CN104573876A - Wind power plant short-period wind speed prediction method based on time sequence long memory model - Google Patents

Abstract

A wind power plant short-period wind speed prediction method based on a time sequence long memory model includes: acquiring years of historical wind speed data of a wind power plant, and preprocessing the historical data to form a wind speed time sequence; inputting the processed data, and creating a long memory time sequence ARFIMA model for the wind speed time sequence by a rescaled range analysis method, namely an R/S analysis method, to acquire a preliminary predicted wind speed collection; further optimizing preliminary predicted wind speeds by a Kalman filtering algorithm to acquire a final predicted wind speed value. The wind power plant short-period wind speed prediction method has the advantages that years of actually-measured historical wind speed data of the wind power plant are combined to serve as a prediction model source, the preliminary predicted wind speed collection is acquired through creation of the long memory time sequence ARFIMA model, and prediction errors are reduced through Kalman filtering to acquire a final predicted wind speed result, so that accuracy in wind speed prediction is improved greatly.

Description

Based on the short-term wind speed forecasting method of wind farm of sequential Long memory model

Technical field

The present invention relates to wind energy turbine set technical field of power generation, particularly a kind of short-term wind speed forecasting method of wind farm.

Background technology

Along with becoming increasingly conspicuous of environmental problem, how solving the protection that energy problem takes into account environment simultaneously, is the significant problem that people are necessary and right.Wind energy is a kind of renewable, free of contamination green energy resource, has good economic benefit and environmental benefit.These outstanding advantages of wind energy, make people consider wind energy more and more as alleviating energy starved present situation and realizing the important means of environmental protect quality.

Wind energy comes from the motion of air, there is very large randomness, intermittence and uncontrollability, and blower fan is exerted oneself and is approximated to direct ratio with the cube of wind speed, the two has direct relation, therefore wind power output power also has very large undulatory property and very strong randomness, the fluctuation range of output power is usually comparatively large, and velocity variations is very fast.These characteristics of wind-power electricity generation, result in wind power generating set switching related frequency, there is randomness, not only impact electrical network, and add the difficulty of dispatching of power netwoks when wind energy turbine set and system carry out energy exchange.Along with developing rapidly of wind-powered electricity generation, after the installed capacity of wind energy turbine set exceedes wind power penetration limit (system under the normal prerequisite run the ratio of the maximum installed capacity of receptible wind energy turbine set and system maximum carrying capacity), Large Scale Wind Farm Integration is incorporated into the power networks and all brings inevitable impact to the power supply quality of electric system and operational reliability, and dispatching of power netwoks department plan arranges difficulty; In addition, in order to ensure electric power netting safe running, wind-powered electricity generation phenomenon of rationing the power supply becomes increasingly conspicuous, and causes clean wind energy resources serious waste.For overcoming above-mentioned defect, wind energy turbine set adopts to utilize the dispatching management of electric power system mode based on wind power prediction technology, and this way to manage is the most cost-effective way that solves the problem at present.In this way to manage, wind farm wind velocity and power prediction are the problems primarily solved, and are also current important research topics.

The free serial method of traditional wind speed forecasting method, neural network, Kalman filtering method etc.Wherein, time series method has that lower-order model precision of prediction is low, high-order model parameter fixes the large deficiency of difficulty; Neural network also exists that speed of convergence is slow, choosing of hidden node lacks the defects such as theoretical direction, training data be huge; Kalman filter method exists again to be set up Kalman state equation and measures the more difficult deficiency of equation, and is difficult to Accurate Prediction to the nonlinear system of complexity.The accuracy of therefore traditional wind speed forecasting method prediction is not high.

Summary of the invention

The present invention is directed to the deficiencies in the prior art, a kind of high-precision short-term wind speed forecasting method of wind farm is provided.

For solving the problems of the technologies described above, the technical solution used in the present invention is as follows.

Based on the short-term wind speed forecasting method of wind farm of sequential Long memory model, said method comprising the steps of:

A. obtain wind energy turbine set wind speed historical data for many years, pre-service is carried out to historical data, form wind speed time series;

B. data after input processing, adopt Rescaled range analysis and R/S analytical approach to set up long-memory time series ARFIMA model to wind speed time series, obtain the set of tentative prediction wind speed;

C. optimize tentative prediction wind speed further by Kalman filtering algorithm, obtain final forecasting wind speed value.

The above-mentioned short-term wind speed forecasting method of wind farm based on sequential Long memory model, carries out pre-service to historical data described in steps A and comprises the misdata removed and exceed actual wind speed scope and the step adopting EM algorithm to supplement missing data, specific as follows:

A1. analyze data, make initial division, remove the misdata exceeding actual wind speed scope;

A2. initialization: the parameter Θ initial value to be estimated to data set Density Distribution is arranged, and comprises ratio α of all categories _j, mean vector e _jwith covariance matrix ∑ _j; Valid data in data are divided into the n group of equal Gaussian distributed, the given initial weight often organized, i.e. α _j=1/n, j=1,2,3,4 ... n;

A3. E step is calculated: the expectation value calculating implicit variable data;

A4. calculate M step: by solving log-likelihood equation, calculation expectation value arrives the stylish average e of maximum point _jand weight α _j,

Whether A5, test meet the loop iteration condition of initial setting up | α _j ⁽ⁱ⁺¹⁾-α _j ⁽ⁱ⁾|≤ε, if do not satisfied condition, then goes to step A3; If satisfied condition, then iteration stopping, exports final argument α _j, e _j, j=1,2,3,4 ... n, now e _j, j=1,2,3,4 ... n is namely as the interpolation value of data centralization missing point.

The above-mentioned short-term wind speed forecasting method of wind farm based on sequential Long memory model, the step setting up long-memory time series ARFIMA (p, η, d, q) model in step B is specific as follows:

Long-term memory factor in B1, analytical sequence, by R/S analytic approach, d=H-0.5, determines d value;

B2, carry out fractional order difference, obtain ARMA (p, q) sequence;

B3, to ARFIMA (p, η, d, q)) carry out determining rank, determine p and q value.

The above-mentioned short-term wind speed forecasting method of wind farm based on sequential Long memory model, step B also comprises the step generating forecasting wind speed model equation, namely according to the air speed data before t, inferred by Bayesian statistics and model parameter is estimated, generate forecasting wind speed model equation.

The above-mentioned short-term wind speed forecasting method of wind farm based on sequential Long memory model, step C specifically comprises the following steps:

C1, obtained the state equation of Ensemble Kalman Filter by forecasting wind speed model equation;

C2, using forecasting wind speed data as observation data, and white Gaussian noise is joined in observation data, obtains the data of independent observation;

C3, each value upgraded in the data acquisition of independent observation, adopt following formula to obtain state value:

x _k＝A _kx _k-1+H _k(y _k-C _kA _kx _k-1)，

Wherein: x _kfor the updated value of k moment state value; x _k-1for the k-1 moment gather in predicted value, H _kfor kalman gain matrix, y _kfor the data vector of independent observation, C _kfor observing matrix;

C4. carry out successive ignition, obtain final forecasting wind speed value.

The above-mentioned short-term wind speed forecasting method of wind farm based on sequential Long memory model, described kalman gain matrix H _kcomputing formula be:

$H_k=P_k^t{C}_k^I{(C_k{P}_k^t{C}_k^I+R_k)}^{-1},$

Wherein, R _kthe covariance of observed reading error, P _kfor the covariance of data acquisition,

P _kcomputing formula be:

P _k＝(I-H _kC _k)P _k′

$P_k^{\′}=A_k{P}_{k-1}{A}_k^{\mathrm{\τ}}+Q_{k-1},$

Owing to have employed above technical scheme, the technical progress effect acquired by the present invention is as follows.

The present invention be integrated with wind energy turbine set for many years wind speed actual measurement historical data originate as forecast model, by setting up long-memory time series ARFIMA model, obtain the set of tentative prediction wind speed, predicated error is reduced again by Kalman filtering, obtain final forecasting wind speed result, substantially increase the precision of forecasting wind speed.

Accompanying drawing explanation

Fig. 1 is overview flow chart of the present invention;

Fig. 2 is the process flow diagram of data prediction in steps A of the present invention;

Fig. 3 is the process flow diagram building long memory Time series forecasting model in step B of the present invention;

Fig. 4 is the process flow diagram of Kalman filtering method in step C of the present invention.

Embodiment

Below in conjunction with the drawings and specific embodiments, the present invention is further elaborated.

Sequential Long memory model ARFIMA model of the present invention is a frontier in time series analysis, its complicated structure, and there is the difference of internal with ARIMA model, ARFIMA model does not need hypothesis seasonal effect in time series development model being made to priori, while method itself ensure that by repeatedly identifying amendment, until obtain satisfied model, therefore be suitable for various types of time series data, be included in the prediction under the very difficult and complex situations of the characteristic feature debating other time series data.This method not only investigates past value and the currency of predictive variable, also enters model as key factor to the error that model produces with past value matching simultaneously, and being conducive to the degree of accuracy improving model, is the Forecasting Methodology that a kind of degree of accuracy is quite high.By ARFIMA models applying in the forecasting wind speed field of wind energy turbine set, will greatly improve the precision of forecasting wind speed.

Namely short-term wind speed forecasting method of wind farm of the present invention is that the process flow diagram of the method as shown in Figure 1, specifically comprises the following steps based on sequential long memory ARFIMA model:

A. obtain wind energy turbine set wind speed historical data for many years, pre-service is carried out to historical data, form wind speed time series.Wind speed historical data mainly comprises the data such as wind speed, wind direction, temperature, atmospheric pressure.

Carry out pre-service to historical data and mainly comprise the misdata removed and exceed actual wind speed scope and the step adopting EM algorithm to supplement missing data, as shown in Figure 2, concrete steps are as follows for its process flow diagram.

A1. analyze data, make initial division, remove the misdata exceeding actual wind speed scope.

A2. initialization: the parameter Θ initial value to be estimated to data set Density Distribution is arranged, and comprises weight α of all categories _j, mean vector e _jwith covariance matrix ∑ _j; Valid data in data are divided into the n group of equal Gaussian distributed, the given initial weight often organized, i.e. α _j=1/n, j=1,2,3,4 ... n.Because Monitoring Data is one dimension, therefore ∑ _j=1, j=1,2,3,4 ... n, therefore ∑ in following calculating _jall can omit.

A3. E step is calculated: the expectation value calculating implicit variable data, herein, the conditional probability P (j|x) of a demand data.

Adopt stochastic variable C designation data composition, then probability P _ij=P (C=i|x _j) represent data x _jthe probability produced by component i, the probability also namely produced by a jth Gaussian distribution.

Obtained by Bayesian formula, P _ij=P (x _j| C=i) * P (C=i), wherein P (x _j| C=i)), i.e. x _jprobability in i-th Gaussian distribution, P (C=i) is the weight parameter of i-th Gaussian distribution.According to the thought of mixed Gauss model, be expressed as follows:

A4. calculate M step: by solving log-likelihood equation, calculation expectation value arrives the stylish average e of maximum point _jand weight α _j, for next iteration, wherein ∑ _j=1,

$e_j^{(i+1)}=\frac{\underset{i=1}{\mathrm{\Σ}}P\left(j\right|x_i)x_i}{\underset{i=1}{\mathrm{\Σ}}P\left(j\right|x_i)},$

${\mathrm{\α}}_j^{(i+1)}=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\frac{{\mathrm{\α}}_j^{\left(i\right)}G(x,e_j^i,{\mathrm{\Σ}}_j)}{\underset{i-1}{\mathrm{\Σ}}{\mathrm{\α}}_j^{\left(i\right)}G(x,e_j^i,{\mathrm{\Σ}}_j)}=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}P\left(j\right|x_i),$

Wherein P (j|x) calculates according to following formula:

$P\left(j\right|x)=\frac{{\mathrm{\α}}_{j}G(x,e_j^i,{\mathrm{\Σ}}_j)}{P(x,\mathrm{\Θ})},j=\mathrm{1,2,3}...n.$

Whether A5, test meet the loop iteration condition of initial setting up | α _j ⁽ⁱ⁺¹⁾-α _j ⁽ⁱ⁾|≤ε, if do not satisfied condition, then goes to step A3; If satisfied condition, then iteration stopping, exports final argument α _j, e _j, j=1,2,3,4 ... n, now e _j, j=1,2,3,4 ... n is namely as the interpolation value of data centralization missing point.

With an example, its treatment step is described below.

Using Ningxia, China, a certain wind energy turbine set history actual measurement wind speed is as input historical data, setting up sequential Long memory model accurately needs historical data more, therefore the data of zero during 26, zero on February of zero to 2014 when selecting 26 days zero February in 2013, every five minutes actual measurement air speed value, forming dimension is 105120 initial time sequences, because data are various, do not describe in detail at this.The time series x of wind speed is formed after the data processing of step A _t.

B. the time series x of wind speed is inputted _t, adopt Rescaled range analysis and R/S analytical approach to set up long-memory time series ARFIMA model to wind speed time series, obtain the set of tentative prediction wind speed.

Set up the process flow diagram of long-memory time series ARFIMA (p, η, d, q) model in this step as shown in Figure 3, concrete steps are as follows:

B1, analysis time sequence Long Memory, by R/S analytic approach, d=H-0.5, determines d value.By R/S analytic approach: wind speed time series is divided into m the continuous sub-range that length is n by (1).Each sub-range is marked.Each sub-range is designated as A _i, i=1 ..., m.A _iseveral middle every bit can be expressed as a _k,i, k=1 ..., n, i=1 ..., m.

(2) to each length be the sub-range A of n _i, calculating its average is:

(3) its cumulative mean value deviation y is calculated to single sub-range _k,i:

$y_{k,i}=\underset{i=1}{\overset{k}{\mathrm{\Σ}}}(a_{k,i}-e_i),k=\mathrm{1,2},...,n,$

(4) extreme difference (Range) defining single sub-range is R _ai=Max (y _k,i)-Min (y _k,i), k=1,2 ..., n.

(5) standard deviation in each sub-range is calculated:

And with it, rescaling/standardization (R is carried out to extreme difference _ai/ S _ai).Therefore, we can calculate the average rescaling extreme difference in A sub-range:

6) according to the model that Hurst proposes, the relational expression (R/S) of foundation _n=(c*n) ^htaking the logarithm in both sides, thus has

log(R/S) _n＝H*logn+logc

R _anbe the extreme difference that in a time series, n data depart from the accumulated value of its average, be called the extreme difference of n data, S _nrepresent the variation range that time series is maximum; S _nbe seasonal effect in time series standard deviation, representing the degree departing from average, is the side degree of degree of scatter.(R/S) _n, represent that pole extent uses S again _nweigh, Here it is weighs or has the stochastic process of long-term memory effect all applicable.H is Hurst index, and c is constant.Do least square regression, H value and standard deviation thereof can be tried to achieve.Calculate the fractional order difference d in ARFIMA model, can be solved by Hurst index, i.e. d=H-0.5.

Through calculating H=0.48246, so d=H-0.5=-0.01754.Due to-0.5<d<0.5, so the time series x of wind speed _t, there is Long Memory Properties.

B2, carry out fractional order difference, obtain the differentiated time series sequences y of mark _t; Utilize MATLAB software, to wind speed time series x _t, carry out d=-0.01754 mark difference, obtain the differentiated time series sequences y of mark _t.

B3, carry out determining rank to ARFIMA (p, η, d, q), determine p and q value.Utilize statistical analysis software Eviews, calculate, p=2, q=11, η=1.So wind speed sequential Long memory model is that ARFIMA (2,1 ,-0.01754,6) is expressed as (1+0.431L) (1+0.162L) (1-L) (l-L) in this example ^-0.01754y _t=(1-0.982) (1-0.16L) (1-0.095L) (1-0.783L) (1-0.951L) (1-0.875L) μ _t

B4, according to the air speed data before t, inferred by Bayesian statistics and estimate model parameter, generate forecasting wind speed model equation, in this example, wind speed forecast model equation is x _t=-0.431x _t-0.162x _t+μ _t-0.982 μ _t-0.16 μ _t-0.095 μ _t-0.783 μ _t-0.951 μ _t-0.875 μ _t;

Data formation time sequence { y after processing of step A _t,

Autoregressive model (AR (p)) is y _t=Φ ₁y _t-1+ Φ ₂y _t-2+ ... + Φ _py _t-p+ μ _t,

Moving average model(MA model) modeling (MA (q)) is y _t=μ ₁-θ ₁μ _t-1-θ ₂μ _t-2-...-θ _qμ _t-q.

Introduce hysteresis factors L, note L ^kfor k walks lag operator, i.e. L ^ky _t=y _t-k.

Then autoregressive model (AR (p)) is y _t=Φ ₁ly _t+ Φ ₂l ²y _t+ ₊Φ _pl ^py _t+ μ _t,

Make Φ (L)=1-Φ ₁l-Φ ₂l ²-...-Φ _pl ^p, Φ Ly can be write a Chinese character in simplified form into _t=μ _t;

Moving average model(MA model) modeling (MA (q)) is y _t=μ ₁-θ ₁l μ _t-θ ₂l ²μ _t-...-θ _ql ^qμ _t.

Make θ (L)=1-θ ₁l-θ ₂l ²-...-θ _ql ^q, y can be write a Chinese character in simplified form into _t=θ (L) μ _t.

If { y _tstationary process, and meet difference equation:

Φ(L)(1-L) ^η(l-L) ^dy _t＝θ(L)μ _t

Then claim { y _tit is ARFIMA (p, η, d, q) process.

Wherein, L is lag operator, μ _tfor white noise sequence, η is that integer jump divides, and d is mark difference operator, and-0.5<d<0.5;

Φ (L) and θ (L) be respectively the stable delayed polynomial operator in p rank and q rank and root outside unit circle.

Obviously, process { y _tthe sufficient and necessary condition of ARFIMA (p, η, d, q) process, (1-L) ^η(l-L) ^dit is ARMA (p, q) process.

C. optimize tentative prediction wind speed further by Kalman filtering algorithm, obtain final forecasting wind speed value.This flow chart of steps as shown in Figure 4, specifically comprises the following steps.

C1, obtained the state equation of Ensemble Kalman Filter by forecasting wind speed model equation: x _k=A _kx _k-1+ H _k(y _k-C _ka _kx _k-1),

C2, using forecasting wind speed data as observation data, and white Gaussian noise is joined in observation data, obtains independent observation data;

Wherein: x _kthe initial air speed data in k moment,

Φ _ibe auto-regressive parameter, p is Autoregressive,

θ _jbe moving average parameter, q is moving average exponent number, e _t-jfor moving average error,

ε _tit is stochastic error.

C3, init state equation covariance and observation equation covariance.Because state equation covariance and observation equation covariance can be constantly updated with iteration, progressively close to actual value, so initial value arranges and can be set to 2 and be multiplied by unit matrix.Upgrade each value in the data acquisition of independent observation, adopt following computing formula to obtain state value:

x _k＝A _kx _k-1+H _k(y _k-C _kA _kx _k-1)；

Wherein: x _kfor the updated value of k moment state value; x _k-1for the k-1 moment gather in predicted value, H _kfor kalman gain matrix, y _kfor the data vector of independent observation, C _kfor observing matrix.

Wherein, kalman gain matrix H _kcomputing formula be:

$H_k=P_k^t{C}_k^I{(C_k{P}_k^t{C}_k^I+R_k)}^{-1},$

Wherein, R _kthe covariance of observed reading error, P _kfor the covariance of data acquisition.

P _kcomputing formula be:

P _k＝(I-H _kC _k)P _k′

$P_k^{\&prime;}=A_k{P}_{k-1}{A}_k^{\mathrm{\&tau;}}+Q_{k-1},$

Wherein, A _kfor observing matrix, Q _kfor the covariance of state equation.

C4. carry out successive ignition, obtain final forecasting wind speed value.Above example is example, the value going out 00:10-00:30 on the 27th February in 2014 by computational prediction is 1.06,2.26,3.14,3.54, between 3.39 and actual value, error is 7.2%.

Claims (6)

1. based on the short-term wind speed forecasting method of wind farm of sequential Long memory model, it is characterized in that, said method comprising the steps of:

A. obtain wind energy turbine set wind speed historical data for many years, pre-service is carried out to historical data, form wind speed time series;

B. data after input processing, adopt Rescaled range analysis and R/S analytical approach to set up long-memory time series ARFIMA model to wind speed time series, obtain the set of tentative prediction wind speed;

C. optimize tentative prediction wind speed further by Kalman filtering algorithm, obtain final forecasting wind speed value.

2. the short-term wind speed forecasting method of wind farm based on sequential Long memory model according to claim 1, it is characterized in that, described in steps A, pre-service is carried out to historical data and comprise the misdata removed and exceed actual wind speed scope and the step adopting EM algorithm to supplement missing data, specific as follows:

A1. analyze data, make initial division, remove the misdata exceeding actual wind speed scope;

A2. initialization: the parameter Θ initial value to be estimated to data set Density Distribution is arranged, and comprises ratio α of all categories _j, mean vector e _jwith covariance matrix ∑ _j; Valid data in data are divided into the n group of equal Gaussian distributed, the given initial weight often organized, i.e. α _j=1/n, j=1,2,3,4 ... n;

A3. E step is calculated: the expectation value calculating implicit variable data;

A4. calculate M step: by solving log-likelihood equation, calculation expectation value arrives the stylish average e of maximum point _jand weight α _j,

Whether A5, test meet the loop iteration condition of initial setting up | α _j ⁽ⁱ⁺¹⁾-α _j ⁽ⁱ⁾|≤ε, if do not satisfied condition, then goes to step A3; If satisfied condition, then iteration stopping, exports final argument α _j, e _j, j=1,2,3,4 ... n, now e _j, j=1,2,3,4 ... n is namely as the interpolation value of data centralization missing point.

3. the short-term wind speed forecasting method of wind farm based on sequential Long memory model according to claim 2, is characterized in that, in described step B, the step setting up long-memory time series ARFIMA (p, η, d, q) model is specific as follows:

Long-term memory factor in B1, analytical sequence, by R/S analytic approach, d=H-0.5, determines d value;

B2, carry out fractional order difference, obtain zero-mean ARMA (p, q) sequence;

B3, to ARFIMA (p, η, d, q)) carry out determining rank, determine p and q value.

4. the short-term wind speed forecasting method of wind farm based on sequential Long memory model according to claim 3, it is characterized in that, step B also comprises the step generating forecasting wind speed model equation, namely according to the air speed data before t, inferred by Bayesian statistics and model parameter is estimated, generate forecasting wind speed model equation.

5. the short-term wind speed forecasting method of wind farm based on sequential Long memory model according to claim 1, it is characterized in that, step C specifically comprises the following steps:

C1, obtained the state equation of Ensemble Kalman Filter by forecasting wind speed model equation;

C2, using forecasting wind speed data as observation data, and white Gaussian noise is joined in observation data, obtains the data of independent observation;

C3, each value upgraded in the data acquisition of independent observation, adopt following formula to obtain state value:

x _k＝A _kx _k-1+H _k(y _k-C _kA _kx _k-1)；

Wherein: x _kfor the updated value of k moment state value; x _k-1for the k-1 moment gather in predicted value, H _kfor kalman gain matrix, y _kfor the data vector of independent observation, C _kfor observing matrix;

C4. carry out successive ignition, obtain final forecasting wind speed value.

6. the short-term wind speed forecasting method of wind farm based on sequential Long memory model according to claim 5, is characterized in that, described kalman gain matrix H _kcomputing formula be:

$H_k=P_k^{\&prime;}{C}_k^I{(C_k{P}_k^{\&prime;}{C}_k^I+R_k)}^{-1};$

Wherein, R _kthe covariance of observed reading error, P _kfor the covariance of data acquisition,

P _kcomputing formula be:

P _k＝(I-H _kC _k)P′ _k，

$P_k^{\&prime;}=A_k{P}_{k-1}{A}_k^{\mathrm{\&tau;}}+Q_{k-1},$

Wherein, A _kfor observing matrix, Q _kfor the covariance of state equation.