CN102521671B - Ultrashort-term wind power prediction method - Google Patents

Abstract

The invention discloses the technical field of wind power prediction, particularly, relates to an ultrashort-term power prediction method. The method comprises the following steps of: firstly, acquiring the wind speed, the wind direction and the wind power of a wind power farm to form a sample set; then, preprocessing the data of the sample set; reducing dimensions of the preprocessed sample set by a depth autocoder network; and finally, training a relevance vector machine regression model by the sample set with reduced dimensions, and predicting the ultrashort-term wind power through the trained relevance vector machine regression model. The method reduces the training time of a prediction model, satisfies the requirements on precision and real-time property in system status estimation, and enables the prediction method to be more accurate.

Description

A kind of ultrashort-term wind power prediction method

Technical field

The invention belongs to wind power electric powder prediction, relate in particular to a kind of ultrashort-term wind power prediction method.

Background technology

Along with becoming increasingly conspicuous of the energy and environmental problem, wind energy is as a kind of energy with fastest developing speed in renewable energy power generation technology, more and more be subject to people's attention, because the output power of wind energy turbine set has the feature of intermittent and undulatory property, so can produce important impact to the stability of the quality of power supply of whole electrical network and Operation of Electric Systems after wind-powered electricity generation access electrical network, for reducing this adverse effect, the forecasting accuracy of wind power is just seemed to particularly important.< < wind power forecast function standard > > in the power industry standard > > of the < < People's Republic of China (PRC) being issued by National Energy Board of the People's Republic of China (PRC) has pointed out that the prediction of ultra-short term wind power refers to the prediction wind energy turbine set active power of following 0-4 hour.Because dispatching of power netwoks, required time of resource distribution mainly concentrate on 0～3 hours section, therefore, ultra-short term prediction to wind power seems particularly important, especially ultrashort-term wind power prediction can effectively alleviate the impact of wind-electricity integration on whole electrical network, all significant for power grid security economic load dispatching, electricity market and wind energy turbine set operation.

The Forecasting Methodology of wind power can be divided into two classes according to the physical quantity difference of directly prediction at present: a class is directly to predict the output power of wind energy turbine set, another kind of is advanced row forecasting wind speed, then according to the relation of wind-powered electricity generation unit or wind power output power and wind speed, obtains wind power output power.Wherein, direct forecast methods often comprises two kinds of physical model method and statistic laws.Physical model is by carrying out physical modeling to wind energy turbine set location, utilize the numerical weather forecasts such as wind speed, wind direction, temperature and air pressure to predict, the model quality that predicts the outcome and set up of the method has very large relation, if model is too coarse, precision of prediction will be poor.The essence of statistical method is to utilize effective historical data (as numerical value data of weather forecast, historical statistics wind power data) to predict.Common correlation technique has persistence forecasting method, space smoothing method, time series method, Kalman filtering method, grey method, artificial neural network method, wavelet analysis method, the support vector machine Return Law, least square method, fuzzy logic method etc.Wherein persistence forecasting method is the simplest, the method nearest any power observed reading as more lower predicted value.The correlative study of artificial neural network and support vector machine is the method for mainly applying now, but artificial neural network complexity is higher, and the parameter of support vector machine selects that model prediction precision is had to considerable influence.

Summary of the invention

For the existing wind power forecasting method of mentioning in above-mentioned background technology, the complicated and not high deficiency of precision of prediction, the present invention proposes a kind of ultrashort-term wind power prediction method.

Technical scheme of the present invention is that a kind of ultrashort-term wind power prediction method, is characterized in that the method comprises the following steps:

Step 1: gather wind speed, wind direction and the wind power data of wind energy turbine set, form sample set;

Step 2: sample set is carried out to data pre-service;

Step 3: utilize degree of depth autocoder network to carry out dimensionality reduction to pretreated sample set;

Step 4: to the training of Relevance vector machine for regression model, then utilize the Relevance vector machine for regression model after training to predict the wind power of ultra-short term with the sample set after dimensionality reduction.

Described step 2 specifically comprises the following steps:

Step 2.1: the sample set application finite ridgelet transform to gathering, obtain ridge wave system matrix number, every row of ridge wave system matrix number are carried out respectively to one-dimensional discrete multi-scale wavelet transformation, obtain finite ridgelet transform matrix of coefficients;

Step 2.2: each row to finite ridgelet transform matrix of coefficients carry out wavelet adaptive threshold denoising, and low frequency sub-band coefficient is not processed, carries out soft-threshold processing to each high-frequency sub-band coefficient;

Step 2.3: to the finite ridgelet transform matrix of coefficients application finite ridgelet inverse transformation reconstruction signal after processing;

Step 2.4: reconstruction signal is carried out to self-adaptation dimension and receive Wiener filtering, obtain the sample set after denoising.

Described step 3 specifically comprises the following steps:

Step 3.1: adopt restriction Boltzmann machine model continuously to the sample set after denoising is trained in advance, to obtain one group and independently limit continuously Boltzmann machine model and one group of weights;

Step 3.2: the output neuron that each is limited to Boltzmann machine model is continuously merged into one deck with the input neuron of the next Boltzmann machine of restriction continuously model, expansion formation degree of depth autocoder network;

Step 3.3: weights are finely tuned, adopt the BP algorithm based on cross entropy function to carry out network training, adjust weights so that cross entropy function reaches minimum, complete reduction process.

Described step 4 specifically comprises the following steps:

Step 4.1: the sample set after dimensionality reduction is formatd to processing according to the call format of Method Using Relevance Vector Machine instruction input to sample set;

Step 4.2: select gaussian kernel function as the kernel function of Method Using Relevance Vector Machine model, the parameter initialization to kernel function;

Step 4.3: the sample set after processing with format is to the training of Method Using Relevance Vector Machine model;

Step 4.4: on the basis of step 4.3, utilize the wind power of the Method Using Relevance Vector Machine model prediction ultra-short term training.

The computing formula of described finite ridgelet transform is:

${\mathrm{FRAT}}_f[k,l]=\frac{1}{\sqrt{p}}\underset{(a,b)\&Element;L_{k,l}}{\mathrm{\&Sigma;}}f(a,b)$

In formula:

FRAT _f[k, l] is finite ridgelet transform matrix of coefficients;

P is the dimension of matrix of coefficients;

F (a, b) is discrete data signal;

A is horizontal ordinate;

B is ordinate;

L _{k, l}represent grid Z _p ²upper composition take the set of k as the point of slope, the l straight line that is intercept.The computing formula of described soft-threshold is:

$T=\frac{{\mathrm{\&sigma;}}_n^2}{{\mathrm{\&sigma;}}_j^2},(j=\mathrm{0,1},\&CenterDot;\&CenterDot;\&CenterDot;,n)$

Wherein:

T is soft-threshold;

variance for estimating noise;

for estimator band signal variance.

The method that the present invention predicts wind power is first with ridgelet transform, the sample set obtaining to be carried out to denoising, then with degree of depth autocoder network, sample set data is carried out to dimensionality reduction, finally with Relevance vector machine for regression model, predicts.The present invention is suitable for the prediction of wind energy turbine set ultra-short term power, realized the principle of structural risk minimization, the introduction of degree of depth autocoder network, make sample set can in the sample data of low dimension, carry out model training, greatly reduce the training time expense of forecast model, and then finely meet in system state estimation for the requirement of precision and real-time, make forecast model more accurate.

Accompanying drawing explanation

Fig. 1 is ultra-short term wind power forecasting method process flow diagram;

Fig. 2 is the process flow diagram of the method based on ridgelet transform to the sample set denoising obtaining;

Fig. 3 is the process flow diagram with Relevance vector machine for regression model prediction.

Embodiment

Below in conjunction with accompanying drawing, preferred embodiment is elaborated.Should be emphasized that, following explanation is only exemplary, rather than in order to limit the scope of the invention and to apply.

Good forecast model must be considered precision and the Space-time Complexity problem that prediction can reach simultaneously, considers above-mentioned reason, and the thought that has adopted degree of depth autocoder network and associated vector regression model to combine in this method is predicted.The data that directly gather are coarse uneven and have noise, so adopt ridgelet transform to process sample set data.But the training Space-time Complexity under higher-dimension sample set is too large, and degree of depth autocoder network (Deep Autoencoder Network, DAN) method adopts successive value restricted type Boltzmann model (Continouse Restricted Boltzmann Machine, CRBM) network structure, the two-way deep layer neural network by training with a plurality of middle layers converts high dimensional data to low-dimensional nested and reconstruct high dimensional data then.Especially, DAN provides the two-way mapping of high-dimensional data space and low-dimensional nested structure, efficiently solves the not available reverse Mapping problem of most of Method of Nonlinear Dimensionality Reductions.High dimensional data converts the nested and reconstruct high dimensional data method then of low-dimensional to, can, at the prerequisite decline low spatial dimension that keeps overall situation and partial situation's character of observation space as far as possible, can solve and under higher-dimension sample set, train the problem that Space-time Complexity is large.Method Using Relevance Vector Machine (Relevance Vector Machine, RVM) has solved the problem of precision of prediction well.Solution thinking of the present invention is: first, adopt ridgelet transform to process sample set data, then by DAN method, carry out dimension and approximately subtract, make it to be mapped as lower dimensional space data from High dimensional space data, finally with RVM, carry out sample training and prediction.

Fig. 1 is a kind of ultra-short term power forecasting method process flow diagram provided by the invention.In Fig. 1, method provided by the invention comprises following step:

Step 1: gather certain wind energy turbine set wind speed, wind direction and the wind power data in August, form sample set;

These data gathered once every 15 minutes, formed N * T two-dimensional array.

Step 2: sample set is carried out to data pre-service, and the method for employing based on ridgelet transform is to the sample set denoising obtaining; Fig. 2 is method based on the ridgelet transform process flow diagram to the sample set denoising obtaining.Step 2 comprises again the following step:

Step 2.1: to the sample set application finite ridgelet transform gathering, in each direction, will produce a FRAT sequence, therefore the result of finite ridgelet transform is the ridge wave system matrix number that produces a p * (p+1), every row of ridge wave system matrix number are carried out respectively to one-dimensional discrete multi-scale wavelet transformation, obtain finite ridgelet transform matrix of coefficients FRAT _f[k, l], (k, l) ∈ P _{i, j}, whole process is called finite ridgelet transform.

If p is a prime number, definition set Z _p=0,1,2 ..., p-1}, defines discrete data signal f (x) so at limited grid Z _p ²on the matrix of coefficients of finite ridgelet transform be:

${\mathrm{FRAT}}_f[k,l]=\frac{1}{\sqrt{p}}\underset{(a,b)\&Element;L_{k,l}}{\mathrm{\&Sigma;}}f(a,b)---\left(1\right)$

In formula:

FRAT _f[k, l] is finite ridgelet transform matrix of coefficients;

P is the dimension of matrix of coefficients;

F (a, b) is discrete data signal;

A is horizontal ordinate;

B is ordinate.

L _{k, l}represent grid Z _p ²the set of point of take the straight line that k is intercept as slope, l (when k=p, representing the infinitely great or vertical straight line of slope) of upper composition, that is:

L _k，l＝{(i，j)|j＝ki+l(mod)p，i∈Z _p}，k∈Z _p (2)

Wherein:

Z _pfor the grid length of side.

L _p，l＝{(l，j)|j∈Z _P} (3)

Above formula has defined limited grid Z _p ²straight line in the individual direction of upper p * (p+1), wherein all comprises p point on every straight line, for any given direction, have p bar straight line parallel to each other, and they have covered grid Z completely _p ².Its reconstruction formula is:

$F(a,b)=\frac{1}{\sqrt{p}}\underset{(k,l)\&Element;p_{a,b}}{\mathrm{\&Sigma;}}{\mathrm{FRAT}}_f[k,l]---\left(4\right)$

In formula:

P _{a, b}for process point (a, b) ∈ Z _p ²the slope of straight line and the array set that forms of intercept, can be expressed as:

$P_{a,b}=\left\{(k,l)\right|l=b-\mathrm{ka}\left(\mathrm{mod}\right)p,k\&Element;Z_p^2\}\&cup;\{(p,a)\};$

Step 2.2: each row to finite ridgelet transform matrix of coefficients, carry out wavelet adaptive threshold denoising, low frequency sub-band is not processed, and each high-frequency sub-band coefficient is carried out to soft-threshold processing.Threshold value formula is:

$T=\frac{{\mathrm{\&sigma;}}_n^2}{{\mathrm{\&sigma;}}_b^2},(b=\mathrm{0,1},\&CenterDot;\&CenterDot;\&CenterDot;,n)---\left(5\right)$

Wherein:

T is threshold value;

variance for estimating noise;

for estimator band signal variance;

Step 2.3: the ridge wave system matrix number application finite ridgelet inverse transformation reconstruction signal after processing;

Step 2.4: reconstruction signal is carried out to self-adaptation dimension and receive Wiener filtering, to reduce " coiling " effect, obtain the sample set after denoising.

Step 3: utilize degree of depth autocoder network to carry out dimensionality reduction to the sample set obtaining.Step 3 comprises again the following step:

Step 3.1: the sample set N * M array note after processing for ridgelet transform is X={x ₁, x ₂..., x _i∈ R ^m, wherein, i≤M, vector x _idimension be N;

Step 3.2: adopt CRBM to train in advance data, can obtain one group of independently CRBM model and one group of weights;

Wherein the pre-training process of CRBM is:

CRBM is connected to form by a visual layers and a hidden layer and their interlayer.Between visual element and hidden unit, by weight matrix w, be connected, the input of each unit in hidden layer comes from the state of all unit in visual layers, contribution according to each visual element to each Hidden unit, connection between them is composed with corresponding weights, the state value of each hidden unit be connected all visual element weights and.With s, represent the state of visual element and hidden unit, s below _jrepresent that input comes from visual element state set { s _ithe output of hidden unit j, w _ijrepresent the weights between visual element i and hidden unit j, and the bi-phases weights ' between them is equal, i.e. w _ij=w _ji, by following formula, carry out right value update calculating.

$s_j={\mathrm{\&phi;}}_j(\underset{i}{\mathrm{\&Sigma;}}{w}_{\mathrm{ij}}{s}_i+\mathrm{\&sigma;}{N}_j\left(\mathrm{0,1}\right))---\left(6\right)$

Function phi in formula _jexpression formula as follows:

${\mathrm{\&phi;}}_j\left(x_j\right)={\mathrm{\&theta;}}_L+({\mathrm{\&theta;}}_H-{\mathrm{\&theta;}}_L)\frac{1}{1+\exp (-a_j{x}_j)}---\left(7\right)$

Wherein, N _j(0,1) represents the Gaussian random variable of zero-mean, unit variance.Constant σ and N _j(0,1) has produced a noisy input component n jointly _j=σ N _j(0,1), its probability distribution is:

$p\left(n_j\right)=\frac{1}{\mathrm{\&sigma;}\sqrt{2\mathrm{\&pi;}}}\exp \left(\frac{-n_j^2}{2{\mathrm{\&sigma;}}^2}\right)---\left(8\right)$

By (2) formula, can be known φ _j(x) be that asymptotic line is at θ _land θ _hthe sigmoid function at place.Parameter a _jcontrol sigmoid slope of a curve, formula noise control variable, works as a _jwhile changing from small to big, can complete from muting determinacy state seamlessly transitting to scale-of-two random state.

CRBM adopts and minimizes the relaxation search that sdpecific dispersion training criterion has been substituted to the Boltzmann machine that only depends on Gibbs sampling, has greatly reduced calculated amount.Minimize the weights { w that training criterion is used for upgrading CRBM _ijand " noise control " parameter { a _i}:

$\mathrm{\&Delta;}{w}_{\mathrm{ij}}={\mathrm{\&eta;}}_w(<s_i{s}_j>-<{\hat{s}}_i{\hat{s}}_j>)---\left(9\right)$

$\mathrm{\&Delta;}{\hat{a}}_j=\frac{{\mathrm{\&eta;}}_a}{a_j^2}(<s_j^2>-<{\hat{s}}_j^2>)---\left(10\right)$

Wherein:

a step reconstituted state that represents unit j;

<> represents the average of training data;

η _wit is the learning rate of weight w.

When formula (4) and formula (5) show to train in advance with CRBM, only carry out simple addition and multiplication, calculated amount is little, can complete easily the renewal process of weights.

Step 3.3: the output neuron of each CRBM model and the input neuron of next CRBM are merged into one deck, launches to form own coding neural network;

Step 3.4: weights are finely tuned.The BP algorithm of employing based on cross entropy (Cross Entropy) function carries out network training, adjusts weights so that cross entropy function reaches minimum;

Step 3.5: the two-dimensional array that the data set after dimensionality reduction forms is as the input data of carrying out next step data processing.

Step 4: to the training of Relevance vector machine for regression model, then utilize the Relevance vector machine for regression model training to predict the wind power of ultra-short term with the sample set after dimensionality reduction.Step 4 comprises the following step:

Step 4.1: using the data of first 30 days of the sample set after dimensionality reduction as training sample, and according to the call format of Method Using Relevance Vector Machine training sample, input data are formatd to processing;

Step 4.2: select gaussian kernel function as the kernel function of Method Using Relevance Vector Machine, to parameter initialization;

Step 4.3: the input data using the prediction data of first 24 hours constantly as training sample, the prediction wind power of latter two hours is constantly as output data, to the training of Relevance vector machine for regression model;

Step 4.4: on the basis of step 4.3, utilize the Method Using Relevance Vector Machine model training to predict the wind power of following 2 hours.

According to the dimensionality reduction result of step 3, x _ibe i the data after moment dimensionality reduction.

The model output of RVM is defined as

$y(x,w)=\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{w}_{i}K(x,x_i)+w_0=\mathrm{\&Phi;w}---\left(11\right)$

In formula:

X _ifor observation vector;

W _iweights for model;

K (x, x _i) be kernel function;

Φ is the rank matrix of N * (N+1).

The likelihood function of training sample set can be expressed as

$P\left(t\right|w,{\mathrm{\&sigma;}}^2)={\left(2\mathrm{\&pi;}{\mathrm{\&sigma;}}^2\right)}^{-N/2}\exp (-\frac{1}{2{\mathrm{\&sigma;}}^2}{\left|\right|t-\mathrm{\&Phi;w}\left|\right|}^2)---\left(12\right)$

In formula:

N is vector x _idimension;

W is weights;

T is object vector, t=[t ₁, t ₂..., t _n] ^t;

σ ²variance for Gaussian noise.

The posterior probability of calculating weights is distributed as:

$p\left(w\right|t,\mathrm{\&alpha;},{\mathrm{\&sigma;}}^2)=\frac{p\left(t\right|w,{\mathrm{\&sigma;}}^2)p\left(w\right|\mathrm{\&alpha;})}{p\left(t\right|\mathrm{\&alpha;},{\mathrm{\&sigma;}}^2)}=N(\mathrm{\&mu;},\mathrm{\&Sigma;})---\left(13\right)$

In formula:

α is the super parameter of the prior distribution of power to make decision value matrix w, α=[α ₀, α ₁..., α _n] ^t, each super parameter alpha _ia corresponding weight w _i;

μ is weights posteriority statistic average, μ=σ ²∑ Φ ^tt;

∑ is summation symbol, ∑=(σ ²Φ ^tΦ+A) ^-1, A=diag (α wherein ₀, α ₁..., α _n).

By maximizing super parameter likelihood distribution, find its most probable value α _mPwith the employing estimation technique that iterates, to p (t| α, σ ²)=∫ p (t|w, σ ²) p (w| α) dw is about α differentiate, to make it be zero and formula is reset:

${\mathrm{\&alpha;}}_i^{\mathrm{new}}={\mathrm{\&gamma;}}_i/{\mathrm{\&mu;}}_i^2---\left(14\right)$

In formula:

μ _ifor weight w _iposteriority statistic average, μ _i=1-α _i∑ _ii;

∑ _iifor current α and σ ²i diagonal element of posteriority weights covariance matrix.

Equally to noise variance σ ²differentiate, obtains upgrading rule:

(σ ²) ^new＝||t-∑μ|| ²/(N-∑ _iγ _i) (15)

Learning algorithm calculates (4) (5) formula repeatedly, upgrades μ and ∑ simultaneously, knows and meets certain suitable condition of convergence.

Modeling process may be summarized to be the following steps:

1. initialization α _iand σ ²;

2. calculate weights posteriority statistic average μ and posterior variance ∑;

3. calculate all γ _iand reappraise α _iand σ ²;

4. if convergence, to step 5, does not restrain and returns to step 2;

5. delete α _ithe weights of → ∞ and basis function, obtain training pattern.

In the process of wind power prediction, to use a large amount of historical datas, these data all likely break down or are interfered at collection, transmission, conversion links, cause data exception, and wind speed and these data variation amplitudes of power large, there is no regularity, therefore in invention, with ridgelet transform, carry out the denoising of data.The new sample set of data formation after processing, then utilizes degree of depth autocoder network to carry out dimensionality reduction, finally with Relevance vector machine for regression model, carries out power prediction.The present invention is guaranteeing under the prerequisite of precision of prediction, can in the sample data of low dimension, carry out model training, has reduced the training time expense of prediction.

The above; be only the present invention's embodiment preferably, but protection scope of the present invention is not limited to this, is anyly familiar with in technical scope that those skilled in the art disclose in the present invention; the variation that can expect easily or replacement, within all should being encompassed in protection scope of the present invention.Therefore, protection scope of the present invention should be as the criterion with the protection domain of claim.

Claims (3)

1. a ultrashort-term wind power prediction method, is characterized in that the method comprises the following steps:

Step 1: gather wind speed, wind direction and the wind power data of wind energy turbine set, form sample set;

Step 2: sample set is carried out to data pre-service, and concrete steps comprise:

Step 2.1: the sample set application finite ridgelet transform to gathering, obtain ridge wave system matrix number, every row of ridge wave system matrix number are carried out respectively to one-dimensional discrete multi-scale wavelet transformation, obtain finite ridgelet transform matrix of coefficients;

Step 2.2: each row to finite ridgelet transform matrix of coefficients carry out wavelet adaptive threshold denoising, and low frequency sub-band coefficient is not processed, carries out soft-threshold processing to each high-frequency sub-band coefficient;

Step 2.3: to the finite ridgelet transform matrix of coefficients application finite ridgelet inverse transformation reconstruction signal after processing;

Step 2.4: reconstruction signal is carried out to self-adaptation dimension and receive Wiener filtering, obtain the sample set after denoising;

Step 3: utilize degree of depth autocoder network to carry out dimensionality reduction to pretreated sample set, concrete steps comprise:

Step 3.1: the sample set after adopting continuously restriction Boltzmann machine model to denoising is trained in advance, obtains one group and independently limits continuously Boltzmann machine model and one group of weights;

Step 3.2: the output neuron that each is limited to Boltzmann machine model is continuously merged into one deck with the input neuron of the next Boltzmann machine of restriction continuously model, expansion formation degree of depth autocoder network;

Step 3.3: weights are finely tuned, adopt the BP algorithm based on cross entropy function to carry out network training, adjust weights so that cross entropy function reaches minimum, complete reduction process;

Step 4: to the training of Relevance vector machine for regression model, then utilize the Relevance vector machine for regression model after training to predict the wind power of ultra-short term with the sample set after dimensionality reduction, concrete steps comprise:

Step 4.1: the sample set after dimensionality reduction is formatd to processing according to the call format of Method Using Relevance Vector Machine input to sample set;

Step 4.2: select gaussian kernel function as the kernel function of Method Using Relevance Vector Machine model, the parameter initialization to kernel function;

Step 4.3: the sample set after processing with format is to the training of Method Using Relevance Vector Machine model;

Step 4.4: on the basis of step 4.3, utilize the Method Using Relevance Vector Machine model training to carry out the wind power prediction of ultra-short term.

2. a kind of ultrashort-term wind power prediction method according to claim 1, is characterized in that the computing formula of described finite ridgelet transform is:

${\mathrm{FRAT}}_f[k,l]=\frac{1}{\sqrt{p}}\underset{(a,b)\&Element;L_{k,l}}{\mathrm{\&Sigma;}}f(a,b)$

In formula:

FRAT _f[k, l] is finite ridgelet transform matrix of coefficients;

P is the dimension of matrix of coefficients;

F (a, b) is discrete data signal;

A is horizontal ordinate;

B is ordinate;

L _k,lrepresent grid Z _p ²upper composition take the set of k as the point of slope, the l straight line that is intercept.

3. a kind of ultrashort-term wind power prediction method according to claim 1, is characterized in that the computing formula of described soft-threshold is:

$T=\frac{{\mathrm{\&sigma;}}_n^2}{{\mathrm{\&sigma;}}_j^2},(j=\mathrm{0,1},...,n)$

Wherein:

T is soft-threshold;

variance for estimating noise;

for estimator band signal variance.