CN103425867A - Short-term wind power combination prediction method - Google Patents

Abstract

The invention relates to a short-term wind power combination prediction method which comprises steps of 1, extracting wind power sequential data from supervisory control and data acquisition (SCADA) related to a wind power plant; 2, performing sequence analysis on the extracted wind power sequential data through ensemble empirical mode decomposition; 3, reconstructing phase space of sequences obtained through the ensemble empirical mode decomposition; 4, according to the data obtained by reconstructing the phase space of the sequences, training the established wavelet neural network prediction model, and superposing predication results of the sequences to obtain a wind power prediction result; 5, performing error analysis on the wind power prediction result. By means of the short-term wind power combination prediction method, the modeling process is simple and practical, and wind power predication can be quickly and effectively performed. Therefore, the short-term wind power combination prediction method has great significance in safety, stability, management and running of a power system, and has wide popularization and application value.

Description

A kind of short-term wind-electricity power combination forecasting method

Technical field

The present invention relates to a kind of short-term wind-electricity power combination forecasting method.

Background technology

In recent years, along with the rapid growth of installed capacity of wind-driven power, the ratio of wind-powered electricity generation in electrical network increases year by year.Due to undulatory property and the randomness of wind energy itself, when wind-powered electricity generation penetrates power over after certain value, will bring severe challenge to management and running and the quality of power supply of electric system, this has seriously limited the development of wind-powered electricity generation.If can effectively be predicted wind power, not only can reduce margin capacity, the reduction system operation cost of electric system, but also can alleviate the adverse effect that wind-powered electricity generation causes electrical network, and effectively improve the maximum installation ratio of wind-powered electricity generation in electric system, improve the competing power of wind-powered electricity generation.

Short-term wind-electricity power forecast model at present commonly used, generally based on statistical prediction methods, mainly contains lasting method, Kalman filtering method, time series method, the support vector machine Return Law, artificial neural network method etc.Wherein continuing method is a kind of comparatively simple and effective DIRECT FORECASTING METHOD, generally as benchmark, weighs validity and the advance of other method; The essence of other several Forecasting Methodologies is the "black box" models between direct matching wind power and its influence factor, cut down the number of intermediate links on the impact predicted the outcome, although obtained certain effect, but because the characteristic of wind power itself does not obtain degree of depth excavation, prediction effect is overly dependent upon the performance of "black box" model, and the model of foundation often has the shortcomings such as not popularity to the susceptibility of model specification form and information source.So short-term wind power prediction should excavate the wind power characteristic from the degree of depth, based on wind power characteristics, select targetedly many algorithms, have complementary advantages to set up combination forecasting by algorithm.

Non-linear and the non-stationary characteristic of wind power is the main cause of impact prediction effect, and effective decomposition of the wind power signal being carried out to some scale or fluctuation tendency can reduce the non-stationary of signal.The set empirical mode decomposition is the vast improvement to traditional empirical mode decomposition method, efficiently solves the mode Aliasing Problem of traditional empirical mode decomposition, and actual signal is farthest retained.Wavelet neural network is the product that wavelet theory and artificial neural network combine, and combines the advantage of wavelet transformation video localization and the self-learning capability of neural network.Its basic thought is wavelet basis function to be used as to neuronic excitation function set up the contact between wavelet transformation and neural network.Because wavelet neural network has been inherited the nature and characteristic of wavelet decomposition, so can learn in sequence distribution loose region with low resolution, with high resolving power at the densely distributed regional learning of sequence, these characteristics are conducive to the more easily inherent law between " seizures " input and output data of neural network, thus wavelet neural network is more intelligent than traditional neural network, efficient and flexible.

Summary of the invention

The present invention, from research wind power characteristic, has proposed the simple short-term wind-electricity power combination forecasting method of relation, modeling method of a kind of direct consideration relevant historical data and power stage.It is a kind of Novel air power combination forecast model based on set empirical mode decomposition and wavelet neural network.At first use the set empirical mode decomposition to reduce the non-stationary of wind power signal, next adopts phase space reconfiguration to excavate the chaotic characteristic of each subsequence signal, then use wavelet neural network to carry out respectively modeling and forecasting to each subsequence, finally the stack that predicts the outcome of each subsequence is finally predicted the outcome.The combination forecasting that the present invention carries has higher precision of prediction and larger engineering application potential.

Above-mentioned technical matters of the present invention is mainly solved by following technical proposals:

A kind of short-term wind-electricity power combination forecasting method comprises the following steps:

Step 1, from wind energy turbine set SCADA system, carry out data extraction record, gather and preserve wind power sequence data;

Step 2, extracted wind power sequence is gathered to empirical mode decomposition, obtained plural subsequence and a surplus;

Step 3, respectively each subsequence and surplus are adopted to C-C method phase space reconstruction;

Step 4, the phase space of take after reconstruct are learning sample, and the training wavelet-neural network model adopts the forecast model trained to be predicted, and predicting the outcome of each subsequence and surplus superposeed, and obtains the wind power prediction result;

Step 5, wind power is predicted the outcome and carries out error analysis.

In described step 2, the wind power sequence gathered is gathered to empirical mode decomposition and is comprised following four steps,

Step 2.1, aweather in power sequence { x (t) }, add white noise sequence;

Step 2.2, the decomposition of use experience mode will add the wind power sequence after white noise to be decomposed into plural intrinsic modal components c _n(t) and one residual components r _N(t);

Step 2.3, repeating step 2.1 and step 2.2 r time altogether, the amplitude difference of the white noise sequence at every turn added;

Step 2.4, the natural mode of vibration component that r decomposition obtained are asked ensemble average, the final natural mode of vibration component using it as original signal.

In described step 5, adopt three kinds of following wind power error evaluation methods to carry out error analysis,

Normalization root-mean-square error e _NRMSE, normalization absolute average error e _NMAE, maximum absolute error e _MAE, it is defined as follows:

$e_{\mathrm{NRMSE}}=\frac{\sqrt{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{(P_{\mathrm{Mi}}-P_{\mathrm{Pi}})}^2}}{\mathrm{Cap}\·\sqrt{n}}\×100\%$ Formula (1)

$e_{\mathrm{NMAE}}=\frac{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}|P_{\mathrm{Mi}}-P_{\mathrm{Pi}}|}{\mathrm{Cap}\·n}\×100\%$ Formula (2)

E _MAE=max (| P _Mi-P _Pi|) formula (3)

In formula (1), formula (2) and formula (3), P _MiFor i measured power constantly, P _PiFor i predicted power constantly, the rated capacity that Cap is blower fan, the quantity that n is forecast sample.

In described step 2.2, the step of empirical mode decomposition is as follows,

Step 2.2.1, try to achieve all maximum value and minimal values in wind power time series { x (t) }, adopt cubic spline function to carry out interpolation fitting coenvelope line b _maxAnd lower envelope line b (t) _min(t);

Step 2.2.2, calculate upper and lower envelope mean value m (t), wherein m (t)=[b _max(t)+b _min(t)]/2, extract h (t)=x (t)-m (t), judges whether h (t) meets natural mode of vibration component condition, if meet, h (t) is exactly first natural mode of vibration component, fruitless satisfied, using h (t) as original series;

Step 2.2.3, repeating step step 2.2.1 and step 2.2.2, until the difference h after n screening _n(t) meet natural mode of vibration component condition, be called an IMF, be designated as c ₁(t)=h _n(t);

Step 2.2.4, obtain c ₁(t), after, according to formula (4), from signal x (t), obtain residual components r ₁(t);

R ₁(t)=x (t)-c ₁(t) formula (4)

Step 2.2.5, by r ₁(t) repeat above-mentioned steps and obtain remaining IMF component, as surplus r _N(t) during for monotonic quantity, stop.

Compared with prior art, the present invention has following advantage:

1. the relation of the direct consideration of the present invention and relevant historical data and power stage, modeling method is simple.

2. strong adaptability of the present invention, can be used as the power prediction model of general wind energy turbine set.

3. due to a lot of modeling links in the middle of not considering, only consider the relation between input and output, comparatively speaking, computing velocity of the present invention is very fast.

4. of the present inventionly realize that cost is low, easily promote.

The accompanying drawing explanation

Fig. 1 is small echo neural network topology structure figure in the present invention.

Fig. 2 is small echo neural network learning procedure chart in the present invention.

Fig. 3 is whole modeling process flow diagram of the present invention.

Embodiment

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

Embodiment:

At first, introduce the theory basis the present invention relates to.

1. gather the empirical mode decomposition principle

Empirical mode decomposition is a kind of self-adapting signal screening technique in essence; can be by the trend step-sizing that is present in different characteristic in former sequence out; obtain having natural mode of vibration component (the intrinsic mode function of same characteristic features; IMF), the natural mode of vibration component need meet (1) and (2) two conditions: in (1) whole natural mode of vibration vector sequence zero point number and the limit number differ at the most 1; (2) at the envelope of arbitrfary point ，You local minizing point definition and the envelope average of Local modulus maxima definition, be 0.

For certain wind power time series { x (t) }, the step of empirical mode decomposition is as follows:

1) try to achieve all maximum value and minimal values in sequence { x (t) }.Adopt cubic spline function to carry out interpolation fitting coenvelope line b _maxAnd lower envelope line b (t) _min(t).

2) calculate upper and lower envelope mean value m (t), wherein m (t)=[b _max(t)+b _min(t)]/2, extract h (t)=x (t)-m (t), judges whether h (t) meets natural mode of vibration component condition, is that h (t) is exactly first natural mode of vibration component, is not using h (t) as original series.

3) repeating step (1) and step (2), until the difference h after n screening _n(t) meet natural mode of vibration component condition, be called an IMF, be designated as c ₁(t)=h _n(t).

4) obtain c ₁(t), after, according to formula (1), from signal x (t), obtain residual components r ₁(t):

r ₁(t)＝x(t)-c ₁(t) （1）

5) by r ₁(t) repeat above-mentioned steps and can obtain remaining IMF component, as surplus r _N(t) during for monotonic quantity, stop.

Empirical mode decomposition has reduced human factor with respect to traditional wavelet decomposition algorithm decomposition result has been impacted, and have certain advance, but above-mentioned algorithm there will be the mode aliasing in some cases.The set empirical mode decomposition utilizes noisiness can effectively avoid this problem.Set empirical mode decomposition step is as follows:

1) add white noise sequence in sequence { x (t) };

2) use the EMD(empirical mode decomposition) will add the wind power sequence after white noise to be decomposed into several intrinsic modal components c _n(t) and one residual components r _N(t);

3) repeating step 1 and step 2 are r time altogether, the amplitude difference of the white noise sequence at every turn added;

4) decompose by r time the IMF obtained and ask ensemble average, the final IMF component using it as original signal;

The white noise sequence added in above-mentioned steps should obey (0, (α ε) ²) normal distribution, the intensive parameter that wherein α is noise, the standard deviation that ε is signal.From pertinent literature, when being 100, α, r can obtain decomposition result preferably when selecting between [0.1,0.3].Therefore select in this article r to equal 100, α and equal 0.25.

2.C-C method phase space reconfiguration ultimate principle

For wind series { x (t) }, wherein x (t)=(x ₁, x ₂..., x _N), the ultimate principle of { x (t) } being carried out to C-C method phase space reconfiguration is to utilize embedding seasonal effect in time series correlation integral function calculate time delay τ simultaneously and embed dimension m, and then according to τ _w=(m-1) τ obtains embedding dimension m, and correlation integral is defined as follows:

$C(m,N,r,t)=\frac{2}{M(m-1)}\underset{1\≤i\≤j\≤M}{\mathrm{\Σ}}\mathrm{\θ}(r-d_{\mathrm{ij}}),r\≥0---\left(2\right)$

Wherein: the size that N is the data group, M=N-(m-1) τ, for sequence τ=t, d _Ij=‖ X _i-X _j‖, wherein: X _i=(x _i, x _I-τ..., x _{I-(m-1) τ}), X _j=(x _j, x _J-τ..., x _{J-(m-1) τ}), d _IjFor the distance of any two phase points in phase space, r is Control Radius.

If x<0, θ (x)=0; If x >=0, θ (x)=1.

For wind series { x (t) }, be divided into t disjoint time series, define each subsequence and be:

$S(m,N,r,t)=\frac{1}{t}\underset{s=1}{\overset{t}{\mathrm{\&Sigma;}}}[C_s(m,\frac{N}{t},r,t)-{C_s}^m(1,\frac{N}{t},r,t)---\left(3\right)$

According to BDS(Brock-Dechert-Scheinkman) test statistics S (m, N, r, τ) and the residual quantity Δ S (m, N, t) of each subsequence of statistical computation, obtain formula (4) to formula (6).

$\stackrel{\&OverBar;}{S}\left(t\right)=\frac{1}{16}\underset{m=2}{\overset{5}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{4}{\mathrm{\&Sigma;}}}S(m,r_j,t)---\left(4\right)$

$\mathrm{\&Delta;}\stackrel{\&OverBar;}{S}\left(t\right)=\frac{1}{4}\underset{m=2}{\overset{5}{\mathrm{\&Sigma;}}}\left[\max \right\{S(m,r_j,t)\}-\min \{S(m,r_j,t)\left\}\right]---\left(5\right)$

$S_{\mathrm{cor}}\left(t\right)=\mathrm{\&Delta;}\stackrel{\&OverBar;}{S}\left(t\right)+\left|\stackrel{\&OverBar;}{S}\left(t\right)\right|---\left(6\right)$

Wherein, r _jFor Control Radius,

For first zero crossing or

First minimal value as time delay τ=t, simultaneously by S _cor(t) minimum value is as time window τ _w, according to τ _w=(m-1) τ determines the m value.

3. wavelet neural network algorithm and improvement

Wavelet neural network is to take the feed-forward network model of wavelet basis function as the neuron excitation function, wavelet neural network combines wavelet analysis can carry out the advantage of multi-Scale Data analysis as " school microscop ", and variation characteristic that can be potential to sequence by the study to Analysis On Multi-scale Features has more dominant description.

If ψ (t) ∈ is L ²(R) (L ²(R) mean square-integrable real number space, i.e. the signal space of energy) be wavelet basis function, ψ (t) meets the admissibility condition:

${\&Integral;}_{-\&infin;}^{+\&infin;}\frac{{\left|\mathrm{\&psi;}\left(\mathrm{\&omega;}\right)\right|}^2}{\left|\mathrm{\&omega;}\right|}\mathrm{d\&omega;}<+\&infin;---\left(7\right)$

Wherein, the Fourier transform that ψ (ω) is ψ (t).ψ (t) produces one group of wavelet basis function after translation is flexible:

${\text{\&psi;}}_{a,\mathrm{\&tau;}}\left(t\right)=\frac{1}{\sqrt{a}}\mathrm{\&psi;}\left(\frac{t-\mathrm{\&lambda;}}{a}\right)---\left(8\right)$

In formula: a is scale factor, and λ is contraction-expansion factor.For signal x (t) ∈ L ²(R), wavelet transformation is:

$W_x(a,\mathrm{\&tau;})=\frac{1}{\sqrt{a}}{\&Integral;}_{-\&infin;}^{+\&infin;}x\left(t\right)\mathrm{\&psi;}\left(\frac{t-\mathrm{\&lambda;}}{a}\right)\mathrm{dt}---\left(9\right)$

Shown in the topological structure accompanying drawing 1 of three layers of wavelet-neural network model.

In Fig. 1, X _i, i=1,2 ..., t, be the list entries of wavelet neural network, Y is output sequence, w _IjFor the connection weights between input layer and hidden layer, w _jConnection weights for hidden layer and output layer.At input signal sequence, be x _i(i=1,2 ..., in the time of k), the output computing formula of hidden layer is:

$g\left(j\right)={\mathrm{\&psi;}}_j[(\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}{w}_{\mathrm{ij}}{x}_i-{\mathrm{\&lambda;}}_j)/a_j],j=\mathrm{1,2},...,l---\left(10\right)$

Wherein, the output of j the node that g (j) is hidden layer, ψ _jFor wavelet basis function; a _jFor ψ _jContraction-expansion factor; λ _jFor ψ _jShift factor; The computing formula of output layer is:

$Y=\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}{w}_{j}g\left(i\right),j=\mathrm{1,2},...,l---\left(11\right)$

In formula, l is the hidden layer node number;

The elementary tactics of wavelet neural network study is to utilize error function minimization principle, constantly changes shape and the yardstick of wavelet basis.Weights and wavelet basis function coefficient makeover process are as follows:

1) calculate the wavelet neural network error of fitting

$e=\widehat{y\left(k\right)}-Y---\left(12\right)$

In formula,

For desired output, Y is wavelet neural network matching output.

2) according to wavelet neural network weights and the wavelet basis function coefficient of the k time iteration of error e correction

$\left\{\begin{array}{c}{w}_{n,k}^{i+1}=w_{n,k}^i+\mathrm{\&Delta;}{w}_{n,k}^{i+1}\\ a_k^{i+1}=a_k^i+\mathrm{\&Delta;}{a}_k^{i+1}\\ {\mathrm{\&lambda;}}_k^{i+1}={\mathrm{\&lambda;}}_k^i+\mathrm{\&Delta;}{\mathrm{\&lambda;}}_k^{i+1}\end{array}\right.---\left(13\right)$

In formula, That the neural network forecast error e calculates according to formula (14):

$\left\{\begin{array}{c}{\mathrm{\&Delta;w}}_{n,k}^{i+1}=-\mathrm{\&eta;}\frac{\&PartialD;e}{\&PartialD;w_{n,k}^i}\\ {\mathrm{\&Delta;a}}_k^{i+1}=-\mathrm{\&eta;}\frac{\&PartialD;e}{\&PartialD;a_{n,k}^i}\\ {\mathrm{\&Delta;\&lambda;}}_k^{i+1}=-\mathrm{\&eta;}\frac{\&PartialD;e}{\&PartialD;{\mathrm{\&lambda;}}_{n,k}^i}\end{array}\right.---\left(14\right)$

In formula (14), η is learning efficiency.The selection of WNN learning step-length η is most important, η cross conference cause unstable, but fast convergence rate; The little speed of convergence of η is slow, but can avoid unstable.For overcoming this contradiction, can adopt the method that increases momentum term to be improved.The WNN learning process formula that increases momentum term can be done following correction:

$\left\{\begin{array}{c}{w}_{n,k}^{i+1}=w_{n,k}^i\left(i\right)+{\mathrm{\&Delta;w}}_{n,k}^{i+1}+k*(w_{n,k}^i-w_{n,k}^{i-1})\\ a_k^{i+1}=a_k^i+{\mathrm{\&Delta;a}}_k^{i+1}+k*(a_k^i-a_k^{i-1})\\ {\mathrm{\&lambda;}}_k^{i+1}={\mathrm{\&lambda;}}_k^i+{\mathrm{\&Delta;\&lambda;}}_k^{i+1}+k*({\mathrm{\&lambda;}}_k^i-{\mathrm{\&lambda;}}_k^{i-1})\end{array}\right.---\left(15\right)$

As shown in Figure 2, concrete steps are as follows for improved wavelet neural network algorithm predicts model realization:

1) initialization network.Random initializtion a _k, λ _k, and network connection weight w _Ij, w _j.

2) error of fitting is calculated.Training sample input network, the error of output and desired output is exported and is calculated in the computational grid prediction.

3) revise weights and wavelet basis function coefficient.According to formula (13) roll-off network weights and wavelet function parameter.

4) whether evaluation algorithm finishes.When the error function absolute value is less than predefined error ξ or reaches maximum iteration time, be to finish; No, return to step 2.

It is below the modeling process of specific embodiments of the invention.

Modeling overall flow figure of the present invention as shown in Figure 3.Concrete steps are as follows:

1) gather and record wind power sequence historical data by SCADA;

2) use the set empirical mode decomposition to reduce the non-stationary of wind power sequence, obtain each component (IMF1-IMFn) and surplus r _N(t);

3) respectively each component and surplus are set up to improved prediction model based on wavelet neural network; For fear of problems such as the randomness of wavelet-neural network model input dimension and sequence loss, adopt C-C method phase space reconstruction;

4) each component is predicted the outcome stack obtains the wind power prediction value;

5) error analysis; Adopt three kinds of error analysis index analysis predicated errors.

Specific embodiment described herein is only to the explanation for example of the present invention's spirit.Those skilled in the art can make various modifications or supplement or adopt similar mode to substitute described specific embodiment, but can't depart from spirit of the present invention or surmount the defined scope of appended claims.

Claims (4)

1. a short-term wind-electricity power combination forecasting method is characterized in that: comprises the following steps,

Step 1, from wind energy turbine set SCADA system, carry out data extraction record, gather and preserve wind power sequence data;

Step 2, extracted wind power sequence is gathered to empirical mode decomposition, obtained plural subsequence and a surplus;

Step 3, respectively each subsequence and surplus are adopted to C-C method phase space reconstruction;

Step 4, the phase space of take after reconstruct are learning sample, and the training wavelet-neural network model adopts the forecast model trained to be predicted, and predicting the outcome of each subsequence and surplus superposeed, and obtains the wind power prediction result;

Step 5, wind power is predicted the outcome and carries out error analysis.

2. a kind of short-term wind-electricity power combination forecasting method according to claim 1 is characterized in that: in described step 2, the wind power sequence gathered gathered to empirical mode decomposition and comprised following four steps,

Step 2.1, aweather in power sequence { x (t) }, add white noise sequence;

Step 2.2, the decomposition of use experience mode will add the wind power sequence after white noise to be decomposed into plural intrinsic modal components c _n(t) and one residual components r _N(t);

Step 2.3, repeating step 2.1 and step 2.2 r time altogether, the amplitude difference of the white noise sequence at every turn added;

Step 2.4, the natural mode of vibration component that r decomposition obtained are asked ensemble average, the final natural mode of vibration component using it as original signal.

3. a kind of short-term wind-electricity power combination forecasting method according to claim 1 and 2 is characterized in that: in described step 5, adopt three kinds of following wind power error evaluation methods to carry out error analysis,

Normalization root-mean-square error e _NRMSE, normalization absolute average error e _NMAE, maximum absolute error e _MAE, it is defined as follows:

$e_{\mathrm{NRMSE}}=\frac{\sqrt{\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{(P_{\mathrm{Mi}}-P_{\mathrm{Pi}})}^2}}{\mathrm{Cap}\&CenterDot;\sqrt{n}}\&times;100\%$ Formula (1)

$e_{\mathrm{NMAE}}=\frac{\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}|P_{\mathrm{Mi}}-P_{\mathrm{Pi}}|}{\mathrm{Cap}\&CenterDot;n}\&times;100\%$ Formula (2)

E _MAE=max (| P _Mi-P _Pi|) formula (3)

In formula (1), formula (2) and formula (3), P _MiFor i measured power constantly, P _PiFor i predicted power constantly, the rated capacity that Cap is blower fan, the quantity that n is forecast sample.

4. a kind of short-term wind-electricity power combination forecasting method according to claim 2, it is characterized in that: in described step 2.2, the step of empirical mode decomposition is as follows,

Step 2.2.1, try to achieve all maximum value and minimal values in wind power time series { x (t) }, adopt cubic spline function to carry out interpolation fitting coenvelope line b _maxAnd lower envelope line b (t) _min(t);

Step 2.2.2, calculate upper and lower envelope mean value m (t), wherein m (t)=[b _max(t)+b _min(t)]/2, extract h (t)=x (t)-m (t), judges whether h (t) meets natural mode of vibration component condition, if meet, h (t) is exactly first natural mode of vibration component, fruitless satisfied, using h (t) as original series;

Step 2.2.3, repeating step step 2.2.1 and step 2.2.2, until the difference h after n screening _n(t) meet natural mode of vibration component condition, be called an IMF, be designated as c ₁(t)=h _n(t);

Step 2.2.4, obtain c ₁(t), after, according to formula (4), from signal x (t), obtain residual components r ₁(t);

R ₁(t)=x (t)-c ₁(t) formula (4)

Step 2.2.5, by r ₁(t) repeat above-mentioned steps and obtain remaining IMF component, as surplus r _N(t) during for monotonic quantity, stop.

Images