CN110443417A - Multiple-model integration load forecasting method based on wavelet transformation - Google Patents
Multiple-model integration load forecasting method based on wavelet transformation Download PDFInfo
- Publication number
- CN110443417A CN110443417A CN201910698138.6A CN201910698138A CN110443417A CN 110443417 A CN110443417 A CN 110443417A CN 201910698138 A CN201910698138 A CN 201910698138A CN 110443417 A CN110443417 A CN 110443417A
- Authority
- CN
- China
- Prior art keywords
- prediction
- load
- model
- wavelet
- kinds
- 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
- 230000009466 transformation Effects 0.000 title claims abstract description 18
- 230000010354 integration Effects 0.000 title claims abstract description 12
- 238000013277 forecasting method Methods 0.000 title claims abstract description 9
- 238000000034 method Methods 0.000 claims abstract description 28
- 238000012549 training Methods 0.000 claims description 24
- 230000006870 function Effects 0.000 claims description 17
- 238000012360 testing method Methods 0.000 claims description 15
- 238000000354 decomposition reaction Methods 0.000 claims description 14
- 230000000306 recurrent effect Effects 0.000 claims description 5
- 238000003062 neural network model Methods 0.000 claims description 4
- 238000013480 data collection Methods 0.000 claims description 3
- 230000007787 long-term memory Effects 0.000 claims description 2
- 230000002596 correlated effect Effects 0.000 abstract description 2
- 238000000844 transformation Methods 0.000 abstract 1
- 230000000875 corresponding effect Effects 0.000 description 15
- 240000002853 Nelumbo nucifera Species 0.000 description 5
- 235000006508 Nelumbo nucifera Nutrition 0.000 description 5
- 235000006510 Nelumbo pentapetala Nutrition 0.000 description 5
- 230000008569 process Effects 0.000 description 5
- 238000013459 approach Methods 0.000 description 3
- 230000001419 dependent effect Effects 0.000 description 3
- 230000005611 electricity Effects 0.000 description 3
- 230000007774 longterm Effects 0.000 description 3
- 229920001228 polyisocyanate Polymers 0.000 description 3
- 239000005056 polyisocyanate Substances 0.000 description 3
- 238000013528 artificial neural network Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000011156 evaluation Methods 0.000 description 2
- 244000131316 Panax pseudoginseng Species 0.000 description 1
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 description 1
- 235000003140 Panax quinquefolius Nutrition 0.000 description 1
- 230000009471 action Effects 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 230000000295 complement effect Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 235000008434 ginseng Nutrition 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000012423 maintenance Methods 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 210000002569 neuron Anatomy 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000012706 support-vector machine Methods 0.000 description 1
- 238000013519 translation Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/14—Fourier, Walsh or analogous domain transformations, e.g. Laplace, Hilbert, Karhunen-Loeve, transforms
-
- 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—Systems or methods specially adapted for specific business sectors, e.g. utilities or tourism
- G06Q50/06—Electricity, gas or water supply
Abstract
The present invention relates to a kind of multiple-model integration load forecasting method based on wavelet transformation, this method is divided into four-stage: 1, on the basis of considering multiple influence factor, historical load related data being obtained the high feature Candidate Set of correlation through maximum information coefficient characteristics selection technique.2, to obtain stable load sequence, precision of prediction is improved, a variety of wavelet transformations are integrated into more prediction models.3, each load correlated series after wavelet function decomposes are trained by an intelligent predicting submodel, these submodels provide different predictions within same hour.4, to combine the optimum prediction of each period and final prediction result is provided using online secondary study during Integrated Decision.This method can be on the basis of various Individual forecast models, further increasing productivity precision of prediction.The method generalization ability is strong, adapts to have stronger application in a variety of environment, helps to reduce Operation of Electric Systems cost.
Description
Technical field
The present invention relates to a kind of load prediction technology, in particular to a kind of multiple-model integration load based on wavelet transformation is pre-
Survey method.
Background technique
Load prediction is most important for ENERGY PLANNING and safety, can carry out including electric system tune based on load prediction
A series of Operation of Electric Systems operations such as degree, planning, bidding price adjustment and maintenance.High-precision load prediction is to improve generating equipment
The important guarantee of utilization rate and economic load dispatching validity.Current artificial intelligence technology plays increasingly heavier in load prediction field
The effect wanted.Artificial neural network, support vector machines, deep learning etc. are the popular techniques in short-term load forecasting field, are all existed
The precision of load prediction is improved to a certain extent.But the above single model has its specific application range, and generalization ability has
Limit, thereby increases and it is possible to which there are overfitting problems.However a variety of model prediction results can be combined by the integrated study of model, be promoted
The stability and generalization ability of model, compared with single model have more accurate precision of prediction.
In model prediction, a variety of identical single models have different precision of predictions to different forecast samples, no
Same single model also has different precision of predictions to identical forecast sample.In order to reduce the prediction error of single model,
In there are two main classes in load prediction field integrated approach, be respectively using different sampled data sets, different model parameters are set and
Multiple similar model integrateds of building and the heterogeneous model integration for being related to many algorithms are predicted.First kind integrated approach can solve
Single model over-fitting and the limited problem of generalization ability.But due to the limitation of algorithm itself, this method not can solve
Single algorithm application limitation problem.Another kind is integrated to be made of a variety of isomery models, and the complementary advantage of more prediction models overcomes
Single algorithm applies limited disadvantage, and generalization ability with higher.
In the Predicting Technique of short term, data set features analysis selection and prediction model building are currently to study
Two priority research areas.In terms of data set features analyze selection, from traditional regularity for studying load data itself,
The inherent mechanism changed to analysis load has significant progress.Nowadays with the widely available and Internet of Things of intelligence instrument
The development of sensing technology can measure and collect more comprehensive and accurate fine granularity load, Correlative Influence Factors and change outside
Amount relies on data.Consider that a variety of multi-source heterogeneous influence factors are to select the important foundation of best features collection.Wherein meteorologic factor and
Date type is closely related with load variations, and meteorologic factor mainly includes real time temperature, relative humidity, wind speed, rainfall etc..Day
Phase type mainly has working day and day off, great festivals or holidays, time and month etc..Influence of the different factors to load is different
, excessive approximate redundancy feature, which will cause, to be extended and cause prediction effect bad cycle of training.Therefore influence factor is divided
Analysis selection is necessary.The main method of Feature Selection has correlation coefficient process (CA), mutual information (MI) and conditional mutual information (CMI)
Technology etc..Correlation coefficient process mainly includes two kinds of line style relevant function methods of Pearson and Spearman, when two characteristic variable lines
When property correlation, these methods are effective.However, the characteristic variable for influencing load is often nonlinear, it is relatively wet such as temperature
Degree etc..Mutual information is the correlation that the concept based on entropy evaluates qualitative independent variable with qualitative dependent variable, but simultaneously inconvenient directly use
In feature selecting.And maximum information coefficient is that improvement is made that on the basis of mutual information, is more suitable for the negative of non-linear variable
Lotus prediction.
Wavelet transformation is that the sequence that will generate electricity is decomposed into one group of this structure component, becomes stationary sequence, is preferably suitable for
Load prediction.
In integrated model prediction, integrated prediction should be the strategy combination individually predicted, main method has simple average
Secondary study of method, weighting method, linear model, nonlinear model etc..Simple average method be mean value taken to multiple predictions output, but
During submodel prediction output, the prediction output of certain individuals can be more accurate than other individuals, and it is high to weaken precision of prediction
Submodel predictive ability.Due to individually predicting submodel for solving same load forecasting problem, their prediction is exported
It is highly relevant.In this case, line style model decision is likely to result in synteny problem, generates inaccurate submodel weight.
Summary of the invention
It is low and the problem of be of limited application the present invention be directed to single algorithm precision of prediction, it proposes a kind of based on small echo
The multiple-model integration load forecasting method of transformation makes load become stationary components, and utilize polyisocyanate structure based on wavelet transformation
Sub- prediction model further increases the short-term load forecasting precision of model.
The technical solution of the present invention is as follows: a kind of multiple-model integration load forecasting method based on wavelet transformation, specifically includes
Following steps:
1) by the electric load demand data d of equal length L, the corresponding Meteorological Characteristics w for considering meteorological index
With date type data r composition data collection D;Then by maximum information coefficient characteristics selection technique based on mutual information to data set
D carries out feature selecting, gets rid of the small attributive character of correlation in initial data, obtains best features collection X;
2) on the basis of the feature set X selected, feature set X is divided into training set X1, test set X2With forecast set X3, selection
Three kinds carry out wavelet decomposition to feature set X respectively to the good morther wavelet of load decomposition;
3) every kind of morther wavelet obtains training set X after decomposing1, test verifying collection X2With forecast set X3Corresponding Wavelet Component, will
Wavelet Component is sent into the corresponding prediction submodel of every kind of morther wavelet and is trained, verifies and predicts, adjusting parameter, after being trained
Three kinds of prediction submodels;
4) prediction result generated using prediction submodel after three kinds of trainingAgain with before predicting
Actual load y most recentlylastGenerate a new training set Dd;According to objective function, by prediction result and corresponding actual negative
Lotus carries out difference calculating, actual load nearest before prediction and the actual load at next moment is carried out difference calculating, by poor
It is worth and assigns different weights to the individual features in test set, determines all kinds of prediction models and the most recently weight of load;In conjunction with life
At weight and DdRegenerate weight training collection Ddω, by DdωIt is sent into decision model, is trained and tests with objective function minimum
Card, adjusting parameter obtains final prediction model, this learns for Two-level ensemble, final prediction model can directly with being predicted,
The decision model is the optimum prediction model selected in three kinds of prediction submodels.
The morther wavelet that the step 2) selects is the corresponding three kinds of prediction submodels point of db4, coif4, sym4, three kinds of morther wavelets
It Wei not Least square support vector regression model, shot and long term memory Recognition with Recurrent Neural Network model and limit gradient boosted tree recurrence mould
Type.
The beneficial effects of the present invention are: the present invention is based on the multiple-model integration load forecasting methods of wavelet transformation, compare
In existing method, non-stationary load correlated series are converted to stable this structure component letter by wavelet transformation first by this method
Number, it is secondary finally by the test set of Weight then by being predicted in load prediction field prediction model of good performance
The integrated of a variety of prediction results is realized in study.Using this method, the load prediction under a variety of different scenes can adapt to, and in list
Load prediction precision is further increased on the basis of one prediction model, and then reduces Operation of Electric Systems cost under smart grid, tool
There are important realistic meaning and good application prospect.
Detailed description of the invention
Fig. 1 is that the present invention is based on the flow diagrams of the multiple-model integration load forecasting method of wavelet transformation.
Specific embodiment
The present invention is proposed by wavelet transformation that load is related on the basis of existing wavelet transformation, Individual forecast model
Data are decomposed into after stationary sequence by polyisocyanate structure model prediction, polyisocyanate structure model include Least square support vector regression model,
Shot and long term remembers Recognition with Recurrent Neural Network model and limit gradient boosted tree regression model.Machine finally best by above-mentioned performance capabilities
Device model carries out secondary study, using the prediction integrated approach of proposition, determines the prediction output weight of different prediction models and incites somebody to action
The training set of Weight is as test input, so that the objective function loss reduction of final integrated predictive model.
Multiple-model integration prediction technique proposed by the present invention based on wavelet transformation, flow diagram is as shown in Figure 1, include
Following steps:
(1) first stage: by the electric load demand data d of equal length L, corresponding consider meteorological index
Meteorological Characteristics w and date type data r composition data collection D, D=[d1, w1, r1, d2, w2, r2... dL, wL, rL].By data set D
It is pressed through maximum information coefficient characteristics selection technique based on mutual information and is greater than 0.4 progress feature selecting, got rid of in initial data
The small attributive character of correlation obtains best features collection X;
If mutual information between two characteristic variables is bigger, then correlation is stronger.On the contrary, then correlation is weaker.
If Y, Z are stochastic variable, Y refers to that Power system load data, Z refer to Meteorological Characteristics corresponding with load data Y
Either date type characteristic, then mutual information is defined as:
Wherein, p (z, y) is the joint probability density function of two variables, and p (z) and p (y) are respectively the edge of two variables
Density function.What y was represented is each Power system load data in data set, and z refers to each gas corresponding with load data y
As feature either date type characteristic, mutual information is bigger between two variables, then correlation is stronger.And MIC overcomes
Mutual information calculates continuous variable inconvenient disadvantage, and extensive relationship can be captured when possessing enough statistical samples, more can
Embody the correlation degree between attribute.
MIC calculating is broadly divided into three steps:
1) m column n row gridding is carried out to two-dimensional space stochastic variable Z, Y, there is power loads on entire two-dimensional space
Then entire two-dimensional plane is divided into m column n row by lotus data and corresponding Meteorological Characteristics (or date type feature), generate m*n
A grid, thenMaximum mutual information value is found out according to mutual information formula;
2) maximum mutual information value is normalized, association relationship is transformed into (0,1) section;
3) select the maximum value of mutual information under different mesh scales as final MIC (maximum information coefficient) value.MIC's is whole
Body evaluation formula are as follows:
In formula: | Z | * | Y | < B indicates grid dividing sum constraint condition, and (B is 0.6 time of total amount of data to generally less than B
Side).Different mesh scales are to give a variety of (m, n) values to carry out grid dividing.MIC is a kind of normalized maximum mutual trust
Breath has accuracy more higher than mutual information.MIC value is bigger between two variables, then its correlation is stronger;On the contrary, then related
Property is weaker.
It selects the attribute strong with historical load correlation according to maximum information coefficient and historical load feature is further sieved
Choosing generates best features collection X, in which:
X=[W, CIHB, THI, T, h, lh-1,lh-2,lh-H,lh-H-1,lh-2H,lh-2H-1,lh-7H]
Wherein, W is day type belonging to load to be predicted, and definition W=1 is working day, and W=0 is nonworkdays, and CIHB is indicated
Body Comfort Index, THI are comfort index, and T represents temperature, and H indicates that the when number of segment for including in one day, h indicate electricity to be predicted
Period where power load, h=1,2 ..., H, lhIndicate the electricity needs load value of h period;
By before the best features collection X after being selected by maximum information coefficient characteristics 80% as training set X1, after
10% is used as test set X2, last 10% is used as forecast set X3.For simple declaration, historical load, meteorologic factor and date type are still
It is indicated with d, w, r:
X1=[d1,w1,r1,...,dT1,wT1,rT1], X2=[dT1+1,wT1+1,rT1+1,...,dT1+T2,wT1+T2,rT1+T2],
X3=[dT1+T2+1,wT1+T2+1,rT1+T2+1,...,dL,wL,rL]
Wherein, training dataset X1Length be T1 (preceding 80%) of L, test data set X2Length be T2 (for L-T1-
T3), data set X3Length be L-T1-T2 (rear 10%) of L.
(2) second stage: on the basis of feature set X elected, three kinds of selection is to the good morther wavelet of load decomposition
Db4, coif4, sym4 carry out wavelet decomposition (WP).
It, can be by wavelet transform (DWT) to above-mentioned female wave signal (i.e. morther wavelet because load sequence is discrete series
Db4, coif4, sym4) discretization Pan and Zoom obtain, DWT formula are as follows:
Wherein, the load sequence signature of x (t) expression input, the female wave signal of ψ (t) expression, 2m、n2mRespectively scale factor
And translation parameters, L are discrete point number (the length L that the number of discrete point is exactly entire data set).By load sequence signal into
Row two-stage wavelet decomposition:
L (t)=A1(t)+D1(t)=A2(t)+D2(t)+D1(t)
L (t) is the load data value that whole length is L, is decomposed, is first split into low frequency A1For with high frequency D1Signal.So
Afterwards, low frequency A1It is further divided into two components: A2And D2.Low-frequency approximation component A2General trend is reflected, load light is presented
Sliding form.D1And D2Describe the high fdrequency component in load.L1, L2, L3 respectively represent the A after decomposing2, D1, D2Sequence length.
(3) phase III: one group of Wavelet Component A is obtained after the decomposition of each morther wavelet2, D1, D2, every group of small wavelength-division
Amount is all sent into a kind of corresponding prediction submodel of morther wavelet and is trained, tests and predicts.The corresponding three kinds of isomeries of three kinds of morther wavelets
Prediction model, respectively Least square support vector regression (LSSVR) model m1,d, shot and long term remember Recognition with Recurrent Neural Network
(LSTM) model m2,dWith limit gradient boosted tree (Xgboost) regression model m3,d, d=1,2,3.
The detailed process of training three classes isomery prediction model are as follows:
(3-1) forms the independent variable X of 9 kinds of prediction models after three classes wavelet decompositionmi,jInput and dependent variable yi,jOutput:
Xmi,j=[W, CIHB, THI, T, h, lh-1,lh-2,lh-H,lh-H-1,lh-2H,lh-2H-1,lh-7H], yi,j=lh
Wherein, i=1,2,3 indicate different types of prediction model;J=1,2,3 indicate the A2, D1, D2 after wavelet transformation
Component.
(3-2) is in data set X1, X2Established 9 prediction submodels on the basis of wavelet decomposition, i.e., least square support to
Measure regression model m1,d(Ф1,j, Xm1,j, β1,j), shot and long term remember Recognition with Recurrent Neural Network model m2,d(Ф2,j, Xm2,j, β2,j) and pole
It limits gradient and promotes regression model m3,d(Ф3,j, Xm3,j, β3,j), j=1,2,3;Wherein Ф1,j, Ф2,j, Ф3,jFor the super ginseng of model
Number, respectively indicates kernel function type, the hidden layer number of plies and neuron number, CART tree number.β1,j, β2,j, β3,jFor prediction model
Corresponding model parameter.
(3-3) model parameter adjustment: with objective function minimum adjusting parameter.Hyper parameter is adjusted first, after being adjusted
Then hyper parameter trains prediction submodel by obtained hyper parameter, repeat this step 9 times, obtain 9 prediction submodels.
(3-4) is by data set X2Input after decomposing as (3-3) training prediction submodel, the output result that will be obtained
Error with actual value is as prediction submodel evaluation index.In test process, corresponding minimum target function, such as target are calculated
Function error is larger, repeats step (3-3) and continues adjusting parameter, until the objective function of test minimizes.Select target
9 prediction submodels of the smallest 3 class of function.
The objective function of training learning process after wavelet decomposition are as follows:
Wherein, A2t, D1t, D2tFor the load sequence inputting after second level wavelet decomposition, y1t、y2t、y3tIt is passed through for three kinds of morther wavelets
In the electric load actual value of t moment after decomposition.For prediction model t moment load forecast
Value.The parameter that prediction model is obtained by training is β1,1, β1,2, β1,3.It can similarly obtain the parameter difference of other prediction regression models
For β2,1, β2,2, β2,3And β3,1, β3,2, β3,3。
(3-5) is by data set X3It is input to after wavelet decomposition in well-drilled prediction submodel, and passes through small echo weight
Structure obtains prediction output of the three classes prediction submodel in one period t of future
(4) fourth stage: the prediction result for predicting that submodel generates using three kindsBefore prediction
Actual load y most recentlylastGenerate a new training set Dd.According to objective function, by prediction result and corresponding actual negative
Lotus carries out difference calculating, actual load nearest before prediction and the actual load at next moment is carried out difference calculating, by poor
Be worth and assign different weights to the individual features in test set, determine all kinds of prediction models and most recently the weights omega of load=
[ω1,ω2,ω3,ω4].In conjunction with the weight and D of generationdRegenerate weight training collection Ddω, by DdωIt is sent into decision model, with
Objective function minimum is trained and verifies, adjusting parameter, obtains final prediction model, this is Two-level ensemble study, described to determine
Plan model is the optimum prediction model selected in three kinds of prediction submodels.Final prediction model can be directly with being predicted.Tool
Body realize the following steps are included:
(4-1) obtains Electric Load Forecasting measured value using sub- prediction modelWith corresponding power load
Lotus actual value y determines weights omega=[ω of each prediction model1,ω2,ω3] it is as follows:
Wherein,Indicate load forecast weighted average of 3 kinds of prediction models in period t, ωiIndicate integrated
I-th class prediction model m in the sub- prediction model of loadiWeight.
(4-2) is attached to prediction result, actual load most recently and the phase of weight factor according to above-mentioned three classes prediction model
Influence factor is closed as training input it is found that one independent variable X of input1t, export a dependent variable y1t, in which:
y1t=[lh]。
(4-3) shows good model m by above-mentioned trainingi,dAs the prediction model of online secondary study, input respectively certainly
Variable X1t, wherein willAccording to weights omega=[ω1,ω2,ω3,ω4] it is used as Second Decision model
mi,dTraining input, to decision model mi,dIt is trained, verifies, adjusting parameter, decision model m after trainingi,dIt can be calculated
The corresponding final Electric Load Forecasting measured value in time t
Claims (2)
1. a kind of multiple-model integration load forecasting method based on wavelet transformation, which is characterized in that specifically comprise the following steps:
1) by the electric load demand data d of equal length L, the corresponding Meteorological Characteristics w for considering meteorological index and day
Phase categorical data r composition data collection D;Then by maximum information coefficient characteristics selection technique based on mutual information to data set D into
Row feature selecting gets rid of the small attributive character of correlation in initial data, obtains best features collection X;
2) on the basis of the feature set X selected, feature set X is divided into training set X1, test set X2With forecast set X3, select three kinds
Wavelet decomposition is carried out to feature set X respectively to the good morther wavelet of load decomposition;
3) every kind of morther wavelet obtains training set X after decomposing1, test verifying collection X2With forecast set X3Corresponding Wavelet Component, by small echo
Component is sent into the corresponding prediction submodel of every kind of morther wavelet and is trained, verifies and predicts, adjusting parameter, three kinds after being trained
Predict submodel;
4) prediction result generated using prediction submodel after three kinds of trainingAgain with nearest before prediction
The actual load y of phaselastGenerate a new training set Dd;According to objective function, by prediction result and corresponding actual load into
Row difference calculates, and actual load nearest before prediction and the actual load at next moment is carried out difference calculating, by difference pair
Individual features in test set assign different weights, determine all kinds of prediction models and the most recently weight of load;In conjunction with generation
Weight and DdRegenerate weight training collection Ddω, by DdωIt is sent into decision model, is trained and verifies with objective function minimum,
Adjusting parameter obtains final prediction model, this is Two-level ensemble study, and final prediction model can be directly with being predicted, institute
Stating decision model is the optimum prediction model selected in three kinds of prediction submodels.
2. the multiple-model integration load forecasting method based on wavelet transformation according to claim 1, which is characterized in that step 2)
The morther wavelet selected for the corresponding three kinds of prediction submodels of db4, coif4, sym4, three kinds of morther wavelets be respectively least square support to
Measure regression model, shot and long term memory Recognition with Recurrent Neural Network model and limit gradient boosted tree regression model.
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN2019106039154 | 2019-07-05 | ||
CN201910603915 | 2019-07-05 |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110443417A true CN110443417A (en) | 2019-11-12 |
Family
ID=68432486
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910698138.6A Pending CN110443417A (en) | 2019-07-05 | 2019-07-31 | Multiple-model integration load forecasting method based on wavelet transformation |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110443417A (en) |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110990784A (en) * | 2019-11-19 | 2020-04-10 | 湖北中烟工业有限责任公司 | Cigarette ventilation rate prediction method based on gradient lifting regression tree |
CN111311436A (en) * | 2020-03-17 | 2020-06-19 | 上海积成能源科技有限公司 | System and method for analyzing and managing comprehensive energy consumption of complex renewable energy sources based on multivariable control |
CN112183848A (en) * | 2020-09-25 | 2021-01-05 | 合肥工业大学 | Power load probability prediction method based on DWT-SVQR integration |
CN112200391A (en) * | 2020-11-17 | 2021-01-08 | 国网陕西省电力公司经济技术研究院 | Power distribution network edge side load prediction method based on k-nearest neighbor mutual information characteristic simplification |
CN112348168A (en) * | 2020-10-27 | 2021-02-09 | 国网四川省电力公司经济技术研究院 | Ultra-short-term load prediction method and system considering data loss and characteristic redundancy |
CN112669173A (en) * | 2021-01-07 | 2021-04-16 | 云南电网有限责任公司电力科学研究院 | Short-term load prediction method based on multi-granularity features and XGboost model |
CN113408815A (en) * | 2021-07-02 | 2021-09-17 | 湘潭大学 | Deep learning-based traction load ultra-short-term prediction method |
CN113505923A (en) * | 2021-07-08 | 2021-10-15 | 广西大学 | Regional power grid short-term load prediction method and system |
CN114925940A (en) * | 2022-07-20 | 2022-08-19 | 广东电网有限责任公司佛山供电局 | Holiday load prediction method and system based on load decomposition |
-
2019
- 2019-07-31 CN CN201910698138.6A patent/CN110443417A/en active Pending
Cited By (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110990784A (en) * | 2019-11-19 | 2020-04-10 | 湖北中烟工业有限责任公司 | Cigarette ventilation rate prediction method based on gradient lifting regression tree |
CN110990784B (en) * | 2019-11-19 | 2024-01-26 | 湖北中烟工业有限责任公司 | Cigarette ventilation rate prediction method based on gradient lifting regression tree |
CN111311436A (en) * | 2020-03-17 | 2020-06-19 | 上海积成能源科技有限公司 | System and method for analyzing and managing comprehensive energy consumption of complex renewable energy sources based on multivariable control |
CN112183848B (en) * | 2020-09-25 | 2022-08-30 | 合肥工业大学 | Power load probability prediction method based on DWT-SVQR integration |
CN112183848A (en) * | 2020-09-25 | 2021-01-05 | 合肥工业大学 | Power load probability prediction method based on DWT-SVQR integration |
CN112348168A (en) * | 2020-10-27 | 2021-02-09 | 国网四川省电力公司经济技术研究院 | Ultra-short-term load prediction method and system considering data loss and characteristic redundancy |
CN112348168B (en) * | 2020-10-27 | 2023-04-07 | 国网四川省电力公司经济技术研究院 | Ultra-short-term load prediction method and system considering data loss and feature redundancy |
CN112200391A (en) * | 2020-11-17 | 2021-01-08 | 国网陕西省电力公司经济技术研究院 | Power distribution network edge side load prediction method based on k-nearest neighbor mutual information characteristic simplification |
CN112200391B (en) * | 2020-11-17 | 2023-07-04 | 国网陕西省电力公司经济技术研究院 | Power distribution network edge side load prediction method based on k-nearest neighbor mutual information feature simplification |
CN112669173A (en) * | 2021-01-07 | 2021-04-16 | 云南电网有限责任公司电力科学研究院 | Short-term load prediction method based on multi-granularity features and XGboost model |
CN113408815A (en) * | 2021-07-02 | 2021-09-17 | 湘潭大学 | Deep learning-based traction load ultra-short-term prediction method |
CN113505923A (en) * | 2021-07-08 | 2021-10-15 | 广西大学 | Regional power grid short-term load prediction method and system |
CN114925940A (en) * | 2022-07-20 | 2022-08-19 | 广东电网有限责任公司佛山供电局 | Holiday load prediction method and system based on load decomposition |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110443417A (en) | Multiple-model integration load forecasting method based on wavelet transformation | |
CN109754113B (en) | Load prediction method based on dynamic time warping and long-and-short time memory | |
Wang et al. | Forecasting stochastic neural network based on financial empirical mode decomposition | |
CN109063911B (en) | Load aggregation grouping prediction method based on gated cycle unit network | |
Shao et al. | Nickel price forecast based on the LSTM neural network optimized by the improved PSO algorithm | |
CN108171379B (en) | Power load prediction method | |
CN109143408B (en) | Dynamic region combined short-time rainfall forecasting method based on MLP | |
CN106022549A (en) | Short term load predication method based on neural network and thinking evolutionary search | |
CN111027775A (en) | Step hydropower station generating capacity prediction method based on long-term and short-term memory network | |
Lin et al. | Temporal convolutional attention neural networks for time series forecasting | |
CN111260116A (en) | Time-interval refined short-term load prediction method based on BOA-SVR and fuzzy clustering | |
Liu et al. | Heating load forecasting for combined heat and power plants via strand-based LSTM | |
CN109919421A (en) | Short-term power load prediction model establishment method based on VMD-PSO-BPNN | |
CN113554466A (en) | Short-term power consumption prediction model construction method, prediction method and device | |
Akpinar et al. | Forecasting natural gas consumption with hybrid neural networks—Artificial bee colony | |
CN111311001B (en) | Bi-LSTM network short-term load prediction method based on DBSCAN algorithm and feature selection | |
CN115049113A (en) | S2S power load prediction system based on time window and super-parameter adaptive selection | |
Widodo et al. | Model selection using dimensionality reduction of time series characteristics | |
CN113111592A (en) | Short-term wind power prediction method based on EMD-LSTM | |
Gu et al. | Fuzzy time series forecasting based on information granule and neural network | |
CN116167465A (en) | Solar irradiance prediction method based on multivariate time series ensemble learning | |
CN116561569A (en) | Industrial power load identification method based on EO feature selection and AdaBoost algorithm | |
CN117114160A (en) | Short-term photovoltaic power prediction method | |
Irankhah et al. | A parallel CNN-BiGRU network for short-term load forecasting in demand-side management | |
CN115759343A (en) | E-LSTM-based user electric quantity prediction method and device |
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 | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20191112 |