CN109214562A - A kind of power grid scientific research hotspot prediction and method for pushing based on RNN - Google Patents
A kind of power grid scientific research hotspot prediction and method for pushing based on RNN Download PDFInfo
- Publication number
- CN109214562A CN109214562A CN201810970350.9A CN201810970350A CN109214562A CN 109214562 A CN109214562 A CN 109214562A CN 201810970350 A CN201810970350 A CN 201810970350A CN 109214562 A CN109214562 A CN 109214562A
- Authority
- CN
- China
- Prior art keywords
- power grid
- module
- scientific
- period
- 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.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 34
- 238000011160 research Methods 0.000 title claims abstract description 33
- 238000005516 engineering process Methods 0.000 claims abstract description 36
- 238000012549 training Methods 0.000 claims abstract description 22
- 238000000605 extraction Methods 0.000 claims abstract description 9
- 239000011159 matrix material Substances 0.000 claims description 24
- 238000013528 artificial neural network Methods 0.000 claims description 19
- 238000004364 calculation method Methods 0.000 claims description 17
- 230000006870 function Effects 0.000 claims description 7
- 210000002569 neuron Anatomy 0.000 claims description 7
- 230000008569 process Effects 0.000 claims description 4
- 230000004913 activation Effects 0.000 claims description 3
- 239000000284 extract Substances 0.000 claims description 3
- 230000001020 rhythmical effect Effects 0.000 claims description 3
- 230000000007 visual effect Effects 0.000 claims description 3
- 238000010586 diagram Methods 0.000 description 2
- 230000000306 recurrent effect Effects 0.000 description 2
- 230000006399 behavior Effects 0.000 description 1
- 230000015572 biosynthetic process Effects 0.000 description 1
- 230000009193 crawling Effects 0.000 description 1
- 238000007418 data mining Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000009472 formulation Methods 0.000 description 1
- 238000011835 investigation Methods 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000002407 reforming Methods 0.000 description 1
- 238000007363 ring formation reaction Methods 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—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
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/06—Energy or water supply
Landscapes
- Business, Economics & Management (AREA)
- Engineering & Computer Science (AREA)
- Economics (AREA)
- Human Resources & Organizations (AREA)
- Strategic Management (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Marketing (AREA)
- General Physics & Mathematics (AREA)
- General Business, Economics & Management (AREA)
- Tourism & Hospitality (AREA)
- Public Health (AREA)
- General Health & Medical Sciences (AREA)
- Primary Health Care (AREA)
- Water Supply & Treatment (AREA)
- Development Economics (AREA)
- Game Theory and Decision Science (AREA)
- Entrepreneurship & Innovation (AREA)
- Operations Research (AREA)
- Quality & Reliability (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a kind of power grid science and technology hotspot prediction and method for pushing based on RNN, characterized by the following steps: hotspot prediction model crawls module, character representation module, feature extraction module and model training module by data and forms hotspot prediction model；Two, power grid science and technology hotspot predictions and push, A) t period power grid scientific and technical article of acquisition, above-mentioned hotspot prediction model is applied, obtains scientific research hot keyword；B) the power grid scientific and technical terms in database for natural language are clustered, generate the degree of association of scientific and technical terms；C) synthetic time series power grid science and technology hot spot, and it is pushed to user.The present invention can perceive the power grid Technological research hot spot being likely to occur in following a period of time, and for power grid, scientific research personnel provides research direction and Research Thinking；The potential technical application of the correlation of power grid science and technology hot spot can be recommended power grid researcher by the degree of association synthetic time series scientific research hot spot based on power grid scientific research vocabulary.
Description
Technical field
The power grid scientific research hotspot prediction and supplying system and its implementation that the present invention relates to a kind of based on RNN.
Background technique
Scientific and technological information has all played important function to country, society, the strategy of enterprise, the formulation of plan and implementation.Section
Grinding hotspot prediction is the newer application demand in scientific and technological information field.Researcher, scientific research project manager the selected topic, project verification must
Must have it is certain perspective, i.e., based on the contemporary scientific state of the art and social development situation, to the following issuable new reason
By or generate application value new technology judge.
The method of scientific research hotspot prediction depends critically upon this field highly specialized personnel by Literature Consult and market at present
The method of investigation determines the direction that hot spot occurs, in addition, associated application field is also after a new theory is born with technology
It needs largely to work and goes to excavate.Therefore, need to design a kind of scientific research hotspot prediction and supplying system, can to it is one section following when
Interior scientific research hotspot prediction comes out, and pushes it to scientific research clients, assists researcher and scientific research project management
The work of person, especially electrical network field, the field reforming speed and employing new technology is more, thus needs promptly and accurately
It is particularly important to hold current and following scientific research hot spot, power grid research work can be made to be correctly oriented.
Summary of the invention
In view of the above-mentioned problems, the invention proposes a kind of power grid scientific research hotspot prediction and method for pushing based on RNN, the party
Method allows user to obtain following a period of time power grid scientific research hot keyword in time, and carries out accurate prediction and in real time
Push.
To achieve the goals above, the present invention adopts the following technical scheme:
A kind of power grid science and technology hotspot prediction and method for pushing based on RNN, it is characterised in that the following steps are included:
One, hotspot prediction model
Using occur in the past period science and technology news website, bibliographic data base, and crawl module, spy by data
It is pre- to levy representation module, the feature extraction module based on depth Boltzmann machine and the formation hot spot of the model training module based on RNN
Survey model；
Two, power grid science and technology hotspot predictions and push
A t period power grid scientific and technical article) is acquired, by extracting power grid article TF-IDF vector, applying above-mentioned hotspot prediction
Model obtains scientific research hot keyword；
B) the power grid scientific and technical terms in database for natural language are clustered, and generate the pass of scientific and technical terms in cluster process
Connection degree；
C) according to power grid science and technology hot keyword and the scientific and technical terms degree of association degree synthetic time series power grid science and technology predicted
Hot spot, and it is pushed to user.
Hotspot prediction model the following steps are included:
1) data crawl module, crawl power grid scientific and technological information in science and technology news website, bibliographic data base using crawler technology
Article, the article textual that will be crawled, if the scientific and technological information article collection grabbed in a period of time is combined into Tt, wherein t indicates the period
Serial number；
2) character representation module, this module provide a kind of scientific and technological information and technical paper Text Representation method, and right
The text of a cycle carries out character representation, is characterized abstraction module and provides input, obtains T based on weight TF-IDF algorithmt's
Crucial term vector, is denoted asWherein, Q is the quantity of word in scientific and technological dictionary, biFor vocabulary component in corresponding dictionary
Gained weight TF-IDF value, biCalculate specific steps are as follows:
A2 t) is setjIt is TtA this paper, based on standard TF-IDF algorithm obtain tjThe TF-IDF value of i-th of vocabulary, is set asWithIndicate tjIn a keyword, wherein i beThe mark of vocabulary component in corresponding dictionary；
B2 t) is setjDownload or amount of reading be nj, reference amount is mj.SoWherein β is closed
The weight of keyword,WithAll article n in the period respectivelyjWith mjAverage value；
3) feature extraction module, this module, to a cycle Text character extraction, are mould based on depth Boltzmann machine
Type training module, prediction and recommending module provide data input, and depth Boltzmann machine model structure and parameters are provided that
A3) depth Boltzmann machine uses three layers of limitation Boltzmann machine；
B3) first layer is visible element layer, it is seen that elementary layer is Q × B two values matrix.WhereinT is rhythmic set, and [] is to be rounded, and η is a positive integer, and B is all
PeriodMiddle largest component is multiplied by a coefficient；
C3 v) is seti,jFor the unit of a visual layers, within the t period, forIf j=bi* η, then vi,j=1,
D3) second layer is hidden layer, and the second layer isTwo values matrix, βQWith βBFor the coefficient greater than 1.
Third layer is hidden layer, and third layer isTwo values matrix；
E3) depth Boltzmann machine is successively trained using limitation Boltzmann machine training method；
F3 it) is located at period t, the output based on depth Boltzmann machine is Xt；
4) model training module, after hotspot prediction refers to prediction following a period of time e, the research hotspot of appearance, this module
Prediction model is provided with recommending module for prediction, the hotspot prediction model of this module is improved based on RNN structure, knot
Structure is as follows with training method:
A4) hotspot prediction model structure t loop cycle layer and 3 layers of BP neural network as shown in figure 3, be made of, circulation layer
Input cycle data Xt, XtFor the output of t period depth Boltzmann machine,
U is input weight, and W is circulation weight, VlFor the weight of BP neural network l respective layer.VectorIndicate circulation mind
Weighting through member in t moment inputs, and calculation method is as follows:
Wherein St-1For the value of t-1 moment circulation layer, andWherein g () is activation primitive；
VectorIndicate that the BP neural network l layers of weighting in t moment inputs, calculation method is
otFor final output, calculation method is
B4 o) is calculatedtDuring, it is iterated to calculate based on following formula forward:
C4) indicate that model error, error function are with E
Wherein, N indicates sample size, y(n)Indicate the actual value of sample n, o(n)Indicate the output valve of sample n.y(n)With o(n)
It is vocabulary vector, if y(n)=[c1,c2,…,ci,…,cm], m is vocabulary total quantity, if i-th of vocabulary is hot keyword,
ciLabeled as 1, other non-hot vocabulary are labeled as 0, and hot spot vocabulary can have multiple；If o(n)=[d1,d2,…,di,…,dm],
So y(n)logo(n)It is calculated with following formula:
y(n)logo(n)=c1logd1+c2logd2+…+cilogdi+…+cmlogdm。
4) in model training module,
D4 the error term of each layer or each cycle period) is indicated with δ by error back propagation direction again, each error term
Calculation method is as follows:
Wherein, diag [x] indicates to create a diagonal matrix according to vector x,
I) error function E is calculated to any period k weight matrix WkGradient formula it is as follows:
WhereinIt indicatesI-th of component of error item vector, sk-1,jIndicate Sk-1J-th of neuron output valve,
The gradient of circulation layer weight matrix W is the sum of the gradient at each moment, and formula is as follows:
II) weight matrix U gradient calculation formula it is as follows:
Wherein xk,jIndicate XkJ-th of component value,
III) it setsFor the value of l layers of BP neural network layer of t moment, then:
Weight matrix VlGradient calculation formula it is as follows:
Wherein hk-1,jIndicate Hk-1J-th of neuron output valve.
It extracts in power grid article TF-IDF vector, a kind of power grid scientific and technological information and technical paper Text Representation side is provided
Method, and character representation is carried out to the text of a cycle, it is characterized abstraction module and input is provided, obtained based on weight TF-IDF algorithm
Obtain TtCrucial term vector, be denoted asWherein, Q is the quantity of word in power grid science and technology dictionary, biFor corresponding dictionary
Weight TF-IDF value obtained by middle vocabulary component, biCalculate specific steps are as follows:
A1 t) is setjIt is TtA this paper, based on standard TF-IDF algorithm obtain tjThe TF-IDF value of i-th of vocabulary, is set asWithIndicate tjIn a keyword, wherein i beThe mark of vocabulary component in corresponding dictionary；
B1 t) is setjDownload or amount of reading be nj, reference amount is mj.SoWherein β is closed
The weight of keyword,WithAll article n in the period respectivelyjWith mjAverage value.
In this application, TF-IDF (term frequency-inverse document frequency) is a kind of use
In the common weighting technique of information retrieval and data mining, this will not be detailed here.Above-mentioned RNN is Recognition with Recurrent Neural Network,
Recurrent Neural Network, neural network are a kind of artificial neural network of node orientation connection cyclization, this net
The internal state of network can show dynamic time sequence behavior, and different from feedforward neural network, RNN can use its internal note
Recall to handle the list entries of arbitrary sequence, it can be easier to handle handwriting recognition, the speech recognition etc. if not being segmented for this.
Limited Boltzmann machine can also be referred to as by limiting Boltzmann machine, it is a kind of can be distributed by input data set learning probability with
Machine generates neural network, according to the difference of task, is limited the method that supervised learning or unsupervised learning can be used in Bo Ziman machine
It is trained, limitation Boltzmann machine training method is the prior art, and this will not be detailed here.It is poly- in power grid scientific and technical terms cluster
Class uses the prior art, and such as K-means, this will not be detailed here.
The invention has the following advantages that
1) present invention can perceive the power grid Technological research hot spot being likely to occur in following a period of time, be power grid scientific research people
Member provides research direction and Research Thinking；
2) degree of association synthetic time series scientific research hot spot based on power grid scientific research vocabulary, the correlation of power grid science and technology hot spot can be dived
Technical application recommend power grid researcher.
Detailed description of the invention
Fig. 1 is process frame diagram of the invention；
Fig. 2 is scientific research hotspot prediction and push frame diagram in the present invention；
Fig. 3 is the structure chart of depth Boltzmann machine in the present invention；
Fig. 4 is the hotspot prediction model structure in the present invention based on RNN.
Specific embodiment
Refering to what is shown in Fig. 1, the invention discloses a kind of power grid science and technology hotspot prediction and method for pushing based on RNN, this hair
It is bright the following steps are included:
One, hotspot prediction model
With reference to Fig. 2, using occur in the past period science and technology news website, bibliographic data base, and climbed by data
Modulus block, character representation module, the feature extraction module based on depth Boltzmann machine and the model training module shape based on RNN
At hotspot prediction model；
Two, power grid science and technology hotspot predictions and push
A the power grid scientific and technical article for) crawling t period, by extracting, power grid article TF-IDF vector, to apply above-mentioned hot spot pre-
Model is surveyed, obtains scientific research hot keyword；
B) the power grid scientific and technical terms in database for natural language are clustered, and generate the pass of scientific and technical terms in cluster process
Connection degree；
C) according to power grid science and technology hot keyword and the scientific and technical terms degree of association degree synthetic time series power grid science and technology predicted
Hot spot, and it is pushed to user.
Hotspot prediction model the following steps are included:
1) data crawl module, crawl power grid scientific and technological information in science and technology news website, bibliographic data base using crawler technology
Article, the article textual that will be crawled, if the scientific and technological information article collection grabbed in a period of time is combined into Tt, wherein t indicates the period
Serial number；
2) character representation module, this module provide a kind of scientific and technological information and technical paper Text Representation method, and right
The text of a cycle carries out character representation, is characterized abstraction module and provides input, obtains T based on weight TF-IDF algorithmt's
Crucial term vector, is denoted asWherein, Q is the quantity of word in scientific and technological dictionary, biFor vocabulary component in corresponding dictionary
Gained weight TF-IDF value, biCalculate specific steps are as follows:
A2 t) is setjIt is TtA this paper, based on standard TF-IDF algorithm obtain tjThe TF-IDF value of i-th of vocabulary, is set asWithIndicate tjIn a keyword, wherein i beThe mark of vocabulary component in corresponding dictionary；
B2 t) is setjDownload or amount of reading be nj, reference amount is mj.SoWherein β is closed
The weight of keyword,WithAll article n in the period respectivelyjWith mjAverage value；
3) feature extraction module, refering to what is shown in Fig. 3, this module is based on the special to a cycle text of depth Boltzmann machine
Sign is extracted, and provides data input, depth Boltzmann machine model structure and parameters for model training module, prediction and recommending module
It is provided that
A3) depth Boltzmann machine uses three layers of limitation Boltzmann machine；
B3) first layer is visible element layer, it is seen that elementary layer is Q × B two values matrix.WhereinT is rhythmic set, and [] is to be rounded, and η is a biggish positive integer, by
In biValue range is [0,1], so η determines the size of B, in this patent, η setting value is 10000；That is B is in all weeks
PhaseMiddle largest component is multiplied by a coefficient；
C3 v) is seti,jFor the unit of a visual layers, within the t period, forIf j=bi* η, then vi,j=1,
D3) second layer is hidden layer, and the second layer isTwo values matrix, βQWith βBFor the coefficient greater than 1.Third
Layer is hidden layer, and third layer isTwo values matrix；
E3) depth Boltzmann machine is successively trained using limitation Boltzmann machine training method；
F3 it) is located at period t, the output based on depth Boltzmann machine is Xt；
4) model training module, with reference to Fig. 4, the fat full arrows line expression in figure calculates forward direction, for calculating ot, empty
After line arrow line indicates that error back propagation direction, hotspot prediction refer to prediction following a period of time e, the research hotspot of appearance,
This module provides prediction model for prediction and recommending module, and the hotspot prediction model of this module is improved based on RNN structure
, structure is as follows with training method:
A4) hotspot prediction model structure t loop cycle layer and 3 layers of BP neural network as shown in figure 3, be made of, circulation layer
Input cycle data Xt, XtFor the output of t period depth Boltzmann machine,
U is input weight, and W is circulation weight, VlFor the weight of BP neural network l respective layer.VectorIndicate circulation mind
Weighting through member in t moment inputs, and calculation method is as follows:
Wherein St-1For the value of t-1 moment circulation layer, andWherein g () is activation primitive；
VectorIndicate that BP neural network is inputted in the weighting of t moment, calculation method is
otFor final output, calculation method is
B4 o) is calculatedtDuring, it is iterated to calculate based on following formula forward:
C4) indicate that model error, error function are with E
Wherein, N indicates sample size, y(n)Indicate the actual value of sample n, o(n)Indicate the output valve of sample n.y(n)With o(n)
It is vocabulary vector, if y(n)=[c1,c2,…,ci,…,cm], m is vocabulary total quantity, if i-th of vocabulary is hot keyword,
ciLabeled as 1, other non-hot vocabulary are labeled as 0, and hot spot vocabulary can have multiple；If o(n)=[d1,d2,…,di,…,dm],
So y(n)logo(n)It is calculated with following formula:
y(n)logo(n)=c1logd1+c2logd2+…+cilogdi+…+cmlogdm。
In 4) model training module,
D4 the error term of each layer or each cycle period) is indicated with δ by error back propagation direction again, each error term
Calculation method is as follows:
Wherein, diag [x] indicates to create a diagonal matrix according to vector x,
I) error function E is calculated to any period k weight matrix WkGradient formula it is as follows:
WhereinIt indicatesI-th of component of error item vector, sk-1,jIndicate Sk-1J-th of neuron output valve,
The gradient of circulation layer weight matrix W is the sum of the gradient at each moment, and formula is as follows:
II) weight matrix U gradient calculation formula it is as follows:
Wherein xk,jIndicate XkJ-th of component value,
III) it setsFor the value of l layers of BP neural network layer of t moment, then:
Weight matrix VlGradient calculation formula it is as follows:
Wherein hk-1,jIndicate Hk-1J-th of neuron output valve.
It extracts in power grid article TF-IDF vector, a kind of power grid scientific and technological information and technical paper Text Representation side is provided
Method, and character representation is carried out to the text of a cycle, it is characterized abstraction module and input is provided, obtained based on weight TF-IDF algorithm
Obtain TtCrucial term vector, be denoted asWherein, Q is the quantity of word in power grid science and technology dictionary, biFor corresponding dictionary
Weight TF-IDF value obtained by middle vocabulary component, biCalculate specific steps are as follows:
A1 t) is setjIt is TtA this paper, based on standard TF-IDF algorithm obtain tjThe TF-IDF value of i-th of vocabulary, is set asWithIndicate tjIn a keyword, wherein i beThe mark of vocabulary component in corresponding dictionary；
B1 t) is setjDownload or amount of reading be nj, reference amount is mj.SoWherein β is closed
The weight of keyword,WithAll article n in the period respectivelyjWith mjAverage value.
In above-mentioned 4) model training module, error back propagation formula and error term formula based on D4) simultaneously pass through reversed
Propagation Neural Network training algorithm training pattern provides accurate model for power grid science and technology hot spot and prediction.
Although above-mentioned be described and verify to a specific embodiment of the invention and validity in conjunction with attached drawing, not
Limiting the scope of the invention, those skilled in the art should understand that, based on the technical solutions of the present invention, this
Field technical staff does not need to make the creative labor the various modifications or changes that can be made still in protection scope of the present invention
Within.
Claims (4)
1. a kind of power grid science and technology hotspot prediction and method for pushing based on RNN, it is characterised in that the following steps are included:
One, hotspot prediction model
Using occur in the past period science and technology news website, bibliographic data base, and crawl module, mark sheet by data
Show that module, the feature extraction module based on depth Boltzmann machine and the model training module based on RNN form hotspot prediction mould
Type；
Two, power grid science and technology hotspot predictions and push
A t period power grid scientific and technical article) is acquired, by extracting power grid article TF-IDF vector, applying above-mentioned hotspot prediction mould
Type obtains scientific research hot keyword；
B) the power grid scientific and technical terms in database for natural language are clustered, and generate the association of scientific and technical terms in cluster process
Degree；
C the power grid science and technology hot keyword and scientific and technical terms degree of association degree synthetic time series power grid science and technology hot spot that) basis predicts,
And it is pushed to user.
2. the power grid science and technology hotspot prediction and method for pushing according to claim 1 based on RNN, which is characterized in that hot spot
Prediction model the following steps are included:
1) data crawl module, crawl power grid scientific and technological information article in science and technology news website, bibliographic data base using crawler technology,
The article textual that will be crawled, if the scientific and technological information article collection grabbed in a period of time is combined into Tt, wherein t indicates period serial number；
2) character representation module, this module provide a kind of scientific and technological information and technical paper Text Representation method, and to one
The text in period carries out character representation, is characterized abstraction module and provides input, obtains T based on weight TF-IDF algorithmtKey
Term vector is denoted asWherein, Q is the quantity of word in scientific and technological dictionary, biFor obtained by vocabulary component in corresponding dictionary
Weight TF-IDF value, biCalculate specific steps are as follows:
A2 t) is setjIt is TtA this paper, based on standard TF-IDF algorithm obtain tjThe TF-IDF value of i-th of vocabulary, is set as
WithIndicate tjIn a keyword, wherein i beThe mark of vocabulary component in corresponding dictionary；
B2 t) is setjDownload or amount of reading be nj, reference amount is mj.SoWherein β keyword
Weight,WithAll article n in the period respectivelyjWith mjAverage value；
3) feature extraction module, this module, to a cycle Text character extraction, are instructed based on depth Boltzmann machine for model
Practice module, prediction and recommending module and data input is provided, depth Boltzmann machine model structure and parameters are provided that
A3) depth Boltzmann machine uses three layers of limitation Boltzmann machine；
B3) first layer is visible element layer, it is seen that elementary layer is Q × B two values matrix.WhereinT is rhythmic set, and [] is to be rounded, and η is a positive integer, and B is all
PeriodMiddle largest component is multiplied by a coefficient；
C3 v) is seti,jFor the unit of a visual layers, within the t period, forIf j=bi* η, then vi,j=1,
D3) second layer is hidden layer, and the second layer isTwo values matrix, βQWith βBFor the coefficient greater than 1.Third layer is
Hidden layer, third layer areTwo values matrix；
E3) depth Boltzmann machine is successively trained using limitation Boltzmann machine training method；
F3 it) is located at period t, the output based on depth Boltzmann machine is Xt；
4) model training module, after hotspot prediction refers to prediction following a period of time e, the research hotspot of appearance, this module is pre-
Survey with recommending module provide prediction model, the hotspot prediction model of this module is improved based on RNN structure, structure with
Training method is as follows:
A4) hotspot prediction model structure t loop cycle layer and 3 layers of BP neural network as shown in figure 3, be made of, circulation layer input
Cycle data Xt, XtFor the output of t period depth Boltzmann machine,
U is input weight, and W is circulation weight, VlFor the weight of BP neural network l respective layer.VectorIndicate circulation neuron
It is inputted in the weighting of t moment, calculation method is as follows:
Wherein St-1For the value of t-1 moment circulation layer, andWherein g () is activation primitive；
VectorIndicate that the BP neural network l layers of weighting in t moment inputs, calculation method is
otFor final output, calculation method is
B4 o) is calculatedtDuring, it is iterated to calculate based on following formula forward:
C4) indicate that model error, error function are with E
Wherein, N indicates sample size, y(n)Indicate the actual value of sample n, o(n)Indicate the output valve of sample n.y(n)With o(n)It is word
Remittance vector, if y(n)=[c1,c2,…,ci,…,cm], m is vocabulary total quantity, if i-th of vocabulary is hot keyword, ciMark
It is denoted as 1, other non-hot vocabulary are labeled as 0, and hot spot vocabulary can have multiple；If o(n)=[d1,d2,…,di,…,dm], then
y(n)logo(n)It is calculated with following formula:
y(n)logo(n)=c1logd1+c2logd2+…+cilogdi+…+cmlogdm。
3. the power grid science and technology hotspot prediction and method for pushing according to claim 2 based on RNN, which is characterized in that 4) mould
In type training module,
D4 the error term of each layer or each cycle period, the calculating of each error term) are indicated with δ by error back propagation direction again
Method is as follows:
Wherein, diag [x] indicates to create a diagonal matrix according to vector x,
I) error function E is calculated to any period k weight matrix WkGradient formula it is as follows:
WhereinIt indicatesI-th of component of error item vector, sk-1,jIndicate Sk-1J-th of neuron output valve,
The gradient of circulation layer weight matrix W is the sum of the gradient at each moment, and formula is as follows:
II) weight matrix U gradient calculation formula it is as follows:
Wherein xk,jIndicate XkJ-th of component value,
III) it setsFor the value of l layers of BP neural network layer of t moment, then:
Weight matrix VlGradient calculation formula it is as follows:
Wherein hk-1,jIndicate Hk-1J-th of neuron output valve.
4. the power grid science and technology hotspot prediction and method for pushing according to any one of claim 1-3 based on RNN, feature
It is, extracts in power grid article TF-IDF vector, a kind of power grid scientific and technological information and technical paper Text Representation method is provided,
And character representation is carried out to the text of a cycle, it is characterized abstraction module and input is provided, obtained based on weight TF-IDF algorithm
TtCrucial term vector, be denoted asWherein, Q is the quantity of word in power grid science and technology dictionary, biFor in corresponding dictionary
Weight TF-IDF value obtained by vocabulary component, biCalculate specific steps are as follows:
A1 t) is setjIt is TtA this paper, based on standard TF-IDF algorithm obtain tjThe TF-IDF value of i-th of vocabulary, is set as
WithIndicate tjIn a keyword, wherein i beThe mark of vocabulary component in corresponding dictionary；
B1 t) is setjDownload or amount of reading be nj, reference amount is mj.SoWherein β keyword
Weight,WithAll article n in the period respectivelyjWith mjAverage value.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810970350.9A CN109214562A (en) | 2018-08-24 | 2018-08-24 | A kind of power grid scientific research hotspot prediction and method for pushing based on RNN |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810970350.9A CN109214562A (en) | 2018-08-24 | 2018-08-24 | A kind of power grid scientific research hotspot prediction and method for pushing based on RNN |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109214562A true CN109214562A (en) | 2019-01-15 |
Family
ID=64989643
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810970350.9A Pending CN109214562A (en) | 2018-08-24 | 2018-08-24 | A kind of power grid scientific research hotspot prediction and method for pushing based on RNN |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109214562A (en) |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110188263A (en) * | 2019-05-29 | 2019-08-30 | 国网山东省电力公司电力科学研究院 | It is a kind of towards isomery when away from scientific research hotspot prediction method and system |
| CN110378532A (en) * | 2019-07-19 | 2019-10-25 | 中南大学 | A kind of scientific research theme trend prediction method based on random tree |
| CN110472004A (en) * | 2019-08-23 | 2019-11-19 | 国网山东省电力公司电力科学研究院 | A kind of method and system of scientific and technological information data multilevel cache management |
| CN110688477A (en) * | 2019-10-10 | 2020-01-14 | 华夏幸福产业投资有限公司 | Prediction method, device, equipment and storage medium in technical hotspot field |
| CN110705821A (en) * | 2019-08-23 | 2020-01-17 | 上海科技发展有限公司 | Hotspot subject prediction method, device, terminal and medium based on multiple evaluation dimensions |
| CN111832815A (en) * | 2020-07-02 | 2020-10-27 | 山东电力研究院 | Scientific research hotspot prediction method and system |
| CN112187890A (en) * | 2020-09-15 | 2021-01-05 | 卢霞浩 | Information distribution method based on cloud computing and big data and block chain financial cloud center |
| CN112734126A (en) * | 2021-01-18 | 2021-04-30 | 武汉烽火技术服务有限公司 | Hot spot prediction method, device and equipment and readable storage medium |
| CN113449175A (en) * | 2020-03-24 | 2021-09-28 | 北京沃东天骏信息技术有限公司 | Hot data recommendation method and device |
| CN113722424A (en) * | 2021-07-20 | 2021-11-30 | 山东电力研究院 | Scientific research direction recommendation method and system based on news events |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102012929A (en) * | 2010-11-26 | 2011-04-13 | 北京交通大学 | Network consensus prediction method and system |
| JP2013257765A (en) * | 2012-06-13 | 2013-12-26 | Ntt Data Corp | Term extraction device, term extraction method, and program |
| CN105224608A (en) * | 2015-09-06 | 2016-01-06 | 华南理工大学 | The hot news Forecasting Methodology analyzed based on microblog data and system |
| CN106651030A (en) * | 2016-12-21 | 2017-05-10 | 重庆邮电大学 | Method for predicting user participation behavior of hot topic by improved RBF neural network |
| CN106682172A (en) * | 2016-12-28 | 2017-05-17 | 江苏大学 | Keyword-based document research hotspot recommending method |
| CN107038156A (en) * | 2017-04-28 | 2017-08-11 | 北京清博大数据科技有限公司 | A kind of hot spot of public opinions Forecasting Methodology based on big data |
| CN107992976A (en) * | 2017-12-15 | 2018-05-04 | 中国传媒大学 | Much-talked-about topic early-stage development trend predicting system and Forecasting Methodology |
-
2018
- 2018-08-24 CN CN201810970350.9A patent/CN109214562A/en active Pending
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102012929A (en) * | 2010-11-26 | 2011-04-13 | 北京交通大学 | Network consensus prediction method and system |
| JP2013257765A (en) * | 2012-06-13 | 2013-12-26 | Ntt Data Corp | Term extraction device, term extraction method, and program |
| CN105224608A (en) * | 2015-09-06 | 2016-01-06 | 华南理工大学 | The hot news Forecasting Methodology analyzed based on microblog data and system |
| CN106651030A (en) * | 2016-12-21 | 2017-05-10 | 重庆邮电大学 | Method for predicting user participation behavior of hot topic by improved RBF neural network |
| CN106682172A (en) * | 2016-12-28 | 2017-05-17 | 江苏大学 | Keyword-based document research hotspot recommending method |
| CN107038156A (en) * | 2017-04-28 | 2017-08-11 | 北京清博大数据科技有限公司 | A kind of hot spot of public opinions Forecasting Methodology based on big data |
| CN107992976A (en) * | 2017-12-15 | 2018-05-04 | 中国传媒大学 | Much-talked-about topic early-stage development trend predicting system and Forecasting Methodology |
Non-Patent Citations (2)
| Title |
|---|
| 张祎萍: ""基于聚类分析的热点图书排序推荐方法研究"", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
| 胡悦: ""微博舆情热点发现及趋势预测研究"", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110188263A (en) * | 2019-05-29 | 2019-08-30 | 国网山东省电力公司电力科学研究院 | It is a kind of towards isomery when away from scientific research hotspot prediction method and system |
| CN110188263B (en) * | 2019-05-29 | 2021-11-30 | 国网山东省电力公司电力科学研究院 | Heterogeneous time interval-oriented scientific research hotspot prediction method and system |
| CN110378532A (en) * | 2019-07-19 | 2019-10-25 | 中南大学 | A kind of scientific research theme trend prediction method based on random tree |
| CN110378532B (en) * | 2019-07-19 | 2021-12-14 | 中南大学 | Scientific research topic state prediction method based on random tree |
| CN110472004A (en) * | 2019-08-23 | 2019-11-19 | 国网山东省电力公司电力科学研究院 | A kind of method and system of scientific and technological information data multilevel cache management |
| CN110705821A (en) * | 2019-08-23 | 2020-01-17 | 上海科技发展有限公司 | Hotspot subject prediction method, device, terminal and medium based on multiple evaluation dimensions |
| WO2021035975A1 (en) * | 2019-08-23 | 2021-03-04 | 上海科技发展有限公司 | Method and apparatus for predicting hot-topic subject on basis of multiple evaluation dimensions, terminal, and medium |
| CN110688477A (en) * | 2019-10-10 | 2020-01-14 | 华夏幸福产业投资有限公司 | Prediction method, device, equipment and storage medium in technical hotspot field |
| CN110688477B (en) * | 2019-10-10 | 2022-11-15 | 华夏幸福产业投资有限公司 | Prediction method, device, equipment and storage medium in technical hotspot field |
| CN113449175A (en) * | 2020-03-24 | 2021-09-28 | 北京沃东天骏信息技术有限公司 | Hot data recommendation method and device |
| CN111832815B (en) * | 2020-07-02 | 2023-12-05 | 国网山东省电力公司电力科学研究院 | Scientific research hot spot prediction method and system |
| CN111832815A (en) * | 2020-07-02 | 2020-10-27 | 山东电力研究院 | Scientific research hotspot prediction method and system |
| CN112187890A (en) * | 2020-09-15 | 2021-01-05 | 卢霞浩 | Information distribution method based on cloud computing and big data and block chain financial cloud center |
| CN112187890B (en) * | 2020-09-15 | 2021-05-07 | 北京联银通科技有限公司 | Information distribution method based on cloud computing and big data and block chain financial cloud center |
| CN112734126A (en) * | 2021-01-18 | 2021-04-30 | 武汉烽火技术服务有限公司 | Hot spot prediction method, device and equipment and readable storage medium |
| CN113722424A (en) * | 2021-07-20 | 2021-11-30 | 山东电力研究院 | Scientific research direction recommendation method and system based on news events |
| CN113722424B (en) * | 2021-07-20 | 2024-02-02 | 国网山东省电力公司电力科学研究院 | Scientific research direction recommendation method and system based on news event |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109214562A (en) | A kind of power grid scientific research hotspot prediction and method for pushing based on RNN | |
| CN109299396A (en) | Merge the convolutional neural networks collaborative filtering recommending method and system of attention model | |
| CN108733742A (en) | Global normalization's reader system and method | |
| CN104598611B (en) | The method and system being ranked up to search entry | |
| CN113761218B (en) | Method, device, equipment and storage medium for entity linking | |
| Cohen et al. | End to end long short term memory networks for non-factoid question answering | |
| Anupriya et al. | LDA based topic modeling of journal abstracts | |
| Wang et al. | Outfit compatibility prediction and diagnosis with multi-layered comparison network | |
| CN109522553A (en) | Name recognition methods and the device of entity | |
| CN110796190A (en) | Exponential modeling with deep learning features | |
| CN109299976A (en) | Clicking rate prediction technique, electronic device and computer readable storage medium | |
| CN110222163A (en) | A kind of intelligent answer method and system merging CNN and two-way LSTM | |
| CN107391485A (en) | Entity recognition method is named based on the Korean of maximum entropy and neural network model | |
| CN104615767A (en) | Searching-ranking model training method and device and search processing method | |
| CN111222318B (en) | Trigger word recognition method based on double-channel bidirectional LSTM-CRF network | |
| CN110413769A (en) | Scene classification method, device, storage medium and its electronic equipment | |
| CN106970981B (en) | Method for constructing relation extraction model based on transfer matrix | |
| CN111710428B (en) | Biomedical text representation method for modeling global and local context interaction | |
| CN106776808A (en) | Information data offering method and device based on artificial intelligence | |
| CN104182421A (en) | Video clustering method and detecting method | |
| CN109815478A (en) | Medicine entity recognition method and system based on convolutional neural networks | |
| CN111191461B (en) | Remote supervision relation extraction method based on course learning | |
| Munshi et al. | Towards an automated islamic fatwa system: Survey, dataset and benchmarks | |
| CN103593334B (en) | A kind of method and system for being used to judge emotional degree of text | |
| Matsuda et al. | Benchmark for Personalized Federated Learning |
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 | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20190115 |
|
| WD01 | Invention patent application deemed withdrawn after publication |