CN110119846A - A kind of multiple target deepness belief network for mid-term electric load forecasting - Google Patents
A kind of multiple target deepness belief network for mid-term electric load forecasting Download PDFInfo
- Publication number
- CN110119846A CN110119846A CN201910393383.6A CN201910393383A CN110119846A CN 110119846 A CN110119846 A CN 110119846A CN 201910393383 A CN201910393383 A CN 201910393383A CN 110119846 A CN110119846 A CN 110119846A
- Authority
- CN
- China
- Prior art keywords
- data
- prediction model
- individual
- belief network
- prediction
- 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
- 238000000354 decomposition reaction Methods 0.000 claims abstract description 16
- 239000013598 vector Substances 0.000 claims abstract description 9
- 238000002790 cross-validation Methods 0.000 claims abstract description 8
- 238000005457 optimization Methods 0.000 claims abstract description 7
- 238000012549 training Methods 0.000 claims description 30
- 238000000034 method Methods 0.000 claims description 17
- 238000012360 testing method Methods 0.000 claims description 10
- 230000010354 integration Effects 0.000 claims description 4
- 238000005516 engineering process Methods 0.000 claims description 2
- 238000010801 machine learning Methods 0.000 claims description 2
- 230000006870 function Effects 0.000 description 13
- 238000013528 artificial neural network Methods 0.000 description 5
- 238000013461 design Methods 0.000 description 4
- 230000035772 mutation Effects 0.000 description 3
- 238000013459 approach Methods 0.000 description 2
- 238000004364 calculation method Methods 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 241000208340 Araliaceae Species 0.000 description 1
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 description 1
- 235000003140 Panax quinquefolius Nutrition 0.000 description 1
- 210000003484 anatomy Anatomy 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 210000004556 brain Anatomy 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000005611 electricity Effects 0.000 description 1
- 238000013213 extrapolation Methods 0.000 description 1
- 235000008434 ginseng Nutrition 0.000 description 1
- 238000001027 hydrothermal synthesis Methods 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 210000002569 neuron Anatomy 0.000 description 1
- 239000002245 particle Substances 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000000192 social effect Effects 0.000 description 1
Classifications
-
- 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
- G06N3/08—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
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/06—Energy or water supply
Landscapes
- Engineering & Computer Science (AREA)
- Business, Economics & Management (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Economics (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Strategic Management (AREA)
- Human Resources & Organizations (AREA)
- Software Systems (AREA)
- Tourism & Hospitality (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Mathematical Physics (AREA)
- Data Mining & Analysis (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- General Business, Economics & Management (AREA)
- Molecular Biology (AREA)
- Artificial Intelligence (AREA)
- Marketing (AREA)
- Life Sciences & Earth Sciences (AREA)
- Primary Health Care (AREA)
- Water Supply & Treatment (AREA)
- Public Health (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 present invention is using handling mid-term electric load forecasting problem based on the multiple-objection optimization deepness belief network of decomposition, data are decomposed using empirical mode decomposition first, and the data of decomposition and former data are combined into new data set, then it is excavated using Multi-target Data, using single DBN as the candidate solution of multi-objective optimization algorithm, multiple DBN iterative evolution under the driving of MOEA/D algorithm simultaneously, take into account the accuracy and diversity of model, cross validation is being used to sample after circulation terminates, is preventing over-fitting；It proposes two-stage policy later to screen to these prediction models, finally distributes weight vectors to single prediction model using integrated study；The present invention with the parameter of optimum choice prediction model and can effectively improve the efficiency of algorithm and the accuracy of prediction.
Description
Technical field
The present invention relates to mid-term electric load requirement forecasting fields, are exactly using the multiple-objection optimization based on decomposition
Deepness belief network handles electric load needs of problems.
Background technique
Load forecast is to carry out prediction work using electric power big data as research object.From the point of view of predicting object,
Electric load requirement forecasting mainly predicts following electricity demand, and improve workload demand prediction accuracy and
Efficiency promotes electric power enterprise to design reasonable electric power distribution planning, improves the social effect and economy of electric system with this.
Algorithm used in electric load requirement forecasting plays an important role load forecast.Currently, load needs
The method for asking prediction mainly has neural network, returns anatomy method, tends to extrapolation, elastic coefficient method, gray forecast approach etc..
In these methods, neural network can simulate human brain and carry out intelligent processing to data, and in face of a large amount of
Data can be predicted in the correlation in non-supervisory situation between autonomous learning data, and with this.Remember with information
Recall, the features such as autonomous learning.But there is also disadvantages for this method.The parameter of neural network is to need to think setting, and recognize
It needs to be adjusted according to the effect of training set for setting parameter.This not only increases workload, and is possible to generate training
Collect over-fitting and test result is made to obtain error increase；Therefore need a kind of method that neural network can be allow independently to select to join
Number.
Summary of the invention
The present invention mainly utilizes deepness belief network (DBN) Lai Jianli prediction model, captures number to improve prediction model
According to ability and prediction model precision, the present invention decomposes data using empirical mode decomposition (EMD), uses
The theory of " dividing and rule ", the data of decomposition, which are sequentially placed into after deepness belief network (DBN) is trained, makes trained number
According to former Data Integration at a group data set.
Technical staff needs to manually adjust the parameter of prediction model when establishing prediction model, and complicated parameter may sometimes
Staff can be caused to increase workload.Parameter selection is difficult when the present invention is directed to present neural network prediction electric load, is
The time of the artificial adjusting parameter of reduction, devise a kind of novel multiple target deepness belief network (MODBN).
The present invention applies the ginseng of multi-objective Evolutionary Algorithm (MOEA/D) the Lai Youhua deepness belief network (DBN) based on decomposition
Number, is compared multiple prediction models in object space；Provided with 2 objective functions, one of them is prediction model prediction
The root-mean-square error of training set；The other is diversity of the prediction model in object space.Make the two objective functions simultaneously
Minimum can make two targets conflict with each other, and be conducive to select the prediction model that accuracy is high and generalization is strong.
The method that the present invention designs optimizes unlike the method for parameter from single goal evolution algorithm.With difference into
Change the single object optimization algorithm that algorithm (DE), particle swarm algorithm (PSO) etc. are representative, the general root-mean-square error for utilizing training set
As objective function, due to no conflicting other objective functions, the prediction model of generation may be generated to training set
Over-fitting, reduce the generalization ability of prediction model.Root mean square and each depth of the method that the present invention designs using training set
The diversity of belief network (DBN) is spent as objective function, this objective function of diversity is set and is effectively prevented training set mistake
Fitting.Application integration study (Ensemble Learning) integrates multiple deepness belief networks (DBN), improves pre-
Survey the generalization ability of model.
The present invention implements that specific step is as follows:
Step 1: acquiring the key data of power load, and data are standardized.By the data after standardization
Classified using time window, sorted data set is divided into training set and test set later；
Step 2: using training set training prediction model, until meeting certain cycle-index.Wherein, last time recycles
Use the Cross-Validation technique (Cross-validation) that over-fitting is prevented in machine learning.It can by multiple-objection optimization technology
To obtain multiple prediction models；
Step 3: obtained population being selected using non-dominant criterion and crowding distance, abbreviation two-stage policy.
Non-dominated ranking is carried out to the individual in population first, is ranked up later according to the crowding of each individual.According to sequence according to
Secondary selection individual, select number is the 50% of original Population Size, these it is individual for other 50% individual more
It is more preferable close to Pareto forward position or degree of scatter；
Step 4: utilizing integrated study, distribute weight vectors for each individual；
Step 5: carrying out the prediction of data to test set using obtained integrated model.
Detailed description of the invention
In order to preferably be supplemented the contents of the present invention and be illustrated, it is described below in conjunction with attached drawing.
Fig. 1 is prediction technique flow chart of the invention；
Fig. 2 is modified hydrothermal process of the present invention: the flow chart of multiple target deepness belief network (MODBN)；
The working drawing that Fig. 3 is empirical mode decomposition (EMD) to decompose data and be integrated into a group data set；
Specific embodiment
Collected Power system load data is handled first, data are standardized by the present invention using following formula:
Data used in the present invention are per half an hour to acquire the data once generated, will have 48 data daily.Benefit
With time window method, using every 48 data as one group, next group of data always with upper one group of data break half an hour ---
Namely time window is moved with half an hour.The matrix of a n × m can be generated later using time window, wherein row vector n is
Number of samples, column vector m are sample dimension, m i.e. time window length 48.Data are divided using empirical mode decomposition (EMD)
Data set is divided into training set and test set at a data set by solution, the data and former Data Integration after decomposition.
It is described according to implementation steps of the figure two to algorithm:
Step 1: the objective function of algorithm for design and the range of decision variable.
The algorithm using the population diversity of deepness belief network (DBN) in the root mean square of training set and object space as
Objective function.Root mean square formula are as follows:
Wherein, N indicates the number for the training sample being predicted,Indicate i-th of m-th of deepness belief network (DBN)
The predicted value of training sample, RiIndicate the actual value of i-th of training sample.
The calculation formula of population diversity are as follows:
In above formula,WithRespectively indicate m-th of deepness belief network (DBN) and j-th of deepness belief network (DBN)
The predicted value of i-th of training sample.Multiple deepness belief networks (DBN) are in i-th of training sample after indicating circulation every time
The average of predicted value.(3) result of formula must be a negative, and the absolute value of the negative is bigger, single deepness belief network
(DBN) bigger with the negative correlation of other deepness belief networks (DBN), diversity is also better.
By the initial weight { λ of deepness belief network1,…λM, the quantity n of neuron, weight and deviation learning rate a make
For 3 decision variables of multiple target decomposition algorithm (MOEA/D), and set gradually the variation range of 3 variables.
Step 2: setting control parameter.
Control parameter of the invention has the number M of the point of multiple target decomposition algorithm (MOEA/D).Each neighborhood of a point T,
In, i-th of neighborhood of a point is B (i)={ i1,…,iT}.And cycle-index G.Z is set*Two stages are arranged in=(Inf, Inf)
The number of individuals of policy selection is the 50% of Population Size.
Step 3: one group of equally distributed weight vectors is generated on object space.It is random to generate decision variable, according to life
At decision variable construction depth belief network (DBN), training set is inputted to each network respectively and establishes prediction model.And root
The mean square deviation and its diversity of each deepness belief network (DBN) training set are found out according to training result.It needs simultaneously to original
The z come*It is updated.For the optimal value of i-th of objective function in population.
The mean square deviation of each prediction model training set and their diversity in object space are subjected to standard respectively
Change.Specially successively by objective function divided by the maximum value of objective function in previous generation population.
Step 4: Gaussian mutation is applied after generating new individual using differential evolution algorithm, specific as follows:
It is updated to i-th of individual: from the vertex neighborhood B (i)={ i1,…,iTIn 2 parents of random selection carry out
Differential evolution, if rndU(0,1)≤0.5, then occur Gaussian mutation.
Wherein, rndU(0,1) equally distributed random number between [0,1] is indicated,Represent i-th of individual of the g times circulation.WithThe point of j and k is designated as under respectively indicating in i individual neighborhood.For the candidate solution finally generated.In general, the value of setting F
It is 0.5.The probability of Gaussian mutation is 0.5 or less.Each element in vector σ is arranged to the two of corresponding decision variable
1/10th.After generating new individual, respectively to zf, z*And target function value is updated.
Step 5: the candidate solution of generation is compared with former individual.Chebyshev's method has been used when relatively herein
(Tchebycheff approach) is used as aggregate function, is defined as follows:
Individual is updated by above formula, if cycle-index is not up to maximum value G at this time, continues to execute step 4.If reaching
To maximum value G, needs that cross validation strategy is added before training, prevent the over-fitting of training set, can reinforce so pre-
Survey the generalization ability of model.
Step 6: population at individual being ranked up using two-stage policy, optimal 50% in selected population individual.Specifically
Are as follows:
(1) non-dominated ranking is carried out to the individual in population；
(2) crowding sequence is carried out to the population after sequence；
(3) individual of selection preceding 50%；
Step 7: integrated study is utilized, distributes weight vectors for each prediction model, weight calculation is as follows:
In above formula, wiFor the weight that i-th of prediction model is assigned, eiIt is the root-mean-square error of i-th of training set.Relatively
In other utilization single goal algorithm come generate weight integrated study for calculate simple, and calculated result is more accurate.
According to S=w1s1+w2s2+…+wnsn, obtain the predicted value of test set.Wherein siIndicate i-th of deepness belief network
(DBN) predicted value of test set.
Claims (5)
1. pre- in order to improve with mid-term electric load needs of problems is handled based on the multiple-objection optimization deepness belief network of decomposition
It surveys model and captures the ability of data and the precision of prediction model；The present invention using empirical mode decomposition (EMD) come to data into
Row decomposes, and using the theory of " dividing and rule ", the data of decomposition, which are sequentially placed into after deepness belief network (DBN) is trained, to be made
Trained data and former Data Integration are at a group data set；
Step 1: acquiring the key data of power load, and data are standardized, the data after standardization are utilized
Time window is classified, and sorted data set is divided into training set and test set later；
Step 2: using training set training prediction model, until meeting certain cycle-index, wherein last time is recycled
The Cross-Validation technique (Cross-validation) that over-fitting is prevented in machine learning, can be obtained by multiple-objection optimization technology
To multiple prediction models；
Step 3: obtained population being selected using non-dominant criterion and crowding distance, abbreviation two-stage policy, first
Non-dominated ranking is carried out to the individual in population, carries out two minor sorts further according to the crowding of each individual；According to ranking results
Successively selection individual, selects number for the 50% of original Population Size；
Step 4: utilizing integrated study, distribute weight vectors for each individual；
Step 5: carrying out the prediction of data to test set using obtained integrated model.
2. the present invention mainly utilizes deepness belief network (DBN) Lai Jianli prediction model as described in claim 1 step 1, in order to
It improves prediction model and captures the ability of data characteristics and the precision of prediction model, decomposed using empirical mode decomposition (EMD) method
Data, and obtained data will be decomposed and merge into new data set with former data.
3. described in step 2 as claimed in claim 1, using training set training prediction model, until meeting certain cycle-index,
In, the Cross-Validation technique (Cross-validation) for preventing over-fitting is recycled in machine study in last time.
4. as described in claim 1, being selected using non-dominant criterion and crowding distance obtained population, referred to as two ranks
Section strategy；Non-dominated ranking is carried out to the individual in population first, is ranked up later according to the crowding of each individual；According to
Successively selection is individual for sequence, selects number for the 50% of original Population Size.
5. described in step 5 as claimed in claim 1, using integrated study, distributing weight vectors, weight meter for each prediction model
It calculates as follows:
In above formula, wiFor the weight that i-th of prediction model is assigned, eiIt is the root-mean-square error of i-th of training set；Relative to it
It is generated using single goal algorithm and is calculated simply for the integrated study of weight, and calculated result is more accurate；
According to S=w1s1+w2s2+…+wnsn, obtain the predicted value of test set, wherein siIndicate i-th of deepness belief network (DBN)
The predicted value of test set.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910393383.6A CN110119846A (en) | 2019-05-10 | 2019-05-10 | A kind of multiple target deepness belief network for mid-term electric load forecasting |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910393383.6A CN110119846A (en) | 2019-05-10 | 2019-05-10 | A kind of multiple target deepness belief network for mid-term electric load forecasting |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110119846A true CN110119846A (en) | 2019-08-13 |
Family
ID=67522235
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910393383.6A Pending CN110119846A (en) | 2019-05-10 | 2019-05-10 | A kind of multiple target deepness belief network for mid-term electric load forecasting |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110119846A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111610394A (en) * | 2020-05-20 | 2020-09-01 | 湘潭大学 | Electric energy quality disturbance positioning and identifying method for electrified railway traction power supply system |
| CN112381271A (en) * | 2020-10-30 | 2021-02-19 | 广西大学 | Distributed multi-objective optimization acceleration method for rapidly resisting deep belief network |
| CN113283179A (en) * | 2021-06-17 | 2021-08-20 | 湘潭大学 | Short-term load prediction method based on multi-target LSTM integrated network |
| CN114760095A (en) * | 2022-03-09 | 2022-07-15 | 西安电子科技大学 | Intention-driven network defense strategy generation method, system and application |
-
2019
- 2019-05-10 CN CN201910393383.6A patent/CN110119846A/en active Pending
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111610394A (en) * | 2020-05-20 | 2020-09-01 | 湘潭大学 | Electric energy quality disturbance positioning and identifying method for electrified railway traction power supply system |
| CN112381271A (en) * | 2020-10-30 | 2021-02-19 | 广西大学 | Distributed multi-objective optimization acceleration method for rapidly resisting deep belief network |
| CN113283179A (en) * | 2021-06-17 | 2021-08-20 | 湘潭大学 | Short-term load prediction method based on multi-target LSTM integrated network |
| CN114760095A (en) * | 2022-03-09 | 2022-07-15 | 西安电子科技大学 | Intention-driven network defense strategy generation method, system and application |
| CN114760095B (en) * | 2022-03-09 | 2023-04-07 | 西安电子科技大学 | Intention-driven network defense strategy generation method, system and application |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Mahmoud et al. | An advanced approach for optimal wind power generation prediction intervals by using self-adaptive evolutionary extreme learning machine | |
| Lu et al. | Review of meta-heuristic algorithms for wind power prediction: Methodologies, applications and challenges | |
| CN110119846A (en) | A kind of multiple target deepness belief network for mid-term electric load forecasting | |
| Heydari et al. | A novel composite neural network based method for wind and solar power forecasting in microgrids | |
| CN108921339B (en) | Quantile regression-based photovoltaic power interval prediction method for genetic support vector machine | |
| CN109558975B (en) | Integration method for multiple prediction results of power load probability density | |
| Deng et al. | A novel parallel hybrid intelligence optimization algorithm for a function approximation problem | |
| CN111929748A (en) | Meteorological element forecasting method and system | |
| CN109214449A (en) | A kind of electric grid investment needing forecasting method | |
| Grimaccia et al. | Neuro-fuzzy predictive model for PV energy production based on weather forecast | |
| Pranolo et al. | Robust LSTM With tuned-PSO and bifold-attention mechanism for analyzing multivariate time-series | |
| CN109919421A (en) | Short-term power load prediction model establishment method based on VMD-PSO-BPNN | |
| Zhou et al. | Deep learning-based rolling horizon unit commitment under hybrid uncertainties | |
| CN109934422A (en) | Neural network wind speed prediction method based on time series data analysis | |
| CN115986728A (en) | Power grid situation prediction method considering uncertainty factors and terminal | |
| CN112949207A (en) | Short-term load prediction method based on improved least square support vector machine | |
| Shehu et al. | Flower pollination–feedforward neural network for load flow forecasting in smart distribution grid | |
| CN115099511A (en) | Photovoltaic power probability estimation method and system based on optimized copula | |
| CN115965132A (en) | Distributed photovoltaic digital twin system power prediction method based on GA-BP neural network | |
| Zhang et al. | Transfer learning featured short-term combining forecasting model for residential loads with small sample sets | |
| Tavares et al. | Comparison of PV power generation forecasting in a residential building using ANN and DNN | |
| Massel et al. | The use of machine learning in situational management in relation to the tasks of the power industry | |
| Smyl et al. | ES-dRNN with dynamic attention for short-term load forecasting | |
| CN113570414A (en) | Electricity price prediction method for optimizing deep neural network based on improved Adam algorithm | |
| CN111027760A (en) | Power load prediction method based on least square vector machine |
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 | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20190813 |