CN105956722A - Short-term wind power prediction method and apparatus - Google Patents

Abstract

The invention provides a short-term wind power prediction method and apparatus. Through obtained sample data, a correlation vector machine prediction model is established, parameters in the correlation vector machine prediction model, obtained according to the correlation vector machine prediction model are optimized, an optimal correlation vector machine prediction model is obtained, prediction errors are reduced from the perspective of prediction model optimization, then according to actual prediction values obtained through the optimal correlation vector machine prediction model and the sample data, a first relative error sequence is calculated, by use of the first relative error sequence, an established GARCH error prediction model is optimized, an optimal GARCH error prediction model is obtained, reduction of the prediction errors from the perspective of the prediction model optimization is realized, afterwards, by use of prediction values obtained after the first relative error sequence is predicted through the optimal GARCH error prediction model, the actual prediction values are corrected, accordingly, the prediction errors are reduced from the perspective of error correction, and the prediction precision is improved.

Description

The Forecasting Methodology of a kind of short-term wind-electricity power and device

Technical field

The present invention relates to technical field of wind power, in particular, relate to a kind of short-term wind-electricity power Forecasting Methodology and device.

Background technology

Wind-power electricity generation, with abundant, the cleanliness without any pollution of himself resource, the advantage such as renewable, becomes electric power Industry substitutes one of new forms of energy of existing chemical energy source.But, due to the random wave that wind-power electricity generation is intrinsic Dynamic property can affect the stable operation of power system, it is therefore desirable to predicts wind power accordingly, from And preferably management and use wind-powered electricity generation.And short-term wind-electricity power prediction is usually following 24 hours-72 hours The active power of blower fan or wind energy turbine set is predicted, but the prediction data a few days ago that in reality, wind energy turbine set reports is by mistake Difference is relatively big, causes the difficulty of electric power system dispatching plan to increase, and then reduces the peace that system is run Full property and economy.

At present, domestic focus primarily upon by improve prediction algorithm reduce forecast error, wherein, support to The Forecasting Methodology of amount machine realizes predicting the most accurately because having the less training sample of utilization, and effectively keeps away Exempt to be absorbed in the advantage such as danger of Local Minimum so that the research in wind power prediction field is with application gradually Increase, and obtain good result.But, in Practical Project uses, support vector machine there is also some not In place of foot, such as the Mercer condition that must is fulfilled for of choosing of kernel function, support that the number of vector is along with training sample The increase of number and linear increase, the choosing of insensitive parameter does not also have a kind of to generally acknowledge unified best method, Cause the meaningless increase of amount of calculation and parameter amount, need further to study and improve, and present stage is the most not Expertly after forecast error Producing reason with prediction, forecast error can be reduced in terms of error correction two, And then raising precision of prediction.

Summary of the invention

In view of this, the invention provides Forecasting Methodology and the device of a kind of short-term wind-electricity power, from optimization Forecast model and error correction two aspect reduce forecast error, improve precision of prediction.

For achieving the above object, the present invention provides following technical scheme:

A kind of Forecasting Methodology of short-term wind-electricity power, including:

Obtaining real data according to preset rules, described real data comprises the power data of wind energy turbine set and right The air speed data answered；

Revise described real data, it is thus achieved that sample data；

Described sample data is carried out the first pretreatment, and generates the first input sample data and the first output Sample data, described first input sample data comprises the power data of wind energy turbine set and the air speed data of correspondence, Described first output sample data comprises the power data of wind energy turbine set；

Calculate the kernel function of Method Using Relevance Vector Machine, and by described kernel function, sample data, the first input sample Data based on data and the first output sample data, set up Method Using Relevance Vector Machine forecast model；

According to described Method Using Relevance Vector Machine forecast model, set up Method Using Relevance Vector Machine training pattern；

Optimize the parameter of Method Using Relevance Vector Machine in described Method Using Relevance Vector Machine training pattern, it is thus achieved that corresponding optimize after The Method Using Relevance Vector Machine training pattern of parameter of described Method Using Relevance Vector Machine, as optimum Method Using Relevance Vector Machine training Model；

Described optimum Method Using Relevance Vector Machine training pattern is carried out the second pretreatment, it is thus achieved that actual prediction value；

Utilize described sample data and actual prediction value, be calculated the first relative error sequence；

Set up GARCH error prediction model；

Utilize GARCH error prediction model described in described first relative error sequence pair to carry out the 3rd to locate in advance Reason, it is thus achieved that optimum GARCH error prediction model；

Utilize described optimum GARCH error prediction model that described first relative error sequence is predicted, Obtain the predictive value of the second relative error sequence；

Utilize actual prediction value described in the predictive value correction of described second relative error sequence, it is thus achieved that the most pre- Measured value.

Preferably, described according to preset rules obtain real data, including:

Obtain the initial real data in time limit stipulated time；

Described time limit stipulated time is divided into n stipulated time section；

M time interval of described initial real data is obtained in setting arbitrary described stipulated time section；

Through arbitrary described time interval, from described initial real data, obtain a data value, As the data value of an acquisition point in described stipulated time section arbitrary in described real data, described in obtain Take a little with described time interval one_to_one corresponding；

Wherein, n Yu m is positive integer.

Preferably, the described real data of described correction, it is thus achieved that sample data, including:

Determine the problem data in arbitrary described stipulated time section in described real data, and described problem Data acquisition point position in arbitrary described stipulated time section, described problem data is missing data or different Regular data；

Described problem data replaces with correction data successively, and described correction data are for being positioned at described problem number According to the data value on the previous acquisition point position of the acquisition point position in arbitrary described stipulated time section；

Obtain revised data, as sample data.

Preferably, the kernel function of described calculating Method Using Relevance Vector Machine, and by described kernel function, sample data, Data based on first input sample data and the first output sample data, set up Method Using Relevance Vector Machine prediction Model, including:

By described sample data and the first input sample of data substitution kernel function formula:

$K(x_i,x_j)=\exp \left(-\frac{\left|\right|x_i-x_j\left|\right|}{{\mathrm{\σ}}^2}\right)$

Calculate the kernel function of Method Using Relevance Vector Machine, wherein x_iFor the input vector of described sample data, x_jFor The input vector of described first input sample data, σ is the kernel function width of Method Using Relevance Vector Machine；

Setting up described Method Using Relevance Vector Machine forecast model, described Method Using Relevance Vector Machine forecast model comprises described relevant The kernel function of vector machine, described sample data, described first input sample data and the first output sample number According to.

Preferably, described set up Method Using Relevance Vector Machine training pattern according to described Method Using Relevance Vector Machine forecast model, Including:

By described first input sample data and the first output sample data classification, generate the first training sample Data and the first test samples data；

Utilize described first training sample data, described Method Using Relevance Vector Machine forecast model is trained, builds Vertical Method Using Relevance Vector Machine training pattern；

Utilize described first test samples data, described Method Using Relevance Vector Machine training pattern is verified, really Fixed described Method Using Relevance Vector Machine training pattern.

Preferably, the parameter of the Method Using Relevance Vector Machine in described optimization described Method Using Relevance Vector Machine training pattern, obtains Obtain optimum Method Using Relevance Vector Machine training pattern, including:

Optimize the kernel function width cs of Method Using Relevance Vector Machine in described Method Using Relevance Vector Machine training pattern, it is thus achieved that The kernel function width cs of excellent Method Using Relevance Vector Machine；

Obtain the Method Using Relevance Vector Machine training mould corresponding with the kernel function width cs of described optimum Method Using Relevance Vector Machine Type, as optimum Method Using Relevance Vector Machine training pattern.

Preferably, the kernel function width of the Method Using Relevance Vector Machine in described optimization described Method Using Relevance Vector Machine training pattern Degree σ, it is thus achieved that the kernel function width cs of optimum Method Using Relevance Vector Machine, including:

The kernel function width cs needing the described Method Using Relevance Vector Machine optimized is mapped as a voxel vector；

Utilize the individual voxel vector being mapped to, set up and initialize population；

Obtain maximum iteration time, TSP question factor F initial value and the adaptive crossover mutation CR set Initial value；

Utilize described TSP question factor F, described initialization population and current individual vector are carried out adaptive Answer mutation operation, it is thus achieved that variation vector；

Utilize adaptive crossover mutation CR, described variation vector current individual vector is carried out self adaptation friendship Fork selects, it is thus achieved that intersection vector；

Described intersection vector and described current individual vector are substituted into fitness function respectively, it is thus achieved that intersect and vow The fitness value of amount and the fitness value of current individual vector；

Compare the size of the fitness value of described intersection vector and the fitness value of described current individual vector, Obtain the vector that fitness value is less；

Select the vector that fitness value is less as the current individual vector of next iteration, return and perform profit By described TSP question factor F, described initialization population and current individual vector are carried out TSP question Operation, it is thus achieved that variation vector, during until arriving maximum iteration time, the minimum fitness value of output；

Obtain the kernel function width cs of the Method Using Relevance Vector Machine corresponding with described minimum fitness value, as optimum The kernel function width cs of Method Using Relevance Vector Machine.

Preferably, described set up GARCH error prediction model, including:

Set up initial ARMA model；

Set up initial GARCH model；

Described initial ARMA model is fitted, determines the exponent number of final arma modeling；

Described initial GARCH model is fitted, determines the exponent number of final GARCH model；

Utilize described final arma modeling and final GARCH model to described first relative error sequence Being fitted, set up GARCH error prediction model, wherein, described final arma modeling corresponds to The exponent number of the described final arma modeling determined, described final GARCH model is corresponding to the institute determined State the exponent number of final GARCH model.

Preferably, described GARCH error prediction model described in described first relative error sequence pair is utilized Carry out the 3rd pretreatment, it is thus achieved that optimum GARCH error prediction model, including:

Described first relative error sequence is classified, generates the second training sample data and the second inspection Sample data；

Utilize described second training sample data, described GARCH error prediction model be fitted, Obtain the GARCH error prediction model optimized；

Utilize described second test samples data, the GARCH error prediction model of described optimization is carried out Checking, it is thus achieved that validation value；

Relatively described validation value and the size of preassigned；

At described validation value less than or equal to described predetermined standard time, it is thus achieved that optimum GARCH error prediction mould Type.

Preferably, after the described comparison described validation value size with preassigned, also include:

At described validation value more than described predetermined standard time, return and described initial ARMA model is intended Close, determine the exponent number of final arma modeling, until described validation value is less than or equal to described preassigned.

Preferably, actual prediction value described in the described predictive value correction utilizing described second relative error sequence, Obtain final predictive value, including:

By the predictive value substitution formula of described second relative error sequence:

$\widehat{RE}={\widehat{RE}}_p\×RP$

It is calculated the residual sequence predictive value of correspondenceWherein,It it is the second relative error sequence Predictive value, RP is the rated power of blower fan；

Described residual sequence predictive value is added with described actual prediction value, it is thus achieved that final predictive value.

A kind of prediction means of short-term wind-electricity power, including:

First acquisition module, for obtaining real data according to preset rules, described real data comprises wind The power data of electric field and the air speed data of correspondence；

First correcting module, is used for revising described real data, it is thus achieved that sample data；

First pretreatment module, for described sample data is carried out the first pretreatment, and it is defeated to generate first Entering sample data and the first output sample data, described first input sample data comprises the power of wind energy turbine set Data and the air speed data of correspondence, described first output sample data comprises the power data of wind energy turbine set；

First model building module, for calculating the kernel function of Method Using Relevance Vector Machine, and by described kernel function, Data based on sample data, the first input sample data and the first output sample data, set up relevant Vector machine forecast model；

Second model building module, for according to described Method Using Relevance Vector Machine forecast model, sets up associated vector Machine training pattern；

First optimizes module, for optimizing the ginseng of the Method Using Relevance Vector Machine in described Method Using Relevance Vector Machine training pattern Number, it is thus achieved that the Method Using Relevance Vector Machine training pattern of the parameter of the described Method Using Relevance Vector Machine after corresponding optimization, as Optimum Method Using Relevance Vector Machine training pattern；

Second pretreatment module, for described optimum Method Using Relevance Vector Machine training pattern is carried out the second pretreatment, Obtain actual prediction value；

First computing module, is used for utilizing described sample data and actual prediction value, is calculated the first phase To error sequence；

3rd model building module, is used for setting up GARCH error prediction model；

3rd pretreatment module, is used for utilizing GARCH error described in described first relative error sequence pair Forecast model carries out the 3rd pretreatment, it is thus achieved that optimum GARCH error prediction model；

First prediction module, is used for utilizing described optimum GARCH error prediction model to described first phase Error sequence is predicted, it is thus achieved that the predictive value of the second relative error sequence；

Second correcting module, actual for utilizing described in the predictive value correction of described second relative error sequence Predictive value, it is thus achieved that final predictive value.

Preferably, described first acquisition module includes:

Second acquisition module, for obtaining the initial real data in time limit stipulated time；

Separating modules, for being divided into n stipulated time section by described time limit stipulated time；

Setting module, obtains the m of described initial real data in setting arbitrary described stipulated time section Individual time interval；

3rd acquisition module, for through arbitrary described time interval, from described initial real data Data value of middle acquisition, as an acquisition in described stipulated time section arbitrary in described real data The data value of point, described acquisition point and described time interval one_to_one corresponding；

Wherein, n Yu m is positive integer.

Preferably, described first correcting module includes:

Determine module, for the problem data determined in described real data in arbitrary described stipulated time section, And the acquisition point position that described problem data is in arbitrary described stipulated time section, described problem data is Missing data or abnormal data；

3rd correcting module, for replacing with correction data, described correction number successively by described problem data According to for being positioned at the previous acquisition in the acquisition point position of arbitrary described stipulated time section of the described problem data Data value on some position；

4th acquisition module, is used for obtaining revised data, as sample data.

Preferably, described first model building module includes:

Kernel function computing module, for described sample data and the first input sample data being substituted into formula:

$K(x_i,x_j)=\exp \left(-\frac{\left|\right|x_i-x_j\left|\right|}{{\mathrm{\σ}}^2}\right)$

Calculate the kernel function of Method Using Relevance Vector Machine, wherein x_iFor the input vector of described sample data, x_jFor The input vector of described first input sample data, σ is the kernel function width of Method Using Relevance Vector Machine；

Submodule set up by first model, is used for setting up described Method Using Relevance Vector Machine forecast model, described be correlated with to Amount machine forecast model comprises the kernel function of described Method Using Relevance Vector Machine, described sample data, described first input Sample data and the first output sample data.

Preferably, described second model building module includes:

First sort module, for described first input sample data is classified with the first output sample data, Generate the first training sample data and the first test samples data；

First training module, for described first training sample utilizing described first sort module to sort out Data, are trained described Method Using Relevance Vector Machine forecast model, set up Method Using Relevance Vector Machine training pattern；

First inspection module, for described first test samples utilizing described first sort module to sort out Data, verify described Method Using Relevance Vector Machine training pattern, determine described Method Using Relevance Vector Machine training pattern.

Preferably, described first optimization module includes:

Parameter optimization module, for optimizing the core of the Method Using Relevance Vector Machine in described Method Using Relevance Vector Machine training pattern Function widths σ, it is thus achieved that the kernel function width cs of optimum Method Using Relevance Vector Machine；

5th acquisition module is corresponding with the kernel function width cs of described optimum Method Using Relevance Vector Machine for obtaining Method Using Relevance Vector Machine training pattern, as optimum Method Using Relevance Vector Machine training pattern.

Preferably, described parameter optimization module includes:

Mapping block, individual for the kernel function width cs needing the described Method Using Relevance Vector Machine optimized is mapped as Voxel vector；

Population foundation module, for utilizing the individual voxel vector being mapped to, sets up and initializes population；

6th acquisition module, for obtaining the maximum iteration time of setting, TSP question factor F initial value With adaptive crossover mutation CR initial value；

TSP question module, is used for utilizing described TSP question factor F, to described initialization population and Current individual vector carries out TSP question operation, it is thus achieved that variation vector；

Self adaptation Cross module, is used for utilizing adaptive crossover mutation CR, current to described variation vector Individual voxel vector carries out self adaptation cross selection, it is thus achieved that intersection vector；

Fitness value acquisition module, for substituting into described intersection vector respectively with described current individual vector Fitness function, it is thus achieved that the fitness value of intersection vector and the fitness value of current individual vector；

First comparison module, for fitness value and the described current individual vector of relatively described intersection vector The size of fitness value, it is thus achieved that the vector that fitness value is less；

Judge module, for judging the number of times of current iteration；

TSP question module, for judging the number of times of current iteration less than maximum repeatedly at described judge module During generation number, vector less for the fitness value of described first comparison module acquisition is chosen as current individual Vector, utilizes described TSP question factor F, carries out described initialization population and current individual vector certainly Adequate variation operates, it is thus achieved that variation vector；

Output module, for judging that at described judge module the number of times of current iteration is equal to maximum iteration time Time, the minimum fitness value of output；

7th acquisition module, for obtaining the core letter of the Method Using Relevance Vector Machine corresponding with described minimum fitness value Number width cs, as the kernel function width cs of optimum Method Using Relevance Vector Machine_。

Preferably, described 3rd model building module includes:

Arma modeling sets up module, is used for setting up initial ARMA model；

GARCH model building module, is used for setting up initial GARCH model；

ARMA fitting module, for being fitted described initial ARMA model, determines described The exponent number of arma modeling；

GARCH fitting module, for being fitted described initial GARCH model, determines described The exponent number of GARCH model；

Submodule set up by 3rd model, is used for utilizing described final arma modeling and final GARCH mould Described first relative error sequence is fitted by type, sets up GARCH error prediction model, wherein, Described final arma modeling corresponding to the exponent number of described final arma modeling determined, described finally The exponent number of the GARCH model described final GARCH model corresponding to determining.

Preferably, described 3rd pretreatment module includes:

Second sort module, for described first relative error sequence being classified, generates the second training Sample data and the second test samples data；

Error prediction models fitting module, is used for utilize described second sort module to sort out described second Training sample data, are fitted described GARCH error prediction model, it is thus achieved that the GARCH of optimization Error prediction model；

Second authentication module, for described second test samples utilizing described second sort module to sort out, The GARCH error prediction model of described optimization is verified, it is thus achieved that validation value；

Second comparison module, for the size of relatively described validation value with preassigned；

8th acquisition module, for being less than or equal to described predetermined standard time at described validation value, it is thus achieved that optimum GARCH error prediction model.

Preferably, after the described second more described validation value of the comparison module size with preassigned, Described ARMA fitting module is additionally operable to:

At described validation value more than described predetermined standard time, described initial ARMA model is fitted, Determine the exponent number of described arma modeling.

Preferably, described second correcting module includes:

Second computing module, for the predictive value of described second relative error sequence being substituted into formula:

$\widehat{RE}={\widehat{RE}}_p\×RP$

It is calculated the residual sequence predictive value of correspondenceWherein,It it is the second relative error sequence Predictive value, RP is the rated power of blower fan；

4th correcting module, for being added with described actual prediction value by described residual sequence predictive value, obtains Obtain final predictive value.

Understand via above-mentioned technical scheme, compared with prior art, the invention provides a kind of short-term wind The Forecasting Methodology of electrical power and device, by obtaining and revising described real data, it is thus achieved that sample data, Utilize described sample data to set up Method Using Relevance Vector Machine forecast model, and predict mould according to described Method Using Relevance Vector Machine Type obtains Method Using Relevance Vector Machine training pattern, optimizes the Method Using Relevance Vector Machine in described Method Using Relevance Vector Machine training pattern Parameter obtain optimum Method Using Relevance Vector Machine training pattern, and then reduce pre-in terms of optimal prediction model Survey error, afterwards, described optimum Method Using Relevance Vector Machine training pattern is carried out the second pretreatment, it is thus achieved that actual Predictive value, utilizes described sample data and actual prediction value, calculates the first relative error sequence, and utilize The GARCH error prediction model that described first relative error sequence pair is set up carries out the 3rd pretreatment, it is thus achieved that Optimum GARCH error prediction model, realizes reducing forecast error in terms of optimal prediction model again, Utilize optimum GARCH error prediction model that described first relative error sequence is predicted afterwards, it is thus achieved that The predictive value of the second relative error sequence, in order to described actual prediction value is modified, it is thus achieved that the most pre- Measured value, and then in terms of error correction, reduce forecast error, improve precision of prediction.

Accompanying drawing explanation

In order to be illustrated more clearly that the embodiment of the present invention or technical scheme of the prior art, below will be to reality Execute the required accompanying drawing used in example or description of the prior art to be briefly described, it should be apparent that below, Accompanying drawing in description is only embodiments of the invention, for those of ordinary skill in the art, not On the premise of paying creative work, it is also possible to obtain other accompanying drawing according to the accompanying drawing provided.

The Forecasting Methodology flow chart of a kind of short-term wind-electricity power that Fig. 1 provides for the embodiment of the present invention；

A kind of method flow diagram obtaining real data that Fig. 2 provides for the embodiment of the present invention；

A kind of method flow diagram revising real data that Fig. 3 provides for the embodiment of the present invention；

A kind of method flow diagram setting up Method Using Relevance Vector Machine forecast model that Fig. 4 provides for the embodiment of the present invention；

A kind of method flow diagram setting up Method Using Relevance Vector Machine training pattern that Fig. 5 provides for the embodiment of the present invention；

A kind of associated vector optimized in Method Using Relevance Vector Machine training pattern that Fig. 6 provides for the embodiment of the present invention The method flow diagram of the parameter of machine；

A kind of method stream obtaining optimum GARCH error prediction model that Fig. 7 provides for the embodiment of the present invention Cheng Tu；

A kind of method flow diagram revising actual prediction value that Fig. 8 provides for the embodiment of the present invention；

The structural representation of the prediction means of a kind of short-term wind-electricity power that Fig. 9 provides for the embodiment of the present invention；

A kind of associated vector optimized in Method Using Relevance Vector Machine training pattern that Figure 10 provides for the embodiment of the present invention The structural representation of the parameter of machine.

Detailed description of the invention

Below in conjunction with the accompanying drawing in the embodiment of the present invention, the technical scheme in the embodiment of the present invention is carried out Clearly and completely describe, it is clear that described embodiment is only a part of embodiment of the present invention, and It is not all, of embodiment.Based on the embodiment in the present invention, those of ordinary skill in the art are not doing Go out the every other embodiment obtained under creative work premise, broadly fall into the scope of protection of the invention.

Embodiments provide the Forecasting Methodology of a kind of short-term wind-electricity power, refer to accompanying drawing 1, described Method specifically includes following steps:

Step 101: obtaining real data according to preset rules, described real data comprises the power of wind energy turbine set Data and the air speed data of correspondence.

Step 102: revise described real data, it is thus achieved that sample data.

Step 103: described sample data is carried out the first pretreatment, and generate the first input sample of data with First output data sample, described first input sample data comprises the power data of wind energy turbine set and correspondence Air speed data, described first output sample data comprises the power data of wind energy turbine set；

Concrete, the described sample data obtained first is normalized, i.e. described sample data Being mapped in a less interval, described mapping range can be between [-1,1], and described normalization is public Formula is:

$x_{1i}^{*}=\frac{(y_{1max}-y_{1min})\×(x_{1i}-x_{1min})}{x_{1\max }-x_{1\min }}+y_{1\min }$

Wherein, x_1iFor the value before described sample data normalization,After described sample data normalization Value, y_1maxThe maximum of the mapping range for being normalized, y_1minFor the map section being normalized Between minima, x_1maxFor the maximum in described sample data, x_1minFor in described sample data Minima；

Afterwards, then the data value obtained after described sample data normalization is classified, generate first defeated Entering data sample and the first output data sample, wherein, described first input sample of data is wind energy turbine set Power data and the air speed data of correspondence, described first output data sample is the power data of wind energy turbine set.

Step 104: calculate Method Using Relevance Vector Machine kernel function, and by described kernel function, sample data, first Data based on input sample data and the first output sample data, set up Method Using Relevance Vector Machine forecast model.

Step 105: according to described Method Using Relevance Vector Machine forecast model, set up Method Using Relevance Vector Machine training pattern.

Step 106: optimize the parameter of Method Using Relevance Vector Machine in described Method Using Relevance Vector Machine training pattern, it is thus achieved that right The Method Using Relevance Vector Machine training pattern of the parameter of the described Method Using Relevance Vector Machine after should optimizing, as optimum be correlated with to Amount machine training pattern；

Concrete, based on forecast error Producing reason, the parameter of described Method Using Relevance Vector Machine is optimized, And then obtain the optimum Method Using Relevance Vector Machine training mould corresponding with the parameter of the described Method Using Relevance Vector Machine after optimization Type.

Step 107: described optimum Method Using Relevance Vector Machine training pattern is carried out the second pretreatment, it is thus achieved that actual pre- Measured value；

Concrete, export data sample to described optimum first with described first input sample of data and first Method Using Relevance Vector Machine training pattern is trained and checks, and obtains initial prediction, then to described initial predicted Value carries out renormalization process, obtains actual prediction value, and wherein, described renormalization processes formula and is:

$x_{1r}=\frac{x_{1r}^{*}\×(x_{1max}-x_{1min})-y_{1min}}{y_{1max}-y_{1min}}+x_{1\min }$

Wherein,For the data value before the described initial prediction renormalization after processing, x_1rAfter processing Described initial prediction renormalization after data value, y_1maxIt is normalized place for described sample data The maximum 1, y of mapping range during reason_1minMapping when being normalized for described sample data Interval minima-1, x_1maxFor the maximum in described sample data, x_1minFor in described sample data Minima.

Step 108: utilize described sample data and actual prediction value, be calculated the first relative error sequence；

Concrete, by described sample data and actual prediction value substitution relative error magnitudes RE computing formula:

$RE=\frac{y\left(t\right)-\hat{y}\left(t\right)}{RP}$

Being calculated initial first relative error sequence, described initial first relative error sequence is by multiple phases Forming error amount RE, wherein, y (t) is described sample data,For described actual prediction value, RP Rated power for blower fan；

Wherein, described initial first relative error sequence is that is 2 middle of the month second of described time limit stipulated time Individual month remove last day residue corresponding to natural law multiple relative error magnitudes RE composition sequence, and When calculating the sequence that the multiple relative error magnitudes RE corresponding to arbitrary sky in second month form, need every time Will be using the sample data corresponding to all natural law before this day in first month and second month as one Individual training sample data substitute into described optimum Method Using Relevance Vector Machine training pattern and carry out the second pretreatment, it is thus achieved that right The actual prediction value answered, afterwards, by sample data corresponding for the described actual prediction value that obtains together Substitute into relative error magnitudes RE computing formula, finally after double counting repeatedly relative error magnitudes RE, it is thus achieved that Described initial first relative error sequence；

Afterwards, calculated described initial first relative error sequence is carried out time series tranquilization inspection Survey, and showing that the testing result of described initial first relative error sequence is non-stationary relative error sequence Time, the relative error sequence of non-stationary is carried out difference processing and unit root test, if each unit The assay that root inspection obtains is still the relative error sequence of non-stationary, then repeat non-stationary Relative error sequence carries out difference processing and unit root test, until assay is the relative of tranquilization Error sequence, just output result, and using the relative error sequence of tranquilization as described first relative error Sequence.

Step 109: set up GARCH error prediction model；

Concrete, owing to generilized auto regressive conditional heteroskedastic (GARCH) can be portrayed the time more meticulously The time dependent characteristic of variance of sequence, and information possible in residual error item is excavated, therefore build Vertical GARCH error prediction model, reduces forecast error for follow-up after prediction in terms of error correction, Improve forecast error precision.

Step 1010: utilize GARCH error prediction model described in described first relative error sequence pair to enter Row the 3rd pretreatment, it is thus achieved that optimum GARCH error prediction model；

Concrete, based on forecast error Producing reason, described GARCH error prediction model is carried out Optimization processes, it is thus achieved that corresponding optimum GARCH error prediction model.

Step 1011: utilize described optimum GARCH error prediction model to described first relative error sequence It is predicted, it is thus achieved that the predictive value of the second relative error sequence；

Concrete, after obtaining optimum GARCH error prediction model, recycle described optimum Described first relative error sequence is predicted by GARCH error prediction model, and then can be the most accurate Data value institute in m acquisition point in the last day in ground that is 2 months described time limit stipulated time of prediction is right The relative error predictive value answeredPredictive value as the second relative error sequence.

Step 1012: utilize actual prediction value described in the predictive value correction of described second relative error sequence, Obtain final predictive value；

Concrete, use described the obtained after described optimum GARCH error prediction model prediction The predictive value of two relative error sequences goes to revise described actual prediction value, it is possible to error correction side after prediction Face reduces forecast error, improves forecast error precision.

In the Forecasting Methodology of short-term wind-electricity power disclosed in the embodiment of the present invention, described by obtaining and revising Real data, it is thus achieved that sample data, utilizes described sample data to set up Method Using Relevance Vector Machine forecast model, and Obtain Method Using Relevance Vector Machine training pattern according to described Method Using Relevance Vector Machine forecast model, optimize described associated vector The parameter of the Method Using Relevance Vector Machine in machine training pattern obtains optimum Method Using Relevance Vector Machine training pattern, Jin Ercong Optimal prediction model aspect reduces forecast error, afterwards, to described optimum Method Using Relevance Vector Machine training pattern Carry out the second pretreatment, it is thus achieved that actual prediction value, utilize described sample data and actual prediction value, calculate First relative error sequence, and utilize the GARCH error prediction of described first relative error sequence pair foundation Model carries out the 3rd pretreatment, it is thus achieved that optimum GARCH error prediction model, again realizes from Optimization Prediction Model aspect reduces forecast error, utilizes optimum GARCH error prediction model to described first phase afterwards Error sequence is predicted, it is thus achieved that the predictive value of the second relative error sequence, in order to pre-to described reality Measured value is modified, it is thus achieved that final predictive value, and then reduces forecast error in terms of error correction, carries High precision of prediction.

Referring to accompanying drawing 2, the process that implements of the step 101 provided in above-described embodiment includes following step Rapid:

Step 201: obtain the initial real data in time limit stipulated time；

Concrete, described initial real data be the power data comprising wind energy turbine set in time limit stipulated time and The measured data of corresponding air speed data, wherein, described time limit stipulated time can be 2 months.

Step 202: described time limit stipulated time is divided into n stipulated time section, n is positive integer；

Concrete, arbitrary described stipulated time section is any corresponding to i.e. 2 middle of the month in described time limit stipulated time One day.

Step 203: obtain m the time of described initial real data in setting arbitrary described stipulated time section Interval, m is positive integer；

Concrete, within arbitrary described stipulated time section i.e. every day, all set identical acquisition described at the beginning of M time interval of beginning real data, arbitrary described time interval can be single according to times such as minute, hours The data value in corresponding moment in described initial real data is removed to obtain in position.

Step 204: through arbitrary described time interval, obtain once from described initial real data Data value, as the data value of an acquisition point in described stipulated time section arbitrary in described real data, Described acquisition point and described time interval one_to_one corresponding；

Concrete, within arbitrary described stipulated time section i.e. every day, often through arbitrary described time interval Time, from described initial real data, just obtain once the data value in corresponding moment, as this stipulated time In section i.e. every day, the data value in acquisition point corresponding to current time interval, gets this most according to this The data value in m acquisition point corresponding to m time interval in stipulated time section i.e. every day, it After, the data value obtained in n described stipulated time section is gathered, obtains real data.

In the disclosed method obtaining real data of the embodiment of the present invention, by by described time limit stipulated time It is divided into n stipulated time section, and sets the described initial actual number of acquisition in arbitrary described stipulated time section According to m time interval, obtain described initial real data successively, finally give described real data, Can refine to for the data value in each acquisition point in the every day in described time limit stipulated time It is predicted, is predicted providing basic data accurately for the follow-up power data to wind energy turbine set.

Referring to accompanying drawing 3, the process that implements of the step 102 provided in above-described embodiment includes following step Rapid:

Step 301: determine the problem data in arbitrary described stipulated time section in described real data, and Described problem data acquisition point position in arbitrary described stipulated time section, described problem data is disappearance Data or abnormal data；

Concrete, after acquiring described real data, in addition it is also necessary to described real data is carried out exception Data identification and missing data identification, wherein, described abnormal data includes that the power data of wind energy turbine set is negative The data of value；

Step 302: described problem data replaces with correction data successively, and described correction data are for being positioned at State on the previous acquisition point position of problem data acquisition point position in arbitrary described stipulated time section Data value；

Concrete, arbitrary described in the arbitrary described stipulated time section in determining described real data When the data value in acquisition point corresponding to time interval is abnormal data, then by the exception in this acquisition point Data are rejected, and are substituted in this acquisition point by the data value in an acquisition point before it；? Determine corresponding to the arbitrary described time interval in the arbitrary described stipulated time section in described real data Acquisition point on data value when being missing data, then directly by the data in an acquisition point before it Value is substituted in this acquisition point.

Step 303: obtain revised data, as sample data；

Concrete, after the whole issue determined in described real data data are all modified, it is thus achieved that Revised data be described sample data.

In the disclosed method revising real data of the embodiment of the present invention, by will be from described real data The problem data determined replaces with the data value in the previous acquisition point of this problem data place acquisition point, Achieve the correction to described real data, and then be predicted carrying for the follow-up power data to wind energy turbine set For basic data accurately.

Referring to accompanying drawing 4, the process that implements of the step 104 provided in above-described embodiment includes following step Rapid:

Step 401: by described sample data and the first input sample of data substitution kernel function formula:

$K(x_i,x_j)=\exp \left(-\frac{\left|\right|x_i-x_j\left|\right|}{{\mathrm{\σ}}^2}\right)$

Calculate the kernel function of Method Using Relevance Vector Machine, wherein x_iFor the input vector of described sample data, x_jFor The input vector of described first input sample of data, σ is the kernel function width of Method Using Relevance Vector Machine；

Concrete, typically will there is the Radial basis kernel function of stronger nonlinear fitting ability and learning capacity Elect the kernel function of described Method Using Relevance Vector Machine as.

Step 402: by the kernel function of described Method Using Relevance Vector Machine, sample data, first input sample data and Data based on first output sample data, set up described Method Using Relevance Vector Machine forecast model；

Concrete, described Method Using Relevance Vector Machine forecast model is including but not limited to the core letter of described Method Using Relevance Vector Machine Sample data several, described, described first input sample of data and the first output data sample.

Set up in the method for Method Using Relevance Vector Machine forecast model disclosed in the embodiment of the present invention, by selecting and counting Calculate the kernel function of described Method Using Relevance Vector Machine, and by described kernel function, sample data, the first input sample Data based on data and the first output sample data, establish Method Using Relevance Vector Machine forecast model, and then The Method Using Relevance Vector Machine forecast model that can utilize foundation selects the kernel function of Method Using Relevance Vector Machine neatly, with Time use extremely management loading method due to Method Using Relevance Vector Machine forecast model, reduce answering of calculating Miscellaneous degree, decreases required data volume, finally shortens the time of calculating.

Referring to accompanying drawing 5, the process that implements of the step 105 provided in above-described embodiment includes following step Rapid:

Step 501: by described first input sample data and the first output sample data classification, generate first Training sample data and the first test samples data；

Concrete, the sample data that will comprise in described Method Using Relevance Vector Machine forecast model is classified, and generates first Training sample data and the first test samples data, described first training sample data are mainly used in training institute State the process of Method Using Relevance Vector Machine forecast model learning training sample, and described first training sample data are institute Stating the data value that time limit stipulated time that is 2 months are corresponding, described first test samples data are mainly used in testing Card passes through the accuracy of the predictive ability of the forecast model after training, and described first test samples number According to the data value corresponding to that is 2 last daies in the middle of the month of described time limit stipulated time.

Step 502: utilize described first training sample data, described Method Using Relevance Vector Machine forecast model is carried out Training, sets up Method Using Relevance Vector Machine training pattern；

Concrete, first it is calculated variance Σ of the Posterior distrbutionp of the first training sample data, described posteriority The computing formula of variance Σ of distribution is:

Σ=(σ^-2Φ^TΦ+A)^-1

Wherein, A is hyper parameter diagonal entry；And A=diag (a₀, a₁... a_N)；σ is relevant The kernel function width of vector machine；Φ is basis function vector；

It is calculated the mean μ of the Posterior distrbutionp of the first training sample data again, described Posterior distrbutionp equal The computing formula of value μ is:

μ=σ^-2ΣΦ^Tt

Wherein, t is the target component number of described first training sample data, and σ is the core of Method Using Relevance Vector Machine Function widths, Φ is basis function vector, and Σ is the variance of the Posterior distrbutionp of the first training sample data；

Afterwards, hyper parameter α it is calculated_iWith σ²Likelihood estimator L (α), described likelihood estimator The computing formula of L (α) is:

L (α)=-0.5 [Nlog2 π+log | C |+t^TC^-1t]

Wherein, C=σ²I+ΦA^-1Φ^TAnd C is covariance matrix, N is described sample data Number, t is the target component number of described first training sample data；

By described hyper parameter α_iWith σ²Likelihood estimator L (α) maximization process, obtain hyper parameter α_iWith σ²Maximization prior distribution:

$\begin{array}{cc}{\mathrm{\α}}_i=\frac{{\mathrm{\γ}}_i}{u_i}& {\mathrm{\σ}}^2=\frac{\left|\right|t-\mathrm{\Φ}u|{|}^2}{N-{\mathrm{\Σ}}_{i=1}^N{\mathrm{\γ}}_i}\end{array}$

Wherein, γ_i=1-α_iΣ_iiAnd Σ_iiVariance for the Posterior distrbutionp of described first training sample data The i-th diagonal entry of Σ, N is the number of described sample data, α_iFor hyper parameter；

Finally, initial prediction y is calculated_*And varianceCalculate initial prediction y_*And variance's Formula is respectively as follows:

y_*=μ^TΦ(x_*)

${\mathrm{\σ}}_{*}^2={\mathrm{\σ}}_{MP}^2+{\mathrm{\Φ}}^T\mathrm{\Σ}\mathrm{\Φ}$

Wherein, μ is the average of Posterior distrbutionp, and Φ is basis function vector, and Σ is the first training sample data The variance of Posterior distrbutionp,For hyper parameter σ²Maximization prior distribution, Φ (x_*) for calculating The kernel function of Method Using Relevance Vector Machine be mapped to higher dimensional space after the vector matrix that obtains, and

$\mathrm{\Φ}\left(x{*}_{}\right)=\left[\begin{array}{ccccc}1& K(x_1,x_1)& K(x_1,x_2)& \mathrm{...}& K(x_1,x_N)\\ 1& K(x_2,x_1)& K(x_2,x_2)& \mathrm{...}& K(x_2,x_N)\\ .& .& .& .& .\\ .& .& .& .& .\\ .& .& .& .& .\\ 1& K(x_N,x_1)& K(x_N,x_2)& \mathrm{...}& K(x_N,x_N)\end{array}\right];$

By data based on described initial prediction and variance, set up described Method Using Relevance Vector Machine training pattern.

Step 503: utilize described first test samples data, described Method Using Relevance Vector Machine training pattern is carried out Checking, determines described Method Using Relevance Vector Machine training pattern；

Concrete, after utilizing described first training sample data to establish Method Using Relevance Vector Machine training pattern, Also need to the predictive ability by Method Using Relevance Vector Machine training pattern to setting up of described first test samples data Accuracy is verified, further determines that described Method Using Relevance Vector Machine training pattern.

Set up in the method for Method Using Relevance Vector Machine training pattern disclosed in the embodiment of the present invention, by by described phase The sample data that pass vector machine forecast model comprises is classified, and generates the first training sample data and first Test samples data, and it is utilized respectively described first training sample data to described Method Using Relevance Vector Machine prediction mould Type is trained, and utilizes described first test samples data to test described Method Using Relevance Vector Machine forecast model Card, the final Method Using Relevance Vector Machine training pattern that obtains, and then be follow-up acquisition optimum Method Using Relevance Vector Machine training mould Type provides basic model.

Referring to accompanying drawing 6, the process that implements of the step 106 provided in above-described embodiment includes following step Rapid:

Step 601: the kernel function width cs needing the described Method Using Relevance Vector Machine optimized is mapped as individual arrow Amount.

Step 602: utilize the individual voxel vector being mapped to, sets up and initializes population；

Concrete, described initialization population X_i,G(i=1,2 ..., N) in comprise individual voxel vector Number N is the number of the kernel function width cs of the described Method Using Relevance Vector Machine needing optimization.

Step 603: obtain maximum iteration time, TSP question factor F initial value and the self adaptation set and hand over Fork probability CR initial value.

Step 604: utilize described TSP question factor F, vows described initialization population and current individual Amount carries out TSP question operation, it is thus achieved that variation vector；

Concrete, first find, from described initialization population, the base that two different individual voxel vectors are corresponding at random Because of position, and the gene position that current individual vector is corresponding, and substituted into TSP question computing formula:

V_i,G=X_r1,G+F(X_r2,G-X_r3,G), r1 ≠ r2 ≠ r3 ≠ i

Wherein, r1, r2 and r3 be random selected from described initialization population scope 1,2 ..., N}, and r1, r2, Tetra-its values of constant of r3 with i are different；F is the TSP question factor, V_i,GFor newborn after performing mutation operation The variation vector become, X_i,GFor belonging to the current individual vector initialized in population.

Wherein, described in the initial stage, the value of TSP question factor F is bigger, it is possible to ensure described initialization population Multiformity, afterwards, the value of described TSP question factor F will gradually can subtract along with the increase of iterations Little so that described in the later stage, the value of TSP question factor F is less, it is possible to retain excellent individual voxel vector, institute The computing formula stating TSP question factor F is:

$F=F_{\min }+(F_{max}-F_{min})\×e^{1-\frac{Mgen}{Mgen-G+1}}$

Wherein, F_minFor the minima of mutagenic factor, F_maxFor the maximum of mutagenic factor, Mgen is Big iteration algebraically, G is the algebraically when evolution.

Step 605: utilize adaptive crossover mutation CR, is carried out described variation vector current individual vector Self adaptation cross selection, it is thus achieved that intersection vector；

Concrete, by described variation vector V_i,GWith the current individual vector X belonging to described initialization population_i,G Carrying out self adaptation cross selection, described self adaptation cross selection formula is:

$U_{i,G}=\left\{\begin{array}{c}{V}_{i,G},rand\≤CR\\ X_{i,G},otherwise\end{array}\right.$

Wherein, CR is adaptive crossover mutation, and CR ∈ [0,1]；Rand is the random number between 0-1；V_i,G For variation vector newly-generated after performing mutation operation, X_i,GFor belonging to the current individual initialized in population Vector, U_i,GFor performing the intersection vector formed after intersection operation；

Wherein, the value of the described adaptive crossover mutation CR at initial stage is bigger, it is possible to ensure the change of global scope Different situation, afterwards, the value of described adaptive crossover mutation CR can be along with the increase of iterations, will gradually Reduce so that the value of the described adaptive crossover mutation CR in later stage is less, it is possible to pay close attention to the convergence feelings of local Condition, the computing formula of described adaptive crossover mutation CR is:

$CR={\mathrm{CR}}_{max}-\frac{G({\mathrm{CR}}_{max}-{\mathrm{CR}}_{min})}{Mgen}$

Wherein, CR_minFor the minima of adaptive crossover mutation, CR_maxMaximum for adaptive crossover mutation Value, Mgen is maximum iteration algebraically,.

Step 606: described intersection vector and described current individual vector are substituted into fitness function respectively, obtains Must intersect the fitness value of vector and the fitness value of current individual vector；

Concrete, by absolute average percent error e_MAPEAs fitness function, it is used for optimizing described phase The parameter of the Method Using Relevance Vector Machine in the vector machine training pattern of pass.

Step 607: compare the fitness value of described intersection vector and the fitness value of described current individual vector Size, it is thus achieved that the vector that fitness value is less.

Step 608: judge the number of times of current iteration, if the number of times of current iteration is equal to maximum iteration time, Then perform step 609, if the number of times of current iteration is less than maximum iteration time, then perform step 604；

Concrete, when the number of times of current iteration is less than maximum iteration time, select fitness value less Vector is as the current individual vector of next iteration, and returns execution step 604, until current iteration Number of times is equal to maximum iteration time.

Step 609: the minimum fitness value of output, and perform step 6010；

Concrete, when the number of times of current iteration is equal to maximum iteration time, the result of calculation of output is this Minimum fitness value in iterative computation.

Step 6010: obtain the kernel function width cs of the Method Using Relevance Vector Machine corresponding with described minimum fitness value, Kernel function width cs as optimum Method Using Relevance Vector Machine.

Step 6011: obtain the associated vector corresponding with the kernel function width cs of described optimum Method Using Relevance Vector Machine Machine training pattern, as optimum Method Using Relevance Vector Machine training pattern.

The parameter of the disclosed Method Using Relevance Vector Machine optimized in Method Using Relevance Vector Machine training pattern of the embodiment of the present invention In method, the differential evolution algorithm (IDE) improved by employing optimizes described Method Using Relevance Vector Machine training pattern In the kernel function width cs of Method Using Relevance Vector Machine, it is thus achieved that the kernel function width cs of optimum Method Using Relevance Vector Machine, and By the Method Using Relevance Vector Machine training pattern corresponding with the kernel function width cs of described optimum Method Using Relevance Vector Machine, as Optimum Method Using Relevance Vector Machine training pattern, so achieve reduce in terms of forecast error Producing reason pre- Survey error, improve precision of prediction.

Refer to accompanying drawing 7, the optimum GARCH error prediction mould of a kind of acquisition disclosed in the embodiment of the present invention The method of type, described method specifically includes following steps:

Step 701: set up initial ARMA model；

Concrete, set up average equation, described initial ARMA mould by described initial ARMA model The expression formula of type is:

Wherein,For auto-regressive parameter to be estimated；θ_jFor rolling average parameter；R is Autoregressive；m Exponent number for rolling average；ε (t) is the residual error of t；C is constant.

Step 702: set up initial GARCH model；

Concrete, owing to described initial ARMA model is unsuitable for processing described first relative error sequence Variance, accordingly, it would be desirable to use EC GARCH i.e. by described initial GARCH Variance equation set up by model, and the expression formula of described initial GARCH model is:

${\mathrm{\σ}}_t^2=k+\underset{i=1}{\overset{p}{\Σ}}{G}_i{\mathrm{\σ}}_{t-i}^2+\underset{j=1}{\overset{q}{\Σ}}{A}_j{\mathrm{\ϵ}}_{t-j}^2$

Wherein, G_iThe parameter to be estimated affected for initial GARCH model, and more than 0；A_jFor initial ARMA The parameter to be estimated of model impact, and more than 0；P Yu q is the order of initial ARMA model,For condition side Difference；K is constant.

Step 703: be fitted described initial ARMA model, determines the exponent number of final arma modeling；

Concrete, use described initial ARMA model to described the after multi-difference tranquilization processes One relative error sequence is fitted, the autocorrelation of the inspection the first relative error sequence after over-fitting With partial autocorrelation, and calculate its autocorrelation coefficient and PARCOR coefficients, as weighing this matching effect The standard that fruit is good and bad, to determine the exponent number of final arma modeling.

Step 704: described initial GARCH model is fitted, determines the rank of final GARCH model Number；

Concrete, use red pond information criterion (AIC) to weigh the quality of described initial GARCH models fitting Property, to determine the exponent number of final GARCH model.

Step 705: utilize described final arma modeling and final GARCH model to described first relative Error sequence is fitted, and sets up GARCH error prediction model, wherein, described final arma modeling The exponent number of the described final arma modeling corresponding to determining, described final GARCH model is corresponding to determining The exponent number of described final GARCH model；

Concrete, utilize described final arma modeling and described in the difference matching of final GARCH model first Relative error sequence, and then determine the auto-regressive parameter to be estimated in final arma modelingMobile flat All parameter θ_jWith constant c, determine simultaneously in final GARCH model initial ARMA model impact wait estimate Parameter A_jWith constant k, finally give and contain the average equation corresponding with determining parameter and variance equation GARCH error prediction model.

Step 706: described first relative error sequence is classified, generate the second training sample data and Second test samples data；

Concrete, described second training sample data are second month in the middle of the month in described time limit stipulated time that is 2 Removing the data value that the All Time of this month last day is corresponding, described second test samples data are described The data value of the previous day of time limit stipulated time that is 2 second month in middle of the month last day.

Step 707: utilize described second training sample data, enters described GARCH error prediction model Row matching, it is thus achieved that the GARCH error prediction model of optimization；

Concrete, utilize described second training sample data that described GARCH error prediction model is intended Close, and then determine auto-regressive parameter to be estimated relevant in described GARCH error prediction modelMobile Mean parameter θ_j, constant c, and initial ARMA model impact parameter A to be estimated_jWith constant k, final GARCH error prediction to the optimization containing the average equation corresponding with determining parameter and variance equation Model.

Step 708: utilize described second test samples data, the GARCH error prediction to described optimization Model is verified, it is thus achieved that validation value；

Concrete, by absolute average percent error e_MAPEAs validation value, and by described second inspection The data of the previous day of second month last day in sample data described time limit stipulated time of that is 2 months Value substitutes into absolute average percent error e_MAPEComputing formula:

$e_{MAPE}=\frac{1}{n}\underset{i=1}{\overset{n}{\Σ}}\frac{|Y_i-{\hat{Y}}_i|}{\left|Y_i\right|}\×100\%$

Obtain validation value, wherein, Y_iFor the power of wind energy turbine set corresponding to described second test samples data,For the actual prediction value that described second test samples data are corresponding, n is that described second test samples data obtain Quantity；

Step 709: relatively described validation value and the size of preassigned, if described validation value is less than or equal to Preassigned, then perform step 7010, if described validation value is more than preassigned, then returns and performs step 703；

Described absolute average percent error e that is concrete, that will calculate_MAPECompare with preassigned Relatively, and then judge whether described validation value meets preassigned；

In the described absolute average percent error e calculated_MAPEMore than predetermined standard time, then return step Rapid 703, again described initial ARMA model is fitted, determines the exponent number of final arma modeling, Until described absolute average percent error e_MAPELess than or equal to preassigned.

Step 7010: obtain optimum GARCH error prediction model；

Concrete, in the described absolute average percent error e calculated_MAPELess than or equal to pre-calibration Accurate, it is determined that the GARCH error prediction model of the optimization now obtained is optimum GARCH error prediction Model.

In the disclosed method obtaining optimum GARCH error prediction model of the embodiment of the present invention, by building Vertical initial ARMA model and initial GARCH model, determine corresponding final arma modeling Exponent number and the exponent number of final GARCH model, and then set up GARCH error prediction model, afterwards, Described first relative error sequence is classified, utilizes the second training sample data generated and the second inspection Test sample data successively described GARCH error prediction model is fitted and is verified, finally obtain Excellent GARCH error prediction model, and then utilize after generilized auto regressive conditional heteroskedastic (GARCH) is Continuous realization reduces forecast error in terms of error correction, and improving precision of prediction provides optimal models.

Referring to accompanying drawing 8, the process that implements of the step 1012 provided in above-described embodiment includes following step Rapid:

Step 801: by the predictive value substitution formula of described second relative error sequence:

$\widehat{RE}={\widehat{RE}}_p\×RP$

It is calculated the residual sequence predictive value of correspondenceWherein,It it is the second relative error sequence Predictive value, RP is the rated power of blower fan；

Concrete, being calculated described residual sequence predictive value is that is 2 middle of the month of described time limit stipulated time M residual sequence predictive value corresponding to data value in m acquisition point in last dayFormed Combination.

Step 802: described residual sequence predictive value is added with described actual prediction value, it is thus achieved that finally predict Value；

Concrete, that is 2 middle of the month of described time limit stipulated time are appointed data in intraday m acquisition point Actual prediction value corresponding to value respectively be calculated described residual sequence predictive valueIn the m that comprises Individual residual sequence predictive valueCarry out addition calculation, finally draw that is 2 middle of the month of described time limit stipulated time The final predictive value of arbitrary day, after addition calculation is repeated several times, draws described time limit stipulated time that is 2 The final predictive value of whole natural law of individual month.

In the disclosed method revising actual prediction value of the embodiment of the present invention, described second relative by utilizing The predictor calculation of error sequence goes out the residual sequence predictive value of correspondence, is predicted by described residual sequence afterwards Value is added with described actual prediction value, it is thus achieved that final predictive value, and then reduces in terms of error correction Forecast error, improves precision of prediction.

Embodiments provide the prediction means of a kind of short-term wind-electricity power, refer to accompanying drawing 9, described Device includes:

First acquisition module 901, for obtaining real data according to preset rules, described real data comprises The power data of wind energy turbine set and the air speed data of correspondence；

First correcting module 902, is used for revising described real data, it is thus achieved that sample data；

First pretreatment module 903, for described sample data is carried out the first pretreatment, and generates first Input sample data and the first output sample data, described first input sample data comprises the merit of wind energy turbine set Rate data and the air speed data of correspondence, described first output sample data comprises the power data of wind energy turbine set；

First model building module 904, for calculating the kernel function of Method Using Relevance Vector Machine, and by described kernel function, Data based on sample data, the first input sample data and the first output sample data, set up relevant Vector machine forecast model；

Second model building module 905, for according to described Method Using Relevance Vector Machine forecast model, set up be correlated with to Amount machine training pattern；

First optimizes module 906, for optimizing Method Using Relevance Vector Machine in described Method Using Relevance Vector Machine training pattern Parameter, it is thus achieved that the Method Using Relevance Vector Machine training pattern of the parameter of the described Method Using Relevance Vector Machine after corresponding optimization, makees For optimum Method Using Relevance Vector Machine training pattern；

Second pretreatment module 907, locates in advance for described optimum Method Using Relevance Vector Machine training pattern is carried out second Reason, it is thus achieved that actual prediction value；

First computing module 908, is used for utilizing described sample data and actual prediction value, is calculated first Relative error sequence；

3rd model building module 909, is used for setting up GARCH error prediction model；

3rd pretreatment module 910, is used for utilizing GARCH described in described first relative error sequence pair by mistake Difference forecast model carries out the 3rd pretreatment, it is thus achieved that optimum GARCH error prediction model；

First prediction module 911, is used for utilizing described optimum GARCH error prediction model to described first Relative error sequence is predicted, it is thus achieved that the predictive value of the second relative error sequence；

Second correcting module 912, real for utilizing described in the predictive value correction of described second relative error sequence Border predictive value, it is thus achieved that final predictive value.

In the prediction means of short-term wind-electricity power disclosed in the embodiment of the present invention, by the first acquisition module 901 And first correcting module 902 obtain and revise described real data, it is thus achieved that sample data, first model set up Module 904 utilizes described sample data to set up Method Using Relevance Vector Machine forecast model, the second model building module 905 Obtaining Method Using Relevance Vector Machine training pattern according to described Method Using Relevance Vector Machine forecast model, the first optimization module 906 is excellent The parameter changing the Method Using Relevance Vector Machine in described Method Using Relevance Vector Machine training pattern obtains optimum Method Using Relevance Vector Machine instruction Practice model, and then in terms of optimal prediction model, reduce forecast error, afterwards, the second pretreatment module 907 obtain actual prediction value, and the first computing module 908 is then calculated the first relative error sequence, utilize The GARCH error prediction model that 3rd model building module 909 is set up by the 3rd pretreatment module 910 enters Row pretreatment, it is thus achieved that optimum GARCH error prediction model, realizes coming in terms of optimal prediction model again Reduce forecast error, afterwards the second correcting module 912 according to the first prediction module 911 obtain second relative The predictive value of error sequence, to revising described actual prediction value, it is thus achieved that final predictive value, and then from error Correction aspect reduces forecast error, improves precision of prediction.

The work process of the modules that the embodiment of the present invention provides, refer to the flow process corresponding to accompanying drawing 1 Figure, specific works process repeats no more.

Described first acquisition module 901 provided in above-described embodiment includes:

Second acquisition module, for obtaining the initial real data in time limit stipulated time；

Separating modules, for described time limit stipulated time is divided into n stipulated time section, n is positive integer；

Setting module, obtains the m of described initial real data in setting arbitrary described stipulated time section Individual time interval, m is positive integer；

3rd acquisition module, for through arbitrary described time interval, from described initial real data Data value of middle acquisition, as an acquisition in described stipulated time section arbitrary in described real data The data value of point, described acquisition point and described time interval one_to_one corresponding.

In the embodiment of the present invention, by described separating modules, described time limit stipulated time is divided into n rule The section of fixing time, and described setting module sets, and arbitrary described stipulated time section is interior obtains described initial reality M time interval of data so that described 3rd acquisition module obtains from described initial real data successively Corresponding data value, finally gives described real data, it is possible to refine to for described time limit stipulated time The interior data value in each acquisition point in every day is predicted, for the follow-up power to wind energy turbine set Data are predicted providing basic data accurately.

The work process of the modules that the embodiment of the present invention provides, refer to the flow process corresponding to accompanying drawing 2 Figure, specific works process repeats no more.

Described first correcting module 902 provided in above-described embodiment includes:

Determine module, for the problem data determined in described real data in arbitrary described stipulated time section, And the acquisition point position that described problem data is in arbitrary described stipulated time section, described problem data is Missing data or abnormal data；

3rd correcting module, for replacing with correction data, described correction number successively by described problem data According to for being positioned at the previous acquisition in the acquisition point position of arbitrary described stipulated time section of the described problem data Data value on some position；

4th acquisition module, is used for obtaining revised data, as sample data.

In the embodiment of the present invention, will determine that determine in module asks from described by described 3rd correcting module Topic data replace with the data value in the previous acquisition point of this problem data place acquisition point, it is achieved that right The correction of described real data, and then be predicted providing accurately for the follow-up power data to wind energy turbine set Basic data.

The work process of the modules that the embodiment of the present invention provides, refer to the flow process corresponding to accompanying drawing 3 Figure, specific works process repeats no more.

Described first model building module 904 provided in above-described embodiment includes:

Kernel function computing module, for described sample data and the first input sample data being substituted into formula:

$K(x_i,x_j)=\exp \left(-\frac{\left|\right|x_i-x_j\left|\right|}{{\mathrm{\σ}}^2}\right)$

Calculate the kernel function of Method Using Relevance Vector Machine, wherein x_iFor the input vector of described sample data, x_jFor The input vector of described first input sample data, σ is the kernel function width of Method Using Relevance Vector Machine；

Submodule set up by first model, is used for setting up described Method Using Relevance Vector Machine forecast model, described be correlated with to Amount machine forecast model comprises the kernel function of described Method Using Relevance Vector Machine, described sample data, described first input Sample data and the first output sample data.

In the embodiment of the present invention, select and calculate described associated vector by described kernel function computing module The kernel function of machine, described first model sets up submodule by described kernel function, sample data, the first input Data based on sample data and the first output sample data, establish Method Using Relevance Vector Machine forecast model, And then the Method Using Relevance Vector Machine forecast model of foundation can be utilized to select the core letter of Method Using Relevance Vector Machine neatly Number, simultaneously because Method Using Relevance Vector Machine forecast model uses extremely management loading method, reduces meter The complexity calculated, decreases required data volume, finally shortens the time of calculating.

The work process of the modules that the embodiment of the present invention provides, refer to the flow process corresponding to accompanying drawing 4 Figure, specific works process repeats no more.

Described second model building module 905 provided in above-described embodiment includes:

First sort module, for described first input sample data is classified with the first output sample data, Generate the first training sample data and the first test samples data；

First training module, for described first training sample utilizing described first sort module to sort out Data, are trained described Method Using Relevance Vector Machine forecast model, set up Method Using Relevance Vector Machine training pattern；

First inspection module, for described first test samples utilizing described first sort module to sort out Data, verify described Method Using Relevance Vector Machine training pattern, determine described Method Using Relevance Vector Machine training pattern.

In the embodiment of the present invention, by described first sort module by described Method Using Relevance Vector Machine forecast model bag The sample data contained is classified, and generates the first training sample data and the first test samples data, described First training module and described first inspection module are utilized respectively described first training sample data to described phase Close vector machine forecast model be trained, and utilize described first test samples data to described be correlated with to Amount machine forecast model is verified, the final Method Using Relevance Vector Machine training pattern that obtains, so be follow-up acquisition Excellent Method Using Relevance Vector Machine training pattern provides basic model.

The work process of the modules that the embodiment of the present invention provides, refer to the flow process corresponding to accompanying drawing 5 Figure, specific works process repeats no more.

Described first provided in above-described embodiment optimizes module 906 and includes:

Parameter optimization module, for optimizing the core of the Method Using Relevance Vector Machine in described Method Using Relevance Vector Machine training pattern Function widths σ, it is thus achieved that the kernel function width cs of optimum Method Using Relevance Vector Machine；

5th acquisition module is corresponding with the kernel function width cs of described optimum Method Using Relevance Vector Machine for obtaining Method Using Relevance Vector Machine training pattern, as optimum Method Using Relevance Vector Machine training pattern.

In the embodiment of the present invention, optimize described Method Using Relevance Vector Machine training pattern by described parameter optimization module In the kernel function width cs of Method Using Relevance Vector Machine, described 5th acquisition module obtains optimum Method Using Relevance Vector Machine Kernel function width cs, and by the associated vector corresponding with the kernel function width cs of described optimum Method Using Relevance Vector Machine Machine training pattern, as optimum Method Using Relevance Vector Machine training pattern, and then achieves from forecast error generation Reason aspect reduces forecast error, improves precision of prediction.

The work process of the modules that the embodiment of the present invention provides, refer to the flow process corresponding to accompanying drawing 6 Figure, specific works process repeats no more.

Refer to accompanying drawing 10, the described parameter optimization module provided in above-described embodiment includes:

Mapping block 1001, for mapping the kernel function width cs needing the described Method Using Relevance Vector Machine optimized For individual voxel vector；

Population foundation module 1002, for utilizing the individual voxel vector being mapped to, sets up and initializes population；

6th acquisition module 1003, for obtaining the maximum iteration time of setting, TSP question factor F Initial value and adaptive crossover mutation CR initial value；

TSP question module 1004, is used for utilizing described TSP question factor F, to described initialization Population and current individual vector carry out TSP question operation, it is thus achieved that variation vector；

Self adaptation Cross module 1005, is used for utilizing adaptive crossover mutation CR, to described variation vector Current individual vector carries out self adaptation cross selection, it is thus achieved that intersection vector；

Fitness value acquisition module 1006, for by described intersection vector with described current individual vector respectively Substitute into fitness function, it is thus achieved that the fitness value of intersection vector and the fitness value of current individual vector；

First comparison module 1007, for fitness value and the described current individual of relatively described intersection vector The size of the fitness value of vector, and obtain the vector that fitness value is less；

Judge module 1008, for judging the number of times of current iteration；

At described judge module 1008, TSP question module 1004, for judging that the number of times of current iteration is little When maximum iteration time, the vector choosing that the fitness value that obtained by described first comparison module 1007 is less It is selected as current individual vector, utilizes described TSP question factor F, to described initialization population with when the one before Voxel vector carries out TSP question operation, it is thus achieved that variation vector；

Output module 1009, for judging that at described judge module the number of times of current iteration is equal to greatest iteration During number of times, the minimum fitness value of output；

7th acquisition module 1010, for obtaining the Method Using Relevance Vector Machine corresponding with described minimum fitness value Kernel function width cs, as the kernel function width cs of optimum Method Using Relevance Vector Machine.

In the embodiment of the present invention, used the differential evolution algorithm (IDE) improved by described parameter optimization module Optimizing the kernel function width cs of Method Using Relevance Vector Machine in described Method Using Relevance Vector Machine training pattern, the described 7th obtains Delivery block 1010 obtains the kernel function width cs of optimum Method Using Relevance Vector Machine, and then achieves from forecast error product Raw reason aspect reduces forecast error, improves precision of prediction.

The work process of the modules that the embodiment of the present invention provides, refer to the flow process corresponding to accompanying drawing 6 Figure, specific works process repeats no more.

Described 3rd model building module 909 provided in above-described embodiment includes:

Arma modeling sets up module, is used for setting up initial ARMA model；

GARCH model building module, is used for setting up initial GARCH model；

ARMA fitting module, for being fitted described initial ARMA model, determines described The exponent number of arma modeling；

GARCH fitting module, for being fitted described initial GARCH model, determines described The exponent number of GARCH model；

Submodule set up by 3rd model, is used for utilizing described final arma modeling and final GARCH mould Described first relative error sequence is fitted by type, sets up GARCH error prediction model, wherein, Described final arma modeling corresponding to the exponent number of described final arma modeling determined, described finally The exponent number of the GARCH model described final GARCH model corresponding to determining.

In the embodiment of the present invention, by utilizing described arma modeling to set up module and described GARCH model Set up module, set up initial ARMA model and initial GARCH model, described ARMA matching mould respectively Block and described GARCH fitting module, determine that the exponent number of corresponding final arma modeling is with final The exponent number of GARCH model, and then making described 3rd model set up submodule, to set up GARCH error pre- Survey model, so as to compared with portraying the relative error time dependent characteristic of seasonal effect in time series variance meticulously, And provide basic model for obtaining optimum GARCH error prediction model.

The work process of the modules that the embodiment of the present invention provides, refer to the flow process corresponding to accompanying drawing 7 Figure, specific works process repeats no more.

Described 3rd pretreatment module 910 provided in above-described embodiment includes:

Second sort module, for described first relative error sequence being classified, generates the second training Sample data and the second test samples data；

Error prediction models fitting module, is used for utilize described second sort module to sort out described second Training sample data, are fitted described GARCH error prediction model, it is thus achieved that the GARCH of optimization Error prediction model；

Second authentication module, for described second test samples utilizing described second sort module to sort out, The GARCH error prediction model of described optimization is verified, it is thus achieved that validation value；

Second comparison module, for the size of relatively described validation value with preassigned；

8th acquisition module, for being less than or equal to described at the described second more described validation value of comparison module Predetermined standard time, it is thus achieved that optimum GARCH error prediction model；

Described 3rd model building module, for being more than at the described second more described validation value of comparison module Described predetermined standard time, sets up GARCH error prediction model；

Concrete, at the described second more described validation value of comparison module more than described predetermined standard time, institute State the ARMA fitting module in the 3rd model building module, again described initial ARMA model is entered Row matching, determines the exponent number of described arma modeling, until the described second more described checking of comparison module Value is less than or equal to described preassigned.

In the embodiment of the present invention, by described second sort module, described first relative error sequence is carried out Classification, described error prediction models fitting module and described second authentication module utilize the second training generated Sample data and the second test samples data successively described GARCH error prediction model is fitted and Checking, finally makes described 8th acquisition module obtain optimum GARCH error prediction model, and then utilizes wide Justice ARCH (GARCH) is that follow-up realization reduces prediction by mistake in terms of error correction Difference, improving precision of prediction provides optimal models.

The work process of the modules that the embodiment of the present invention provides, refer to the flow process corresponding to accompanying drawing 7 Figure, specific works process repeats no more.

Described second correcting module 912 provided in above-described embodiment includes:

Second computing module, for the predictive value of described second relative error sequence being substituted into formula:

$\widehat{RE}={\widehat{RE}}_p\×RP$

It is calculated the residual sequence predictive value of correspondenceWherein,It it is the second relative error sequence Predictive value, RP is the rated power of blower fan；

4th correcting module, for being added with described actual prediction value by described residual sequence predictive value, obtains Obtain final predictive value.

In the embodiment of the present invention, utilize described second relative error sequence by described second computing module Predictive value calculates, it is thus achieved that corresponding residual sequence predictive value, the most described 4th correcting module is by institute State residual sequence predictive value to be added with described actual prediction value, it is thus achieved that final predictive value, and then from error school Positive aspect reduces forecast error, improves precision of prediction.

The work process of the modules that the embodiment of the present invention provides, refer to the flow process corresponding to accompanying drawing 8 Figure, specific works process repeats no more.

Between each embodiment disclosed in this invention, identical similar part can be with cross-reference.

Described above to the disclosed embodiments, makes professional and technical personnel in the field be capable of or uses The present invention.Multiple amendment to these embodiments will be aobvious and easy for those skilled in the art See, generic principles defined herein can without departing from the spirit or scope of the present invention, Realize in other embodiments.Therefore, the present invention is not intended to be limited to the embodiments shown herein, And it is to fit to the widest scope consistent with principles disclosed herein and features of novelty.

Claims (22)

1. the Forecasting Methodology of a short-term wind-electricity power, it is characterised in that described method includes:

Obtaining real data according to preset rules, described real data comprises the power data of wind energy turbine set and right The air speed data answered；

Revise described real data, it is thus achieved that sample data；

Described sample data is carried out the first pretreatment, and generates the first input sample data and the first output Sample data, described first input sample data comprises the power data of wind energy turbine set and the air speed data of correspondence, Described first output sample data comprises the power data of wind energy turbine set；

Calculate the kernel function of Method Using Relevance Vector Machine, and by described kernel function, sample data, the first input sample Data based on data and the first output sample data, set up Method Using Relevance Vector Machine forecast model；

According to described Method Using Relevance Vector Machine forecast model, set up Method Using Relevance Vector Machine training pattern；

Optimize the parameter of Method Using Relevance Vector Machine in described Method Using Relevance Vector Machine training pattern, it is thus achieved that corresponding optimize after The Method Using Relevance Vector Machine training pattern of parameter of described Method Using Relevance Vector Machine, as optimum Method Using Relevance Vector Machine training Model；

Described optimum Method Using Relevance Vector Machine training pattern is carried out the second pretreatment, it is thus achieved that actual prediction value；

Utilize described sample data and actual prediction value, be calculated the first relative error sequence；

Set up GARCH error prediction model；

Utilize GARCH error prediction model described in described first relative error sequence pair to carry out the 3rd to locate in advance Reason, it is thus achieved that optimum GARCH error prediction model；

Utilize described optimum GARCH error prediction model that described first relative error sequence is predicted, Obtain the predictive value of the second relative error sequence；

Utilize actual prediction value described in the predictive value correction of described second relative error sequence, it is thus achieved that the most pre- Measured value.

Method the most according to claim 1, it is characterised in that described real according to preset rules acquisition Border data, including:

Obtain the initial real data in time limit stipulated time；

Described time limit stipulated time is divided into n stipulated time section；

M time interval of described initial real data is obtained in setting arbitrary described stipulated time section；

Through arbitrary described time interval, from described initial real data, obtain a data value, As the data value of an acquisition point in described stipulated time section arbitrary in described real data, described in obtain Take a little with described time interval one_to_one corresponding；

Wherein, n Yu m is positive integer.

Method the most according to claim 2, it is characterised in that the described real data of described correction, Obtain sample data, including:

Determine the problem data in arbitrary described stipulated time section in described real data, and described problem Data acquisition point position in arbitrary described stipulated time section, described problem data is missing data or different Regular data；

Described problem data replaces with correction data successively, and described correction data are for being positioned at described problem number According to the data value on the previous acquisition point position of the acquisition point position in arbitrary described stipulated time section；

Obtain revised data, as sample data.

Method the most according to claim 1, it is characterised in that the core of described calculating Method Using Relevance Vector Machine Function, and by described kernel function, sample data, the first input sample data and the first output sample data Based on data, set up Method Using Relevance Vector Machine forecast model, including:

By described sample data and the first input sample of data substitution kernel function formula:

$K(x_i,x_j)=\exp (-\frac{\left|\right|x_i-x_j\left|\right|}{{\mathrm{\σ}}^2})$

Calculate the kernel function of Method Using Relevance Vector Machine, wherein x_iFor the input vector of described sample data, x_jFor The input vector of described first input sample data, σ is the kernel function width of Method Using Relevance Vector Machine；

Setting up described Method Using Relevance Vector Machine forecast model, described Method Using Relevance Vector Machine forecast model comprises described relevant The kernel function of vector machine, described sample data, described first input sample data and the first output sample number According to.

Method the most according to claim 4, it is characterised in that described according to described Method Using Relevance Vector Machine Forecast model, sets up Method Using Relevance Vector Machine training pattern, including:

By described first input sample data and the first output sample data classification, generate the first training sample Data and the first test samples data；

Utilize described first training sample data, described Method Using Relevance Vector Machine forecast model is trained, builds Vertical Method Using Relevance Vector Machine training pattern；

Utilize described first test samples data, described Method Using Relevance Vector Machine training pattern is verified, really Fixed described Method Using Relevance Vector Machine training pattern.

Method the most according to claim 1, it is characterised in that the described Method Using Relevance Vector Machine of described optimization The parameter of the Method Using Relevance Vector Machine in training pattern, it is thus achieved that optimum Method Using Relevance Vector Machine training pattern, including:

Optimize the kernel function width cs of Method Using Relevance Vector Machine in described Method Using Relevance Vector Machine training pattern, it is thus achieved that The kernel function width cs of excellent Method Using Relevance Vector Machine；

Obtain the Method Using Relevance Vector Machine training mould corresponding with the kernel function width cs of described optimum Method Using Relevance Vector Machine Type, as optimum Method Using Relevance Vector Machine training pattern.

Method the most according to claim 6, it is characterised in that the described Method Using Relevance Vector Machine of described optimization The kernel function width cs of the Method Using Relevance Vector Machine in training pattern, it is thus achieved that the kernel function width of optimum Method Using Relevance Vector Machine Degree σ, including:

The kernel function width cs needing the described Method Using Relevance Vector Machine optimized is mapped as a voxel vector；

Utilize the individual voxel vector being mapped to, set up and initialize population；

Obtain maximum iteration time, TSP question factor F initial value and the adaptive crossover mutation CR set Initial value；

Utilize described TSP question factor F, described initialization population and current individual vector are carried out adaptive Answer mutation operation, it is thus achieved that variation vector；

Utilize adaptive crossover mutation CR, described variation vector current individual vector is carried out self adaptation friendship Fork selects, it is thus achieved that intersection vector；

Described intersection vector and described current individual vector are substituted into fitness function respectively, it is thus achieved that intersect and vow The fitness value of amount and the fitness value of current individual vector；

Compare the size of the fitness value of described intersection vector and the fitness value of described current individual vector, Obtain the vector that fitness value is less；

Select the vector that fitness value is less as the current individual vector of next iteration, return and perform profit By described TSP question factor F, described initialization population and current individual vector are carried out TSP question Operation, it is thus achieved that variation vector, during until arriving maximum iteration time, the minimum fitness value of output；

Obtain the kernel function width cs of the Method Using Relevance Vector Machine corresponding with described minimum fitness value, as optimum The kernel function width cs of Method Using Relevance Vector Machine.

Method the most according to claim 1, it is characterised in that described to set up GARCH error pre- Survey model, including:

Set up initial ARMA model；

Set up initial GARCH model；

Described initial ARMA model is fitted, determines the exponent number of final arma modeling；

Described initial GARCH model is fitted, determines the exponent number of final GARCH model；

Utilize described final arma modeling and final GARCH model to described first relative error sequence Being fitted, set up GARCH error prediction model, wherein, described final arma modeling corresponds to The exponent number of the described final arma modeling determined, described final GARCH model is corresponding to the institute determined State the exponent number of final GARCH model.

Method the most according to claim 8, it is characterised in that described utilize described first relatively to miss Difference sequence carries out the 3rd pretreatment to described GARCH error prediction model, it is thus achieved that optimum GARCH is by mistake Difference forecast model, including:

Described first relative error sequence is classified, generates the second training sample data and the second inspection Sample data；

Utilize described second training sample data, described GARCH error prediction model be fitted, Obtain the GARCH error prediction model optimized；

Utilize described second test samples data, the GARCH error prediction model of described optimization is carried out Checking, it is thus achieved that validation value；

Relatively described validation value and the size of preassigned；

At described validation value less than or equal to described predetermined standard time, it is thus achieved that optimum GARCH error prediction mould Type.

Method the most according to claim 9, it is characterised in that at the described validation value of described comparison After the size of preassigned, also include:

At described validation value more than described predetermined standard time, return and described initial ARMA model is intended Close, determine the exponent number of final arma modeling, until described validation value is less than or equal to described preassigned.

11. methods according to claim 1, it is characterised in that described utilize described second relative Actual prediction value described in the predictive value correction of error sequence, it is thus achieved that final predictive value, including:

By the predictive value substitution formula of described second relative error sequence:

$\widehat{RE}={\widehat{RE}}_p\×RP$

It is calculated the residual sequence predictive value of correspondenceWherein,It it is the second relative error sequence Predictive value, RP is the rated power of blower fan；

Described residual sequence predictive value is added with described actual prediction value, it is thus achieved that final predictive value.

The prediction means of 12. 1 kinds of short-term wind-electricity powers, it is characterised in that described device includes:

First acquisition module, for obtaining real data according to preset rules, described real data comprises wind The power data of electric field and the air speed data of correspondence；

First correcting module, is used for revising described real data, it is thus achieved that sample data；

First pretreatment module, for described sample data is carried out the first pretreatment, and it is defeated to generate first Entering sample data and the first output sample data, described first input sample data comprises the power of wind energy turbine set Data and the air speed data of correspondence, described first output sample data comprises the power data of wind energy turbine set；

First model building module, for calculating the kernel function of Method Using Relevance Vector Machine, and by described kernel function, Data based on sample data, the first input sample data and the first output sample data, set up relevant Vector machine forecast model；

Second model building module, for according to described Method Using Relevance Vector Machine forecast model, sets up associated vector Machine training pattern；

First optimizes module, for optimizing the ginseng of the Method Using Relevance Vector Machine in described Method Using Relevance Vector Machine training pattern Number, it is thus achieved that the Method Using Relevance Vector Machine training pattern of the parameter of the described Method Using Relevance Vector Machine after corresponding optimization, as Optimum Method Using Relevance Vector Machine training pattern；

Second pretreatment module, for described optimum Method Using Relevance Vector Machine training pattern is carried out the second pretreatment, Obtain actual prediction value；

First computing module, is used for utilizing described sample data and actual prediction value, is calculated the first phase To error sequence；

3rd model building module, is used for setting up GARCH error prediction model；

3rd pretreatment module, is used for utilizing GARCH error described in described first relative error sequence pair Forecast model carries out the 3rd pretreatment, it is thus achieved that optimum GARCH error prediction model；

First prediction module, is used for utilizing described optimum GARCH error prediction model to described first phase Error sequence is predicted, it is thus achieved that the predictive value of the second relative error sequence；

Second correcting module, actual for utilizing described in the predictive value correction of described second relative error sequence Predictive value, it is thus achieved that final predictive value.

13. devices according to claim 12, it is characterised in that described first acquisition module includes:

Second acquisition module, for obtaining the initial real data in time limit stipulated time；

Separating modules, for being divided into n stipulated time section by described time limit stipulated time；

Setting module, obtains the m of described initial real data in setting arbitrary described stipulated time section Individual time interval；

3rd acquisition module, for through arbitrary described time interval, from described initial real data Data value of middle acquisition, as an acquisition in described stipulated time section arbitrary in described real data The data value of point, described acquisition point and described time interval one_to_one corresponding；

Wherein, n Yu m is positive integer.

14. devices according to claim 13, it is characterised in that described first correcting module includes:

Determine module, for the problem data determined in described real data in arbitrary described stipulated time section, And the acquisition point position that described problem data is in arbitrary described stipulated time section, described problem data is Missing data or abnormal data；

3rd correcting module, for replacing with correction data, described correction number successively by described problem data According to for being positioned at the previous acquisition in the acquisition point position of arbitrary described stipulated time section of the described problem data Data value on some position；

4th acquisition module, is used for obtaining revised data, as sample data.

15. devices according to claim 12, it is characterised in that described first model building module Including:

Kernel function computing module, for described sample data and the first input sample data being substituted into formula:

$K(x_i,x_j)=\exp (-\frac{\left|\right|x_i-x_j\left|\right|}{{\mathrm{\σ}}^2})$

Calculate the kernel function of Method Using Relevance Vector Machine, wherein x_iFor the input vector of described sample data, x_jFor The input vector of described first input sample data, σ is the kernel function width of Method Using Relevance Vector Machine；

Submodule set up by first model, is used for setting up described Method Using Relevance Vector Machine forecast model, described be correlated with to Amount machine forecast model comprises the kernel function of described Method Using Relevance Vector Machine, described sample data, described first input Sample data and the first output sample data.

16. devices according to claim 15, it is characterised in that described second model building module Including:

First sort module, for described first input sample data is classified with the first output sample data, Generate the first training sample data and the first test samples data；

First training module, for described first training sample utilizing described first sort module to sort out Data, are trained described Method Using Relevance Vector Machine forecast model, set up Method Using Relevance Vector Machine training pattern；

First inspection module, for described first test samples utilizing described first sort module to sort out Data, verify described Method Using Relevance Vector Machine training pattern, determine described Method Using Relevance Vector Machine training pattern.

17. devices according to claim 12, it is characterised in that described first optimizes module includes:

Parameter optimization module, for optimizing the core of the Method Using Relevance Vector Machine in described Method Using Relevance Vector Machine training pattern Function widths σ, it is thus achieved that the kernel function width cs of optimum Method Using Relevance Vector Machine；

5th acquisition module is corresponding with the kernel function width cs of described optimum Method Using Relevance Vector Machine for obtaining Method Using Relevance Vector Machine training pattern, as optimum Method Using Relevance Vector Machine training pattern.

18. devices according to claim 17, it is characterised in that described parameter optimization module includes:

Mapping block, individual for the kernel function width cs needing the described Method Using Relevance Vector Machine optimized is mapped as Voxel vector；

Population foundation module, for utilizing the individual voxel vector being mapped to, sets up and initializes population；

6th acquisition module, for obtaining the maximum iteration time of setting, TSP question factor F initial value With adaptive crossover mutation CR initial value；

TSP question module, is used for utilizing described TSP question factor F, to described initialization population and Current individual vector carries out TSP question operation, it is thus achieved that variation vector；

Self adaptation Cross module, is used for utilizing adaptive crossover mutation CR, current to described variation vector Individual voxel vector carries out self adaptation cross selection, it is thus achieved that intersection vector；

Fitness value acquisition module, for substituting into described intersection vector respectively with described current individual vector Fitness function, it is thus achieved that the fitness value of intersection vector and the fitness value of current individual vector；

First comparison module, for fitness value and the described current individual vector of relatively described intersection vector The size of fitness value, it is thus achieved that the vector that fitness value is less；

Judge module, for judging the number of times of current iteration；

TSP question module, for judging the number of times of current iteration less than maximum repeatedly at described judge module During generation number, vector less for the fitness value of described first comparison module acquisition is chosen as current individual Vector, utilizes described TSP question factor F, carries out described initialization population and current individual vector certainly Adequate variation operates, it is thus achieved that variation vector；

Output module, for judging that at described judge module the number of times of current iteration is equal to maximum iteration time Time, the minimum fitness value of output；

7th acquisition module, for obtaining the core letter of the Method Using Relevance Vector Machine corresponding with described minimum fitness value Number width cs, as the kernel function width cs of optimum Method Using Relevance Vector Machine.

19. devices according to claim 12, it is characterised in that described 3rd model building module Including:

Arma modeling sets up module, is used for setting up initial ARMA model；

GARCH model building module, is used for setting up initial GARCH model；

ARMA fitting module, for being fitted described initial ARMA model, determines described The exponent number of arma modeling；

GARCH fitting module, for being fitted described initial GARCH model, determines described The exponent number of GARCH model；

Submodule set up by 3rd model, is used for utilizing described final arma modeling and final GARCH mould Described first relative error sequence is fitted by type, sets up GARCH error prediction model, wherein, Described final arma modeling corresponding to the exponent number of described final arma modeling determined, described finally The exponent number of the GARCH model described final GARCH model corresponding to determining.

20. devices according to claim 19, it is characterised in that described 3rd pretreatment module bag Include:

Second sort module, for described first relative error sequence being classified, generates the second training Sample data and the second test samples data；

Error prediction models fitting module, is used for utilize described second sort module to sort out described second Training sample data, are fitted described GARCH error prediction model, it is thus achieved that the GARCH of optimization Error prediction model；

Second authentication module, for described second test samples utilizing described second sort module to sort out, The GARCH error prediction model of described optimization is verified, it is thus achieved that validation value；

Second comparison module, for the size of relatively described validation value with preassigned；

8th acquisition module, for being less than or equal to described predetermined standard time at described validation value, it is thus achieved that optimum GARCH error prediction model.

21. devices according to claim 20, it is characterised in that at described second comparison module ratio After the size of more described validation value and preassigned, described ARMA fitting module is additionally operable to:

At described validation value more than described predetermined standard time, described initial ARMA model is fitted, Determine the exponent number of described arma modeling.

22. devices according to claim 12, it is characterised in that described second correcting module includes:

Second computing module, for the predictive value of described second relative error sequence being substituted into formula:

$\widehat{RE}={\widehat{RE}}_p\×RP$

It is calculated the residual sequence predictive value of correspondenceWherein,It it is the second relative error sequence Predictive value, RP is the rated power of blower fan；

4th correcting module, for being added with described actual prediction value by described residual sequence predictive value, obtains Obtain final predictive value.