CN104820869A - Model-switching-mechanism-contained prediction method of wind power ramp event - Google Patents

Abstract

The invention relates to a wind power prediction method, especially to a model-switching-mechanism-contained prediction method of a wind power ramp event. According to the model-switching-mechanism-contained prediction method and the ramp persisting method considering a power grid operation state, the method comprises the following steps: step one, completing pretreatment of wind power data and calculation of an applicable length of a local model; step two, on the basis of the local segment division, completing identification of an optimized prediction model and completing training of a model switching mechanism based on the identification result; step three, completing long-term wind power prediction based on combination of a prediction model base and the model switching mechanism; and step four, on the basis of combination of ramp event definition and an actual power grid operation state, carrying out ramp event determination on a predicted wind power result. According to the invention, accuracy of ramp event prediction can be guaranteed based on wind power prediction with high precision; and thus the reliable guarantee is provided for a power system to formulate correct ramp control measures.

Description

A kind of climbing of the wind power containing models switching mechanism event prediction method

Technical field

Patent of the present invention relates to a kind of wind power forecasting method, especially relates to a kind of wind power containing forecast model handover mechanism climbing event prediction method.

Background technology

Day by day serious along with energy crisis, is all greatly developing regenerative resource in worldwide.At present, wind-powered electricity generation is as a kind of abundant, and the resource that can develop, progressively improves in the permeability of electrical network.But the impact being subject to air motion due to wind energy has randomness and undulatory property, especially in the Wind Power Development pattern of extensive, high concentration degree, the amplitude wind-powered electricity generation fluctuation in the short time is huge to electrical network harm, to climb event the most very.In order to instruct traffic department to formulate effective generation schedule in advance, the wind-powered electricity generation event of the significant threat such as prevention climbing, and immediately take to control hazard mitigation measure, and ensure the safety of electric system, stable operation, very important to the look-ahead of wind power climbing event.

Foreign scholar to the research starting of wind power climbing event comparatively early.Extensive, harm is a larger power descending event is there occurs at texas,U.S due to 2008, cause the electric reliability council (Electric Reliability Council of Texas, ERCOT) serious economic loss, therefore this problem is obtained and payes attention to widely.At present the definition of climbing event, classification and forecasting research are mainly concentrated on to the research of wind power climbing event.Wherein forecasting research still Shortcomings, the method for main employing on the basis of wind power prediction, carries out climbing detect and differentiate, the definition research therefore for climbing event is the prerequisite of carrying out climbing event prediction.Several numerical value definition conventional at present, relate generally to the key characters such as the change amplitude of climbing event, duration and climbing rate.Doctor Sevlian of Stanford University proposes the climbing incident Detection Algorithm adopting marking mode on this basis.And the object of event classification of climbing is to predict the outcome according to climbing to judge climbing event type, thus select suitable effective control, control measures.

Climbing event is the changed power event be defined in a period of time.In order to the prediction realizing climbing event needs the problem of consideration two aspect, one is the realization of long event prediction, and two is how effectively from the middle extraction wind power climbing affair character amount that predicts the outcome.Mainly comprise physical model and statistical model two class about wind power method at present, wherein physical model can realize longer prediction, but precision of prediction is not high; And statistical model has higher precision of prediction for the wind power prediction of short-term, but error can increase along with the increase of prediction duration.In order to solve the forecasting problem of climbing event, the power prediction result of long period is needed to meet the requirement of climbing event prediction, effectively catch climbing affair character for ensureing simultaneously, the precision of prediction of wind power prediction need be guaranteed, thus ensures effectively to realize the research of climbing event prediction.

Summary of the invention

The object of patent of the present invention is for above-mentioned present situation, propose a kind of can by long-term wind power prediction by climbing event demand convert to continuous print local wind power prediction, and the wind power prediction of degree of precision is completed by choosing optimum Short-term Forecasting Model, the new method of wind power climbing event prediction is completed in conjunction with climbing definition and electric network state.

Above-mentioned technical matters of the present invention is achieved through the following technical solutions:

Containing a wind power climbing event prediction method for forecast model handover mechanism, it is characterized in that: comprise the following steps,

Step 1: with the wind power data of history for research object, detect historied climbing event, and its duration is added up, find out can meet most climbing duration time window as the applicable duration of local increment; , raw data is analyzed meanwhile, completes the work such as data prediction and phase space reconfiguration, comprise following sub-step,

Step 1.1, first carries out the process such as bad value, missing value to original wind power data sequence, to ensure the completeness and efficiency of data, eliminates unnecessary noise by filtering process simultaneously, obtains clean sequence; On the other hand, by the statistical study to history climbing incident duration, the applicable duration of local increment is provided;

Step 1.2, the conveniently foundation of forecast model, first carry out Space Reconstruction to data; If original data sequence is { x _n, then get two observed quantity x of interval time interval τ _nand x _{n+ τ}, through type one calculates association relationship, gets Mutual information entropy first minimum point as the time delay reconstructed;

$I\left(\mathrm{\τ}\right)=\underset{n=1}{\overset{N}{\mathrm{\Σ}}}P(x_n,x_{n+\mathrm{\τ}})\log \left[\frac{P(x_n,x_{n+\mathrm{\τ}})}{P\left(x_n\right)P\left(x_{n+\mathrm{\τ}}\right)}\right]$ Formula one

Step 1.3, supposes that the reconstruct Embedded dimensions of original series is m, then calculate distance between phase point to the sequence after reconstruct according to formula two, if meet formula three, four, then gets the Embedded dimensions of m as reconstruct;

${\left|\right|x_{\mathrm{\η}\left(n\right)}-x_n\left|\right|}_2^{(m+1)}=\sqrt{{\left({\left|\right|x_{\mathrm{\η}\left(n\right)}-x_n\left|\right|}_2^{\left(m\right)}\right)}^2+{(x_{\mathrm{\η}\left(n\right)+\mathrm{m\τ}}-x_{n+\mathrm{m\τ}})}^2}$ Formula two

$\frac{|x_{\mathrm{\η}\left(n\right)+\mathrm{m\τ}}-x_{n+\mathrm{m\τ}}|}{{\left|\right|x_{\mathrm{\η}\left(n\right)}-x_n\left|\right|}_2^{(m+1)}}\≥R_{\mathrm{th}}$ Formula three

$\frac{{\left|\right|x_{\mathrm{\η}\left(n\right)}-x_n\left|\right|}_2^{(m+1)}}{\mathrm{\σ}}>A_{\mathrm{th}}$ Formula four

Step 1.4, according to the time delay obtained in step 1.2 and step 1.3 and Embedded dimensions, completes the phase space reconfiguration of sequence according to formula five; With the data after reconstruct for object, the Data Placement of history is continuous print event section one by one by the partial model prove-in length of trying to achieve according to step 1.1, to study as base unit when carrying out model training and prediction;

X _n=(x _n, x _{n+ τ}... x _{n+ (m-1) τ}) ∈ R ^m, n=1,2..., N ₀=N-(m-1) τ formula five

Step 2: based on multivariable chaotic forecast model, in order to realize longer power prediction, the weather data provided in conjunction with numerical weather forecast sets up multivariable chaotic prediction model; Consider that the physical background of different climbing event is different simultaneously, and be the precision of prediction improving model further, Classical forecast model basis adds different IPs function and forms forecasting model database, complete the Model Identification to historical events section according to given model bank, training obtains optimum local increment; Comprise following sub-step,

Step 2.1, according to the Local-region Linear Prediction model of traditional chaos time sequence, introduces Nonlinear Mapping, constructs such as formula the fundamental forecasting model shown in six;

$y_{n+1}=\underset{i=0}{\mathrm{\Σ}}{\mathrm{\ω}}^T\mathrm{\Φ}\left(x_{n-i}\right)+e$ Formula six

In formula, Φ represents nonlinear transformation, and e represents residual error item;

Step 2.2, in order to the singularity of event changed power of climbing under reflecting different physical background, different kernel functions is adopted to represent to nonlinear transformation Φ, choose 3 kinds of conventional kernel functions: Polynomial kernel function, Radial basis kernel function and Sigmoid kernel function, such as formula seven, thus form the model bank of local increment;

$\left\{\begin{array}{c}{K}_{\mathrm{ploy}}(x,y)={[<x,y>+1]}^q\\ K_{\mathrm{RBF}}(x,y)=\exp (-{\left|\right|x-y\left|\right|}^2/{\mathrm{\&sigma;}}^2)\\ K_{\mathrm{Sigmoid}}(x,y)=\tanh (v<x,y>+e)\end{array}\right.$ Formula seven

Step 3: according to model bank obtained above, respectively model training is carried out to each independently event section, obtain the optimum prediction model in each section, then train according to the corresponding relation of raw data and optimization model, the handover mechanism of different forecast model can be obtained;

Step 4: the model bank in integrating step 2 and the mechanism of the models switching in step 3, complete the prediction of longer wind power prediction; The climbing event detection of predicted data has finally been defined according to the numerical value of climbing event, consider that actual that climbing harm occurs is relevant with the state of electrical network, therefore complete further combined with the actual motion state of electrical network and wind power is climbed the prediction of event and judgement;

$\left\{\begin{array}{c}{K}_{\mathrm{ploy}}(x,y)={[<x,y>+1]}^q\\ K_{\mathrm{RBF}}(x,y)=\exp (-{\left|\right|x-y\left|\right|}^2/{\mathrm{\&sigma;}}^2)\\ K_{\mathrm{Sigmoid}}(x,y)=\tanh (v<x,y>+e)\end{array}\right.$ Formula 11.

In the above-mentioned wind power containing forecast model handover mechanism climbing event prediction method, described step 3 mainly comprises following sub-step,

Step 3.1, is numbered the model in model bank in step 2, when carrying out model training to continuous events section, can obtain the numbered sequence { c of optimization model _n, in conjunction with original event segment data, if the training set of formula eight can be set up;

$T=\{(x_1,c_1),(x_2,c_2),...(x_{N_0},c_{N_0})\}\&Element;{(R^m\&times;C)}^{N_0}$ Formula eight

Step 3.2, above-mentioned training set can be regarded as containing multi-class categorized data set, and wherein a kind represents a kind of forecast model, can set up multi-class support vector cassification model training sorter to this; By classification method, the individual basic SVM model of k-1 need be set up for the data containing k kind, as shown in the formula:

$\begin{array}{cc}\underset{\mathrm{\&omega;},b}{\min }& \frac{1}{2}{\left|\right|\mathrm{\&omega;}\left|\right|}^2+C\underset{i=1}{\overset{N_0}{\mathrm{\&Sigma;}}}{\mathrm{\&xi;}}_i\end{array}$

s.t.y _i(<ω·x _i>+b)≥1-ξ _i

ξ _i>=0, i=1,2...N ₀formula nine

In formula, ω, b are classifier parameters; ξ is slack variable; y _nrepresent the class variable redefined, when a certain classification is main study subject, then y _n=1, remaining class label y _n=-1; Solve the sorter that can obtain 3 class models storehouses in step 2 by SVM model, this sorter is the handover mechanism of local increment.

In the above-mentioned wind power containing forecast model handover mechanism climbing event prediction method, described step 4 mainly comprises following sub-step,

Step 4.1, in units of the partial model prove-in length obtained, in conjunction with the result of numerical weather forecast, completes the local wind power prediction in the short time in step 1; Simultaneously according to models switching mechanism, carry out the switching of different event section forecast model, the prediction of comprehensive multiple continuous events section, completes long-term wind power prediction;

Step 4.2, according to given wind power climbing event definition, such as formula ten, detects in predicting the outcome the climbing event meeting the definition of climbing numerical value; Consider the plan of electrical network actual schedule, as situations such as unit shut algorithms, on the basis adding operation of power networks state, detect the climbing event with harmfulness, to instruct electric system to carry out scheduling and controlling;

$\frac{|P(t+\mathrm{\&Delta;t})-P\left(t\right)|}{\mathrm{\&Delta;t}}>R_{\mathrm{val}}$ Formula ten

This definition thinks that the be separated by Error Absolute Value of wind power amplitude in two moment of length Δ t and the ratio of time interval Δ t is greater than a certain given peak power climbing rate R _val, then a wind power climbing is called.

Therefore, the present invention can reach following beneficial effect: 1, compared with traditional prediction method, and due to the singularity of event of climbing, in the process considering event prediction, the present invention, in conjunction with the architectural feature of climbing event, proposes the Forecasting Methodology in units of event section.Simultaneously by the wind power change in predicted events section and then the method analyzing climbing event, traditional data point prediction is become to solve the predictive conversion of climbing event; 2, in the process realizing wind power long-term forecasting, the short-term forecasting precision not only considering statistical model is high, the advantage of the simultaneously practicable long-term forecasting of Integrative Logistic System Model, in conjunction with both the long-term Forecasting Methodology containing models switching mechanism is proposed, the wind power prediction of degree of precision can be realized; 3, effectively to climb event information to electric system to improve, the present invention not only needs to define in conjunction with the numerical value of climbing event, considers the running status of actual electric network simultaneously, comprehensively provides the following possibility that climbing event occurs.

Accompanying drawing explanation

Fig. 1 is the wind power containing models switching mechanism provided by the invention climbing predicted entire thinking figure.

Fig. 2 a is history provided by the invention climbing incident duration probability distribution graph.

Fig. 2 b is history provided by the invention climbing event accumulated time probability distribution graph.

Fig. 3 is the long-term wind power prediction result figure containing models switching mechanism provided by the invention.

Fig. 4 is wind power provided by the invention climbing event prediction result figure.

Embodiment

Below by embodiment, and by reference to the accompanying drawings, technical scheme of the present invention is further described.

The present invention relates to a kind of wind power containing forecast model handover mechanism climbing Forecasting Methodology.Because the safety and stability of wind power climbing event to electric system brings serious threat, in order to prevention in advance and effective control harmfulness climbing event, very important to the look-ahead of climbing event.

Embodiment

Technical scheme of the present invention mainly comprises two parts, namely utilize complete the long-term forecasting of wind power containing the Forecasting Methodology of models switching mechanism and combine climbing definition, prediction that operation of power networks state carries out climbing event.

One, principle of the present invention is first introduced.

Step 1: with the wind power data of history for research object, detect historied climbing event, and its duration is added up, find out can meet most climbing duration time window as the applicable duration of local increment., raw data is analyzed meanwhile, complete the work such as data prediction and phase space reconfiguration.

Carrying out in historical data analysis process, above-mentioned steps 1 mainly comprises following sub-step,

Step 1.1, first carries out the process such as bad value, missing value to original wind power data sequence, to ensure the completeness and efficiency of data, eliminates unnecessary noise by filtering process simultaneously, obtains clean sequence.On the other hand, by the statistical study to history climbing incident duration, the applicable duration of local increment is provided.

Step 1.2, the conveniently foundation of forecast model, first carry out Space Reconstruction to data.If original data sequence is { x _n, then get two observed quantity x of interval time interval τ _nand x _{n+ τ}, through type (1) calculates association relationship, gets Mutual information entropy first minimum point as the time delay reconstructed.

$I\left(\mathrm{\&tau;}\right)=\underset{n=1}{\overset{N}{\mathrm{\&Sigma;}}}P(x_n,x_{n+\mathrm{\&tau;}})\log \left[\frac{P(x_n,x_{n+\mathrm{\&tau;}})}{P\left(x_n\right)P\left(x_{n+\mathrm{\&tau;}}\right)}\right]---\left(1\right)$

Step 1.3, supposes that the reconstruct Embedded dimensions of original series is m, then calculate distance between phase point to the sequence after reconstruct according to formula (2), if meet formula (3-4), then get the Embedded dimensions of m as reconstruct.

${\left|\right|x_{\mathrm{\&eta;}\left(n\right)}-x_n\left|\right|}_2^{(m+1)}=\sqrt{{\left({\left|\right|x_{\mathrm{\&eta;}\left(n\right)}-x_n\left|\right|}_2^{\left(m\right)}\right)}^2+{(x_{\mathrm{\&eta;}\left(n\right)+\mathrm{m\&tau;}}-x_{n+\mathrm{m\&tau;}})}^2}---\left(2\right)$

$\frac{|x_{\mathrm{\&eta;}\left(n\right)+\mathrm{m\&tau;}}-x_{n+\mathrm{m\&tau;}}|}{{\left|\right|x_{\mathrm{\&eta;}\left(n\right)}-x_n\left|\right|}_2^{(m+1)}}\&GreaterEqual;R_{\mathrm{th}}---\left(3\right)$

$\frac{{\left|\right|x_{\mathrm{\&eta;}\left(n\right)}-x_n\left|\right|}_2^{(m+1)}}{\mathrm{\&sigma;}}>A_{\mathrm{th}}---\left(4\right)$

Step 1.4, according to the time delay obtained in step 1.2 and step 1.3 and Embedded dimensions, completes the phase space reconfiguration of sequence according to formula (5).With the data after reconstruct for object, the Data Placement of history is continuous print event section one by one by the partial model prove-in length of trying to achieve according to step 1.1, so that when carrying out model training and prediction to study as base unit.

x _n＝(x _n,x _n+τ,...x _n+(m-1)τ)∈R ^m,n＝1,2...,N ₀＝N-(m-1)τ (5)

Step 2: based on multivariable chaotic forecast model, in order to realize longer power prediction, the weather data provided in conjunction with numerical weather forecast sets up multivariable chaotic prediction model.Consider that the physical background of different climbing event is different simultaneously, and be the precision of prediction improving model further, Classical forecast model basis adds different IPs function and forms forecasting model database, complete the Model Identification to historical events section according to given model bank, training obtains optimum local increment.

In the process of establishing of forecast model, described step 2 comprises following sub-step,

Step 2.1, according to the Local-region Linear Prediction model of traditional chaos time sequence, introduces Nonlinear Mapping, constructs such as formula the fundamental forecasting model shown in (6).

$y_{n+1}=\underset{i=0}{\mathrm{\&Sigma;}}{\mathrm{\&omega;}}^T\mathrm{\&Phi;}\left(x_{n-i}\right)+e---\left(6\right)$

In formula, Φ represents nonlinear transformation, and e represents residual error item.

Step 2.2, in order to the singularity of event changed power of climbing under reflecting different physical background, different kernel functions is adopted to represent to nonlinear transformation Φ, choose 3 kinds of conventional kernel functions: Polynomial kernel function, Radial basis kernel function (RBF) and Sigmoid kernel function, such as formula (7), thus form the model bank of local increment.

$\left\{\begin{array}{c}{K}_{\mathrm{ploy}}(x,y)={[<x,y>+1]}^q\\ K_{\mathrm{RBF}}(x,y)=\exp (-{\left|\right|x-y\left|\right|}^2/{\mathrm{\&sigma;}}^2)\\ K_{\mathrm{Sigmoid}}(x,y)=\tanh (v<x,y>+e)\end{array}\right.---\left(7\right)$

Step 3: according to model bank obtained above, respectively model training is carried out to each independently event section, obtain the optimum prediction model in each section, then train according to the corresponding relation of raw data and optimization model, the handover mechanism of different forecast model can be obtained.

Ask in process at handover mechanism, above-mentioned steps 3 mainly comprises following sub-step,

Step 3.1, is numbered the model in model bank in step 2, when carrying out model training to continuous events section, can obtain the numbered sequence { c of optimization model _n, in conjunction with original event segment data, if the training set of formula (8) can be set up.

$T=\{(x_1,c_1),(x_2,c_2),...(x_{N_0},c_{N_0})\}\&Element;{(R^m\&times;C)}^{N_0}---\left(8\right)$

Step 3.2, above-mentioned training set can be regarded as containing multi-class categorized data set, and wherein a kind represents a kind of forecast model, can set up multi-class support vector machine (SVM) disaggregated model training classifier to this.By classification method, the individual basic SVM model of k-1 need be set up for the data containing k kind, as shown in the formula:

$\begin{array}{cc}\underset{\mathrm{\&omega;},b}{\min }& \frac{1}{2}{\left|\right|\mathrm{\&omega;}\left|\right|}^2+C\underset{i=1}{\overset{N_0}{\mathrm{\&Sigma;}}}{\mathrm{\&xi;}}_i\end{array}$

s.t.y _i(<ω·x _i>+b)≥1-ξ _i

ξ _i≥0,i＝1,2...N ₀(9)

In formula, ω, b are classifier parameters; ξ is slack variable; y _nrepresent the class variable redefined, when a certain classification is main study subject, then y _n=1, remaining class label y _n=-1.Solve the sorter that can obtain 3 class models storehouses in step 2 by SVM model, this sorter is the handover mechanism of local increment.

Step 4: the model bank in integrating step 2 and the mechanism of the models switching in step 3, complete the prediction of longer wind power prediction.The climbing event detection of predicted data has finally been defined according to the numerical value of climbing event, consider that actual that climbing harm occurs is relevant with the state of electrical network, therefore complete further combined with the actual motion state of electrical network and wind power is climbed the prediction of event and judgement.

In above-mentioned wind power climbing forecasting process, step 4 mainly comprises following sub-step,

Step 4.1, in units of the partial model prove-in length obtained, in conjunction with the result of numerical weather forecast, completes the local wind power prediction in the short time in step 1.Simultaneously according to models switching mechanism, carry out the switching of different event section forecast model, the prediction of comprehensive multiple continuous events section, completes long-term wind power prediction.

Step 4.2, according to given wind power climbing event definition, such as formula (10), detects in predicting the outcome the climbing event meeting the definition of climbing numerical value.Consider the plan of electrical network actual schedule, as situations such as unit shut algorithms, on the basis adding operation of power networks state, detect the climbing event with harmfulness, to instruct electric system to carry out scheduling and controlling.

$\frac{|P(t+\mathrm{\&Delta;t})-P\left(t\right)|}{\mathrm{\&Delta;t}}>R_{\mathrm{val}}---\left(10\right)$

This definition thinks that the be separated by Error Absolute Value of wind power amplitude in two moment of length Δ t and the ratio of time interval Δ t is greater than a certain given peak power climbing rate R _val, then a wind power climbing is called.

Two, choosing the U.S. BPA control area annual wind power output data of 2013 is below sample set, and wherein the temporal resolution of wind power data is every 5 minutes sampled points, describes technical solution of the present invention in detail by reference to the accompanying drawings with case study on implementation.

With original wind power data for object, the pretreatment work of complete paired data, as process such as bad value, missing value and denoisings, then completes choosing of Parameters for Phase Space Reconstruction to clean effectively carrying out according to formula (1-4), and completes phase space reconfiguration.On the other hand, detect the climbing event occurred in historical process, and statistical study is carried out to time of climb, as accompanying drawing 2.The time window that the time of climb getting satisfied 95% requires is as local increment prove-in length.

Set up nonlinear forecast model according to traditional chaotic prediction model, and get the base configuration model bank of 3 kinds of typical kernel functions as nonlinear transformation.Raw data is divided according to model prove-in length, obtains continuous print event section sequence, optimum prediction model is identified to each data segment.Regard the corresponding relation of forecast model and data as assorting process, then can be obtained the handover mechanism of model by SVM model training.Unified model storehouse and models switching mechanism are predicted test set data, and result as shown in Figure 3.

Predicting the outcome according to wind power, foundation is defined as with the climbing of formula (10), total volume due to sample set data is 4000MW, then the threshold value that the changed power getting 50% installed capacity in 4 hours judges as climbing, completes climbing event Preliminary detection.In order to adapt to the demand of practical power systems, consider that operation of power networks state completes the judgement to climbing event further, thus complete the prediction to wind power climbing event.

Above embodiment is only in order to illustrate that technical scheme of the present invention is not intended to limit, those skilled in the art to saying that the instantiation example described makes various amendment or compensation, but can can't depart from the scope that claims of the present invention define.

Claims (3)

1., containing a wind power climbing event prediction method for forecast model handover mechanism, it is characterized in that: comprise the following steps,

Step 1: with the wind power data of history for research object, detect historied climbing event, and its duration is added up, find out can meet most climbing duration time window as the applicable duration of local increment; , raw data is analyzed meanwhile, completes the work such as data prediction and phase space reconfiguration, comprise following sub-step,

Step 1.1, first carries out the process such as bad value, missing value to original wind power data sequence, to ensure the completeness and efficiency of data, eliminates unnecessary noise by filtering process simultaneously, obtains clean sequence; On the other hand, by the statistical study to history climbing incident duration, the applicable duration of local increment is provided;

Step 1.2, the conveniently foundation of forecast model, first carry out Space Reconstruction to data; If original data sequence is { x _n, then get two observed quantity x of interval time interval τ _nand x _{n+ τ}, through type one calculates association relationship, gets Mutual information entropy first minimum point as the time delay reconstructed;

$I\left(\mathrm{\&tau;}\right)=\underset{n=1}{\overset{N}{\mathrm{\&Sigma;}}}P(x_n,x_{n+\mathrm{\&tau;}})\log \left[\frac{P(x_n,x_{n+\mathrm{\&tau;}})}{P\left(x_n\right)P\left(x_{n+\mathrm{\&tau;}}\right)}\right]$ Formula one

Step 1.3, supposes that the reconstruct Embedded dimensions of original series is m, then calculate distance between phase point to the sequence after reconstruct according to formula two, if meet formula three to formula four, then gets the Embedded dimensions of m as reconstruct;

${\left|\right|x_{\mathrm{\&eta;}\left(n\right)}-x_n\left|\right|}_2^{(m+)}=\sqrt{{\left({\left|\right|x_{\mathrm{\&eta;}\left(n\right)}-x_n\left|\right|}_2^{\left(m\right)}\right)}^2+{(x_{\mathrm{\&eta;}\left(n\right)+\mathrm{m\&tau;}}-x_{n+\mathrm{m\&tau;}})}^2}$ Formula two

$\frac{|x_{\mathrm{\&eta;}\left(n\right)+\mathrm{m\&tau;}}-x_{n+\mathrm{m\&tau;}}|}{{\left|\right|x_{\mathrm{\&eta;}\left(n\right)}-x_n\left|\right|}_2^{(m+1)}}\&GreaterEqual;R_{\mathrm{th}}$ Formula three

$\frac{{\left|\right|x_{\mathrm{\&eta;}\left(n\right)}-x_n\left|\right|}_2^{(m+1)}}{\mathrm{\&sigma;}}>A_{\mathrm{th}}$ Formula four

Step 1.4, according to the time delay obtained in step 1.2 and step 1.3 and Embedded dimensions, completes the phase space reconfiguration of sequence according to formula five; With the data after reconstruct for object, the Data Placement of history is continuous print event section one by one by the partial model prove-in length of trying to achieve according to step 1.1, to study as base unit when carrying out model training and prediction;

X _n=(x _n, x _{n+ τ}... x _{n+ (m-1) τ}) ∈ R ^m, n=1,2 ..., N ₀=N-(m-1) τ formula five

Step 2: based on multivariable chaotic forecast model, in order to realize longer power prediction, the weather data provided in conjunction with numerical weather forecast sets up multivariable chaotic prediction model; Consider that the physical background of different climbing event is different simultaneously, and be the precision of prediction improving model further, Classical forecast model basis adds different IPs function and forms forecasting model database, complete the Model Identification to historical events section according to given model bank, training obtains optimum local increment; Comprise following sub-step,

Step 2.1, according to the Local-region Linear Prediction model of traditional chaos time sequence, introduces Nonlinear Mapping, constructs such as formula the fundamental forecasting model shown in six;

$y_{n+1}=\underset{i=0}{\mathrm{\&Sigma;}}{\mathrm{\&omega;}}^T\mathrm{\&Phi;}\left(x_{n-i}\right)+e$ Formula six

In formula, Φ represents nonlinear transformation, and e represents residual error item;

Step 2.2, in order to the singularity of event changed power of climbing under reflecting different physical background, different kernel functions is adopted to represent to nonlinear transformation Φ, choose 3 kinds of conventional kernel functions: Polynomial kernel function, Radial basis kernel function and Sigmoid kernel function, such as formula seven, thus form the model bank of local increment;

$\left\{\begin{array}{c}{K}_{\mathrm{ploy}}(x,y)={[\&lang;x,y\&rang;+1]}^q\\ K_{\mathrm{RBF}}(x,y)=\exp (-{\left|\right|x-y\left|\right|}^2/{\mathrm{\&sigma;}}^2)\\ K_{\mathrm{Sigmoid}}(x,y)=\tanh (v\&lang;x,y\&rang;+e)\end{array}\right.$ Formula seven

Step 3: according to model bank obtained above, respectively model training is carried out to each independently event section, obtain the optimum prediction model in each section, then train according to the corresponding relation of raw data and optimization model, the handover mechanism of different forecast model can be obtained;

Step 4: the model bank in integrating step 2 and the mechanism of the models switching in step 3, complete the prediction of longer wind power prediction; The climbing event detection of predicted data has finally been defined according to the numerical value of climbing event, consider that actual that climbing harm occurs is relevant with the state of electrical network, therefore complete further combined with the actual motion state of electrical network and wind power is climbed the prediction of event and judgement;

$\left\{\begin{array}{c}{K}_{\mathrm{ploy}}(x,y)={[\&lang;x,y\&rang;+1]}^q\\ K_{\mathrm{RBF}}(x,y)=\exp (-{\left|\right|x-y\left|\right|}^2/{\mathrm{\&sigma;}}^2)\\ K_{\mathrm{Sigmoid}}(x,y)=\tanh (v\&lang;x,y\&rang;+e)\end{array}\right.$ Formula 11.

2. a kind of climbing of the wind power containing forecast model handover mechanism event prediction method according to claim 1, is characterized in that: described step 3 mainly comprises following sub-step,

Step 3.1, is numbered the model in model bank in step 2, when carrying out model training to continuous events section, can obtain the numbered sequence { c of optimization model _n, in conjunction with original event segment data, if the training set of formula eight can be set up;

$T\{(x_1,c_1),(x_2,c_2),...(x_{N_0},c_{N_0})\}\&Element;{(R^m\&times;C)}^{N_0}$ Formula eight

Step 3.2, above-mentioned training set can be regarded as containing multi-class categorized data set, and wherein a kind represents a kind of forecast model, can set up multi-class support vector cassification model training sorter to this; By classification method, the individual basic SVM model of k-1 need be set up for the data containing k kind, as shown in the formula:

$\begin{array}{cc}\underset{\mathrm{\&omega;},b}{\min }& \frac{1}{2}{\left|\right|\mathrm{\&omega;}\left|\right|}^2+C\underset{i=1}{\overset{N_0}{\mathrm{\&Sigma;}}}{\mathrm{\&xi;}}_i\end{array}$

$\begin{array}{cc}s.t.& \begin{array}{c}{y}_i(\&lang;\mathrm{\&omega;}\&CenterDot;x_i\&rang;+b)\&GreaterEqual;1-{\mathrm{\&xi;}}_i\\ {\mathrm{\&xi;}}_i\&GreaterEqual;0,i=\mathrm{1,2}...N_0\end{array}\end{array}$ Formula nine

In formula, ω, b are classifier parameters; ξ is slack variable; y _nrepresent the class variable redefined, when a certain classification is main study subject, then y _n=1, remaining class label y _n=-1; Solve the sorter that can obtain 3 class models storehouses in step 2 by SVM model, this sorter is the handover mechanism of local increment.

3. a kind of climbing of the wind power containing forecast model handover mechanism event prediction method according to claim 1, is characterized in that: described step 4 mainly comprises following sub-step,

Step 4.1, in units of the partial model prove-in length obtained, in conjunction with the result of numerical weather forecast, completes the local wind power prediction in the short time in step 1; Simultaneously according to models switching mechanism, carry out the switching of different event section forecast model, the prediction of comprehensive multiple continuous events section, completes long-term wind power prediction;

Step 4.2, according to given wind power climbing event definition, such as formula ten, detects in predicting the outcome the climbing event meeting the definition of climbing numerical value; Consider the plan of electrical network actual schedule, as situations such as unit shut algorithms, on the basis adding operation of power networks state, detect the climbing event with harmfulness, to instruct electric system to carry out scheduling and controlling;

$\frac{|P(t+\mathrm{\&Delta;t})-P\left(t\right)|}{\mathrm{\&Delta;t}}>R_{\mathrm{val}}$ Formula ten

This definition thinks that the be separated by Error Absolute Value of wind power amplitude in two moment of length Δ t and the ratio of time interval Δ t is greater than a certain given peak power climbing rate R _val, then a wind power climbing is called.