CN105160434A - Wind power ramp event prediction method by adopting SVM to select forecasting model - Google Patents

Abstract

The invention discloses a wind power ramp event prediction method by utilizing an SVM to select a forecasting model. The prediction method comprises the following steps: dividing a first sample set into local data segments according to window length; establishing short-period wind power forecasting models of the local data segments by utilizing the SVM, and obtaining a short-period wind power forecasting model library; carrying out cluster fusion on the models in the short-period wind power forecasting model library by utilizing a Ward clustering method, adopting an SVM classification model to express a forecasting model selection mechanism, and training the forecasting model selection mechanism by utilizing a second sample set; and obtaining short-period wind power forecast data according to the forecasting model selection mechanism, wherein the time-continuous short-period wind power forecast data forms long-period wind power forecast data, and carrying out ramp event prediction according to the long-period wind power forecast data. The wind power ramp event prediction method can ensure higher-precision wind power long-period forecast, and furthermore, ensures accuracy of the ramp event prediction.

Description

SVM is adopted to choose the wind power climbing event prediction method of forecast model

Technical field

The invention belongs to wind power prediction technical field, particularly relate to a kind of wind power climbing event prediction method that SVM of employing chooses forecast model.

Background technology

In recent years, under the wind energy turbine set development trend of extensive in China, high concentration degree, the adverse effect that the randomness of wind-resources and undulatory property are brought to electric system is also more and more obvious, wherein endangering maximum is wind power climbing event, there occurs an extensive power descending event as 2008 at texas,U.S.Climbing event refers to a generation class at short notice significantly wind power change events, and causes potential threat to the safety and stability of electric system and economical operation.In order to safeguard the safety and stability of electrical network, the harm reducing wind-powered electricity generation climbing event will be very necessary, wherein predict it is the important step of carrying out climbing Risk health behavior to wind power climbing event under a large amount of wind-powered electricity generation injects the situation of electrical network in future.

At present, for the prediction of climbing event, the main wind power prediction method adopted comprises physical model predicted method and statistical model predicted method two class.Wherein often adopt time series models, neural network model and supporting vector machine model etc. in statistical model predicted method, statistical model predicted method is higher for the wind power prediction precision of short-term, but long-term forecasting precision is poor.Physical model predicted method is usually based on numerical weather forecast (numericalweatherprediction; NWP) system; the method changes according to the wind speed in wind field overhead; go out following possible air speed value in conjunction with atmospheric physics equation inference, then obtain following wind power prediction value according to wind power curve.What NWP obtained predicts the outcome very useful for the long-run development trend holding wind-powered electricity generation, but for assurance wind power local feature Shortcomings.

Because climbing event is defined as long changed power event, its forecasting process not only will relate to long-term wind power prediction, simultaneously detail signal in a short time on assurance climbing feature, to analyze climbing impact very important.Therefore, consider statistical model predicted method and physical model predicted method advantage separately, very important for the wind power climbing event prediction realizing better estimated performance.

Summary of the invention

For the deficiency that prior art exists, the invention provides a kind of wind power climbing event prediction method that SVM of employing chooses forecast model.

Long-term wind power prediction is converted to the short-term wind-electricity power prediction of continuous print multiple local datas section by the present invention, the weather data provided based on NWP sets up support vector machine (SVM) Short-term Forecasting Model of wind power, adopt support vector machine training pattern selection mechanism simultaneously, instruct choosing of the support vector machine Short-term Forecasting Model of local data's section.

For solving the problems of the technologies described above, the present invention adopts following technical scheme:

Adopt SVM to choose the wind power climbing event prediction method of forecast model, comprising:

First sample set is divided into local data's section by S1:

Gather the first sample set and pre-service is carried out to the first sample set, according to window length, the first sample set after pre-service is divided into local data's section of Time Continuous, wherein, first sample set comprises the on-site history weather data of electrical network and the history wind power data with the time, and window length is the mean value of the history climbing incident duration of statistics;

S2 builds short-term wind-electricity power forecasting model database:

With weather data be input, wind power be export, adopt SVM represent short-term wind-electricity power forecast model; Adopt local data's section training short-term wind-electricity power forecast model, obtain the short-term wind-electricity power forecast model of local data's section, the short-term wind-electricity power forecast model of all local data segment forms short-term wind-electricity power forecasting model database;

S3 trains forecast model selection mechanism:

According to model parameter, adopting Ward clustering procedure to carry out Cluster-Fusion to model in short-term wind-electricity power forecasting model database, and mark the class number of each model, is the second sample set with history weather data in the class number of each model and each local data section; With class number be export, weather data is input, adopts svm classifier model representation forecast model selection mechanism, and adopts the second sample set training forecast model selection mechanism;

S4 climbs event prediction:

Adopt the window length described in S1 following weather data to be divided into local data's section of Time Continuous, with local data's section for input, obtain the short-term wind-electricity power predicted data of local data's section based on forecast model selection mechanism; The short-term wind-electricity power predicted data of local data's section of Time Continuous forms long-term wind power prediction data, carries out climbing event prediction according to long-term wind power prediction data.

In S2, the short-term wind-electricity power forecast model adopting SVM to represent is y=< ω x>+b, and wherein, ω and b is model parameter, and x represents meteorological index, and y represents wind power.

In S2, adopt local data's section training short-term wind-electricity power forecast model, be specially:

Structure based principle of minimization risk, the support vector regression be with by constructing hard ε solves the parameter of short-term wind-electricity power forecast model.

In S3, adopt Ward clustering procedure to carry out Cluster-Fusion to model in short-term wind-electricity power forecasting model database, be specially:

With the model parameter of model each in short-term wind-electricity power forecasting model database for cluster inputs, with Ward distance for cluster basis for estimation, Ward Cluster-Fusion is carried out to model each in short-term wind-electricity power forecasting model database, obtain the short-term wind-electricity power forecasting model database only comprising N number of representative model, N is cluster numbers.

In S3, the forecast model selection mechanism adopting svm classifier model representation is f (x)=sign (< ω x _i>+b), wherein, ω and b represents the parameter of each model in short-term wind-electricity power forecasting model database, x _irepresent the meteorological index of local data's section, f (x) represents class number.

Compared with prior art, the present invention has following features and beneficial effect:

(1) the present invention's weather data of utilizing NWP to provide and history wind power data set up SVM forecast model, substantially envisage the impact of meteorologic factor, effectively can improve precision of prediction.

(2) for meeting the long-term wind power requirement needed for climbing event, the present invention converts long-term wind power prediction the short-term forecasting of the many local segments of continuous print to, not only can realize long-term forecasting, compare traditional long-range forecast method, also significantly can improve precision of prediction.

(3) the present invention is in conjunction with weather data and the analysis of local segment forecast model, is trained the forecast model selection mechanism of local segment by SVM, thus the long-run development pattern of reflection wind-powered electricity generation, effectively can instruct the prediction of climbing event.

Accompanying drawing explanation

Fig. 1 is entire block diagram of the present invention;

Fig. 2 is the schematic diagram of SVM prediction model;

Fig. 3 is the basic svm classifier model schematic containing two kinds;

Fig. 4 is the Ward cluster analysis result figure of local segment forecast model;

Fig. 5 is wind power climbing event prediction result figure in embodiment, and wherein, figure (a) is wind power prediction curve, and figure (b) is climbing event prediction result.

Embodiment

Because the safety and stability of wind power climbing event to electric system brings serious threat, endanger electrical network for reducing wind power climbing event, the look-ahead of climbing event is very important.

The specific embodiment of the present invention will be described in detail below.

See Fig. 1, the present invention includes training part and predicted portions, concrete steps are as follows:

Step 1: the first sample set is divided local data's section.

This step comprises further:

The Preprocessing of 1.1 first sample sets.

Gather the first sample set, comprise history weather data that 1. NWP provides and 2. identical with history weather data time, place history wind power data.Preprocessing is carried out to the first sample intensive data, is specially: consider the deviation of acquisition system and artificial cut the operations such as machine, bad value, missing value process being carried out to the first sample set, to ensure the integrality of data.Meanwhile, consider error effect, denoising is carried out to the first sample set, to ensure the validity of data.

The 1.2 local data's sections first sample set being divided into Time Continuous.

Consider that climbing event refers to the significantly wind power change in a period of time, and a climbing event produces under specific meteorological scene, then can think that the applicable models of its inner wind power change is consistent.For this reason, using the mean value of the history of statistics climbing incident duration as window length, according to window length, the first sample set is divided into local data's section of Time Continuous, namely the start time of next local data's section goes up the end time of local data's section.

Step 2: build short-term wind-electricity power forecasting model database.

This step comprises further:

2.1 adopt SVM to represent short-term wind-electricity power forecast model.

For realizing high-precision long-term wind power prediction, SVM is adopted to build multivariable short-term wind-electricity power forecast model respectively for each local data section.Simultaneously, consider that different local datas section wind power variation characteristic may be different, for ensureing predictive validity, and improve precision of prediction, the identification of optimization model parameter is carried out respectively to the short-term wind-electricity power forecast model of different local datas section, and builds the short-term wind-electricity power forecasting model database that can instruct climbing prediction.

Because SVM is suitable for small-sample learning, and there is stronger generalization ability etc., therefore, SVM can be applied in the short-term wind-electricity power prediction of local data's section.

Suppose that the training set T of SVM is as follows:

T＝{(x ₁,y ₁),(x ₂,y ₂),…(x _l,y _l)}∈(R ⁿ×Y) ^l(1)

In formula (1), x ₁, x ₂... x _lrepresent input quantity, field of definition R ⁿfor n ties up real number space, input quantity and history weather data in this step; y ₁, y ₂... y _lrepresent output quantity, codomain Y is real number field, output quantity and history wind power data in this step.

Formula (2) is shown in by SVM basic model:

y＝<ω·x>+b(2)

In formula (2), ω, b are model parameter, and <> represents inner product.

The sample data collection of portion data segment i of setting a trap is { (X _i) _{l × q}, (P _i) _{l × 1}, (X _i) _{l × q}represent the weather data matrix in local data section i, size is that l × q, q represent meteorological index number, and l is window length; (P _i) _{l × 1}for the wind power data matrix in local data section i, size is l × 1, and l is window length.

Adopt SVM basic model shown in formula (2) to represent the short-term wind-electricity power forecast model of local data section i, see formula (3):

P _i,j＝<ω·X _i,j>+b(3)

In formula (3), subscript j represents row vector numbering in local data's section.

2.2 build short-term wind-electricity power forecasting model database.

Adopt the data short-term wind-electricity power forecast model shown in training type (3) respectively of each local data section, obtain parameter ω, the b of the short-term wind-electricity power forecast model of each local data section.

The training of short-term wind-electricity power forecast model is specially:

Structure based principle of minimization risk, the support vector regression be with by constructing hard ε solves model parameter ω, b shown in formula (3), sees formula (4):

$\begin{array}{cc}\underset{\mathrm{\&omega;},\mathrm{\&eta;},b}{min}& \frac{1}{2}\left|\right|\widehat{\mathrm{\&omega;}}|{|}^2+\frac{1}{2}{\mathrm{\&eta;}}^2\end{array}$

$\begin{array}{cc}s.t.& \begin{array}{c}<\widehat{\mathrm{\&omega;}}\&CenterDot;X_{i,j}>+\mathrm{\&eta;}(P_{i,j}+\widehat{\mathrm{\&epsiv;}})+\hat{b}\&GreaterEqual;1\\ <\widehat{\mathrm{\&omega;}}\&CenterDot;X_{i,j}>+\mathrm{\&eta;}(P_{i,j}-\widehat{\mathrm{\&epsiv;}})+\hat{b}\&le;-1\end{array}\end{array}---\left(4\right)$

In formula (4), || || represent vector norm, η is unknown variable, and has and introduce multiplier vector a structure Lagrange function, see formula (5):

$L(\mathrm{\&omega;},b,a^{(*)})=\frac{1}{2}\left|\right|\mathrm{\&omega;}|{|}^2-\underset{j=1}{\overset{l}{\&Sigma;}}{a}_j(\mathrm{\&epsiv;}+P_{i,j}-<\mathrm{\&omega;}\&CenterDot;X_{i,j}>-b)-\underset{j=1}{\overset{l}{\&Sigma;}}{a}_j^{*}(\mathrm{\&epsiv;}-P_{i,j}+<\mathrm{\&omega;}\&CenterDot;X_{i,j}>+b)---\left(5\right)$

The dual problem of formula (4) can be obtained by formula (5), see formula (6):

$\begin{array}{cc}\underset{a^{(*)}}{max}& -\frac{1}{2}\underset{j,k=1}{\overset{l}{\&Sigma;}}(a_j^{*}-a_j)(a_k^{*}-a_k)<X_{i,j}\&CenterDot;X_{i,k}>-\mathrm{\&epsiv;}\underset{j}{\overset{l}{\&Sigma;}}(a_j^{*}+a_j)+\underset{j=1}{\overset{l}{\&Sigma;}}{P}_{i,j}(a_j^{*}-a_j)\end{array}$

$\begin{array}{cc}s.t.& \underset{j=1}{\overset{l}{\&Sigma;}}(a_j^{*}-a_j)=0\end{array}---\left(6\right)$

$a_j^{(*)}\&GreaterEqual;0,j=1,...n$

Through type (6) can try to achieve the optimization model parameter of setting models.

In units of local data's section, complete the parameter identification of short-term wind-electricity power forecast model by local data's section, the short-term wind-electricity power forecast model of get Ge local data section, thus form short-term wind-electricity power forecasting model database.

Step 3: the training of forecast model selection mechanism.

Consider that its corresponding physical background (i.e. meteorological background) occurs climbing event relevant, wherein climbing affair character reflects by the change of wind power, the weather data reflection that physical background provides by NWP.Therefore, according to forecast model in meteorological background and short-term wind-electricity power forecasting model database, training forecast model selection mechanism, selects to instruct the forecast model of climbing event.

This step comprises sub-step further:

In 3.1 pairs of short-term wind-electricity power forecasting model database, model carries out cluster.

For ease of selecting forecast model according to forecasting model database, model in step 2 gained short-term wind-electricity power forecasting model database can be carried out cluster analysis, being fused into the short-term wind-electricity power forecasting model database containing limited representative model.Using the parameter of each short-term wind-electricity power forecast model as cluster input quantity, i.e. ξ=(ω, b), carries out Ward cluster using Ward distance as basis for estimation.

$\mathrm{\&Delta;}(A,B)=\underset{k\&Element;A\&cup;B}{\&Sigma;}{({\mathrm{\&xi;}}_k-{\mathrm{\&mu;}}_{A\&cup;B})}^2-\underset{k\&Element;A}{\&Sigma;}{({\mathrm{\&xi;}}_k-{\mathrm{\&mu;}}_A)}^2-\underset{k\&Element;B}{\&Sigma;}{({\mathrm{\&xi;}}_k-{\mathrm{\&mu;}}_B)}^2=\frac{n_A{n}_B}{n_A+n_B}\left|\right|{\mathrm{\&mu;}}_A-{\mathrm{\&mu;}}_B|{|}_2^2---\left(7\right)$

Formula (7) represents the computation process of Ward cluster.Wherein, k represents model parameter sample number, and Δ (A, B) represents the Ward distance of classification A and classification B; μ _aand μ _brepresent the center of classification A and classification B respectively, n _aand n _bnumber of objects in classification A and classification B respectively, A ∪ B represents the classification after classification A and classification B fusion, A and B represents two classifications that each cluster process middle distance is minimum, by merging A and B, and to merge rear center of a sample as new class center, thus complete the cluster analysis of model bank step by step.

During concrete enforcement, with the model parameter of model each in short-term wind-electricity power forecasting model database for cluster inputs, preset cluster numbers N as required, adopt Ward clustering procedure, the short-term wind-electricity power forecasting model database only comprising N number of representative model can be obtained.N is empirical value, and N value is larger, and cluster result is more accurate, but can increase calculated amount, sets N value according to the actual requirements so general with experience.

3.2 training forecast model selection mechanisms.

Be N class by sub-step 3.1 by Model tying in short-term wind-electricity power forecasting model database, the class number of the short-term wind-electricity power forecast model of each local data section is marked.The original short-term wind-electricity power forecast model of all local data segment is attributed to the short-term wind-electricity power forecast model containing N number of representative model.Take class number as output quantity, history weather data is input quantity, adopt svm classifier model representation forecast model selection mechanism.Be the second sample set with history weather data in model class number each in original short-term wind-electricity power forecasting model database and each local data section, adopt the second sample set training forecast model selection mechanism.

For multi-class svm classifier model, normal employing classification method, i.e. a kind of mainly class of each only separation, and build the second sample set shown in formula (1), and wherein, y _i∈ Y={1 ,-1}, i=1 ... l, y in this step _irepresent the class number of model, Y represents only containing the set of two elements, classification value 1 wherein to be separated, residue classification value-1.Second sample set can be expressed as such as formula relation (8) Suo Shi according to svm classifier model.

$\left\{\begin{array}{cc}<\mathrm{\&omega;}\&CenterDot;x_i>+b\&GreaterEqual;1,& y_i=1\\ <\mathrm{\&omega;}\&CenterDot;x_i>+b\&le;-1,& y_i=-1\end{array}\right.,i=1,\mathrm{2...}l---\left(8\right)$

Usually in formula (8), slack variable ξ is introduced _i>=0, then the objective function of svm classifier model can be written as formula (9).

$\begin{array}{cc}\underset{\mathrm{\&omega;},b}{min}& \frac{1}{2}\left|\right|\mathrm{\&omega;}|{|}^2+C\underset{i=1}{\overset{l}{\&Sigma;}}{\mathrm{\&xi;}}_i\end{array}$ (9)

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

Wherein, C is penalty factor.By the optimum solution ω of model ^*, b ^*determine optimal separating hyper plane, and according to Optimal Separating Hyperplane structure such as formula the decision function shown in (10) as forecast model selection mechanism:

f(x)＝sign(<ω·x _i>+b)(10)

In formula (10), f (x) represents the class number of model, x _irepresent the meteorological index of local data section i, ω and b represents the parameter of the short-term wind-electricity power forecast model of local data section i.

Sign (x) represents sign function, sign (x) value 1 when x is greater than 0; When x value is less than 0, sign (x) value-1; Sign (x) value 0 when x value equals 0.Consideration formula only has two classes in (8), then forecast model selection mechanism judges y=1 when f (x) is 0 or 1, judges y=-1 when f (x) is-1.

For the model bank containing N number of element, (N-1) individual Optimal Separating Hyperplane can be obtained according to above-mentioned steps, they form the model selection mechanism of climbing prediction jointly, can be determined the most suitable forecast model of local data's section by the judgement of (N-1) individual forecast model selection mechanism.

Step 4 is climbed event prediction.

This step comprises sub-step further:

The 4.1 following weather datas that NWP is provided according in step 1 adopt window length to be divided into local data's section, by local data's section input step 3 gained forecast model selection mechanism, determine that local data's section is suitable for the class number of forecast model, adopt this class number corresponding choose forecast model, the short-term wind-electricity power of prediction local data section.Namely the short-term wind-electricity power predicted data of local data's section of multiple Time Continuous forms long-term wind power prediction data.

4.2 carry out climbing event prediction according to long-term wind power prediction data.

In conjunction with climbing event definition, climbing event detection is carried out to long-term wind power prediction data; Simultaneously in conjunction with the possibility that actual electric network running state analysis climbing event occurs, comprehensively complete the prediction of climbing event.

Embodiment will be provided below to further illustrate technical solution of the present invention.

Embodiment

In this enforcement, choose Jiuquan, Gansu Province wind power base sampling interval in 2013 be the NWP data of 15 minutes and corresponding wind power data as training set, training set is divided into local data's section.

The data of Qu Ge local data section are research object, adopt SVM to build the short-term wind-electricity power forecast model of each local data section, thus form short-term wind-electricity power forecasting model database.Adopt Ward cluster analysis to be N class by Model tying in short-term wind-electricity power forecasting model database, see Fig. 4, N is 4.With the weather data of each local data section for input quantity, take class number as output quantity, by the thought of classification, set up N-1 basic SVM, based on multi-class svm classifier model training forecast model selection mechanism.Composition graphs 3 and formula (8-10) complete the training of forecast model selection mechanism.

The short-term wind-electricity power of each local data section is predicted according to forecast model selection mechanism and short-term wind-electricity power forecast model, with the short-term wind-electricity power predicted data of local data's section of multiple Time Continuous for test set, based on event definition of climbing formula (11) Suo Shi, complete the prediction of climbing event.

$\frac{|P(t+\mathrm{\&Delta;}t)-P\left(t\right)|}{\mathrm{\&Delta;}t}>{\mathrm{PRP}}_{val}---\left(11\right)$

When wind power rate of change in Preset Time is greater than given climbing rate threshold value PRP _valtime, then think and there occurs climbing event.Usually, climbing rate threshold value is taken as the wind power change of the installed capacity size of in 4 hours 50%.In the present embodiment, installation total volume P _always=603MW, therefore gets PRP _val=75.375MW/h completes the prediction to climbing event, and climbing time prediction the results are shown in Figure 5.

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

Claims (5)

1. adopt SVM to choose the wind power climbing event prediction method of forecast model, it is characterized in that, comprising:

First sample set is divided into local data's section by S1:

Gather the first sample set and pre-service is carried out to the first sample set, according to window length, the first sample set after pre-service is divided into local data's section of Time Continuous, wherein, first sample set comprises the on-site history weather data of electrical network and the history wind power data with the time, and window length is the mean value of the history climbing incident duration of statistics;

S2 builds short-term wind-electricity power forecasting model database:

With weather data be input, wind power be export, adopt SVM represent short-term wind-electricity power forecast model; Adopt local data's section training short-term wind-electricity power forecast model, obtain the short-term wind-electricity power forecast model of local data's section, and form short-term wind-electricity power forecasting model database according to the short-term wind-electricity power forecast model of all local data segment;

S3 trains forecast model selection mechanism:

According to model parameter, adopting Ward clustering procedure to carry out Cluster-Fusion to model in short-term wind-electricity power forecasting model database, and mark the class number of each model, is the second sample set with history weather data in the class number of each model and each local data section; With class number be export, weather data is input, adopts svm classifier model representation forecast model selection mechanism, and adopts the second sample set training forecast model selection mechanism;

S4 climbs event prediction:

Adopt the window length described in S1 following weather data to be divided into local data's section of Time Continuous, with local data's section for input, obtain the short-term wind-electricity power predicted data of local data's section based on forecast model selection mechanism; The short-term wind-electricity power predicted data of local data's section of Time Continuous forms long-term wind power prediction data, carries out climbing event prediction according to long-term wind power prediction data.

2. the wind power climbing event prediction method adopting SVM to choose forecast model as claimed in claim 1, is characterized in that:

In S2, the short-term wind-electricity power forecast model adopting SVM to represent is y=< ω x>+b, and wherein, ω and b is model parameter, and x represents meteorological index, and y represents wind power.

3. the wind power climbing event prediction method adopting SVM to choose forecast model as claimed in claim 1, is characterized in that:

In S2, adopt local data's section training short-term wind-electricity power forecast model, be specially:

Structure based principle of minimization risk, the support vector regression be with by constructing hard ε solves the parameter of short-term wind-electricity power forecast model.

4. the wind power climbing event prediction method adopting SVM to choose forecast model as claimed in claim 1, is characterized in that:

In S3, adopt Ward clustering procedure to carry out Cluster-Fusion to model in short-term wind-electricity power forecasting model database, be specially:

With the model parameter of model each in short-term wind-electricity power forecasting model database for cluster inputs, with Ward distance for cluster basis for estimation, Ward Cluster-Fusion is carried out to model each in short-term wind-electricity power forecasting model database, obtain the short-term wind-electricity power forecasting model database only comprising N number of representative model, N is cluster numbers.

5. the wind power climbing event prediction method adopting SVM to choose forecast model as claimed in claim 1, is characterized in that:

In S3, the forecast model selection mechanism adopting svm classifier model representation is f (x)=sign (< ω x _i>+b), wherein, ω and b represents the parameter of each model in short-term wind-electricity power forecasting model database, x _irepresent the meteorological index of local data's section, f (x) represents class number.