implicitly generate interaction among the features and the crossing network to carry out explicit crossing calculation on the features, so that workload of manually carrying out feature crossing is reduced, vectors learned by the 2 types of networks are fused through the gate structure network, and finally, high-precision electric power financial budget assessment is realized.
2. The model based on the gate structure depth crossing network can predict the power budget allocation data of the coming years, solves the problems of lack of scientificity, urgent need of technical support and the like in the annual budget and allocation process of the power system units and professional departments, and can effectively promote the high efficiency and scientificity of resource allocation.
Drawings
FIG. 1 is a diagram of the overall system framework of the present invention.
Detailed Description
The present invention is further illustrated below in conjunction with specific embodiments, it being understood that these embodiments are meant to be illustrative of the invention only and not limiting the scope of the invention, as various equivalent modifications to the invention will fall within the scope of the claims appended hereto, after reading the invention.
Referring to fig. 1, the standard of deep learning is to represent vectors with the same shape, different colors represent different vectors, conca and ReLu are commonly used function names, and a transducer is a model name. The invention relates to an automatic generation method of a power budget proposal scheme based on big data analysis, which comprises the following steps:
1) And analyzing data of power budget in past year from the first-level profit center and 2 dimensions of the business domain. And mining influence characteristics of the electric power budget allocation data affecting the past year, wherein the influence characteristics comprise a first-level profit center dimension data characteristic and a service domain dimension data characteristic 2 part, preprocessing related operations are carried out on the characteristics, and data sets are divided according to the ratio of training sets to test sets 9:1.
2) And establishing a power budget evaluation model based on the gate structure depth crossing network. Inputting all the features in the training set obtained in the step 2) into the model for training and learning, obtaining a high-order feature expression vector, and establishing a nonlinear relation between the power budget allocation value and the feature value over the years.
3) Optimizing the power budget evaluation designed in the step 3) based on the gate structure depth crossing network by means of an Adam optimizer to minimize the loss function so as to obtain the optimal parameter configuration.
4) And 3) finally, according to the optimized model obtained in the step 3), automatically generating a power budget proposal scheme by inputting related characteristic values.
The specific method for analyzing the annual power budget allocation data and the influencing factors of the step 1) is as follows: the first-level profit center dimension data of the power enterprise is analyzed to find out the features with higher correlation with the power budget allocation data of the past year, the business domain dimension data is analyzed to find out the features with higher correlation with the power budget allocation of the past year, and finally the extracted features are shown in table 1. From the selected characteristics, the characteristics with data in calendar year are selected and reasonably selected, and the data value or accumulated value of the calendar year is analyzed, the proper value is selected for each characteristic, the characteristics in the sets are divided into discrete type and continuous type, and the sets formed by the 2 types of characteristics are recorded asF The vector is the input to the deep learning model. The invention divides the data set according to the ratio of the training set to the test set 9:1.
Table 1 data characterization
The specific operation of training and learning by using the power budget assessment method based on the gate structure depth crossing network in the step 2) is as follows:
first, the class 2 feature sets are grouped into vectorsAnd->Two Embedding network layers are input, which are aimed at converting these features from discrete and continuous to low-dimensional dense characterization vectors, respectively, denoted asThe finally learned class 2 vectors are spliced by means of a Concat function in the neural network and recorded as. Further, the influence of feature interaction of conflict semantics on the prediction result is considered. The invention adopts the multi-head attention network of the Google transform model to capture the interaction among different features and learn the ambiguity brought by diversified feature interaction, and meanwhile, the model has strong parallel computing performance, can efficiently output high-quality feature expression vectors, and is recorded as->。
Again, the characterization vector of the featureInput into the depth network and the crossover network, respectively, is intended to implicitly generate interactions between features using the depth network, while explicitly cross-computing features using the crossover network. Specifically, the Cross Network (Cross Network) is composed of a plurality of Cross layers, and the first layer is calculated by the following method:each of the intersecting layers after the second layer is calculated as follows:
,
here the number of the elements is the number,and->Respectively represent the firstlLayer and the firstlInput vector of +1 layer cross network, +.>And->Respectively the firstlThe weights and offsets of the layer crossing network.
The Deep Network (Deep Network) is a fully connected feed-forward neural Network, and the first layer is calculated by the following method:the calculation mode of each depth network behind the second layer is as follows:
wherein,and->Respectively the firstlLayer and the firstlOutput vector of +1 layer deep network, +.>And->Respectively the firstlWeights and biases for the depth of layer network. The total layer number of the depth network and the cross network isnThe expression vectors of the finally obtained features are denoted +.>And->。
Finally, a gate structure network is designed to measure the vectors of the class 2 network learningAnd->And fusing the importance of the two, and the calculation mode is as follows:
wherein,、/>and->Is a parameter in the door structure fusion network, +.>The excitation function selected for the present invention. Finally, pass the gate structure pair->And->Fusion can be carried out to obtain the final characterization vector +.>The calculation method is as follows:
finally, useAs an excitation function, will->Converting into predicted power budget values:
the sample loss function in the step 3) is a square loss function, and a training set is given，tFor the collection->The corresponding tag (real power budget value) is +.>The final squaring loss function is defined as follows:
(/>
the invention minimizes the above loss function by means of Adam optimizer, thereby tuning the parameters in the predictive model to optimal configuration.
In the step 4), the power budget evaluation model based on the gate structure depth crossing network trained in the step 3) is used for realizing the automatic generation of a power budget proposal by inputting relevant characteristic values in a test set and finally outputting the power budget value.

Claims (5)

1. A power budget assessment method based on a gate structure depth crossing network is characterized by comprising the following steps of: firstly, starting from 2 dimensions of a first-level profit center and a service domain, extracting representative features affecting power budget from data of the 2 dimensions respectively; thirdly, converting the characteristics from discrete and continuous characteristics into low-dimensional dense characterization vectors by adopting an Embedding network layer; then, splicing the characterization vectors, and transmitting the characterization vectors into a model based on a gate structure depth crossing network; finally, training the model to obtain optimal parameter configuration, and calculating an estimated value of the power budget of the future years of the samples in the test set, wherein the method specifically comprises the following steps:
1) Analyzing power budget allocation data of the past year from the first-level profit center and 2 dimensions of the service domain, mining influence characteristics influencing the power budget allocation data of the past year, including the first-level profit center dimension data characteristics and the service domain dimension data characteristics 2 parts, preprocessing related operations on the characteristics, and dividing a training set and a testing set;
2) Establishing a power budget assessment model based on a gate structure depth crossing network, inputting all the features in the training set obtained in the step 1) into the model for training and learning, obtaining a high-order feature expression vector, and establishing a nonlinear relation between a power budget distribution value and a feature value in the past year;
3) In the training stage, optimizing the power budget evaluation model based on the gate structure depth crossing network in the step 2), and minimizing a loss function by means of an Adam optimizer to obtain optimal parameters;
4) According to the optimized power budget assessment model based on the gate structure depth crossing network obtained in the step 3), in a test stage, finally, through inputting relevant characteristic values, automatic generation of a power budget proposal scheme is realized;
in the step 2), the specific operation of training and learning by using the power budget evaluation model based on the gate structure depth crossing network is as follows:
first, willThese 2 class feature set group vectors F 1 ＝(f 1 ,f 2 ,…,f 10 ) And F 2 ＝(f 11 ,f 12 ,f 13 ) Two Embedding network layers are input, which are aimed at converting these features from discrete and continuous to low-dimensional dense token vectors, denoted as e= (E) 1 ,e 2 ,…,e 13 ) The finally learned class 2 vectors are spliced by means of a Concat function in the neural network and recorded as
Secondly, a multi-head attention network of a Google transform model is adopted to capture interaction among different features and learn ambiguity brought by diversified feature interaction, and meanwhile, the model has strong parallel computing performance and can efficiently output high-quality feature expression vectors; input vector R given a transducer model 0 Potential expression vector H of the ith head of a transducer i The attention can be found by scaling the dot product:
wherein Q, K, V respectively represents three vectors of Query, key and Value in a transducer model,and->Is a weight parameter for learning the ith header of the transducer, d k Is the dimension of vector K; hidden feature H i Form an enhanced token vector/>R 1 Storing information inherent to each feature and ambiguous information; combining a feed forward data network and a Relu's excitation function to learn a nonlinear combination of information:
R 1 ＝Relu(FeedF orward(W m ([H 1 ；H 2 ；…；H h ]))),
wherein W is m Is a trainable weight, h is the number of attention heads; splicing representative vectors;
again, the characterization vector R of the feature 1 The method is respectively input into a depth network and a crossover network, and aims to implicitly generate interaction between the features by using the depth network, and simultaneously, the features are subjected to explicit crossover calculation by using the crossover network, so that the workload of manually carrying out feature crossover is reduced; specifically, the crossover network is composed of a plurality of crossover layers, and the first layer is calculated by the following way: each of the intersecting layers after the second layer is calculated as follows:
here, C l And C l+1 Representing the input vectors of the layer i and layer i +1 crossover networks respectively,and->The weight and the bias of the first layer crossing network are respectively;
the deep network is a fully connected feed-forward neural network, the first layerThe calculation mode of (a) is as follows: the calculation mode of each depth network behind the second layer is as follows:
wherein D is l And D l+1 The output vectors of the depth networks of the first layer and the first +1 layer respectively,and->The weight and bias of the first layer depth network are respectively; the total layer number of the depth network and the cross network is n, and finally obtained expression vectors of the characteristics are respectively marked as D n And C n ；
Finally, vector D of the class 2 network learning is measured by a gate structure network n And C n And fusing the importance of the two, and the calculation mode is as follows:
G＝sigmod(w D D n +w C C n +b D ),
wherein w is D 、w C And b D Is a parameter in a door structure fusion network, sigmod is an excitation function selected by the invention; finally, pass gate structure pair D n And C n The final characterization vector O is obtained by fusion, and the calculation method is as follows:
O＝G⊙D n +(1-G)⊙C n，
finally, the ". As indicated above, represents the Hadamard product of the two vectors; using ReLU as an excitation function, O is converted to a predicted power budget value:
2. the method for evaluating the power budget based on the gate structure depth intersection network according to claim 1, wherein in the step 1), from the 2 dimensions of the "primary profit center" and the "business domain", the features influencing the power budget are extracted from the data of the 2 dimensions, the features with data in the calendar year and reasonable are selected, the data value or the accumulated value of the calendar year is analyzed, the proper value is selected for each feature, and the features are divided into discrete type and continuous type and are input into a deep learning model; the data sets are partitioned according to the ratio of training set to test set 9:1.
3. The method for evaluating the power budget based on the gate structure depth crossover network according to claim 1, wherein in the step 2), the characteristics are converted from discrete and continuous characteristics into low-dimensional dense characterization vectors respectively by adopting an Embedding network layer; then, splicing the characterization vectors, and transmitting the characterization vectors into a model based on a gate structure depth crossing network; finally, in the training phase, the model is trained to obtain optimal parameters, and the estimated value of the power budget of the future years of the samples in the test set is calculated.
4. The method for evaluating the power budget based on the gate structure depth intersection network according to claim 3, wherein a multi-head attention network of a google transform model is adopted to capture interaction among different features and learn ambiguity brought by diversified feature interaction, a model based on the gate structure depth intersection network is established, the interaction among the features and the intersection network are implicitly generated by using the depth network to perform explicit intersection calculation on the features, vectors learned by the 2 types of networks are fused through the gate structure network, and finally high-precision power financial budget evaluation is realized.
5. The method according to claim 1, wherein in the step 3), interactions between features are implicitly generated by using the depth network, and explicit cross computation is performed on the features by using the cross network, so that the workload of manually performing feature cross is reduced, and in a model training stage, the square loss function is minimized by means of an Adam optimizer, so that parameters in the prediction model are adjusted to an optimal configuration.
CN202110514430.5A 2021-05-12 2021-05-12 Power budget assessment method based on gate structure depth crossing network Active CN113222722B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110514430.5A CN113222722B (en) 2021-05-12 2021-05-12 Power budget assessment method based on gate structure depth crossing network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110514430.5A CN113222722B (en) 2021-05-12 2021-05-12 Power budget assessment method based on gate structure depth crossing network

Publications (2)

Publication Number Publication Date
CN113222722A CN113222722A (en) 2021-08-06
CN113222722B true CN113222722B (en) 2024-02-02

Family

ID=77095130

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110514430.5A Active CN113222722B (en) 2021-05-12 2021-05-12 Power budget assessment method based on gate structure depth crossing network

Country Status (1)

Country Link
CN (1) CN113222722B (en)

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8280805B1 (en) * 2006-01-10 2012-10-02 Sas Institute Inc. Computer-implemented risk evaluation systems and methods
CN109829587A (en) * 2019-02-12 2019-05-31 国网山东省电力公司电力科学研究院 Zonule grade ultra-short term and method for visualizing based on depth LSTM network
CN112348352A (en) * 2020-11-05 2021-02-09 国网江苏省电力有限公司 Big data analysis-based automatic generation method for electric power budget proposal scheme

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8280805B1 (en) * 2006-01-10 2012-10-02 Sas Institute Inc. Computer-implemented risk evaluation systems and methods
CN109829587A (en) * 2019-02-12 2019-05-31 国网山东省电力公司电力科学研究院 Zonule grade ultra-short term and method for visualizing based on depth LSTM network
CN112348352A (en) * 2020-11-05 2021-02-09 国网江苏省电力有限公司 Big data analysis-based automatic generation method for electric power budget proposal scheme

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
基于模糊DEA的配电网投入产出综合效率评估;王恩创;任玉珑;朱春波;;工业工程与管理(第02期);全文 *

Also Published As

Publication number Publication date
CN113222722A (en) 2021-08-06

Similar Documents

Publication Publication Date Title
CN109033497B (en) High-concurrency-oriented multi-stage data mining algorithm intelligent selection method
CN108280541A (en) Customer service strategies formulating method, device based on random forest and decision tree
CN111475557A (en) Platform construction system in general financial service platform data
CN115423429A (en) Multimode integrated distribution network operation system based on image and sound information
Quest et al. A 3D indicator for guiding AI applications in the energy sector
CN113204329A (en) Control method for driving business application by unified data model and application system thereof
CN112348352B (en) Big data analysis-based automatic generation method for electric power budget proposal scheme
CN113222722B (en) Power budget assessment method based on gate structure depth crossing network
Keramati et al. The impact of information technology maturity on firm performance considering the moderating role of relational maturity: an empirical research
CN112991032A (en) Power budget assessment method based on hierarchical attention network
Vincent et al. Systematic Review on Decentralised Artificial Intelligence and Its Applications
Kuzina et al. The safe solutions system modeling in a digital technology enterprise
Wang A multi-level fuzzy comprehensive assessment for supply chain risks
Ramírez et al. Artificial intelligence and its impact on the prediction of economic indicators
Ma et al. An intelligent price-appraisal algorithm based on grey correlation and fuzzy mathematics
CN117934209B (en) Regional power system carbon emission big data analysis method based on knowledge graph
Kumar et al. AI-Empowered IoT for Sustainable Energy Technologies
Dou Innovation of Trade Union Work in Colleges and Universities Based on Mathematical Modeling and Multivariable Optimal Design
Fang et al. Power grid user asset value evaluation method and application under the background of artificial intelligence and smart grid
Wang et al. Research on Financial Risk Control Model Based on Machine Learning
Zhang et al. Multiple science data-oriented Technology Roadmapping method
Teng et al. Research on power grid investment capacity prediction model based on multiple method combination
Wang Introducing Cloud Data Technology into Economic Statistics
Liu et al. [Retracted] The Effectiveness Evaluation Model of the Conversion of New and Old Kinetic Energy in the Cloud Computing Environment
Qiao et al. Research and Application of Digital Modeling Technology Based on Multi-source Data of Power Grid

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