CN102682207A - Ultrashort combined predicting method for wind speed of wind power plant - Google Patents

Abstract

The invention relates to an ultrashort combined predicting method for wind speed of a wind power plant. Firstly, wind speed time series of the wind power plant is acquired, and data pretreatment is carried out on the series to obtain input data of a prediction system; the input data is predicted respectively by a continuous prediction model, an ARMA (Auto Regressive Moving Average) prediction model and a wavelet-neural network prediction model, and three groups of prediction values are obtained through computation; and finally the three groups of prediction values are adopted to obtain the prediction value of the wind speed in the future 0-1 hour by a combined prediction method, and the last two groups of prediction values are adopted to obtain the prediction value of the wind speed in the future 1-4 hours by the combined prediction method. The combined prediction method is adopted, useful information of each single prediction method is fully utilized, the prediction precision of the wind speed in the future 4 hours of the wind power plant is improved, and a reference is provided for reasonably scheduling a power grid.

Description

The ultrashort phase combination forecasting method of wind farm wind velocity

Technical field

The present invention relates to the ultrashort phase Forecasting Methodology of wind farm wind velocity in the wind generator system, particularly utilize the ultrashort phase combination forecasting method of wind farm wind velocity that continues method, ARMA method and wavelet-neural net method.

Background technology

Along with wind-powered electricity generation is incorporated into the power networks on a large scale, its safe and stable operation to electric system has caused certain influence.Wind power or wind speed predicted to make dispatching of power netwoks department rationally arrange generation schedule and counter-measure, can reduce system's spinning reserve capacity and operating cost, improve wind-powered electricity generation and penetrate the power limit and the market competitiveness.

Owing to receive many-sided influences such as numerical weather forecast precision of prediction and shortage wind energy turbine set high-quality historical data base; Wind power or forecasting wind speed errors medium-term and long-term even that shift to an earlier date 24 hours are bigger; So must carry out 0～4 hour ultrashort phase prediction; Possesses the real-time prediction ability, to the adjustment of rolling of operation of power networks situation.The common method of forecasting wind speed mainly comprises persistence forecasting method, Kalman filtering method, ARMA method, artificial neural network method, fuzzy logic method and spatial coherence method etc. at present.

The persistence forecasting method simply will a nearest moment the wind speed observed reading as time any forecasting wind speed value; Its prediction principle is the principle of inertia according to Atmosphere System; This algorithm is simple, and precision of prediction is higher in the utmost point short-time forecast (less than 1 hour) of wind farm wind velocity.

Kalman filtering method is set up state-space model to wind speed as state variable and is realized prediction.This method is under the known situation of statistical property of supposition noise, to draw, and in fact the statistical property of estimating noise is the difficult point place that this method is used.

The a large amount of historical data of ARMA method utilization is come modeling, confirms the mathematical model that can describe institute's search time sequence through pattern-recognition, parameter estimation, model testing, and then derives forecast model and reach the prediction purpose.This method only need know that the single wind speed time series of wind energy turbine set can set up model and predict.

Artificial neural network has characteristics such as parallel processing, distributed storage and fault-tolerance, and for finding the solution challenge very effectively, what be used for wind farm wind velocity and power prediction at present has BP neural network, RBF neural network, a local recursion's neural network etc.

A little less than the fuzzy logic method learning ability, when being used for forecasting wind speed separately, effect is relatively poor, usually the fuzzy prediction method is used in combination with additive method.

The spatial coherence method need be considered wind energy turbine set and many groups air speed data in close several places with it, uses the correlativity between the wind speed of several places to predict.It need obtain several groups of air speed datas of wind energy turbine set surrounding area, and several long-range stations of testing the speed need be set, and cost is bigger.

In recent years, along with the influence of wind-powered electricity generation to Operation of Electric Systems scheduling increases, wind power and forecasting wind speed research come into one's own at home, and domestic research is in this respect started late, the precision of prediction space that still is greatly improved.

Because the angle of considering, mode difference; The different predicting method includes different important informations, discovers that single Forecasting Methodology receives the region restriction very strong, maybe be very big in different testing location precision of prediction differences with quadrat method; Even all there is very big-difference in the different test duration point prediction precision of same testing location; So carry out combined prediction, can increase the useful information amount, improve precision of prediction.

Summary of the invention

The objective of the invention is to overcome the shortcoming of prior art,, propose the ultrashort phase combination forecasting method of a kind of wind farm wind velocity to the lower present situation of present China's predicting wind speed of wind farm method precision of prediction.

The present invention combines multiple Forecasting Methodology, and the useful information of comprehensive utilization the whole bag of tricks improves precision of prediction.

According to above inventive concept, Forecasting Methodology of the present invention mainly may further comprise the steps:

Step 1: the air speed data of gathering wind energy turbine set; And the data of obtaining are carried out pre-service; Said pre-service comprises to be filled a vacancy, the replacement of abnormal data and the time span of raw data is changed the leakage detection of raw data, and the formation sampling time is 15 minutes a historical wind speed time series.

Step 2: adopt the persistence forecasting model to carry out 0～1 hour prediction of wind speed; Obtaining leading k step forecasting wind speed value sequence

adopts the ARMA forecast model to carry out 0～4 hour prediction of wind speed; Obtain leading k step forecasting wind speed value sequence

and adopt the wavelet-neural net forecast model to carry out 0～4 hour prediction of wind speed, obtain leading k step forecasting wind speed value sequence

Step 3: utilize the combined prediction method that predicting the outcome of three kinds of forecast models made up, obtain 0～1 hour the ultrashort phase of wind speed to predict the outcome.Combined prediction adopts formula

Error is e behind the Combinatorial Optimization _k=ω _Pke _Pk+ ω _Ake _Ak+ ω _Nke _Nk, variance does

To optimize back variance minimum is objective function, at ω _Pk+ ω _Ak+ ω _NkConstruct Lagrange's equation under=1 the constraint condition: $f={\mathrm{\ω}}_{\mathrm{Pk}}^2\mathrm{Var}\left(e_{\mathrm{Pk}}\right)+{\mathrm{\ω}}_{\mathrm{Ak}}^2\mathrm{Var}\left(e_{\mathrm{Ak}}\right)+{\mathrm{\ω}}_{\mathrm{Nk}}^2\mathrm{Var}\left(e_{\mathrm{Nk}}\right)+\mathrm{\λ}(1-{\mathrm{\ω}}_{\mathrm{Pk}}-{\mathrm{\ω}}_{\mathrm{Ak}}-{\mathrm{\ω}}_{\mathrm{Nk}}),$ Former objective function promptly is evolved into the minimum value of asking for the neotectonics Lagrange's equation, through formula: $\left\{\begin{array}{c}\frac{\∂f}{\∂{\mathrm{\ω}}_{\mathrm{Pk}}}2{\mathrm{\ω}}_{\mathrm{Pk}}\mathrm{Var}\left(e_{\mathrm{Pk}}\right)-\mathrm{\λ}=0\\ \frac{\∂f}{\∂{\mathrm{\ω}}_{\mathrm{Ak}}}=2{\mathrm{\ω}}_{\mathrm{Ak}}\mathrm{Var}\left(e_{\mathrm{Ak}}\right)-\mathrm{\λ}=0\\ \frac{\∂f}{\∂{\mathrm{\ω}}_{\mathrm{Nk}}}=2{\mathrm{\ω}}_{\mathrm{Nk}}\mathrm{Var}\left(e_{\mathrm{Nk}}\right)-\mathrm{\λ}=0\\ \frac{\∂f}{\∂\mathrm{\λ}}=1-{\mathrm{\ω}}_{\mathrm{Pk}}-{\mathrm{\ω}}_{\mathrm{Ak}}-{\mathrm{\ω}}_{\mathrm{Nk}}=0\end{array}\right.$ Can try to achieve weight coefficient ω _Pk, ω _AkAnd ω _NkFor:

$\left\{\begin{array}{c}{\mathrm{\ω}}_{\mathrm{pk}}=\frac{\mathrm{Var}\left(e_{\mathrm{ak}}\right)\mathrm{Var}\left(e_{\mathrm{nk}}\right)}{\mathrm{Var}\left(e_{\mathrm{pk}}\right)\mathrm{Var}\left(e_{\mathrm{ak}}\right)+\mathrm{Var}\left(e_{\mathrm{pk}}\right)\mathrm{Var}\left(e_{\mathrm{nk}}\right)+\mathrm{Var}\left(e_{\mathrm{ak}}\right)\mathrm{Var}\left(e_{\mathrm{nk}}\right)}\\ {\mathrm{\ω}}_{\mathrm{ak}}=\frac{\mathrm{Var}\left(e_{\mathrm{pk}}\right)\mathrm{Var}\left(e_{\mathrm{nk}}\right)}{\mathrm{Var}\left(e_{\mathrm{pk}}\right)\mathrm{Var}\left(e_{\mathrm{ak}}\right)+\mathrm{Var}\left(e_{\mathrm{pk}}\right)\mathrm{Var}\left(e_{\mathrm{nk}}\right)+\mathrm{Var}\left(e_{\mathrm{ak}}\right)\mathrm{Var}\left(e_{\mathrm{nk}}\right)}\\ {\mathrm{\ω}}_{\mathrm{nk}}=\frac{\mathrm{Var}\left(e_{\mathrm{pk}}\right)\mathrm{Var}\left(e_{\mathrm{ak}}\right)}{\mathrm{Var}\left(e_{\mathrm{pk}}\right)\mathrm{Var}\left(e_{\mathrm{ak}}\right)+\mathrm{Var}\left(e_{\mathrm{pk}}\right)\mathrm{Var}\left(e_{\mathrm{nk}}\right)+\mathrm{Var}\left(e_{\mathrm{ak}}\right)\mathrm{Var}\left(e_{\mathrm{nk}}\right)}\end{array}\right..$

Predict the outcome obtaining 0～1 hour the ultrashort phase of wind speed in the weight coefficient substitution formula of trying to achieve

.

Wherein: f is a Lagrangian function; λ is Lagrangian coefficient; V _k ^*Be leading k step combined prediction value; ω _PkBe the weight coefficient of persistence forecasting model in leading k step wind speed combined prediction, ω _AkBe the weight coefficient of ARMA forecast model in leading k step combined prediction, ω _NkBe the weight coefficient of wavelet-neural net forecast model in leading k step wind speed combined prediction; e _PkBe the leading k step forecasting wind speed error of persistence forecasting model, e _AkBe the leading k step forecasting wind speed error of ARMA forecast model, e _NkBe the leading k step forecasting wind speed error of wavelet-neural net forecast model, e _kError for leading k step combined prediction; Var (e _Pk) be the variance of the leading k step predicated error of persistence forecasting model, Var (e _Ak) be the variance of the leading k step predicated error of ARMA forecast model, Var (e _Nk) be the variance of the leading k step predicated error of wavelet-neural net forecast model, Var (e _k) be the variance of leading k step combined prediction error; K is the predicted time step-length, k=1,2,3,4.

Step 4: utilize the combined prediction method that ARMA forecast model and predicting the outcome of wavelet-neural net forecast model are made up, obtain 1～4 hour the ultrashort phase of wind speed to predict the outcome.Combined prediction adopts formula promptly $\left\{\begin{array}{c}{V_k}^{*}={\mathrm{\ω}}_{\mathrm{pk}}{V}_{\mathrm{pk}}^{*}+{\mathrm{\ω}}_{\mathrm{ak}}{V}_{\mathrm{ak}}^{*}+{\mathrm{\ω}}_{\mathrm{nk}}{V}_{\mathrm{nk}}^{*}\\ {\mathrm{\ω}}_{\mathrm{pk}}=0\end{array}\right.,$ Similar in form with step 3, because predicted time step-length k does not have influence to asking for weight coefficient, so put aside predicted time step-length k.Can think that step 4 is that step 3 is at ω _PkSo=0 special circumstances are weight coefficient ω _AkAnd ω _NkCan obtain according to the solution procedure of step 3.Because ω _Pk=0, so the variance of persistence forecasting model is infinitely great, brings formula into and try to achieve weight coefficient ω _AkAnd ω _Nk: $\left\{\begin{array}{c}{\mathrm{\ω}}_{\mathrm{ak}}=\underset{\mathrm{Var}\left(e_{\mathrm{pk}}\right)\→\∞}{\lim }\frac{\mathrm{Var}\left(e_{\mathrm{pk}}\right)\mathrm{Var}\left(e_{\mathrm{nk}}\right)}{\mathrm{Var}\left(e_{\mathrm{pk}}\right)\mathrm{Var}\left(e_{\mathrm{ak}}\right)+\mathrm{Var}\left(e_{\mathrm{pk}}\right)\mathrm{Var}\left(e_{\mathrm{nk}}\right)+\mathrm{Var}\left(e_{\mathrm{ak}}\right)\mathrm{Var}\left(e_{\mathrm{nk}}\right)}=\frac{\mathrm{Var}\left(e_{\mathrm{nk}}\right)}{\mathrm{Var}\left(e_{\mathrm{ak}}\right)+\mathrm{Var}\left(e_{\mathrm{nk}}\right)}\\ {\mathrm{\ω}}_{\mathrm{nk}}=\underset{\mathrm{Var}\left(e_{\mathrm{pk}}\right)\→\∞}{\mathrm{lin}}\frac{\mathrm{Var}\left(e_{\mathrm{pk}}\right)\mathrm{Var}\left(e_{\mathrm{ak}}\right)}{\mathrm{Var}\left(e_{\mathrm{pk}}\right)\mathrm{Var}\left(e_{\mathrm{ak}}\right)+\mathrm{Var}\left(e_{\mathrm{pk}}\right)\mathrm{Var}\left(e_{\mathrm{nk}}\right)+\mathrm{Var}\left(e_{\mathrm{ak}}\right)\mathrm{Var}\left(e_{\mathrm{nk}}\right)}=\frac{\mathrm{Var}\left(e_{\mathrm{ak}}\right)}{\mathrm{Var}\left(e_{\mathrm{ak}}\right)+\mathrm{Var}\left(e_{\mathrm{nk}}\right)}\end{array}\right.;$ Predict the outcome obtaining 1～4 hour the ultrashort phase of wind speed in the weight coefficient substitution formula of trying to achieve

.

Wherein: predicted time step-length k=5,6 ... 16.

The effect that the present invention is useful is, has proposed the ultrashort phase combination forecasting method of a kind of wind speed, and predicting the outcome is superior to predicting the outcome of individual event, has improved the precision of the following 4 hours forecasting wind speeds of wind energy turbine set, has strengthened the region applicability of Forecasting Methodology.

Description of drawings

Fig. 1 is the overview flow chart of 0～4 hour ultrashort phase combined prediction of wind farm wind velocity provided by the invention;

Fig. 2 is the process flow diagram of ARMA forecast model;

Fig. 3 is the process flow diagram of wavelet-neural net forecast model.

Embodiment

The essence of ultrashort phase of wind speed prediction is exactly sometime predicted value v after obtaining through several wind speed observed readings ^*(t+k), in order to obtain more accurate forecasting wind speed value, general forecast time step k is unsuitable excessive.

Further describe the present invention below in conjunction with accompanying drawing and embodiment.

Fig. 1 is the overview flow chart of 0～4 hour ultrashort phase combined prediction of wind farm wind velocity provided by the invention.Comprise data input, predict, predict the outcome output three parts.Adopt step of the present invention following:

Step 1: the historical wind speed seasonal effect in time series generates

Gather the air speed data of wind energy turbine set, arrange according to the sequencing of acquisition time point and form the wind speed time series data.For improving precision of prediction and efficient; Need the data that obtain are carried out pre-service; Comprise the leakage detection of raw data filled a vacancy, converted 15 minutes into to the replacement of unusual and misdata with the time span of raw data that form historical wind speed time series V, wherein V can be expressed as:

V＝{v(t)，t＝1，2，…，N} (1)

Wherein: v (t) representes the air speed value of certain sampled point; T representes sampling instant; N is seasonal effect in time series air speed data point number.

Step 2: adopt the persistence forecasting model to carry out 0～1 hour prediction of wind speed

Persistence forecasting is exactly as the forecasting wind speed value of any down, that is: with the wind speed observed reading in a nearest moment

v ^*(t+k)＝v(t) (2)

Wherein: predicted time step-length k=1,2,3,4; v ^*(t+k) be the forecasting wind speed value in leading k step; V (t) is the wind speed observed reading of current time.Form leading k step forecasting wind speed value sequence

Because this forecast model is when predicted time is elongated; It is big that predicated error becomes rapidly, but its precision of prediction when wind speed prediction in leading 0～1 hour is higher, so when leading 0～1 hour combined prediction of wind speed; It is as one of them input quantity; But when leading 1～4 hour combined prediction of wind speed, do not consider this predicted results, promptly its corresponding weight coefficient equals 0.

Step 3: adopt the ARMA forecast model to carry out 0～4 hour prediction of wind speed

Utilize ARMA (p, q) model carries out forecasting wind speed and can be expressed as:

Wherein: v (t) ... v (t-p+1) is respectively corresponding t ... t-p+1 wind speed observed reading constantly; P is the exponent number of autoregression item; Q is the exponent number of running mean item;

And θ ₁, θ ₂... θ _qBe constant; ε is random disturbance or white noise series.

Fig. 2 is a process flow diagram of setting up the ARMA forecast model.At first, utilize the stationarity of unit root test method check wind series, if steadily then get into next step; If steadily then not carrying out first order difference handles; New data sequence is carried out stationary test again, do not handle if steadily then do not carry out second order difference, until obtaining time series stably.

Calculate steady seasonal effect in time series auto-correlation function value and partial correlation functional value; The form of utilizing autocorrelation function graph and partial correlation functional arrangement to come model of cognition; Come judgment models to belong to AR (p), MA (q) and ARMA (p according to table 1; Q) in which kind of, and the exponent number p of autoregression item and the exponent number q of running mean item in definite model.

Table 1

Model (stationary time series)	?AR(p)	?MA(q)	?ARMA(p，q)
				Autocorrelation function	Hangover	The truncation of q rank	Hangover
The partial correlation function	The truncation of p rank	Hangover	Hangover

After the form of model and exponent number p and q confirm, utilize the seasonal effect in time series auto-correlation function value that the model of tentatively choosing is carried out parameter estimation, calculate the parameters of model

And θ ₁, θ ₂... θ _q

Model carries out the test for randomness of significance test and residual error to parameters after confirming, if residual error is a sample sequence of white noise; The model of then being set up is suitable; Otherwise be improper, then model of cognition form and confirm exponent number p and q again meets the requirements until model.Present embodiment wind speed time series demonstrates stationarity after handling through first order difference, is designated as sequence Z={z (t), t=1,2;, N}, the model of foundation are ARMA (6; 1), this shows that predicted value is not only relevant with 6 historical wind speed values, and is also relevant with t random perturbation constantly.

After model is confirmed, promptly utilize recursion formula (4) that first order difference time series Z is carried out the prediction in leading k step, owing to carry out 0～4 hour ultrashort phase prediction, temporal resolution is 15 minutes in the present embodiment, thus k=1,2 ... 16.

When 1≤k≤q

…… (4)

When k＞q

After confirming the prediction in 0～4 hour of time series Z, promptly the counterplot of first order difference capable of using is calculated and is obtained leading k step forecasting wind speed value sequence

Step 4: adopt the wavelet-neural net forecast model to carry out 0～4 hour prediction of wind speed

Fig. 3 is the process flow diagram of wavelet-neural net forecast model.Because there is saturation problem in neural metwork training, so being carried out normalization, the wind speed time series handles, data normalization is arrived [0,1] interval.Normalized The data Mallet algorithm is carried out j layer wavelet decomposition, obtain j detail signal component D ₁(t), D ₂(t) ... D _j(t) and 1 approximation signal component A (t), present embodiment is got j=3.

J+1 component of signal after decomposing set up the BP neural network model respectively to be predicted.The BP neural network selects to comprise 3 layer networks of 1 hidden layer, and network input neuron number is the exponent number p of autoregression item in the ARMA forecast model, and network output neuron number is k, and the number of hidden nodes of network obtains through experience and the examination method of gathering.The hidden neuron transport function is selected logarithmic function for use, and output layer adopts linear transfer function, and training algorithm adopts the L-M algorithm.J+1 network with training predicted its corresponding j+1 component of signal; To predict the outcome and stack up; Carry out anti-normalization and handle, obtain the leading k step forecasting wind speed value sequence

of wavelet-neural net model

The ultrashort phase combined prediction of step 5:0～1 hour wind speed

Adopt combined method to obtain 0～1 hour ultrashort phase prediction of wind speed.Adopt formula (5) that three kinds of forecast models are carried out weighted optimization:

${V_k}^{*}={\mathrm{\ω}}_{\mathrm{pk}}{V}_{\mathrm{pk}}^{*}+{\mathrm{\ω}}_{\mathrm{ak}}{V}_{\mathrm{ak}}^{*}+{\mathrm{\ω}}_{\mathrm{nk}}{V}_{\mathrm{nk}}^{*}$ k＝1，2，3，4 (5)

Optimizing the back error is:

e _k＝ω _pke _pk+ω _ake _ak+ω _nke _nk k＝1，2，3，4 (6)

Variance is:

$\mathrm{Var}\left(e_k\right)={\mathrm{\ω}}_{\mathrm{pk}}^2\mathrm{Var}\left(e_{\mathrm{pk}}\right)+{\mathrm{\ω}}_{\mathrm{ak}}^2\mathrm{Var}\left(e_{\mathrm{ak}}\right)+{\mathrm{\ω}}_{\mathrm{nk}}^2\mathrm{Var}\left(e_{\mathrm{nk}}\right)$ k＝1，2，3，4 (7)

To optimize back variance minimum is objective function, at ω _Pk+ ω _Ak+ ω _NkConstruct Lagrange's equation under=1 the constraint condition:

$f={\mathrm{\ω}}_{\mathrm{pk}}^2\mathrm{Var}\left(e_{\mathrm{pk}}\right)+{\mathrm{\ω}}_{\mathrm{ak}}^2\mathrm{Var}\left(e_{\mathrm{ak}}\right)+{\mathrm{\ω}}_{\mathrm{nk}}^2\mathrm{Var}\left(e_{\mathrm{nk}}\right)+\mathrm{\λ}(1-{\mathrm{\ω}}_{\mathrm{pk}}-{\mathrm{\ω}}_{\mathrm{ak}}-{\mathrm{\ω}}_{\mathrm{nk}})$ k＝1，2，3，4 (8)

Through asking for the minimum value of neotectonics Lagrange's equation, can try to achieve weight coefficient and be:

$\left\{\begin{array}{c}{\mathrm{\ω}}_{\mathrm{pk}}=\frac{\mathrm{Var}\left(e_{\mathrm{ak}}\right)\mathrm{Var}\left(e_{\mathrm{nk}}\right)}{\mathrm{Var}\left(e_{\mathrm{pk}}\right)\mathrm{Var}\left(e_{\mathrm{ak}}\right)+\mathrm{Var}\left(e_{\mathrm{pk}}\right)\mathrm{Var}\left(e_{\mathrm{nk}}\right)+\mathrm{Var}\left(e_{\mathrm{ak}}\right)\mathrm{Var}\left(e_{\mathrm{nk}}\right)}\\ {\mathrm{\ω}}_{\mathrm{ak}}=\frac{\mathrm{Var}\left(e_{\mathrm{pk}}\right)\mathrm{Var}\left(e_{\mathrm{nk}}\right)}{\mathrm{Var}\left(e_{\mathrm{pk}}\right)\mathrm{Var}\left(e_{\mathrm{ak}}\right)+\mathrm{Var}\left(e_{\mathrm{pk}}\right)\mathrm{Var}\left(e_{\mathrm{nk}}\right)+\mathrm{Var}\left(e_{\mathrm{ak}}\right)\mathrm{Var}\left(e_{\mathrm{nk}}\right)}\\ {\mathrm{\ω}}_{\mathrm{nk}}=\frac{\mathrm{Var}\left(e_{\mathrm{pk}}\right)\mathrm{Var}\left(e_{\mathrm{ak}}\right)}{\mathrm{Var}\left(e_{\mathrm{pk}}\right)\mathrm{Var}\left(e_{\mathrm{ak}}\right)+\mathrm{Var}\left(e_{\mathrm{pk}}\right)\mathrm{Var}\left(e_{\mathrm{nk}}\right)+\mathrm{Var}\left(e_{\mathrm{ak}}\right)\mathrm{Var}\left(e_{\mathrm{nk}}\right)}\end{array}\right.$ k＝1，2，3，4(9)

Formula (9) substitution formula (5) is promptly obtained 0～1 hour ultrashort phase of wind speed

k=1 that predicts the outcome; 2; 3,4.Wherein: f is a Lagrangian function; λ is Lagrangian coefficient; V _k ^*Be leading k step combined prediction value; e _PkBe the leading k step forecasting wind speed error of persistence forecasting model, e _AkBe the leading k step forecasting wind speed error of ARMA forecast model, e _NkBe the leading k step forecasting wind speed error of wavelet-neural net forecast model, e _kError for leading k step combined prediction; Var (e _Pk) be the variance of the leading k step predicated error of persistence forecasting model, Var (e _Ak) be the variance of the leading k step predicated error of ARMA forecast model, Var (e _Nk) be the variance of the leading k step predicated error of wavelet-neural net forecast model, Var (e _k) be the variance of leading k step combined prediction error.

The ultrashort phase combined prediction of step 6:1～4 hour wind speed

Adopt combined method to obtain 1～4 hour ultrashort phase prediction of wind speed.Adopt formula (10) that ARMA and two kinds of models of wavelet-neural net are carried out weighted optimization.

${V_k}^{*}={\mathrm{\ω}}_{\mathrm{ak}}{V}_{\mathrm{ak}}^{*}+{\mathrm{\ω}}_{\mathrm{nk}}{V}_{\mathrm{nk}}^{*}$ k＝5，6，…，16 (10)

Formula can be expressed as:

$\left\{\begin{array}{c}{V_k}^{*}={\mathrm{\ω}}_{\mathrm{pk}}{V}_{\mathrm{pk}}^{*}+{\mathrm{\ω}}_{\mathrm{ak}}{V}_{\mathrm{ak}}^{*}+{\mathrm{\ω}}_{\mathrm{nk}}{V}_{\mathrm{nk}}^{*}\\ {\mathrm{\ω}}_{\mathrm{pk}}=0\end{array}\right.$ k＝5，6，…，16 (11)

Formula (11) is similar in form with formula (5), because predicted time step-length k does not have influence to asking for weight coefficient, so put aside predicted time step-length k.Can think that formula (11) is that formula (5) is at ω _PkSo=0 special circumstances are weight coefficient ω _AkAnd ω _NkCan obtain according to the solution procedure of step 5.Because ω _Pk=0, so the variance of persistence forecasting model is infinitely great, substitution formula (9) can be tried to achieve weight coefficient ω _AkWith little ω _Nk:

$\left\{\begin{array}{c}{\mathrm{\ω}}_{\mathrm{ak}}=\underset{\mathrm{Var}\left(e_{\mathrm{pk}}\right)\→\∞}{\lim }\frac{\mathrm{Var}\left(e_{\mathrm{pk}}\right)\mathrm{Var}\left(e_{\mathrm{nk}}\right)}{\mathrm{Var}\left(e_{\mathrm{pk}}\right)\mathrm{Var}\left(e_{\mathrm{ak}}\right)+\mathrm{Var}\left(e_{\mathrm{pk}}\right)\mathrm{Var}\left(e_{\mathrm{nk}}\right)+\mathrm{Var}\left(e_{\mathrm{ak}}\right)\mathrm{Var}\left(e_{\mathrm{nk}}\right)}=\frac{\mathrm{Var}\left(e_{\mathrm{nk}}\right)}{\mathrm{Var}\left(e_{\mathrm{ak}}\right)+\mathrm{Var}\left(e_{\mathrm{nk}}\right)}\\ {\mathrm{\ω}}_{\mathrm{nk}}=\underset{\mathrm{Var}\left(e_{\mathrm{pk}}\right)\→\∞}{\mathrm{lin}}\frac{\mathrm{Var}\left(e_{\mathrm{pk}}\right)\mathrm{Var}\left(e_{\mathrm{ak}}\right)}{\mathrm{Var}\left(e_{\mathrm{pk}}\right)\mathrm{Var}\left(e_{\mathrm{ak}}\right)+\mathrm{Var}\left(e_{\mathrm{pk}}\right)\mathrm{Var}\left(e_{\mathrm{nk}}\right)+\mathrm{Var}\left(e_{\mathrm{ak}}\right)\mathrm{Var}\left(e_{\mathrm{nk}}\right)}=\frac{\mathrm{Var}\left(e_{\mathrm{ak}}\right)}{\mathrm{Var}\left(e_{\mathrm{ak}}\right)+\mathrm{Var}\left(e_{\mathrm{nk}}\right)}\end{array}\right.$ k＝5，6，…，16 (12)

Formula (12) substitution formula (10) is promptly obtained 1～4 hour ultrashort phase of wind speed

k=5 that predicts the outcome; 6;, 16.

Obtain promptly that 0～4 hour the ultrashort phase of wind speed predicts the outcome

with formula (9) substitution formula (7) through above step 1 to step 6 $\mathrm{Var}\left(e_k\right)=\frac{\mathrm{Var}\left(e_{\mathrm{pk}}\right)\mathrm{Var}\left(e_{\mathrm{ak}}\right)\mathrm{Var}\left(e_{\mathrm{nk}}\right)}{\mathrm{Var}\left(e_{\mathrm{pk}}\right)\mathrm{Var}\left(e_{\mathrm{ak}}\right)+\mathrm{Var}\left(e_{\mathrm{pk}}\right)\mathrm{Var}\left(e_{\mathrm{nk}}\right)+\mathrm{Var}\left(e_{\mathrm{ak}}\right)\mathrm{Var}\left(e_{\mathrm{nk}}\right)}\≤\mathrm{Var}\left(e_{\mathrm{pk}}\right),$ Can draw for leading k and go on foot the error variance of the error variance of wind speed combined prediction, improve 0～4 hour ultrashort phase accuracy of predicting of wind speed of wind energy turbine set less than any single prediction.

Claims (2)

1. ultrashort phase combination forecasting method of wind farm wind velocity is characterized in that the step of said method is following:

Step 1: gather the air speed data of wind energy turbine set, and the data of obtaining are carried out pre-service; Said pre-service comprises to be filled a vacancy, the replacement of abnormal data and the time span of raw data is changed the leakage detection of raw data, forms the wind speed time series;

Step 2: adopt the persistence forecasting model to carry out 0～1 hour prediction of wind speed; Temporal resolution is 15 minutes, obtains leading k step forecasting wind speed value sequence

Step 3: adopt the ARMA forecast model to carry out 0～4 hour prediction of wind speed; Temporal resolution is 15 minutes, obtains leading k step forecasting wind speed value sequence

Step 4: adopt the wavelet-neural net forecast model to carry out 0～4 hour prediction of wind speed; Temporal resolution is 15 minutes, obtains leading k step forecasting wind speed value sequence

Step 5: utilize the combined prediction method, predicting the outcome of step 2, step 3 and step 4 carried out combined prediction, obtain 0～1 hour the ultrashort phase of wind speed to predict the outcome;

Step 6: utilize the combined prediction method, predicting the outcome of step 3 and step 4 carried out combined prediction, obtain 1～4 hour the ultrashort phase of wind speed to predict the outcome.

2. the ultrashort phase combination forecasting method of wind farm wind velocity according to claim 1 is characterized in that, the combined prediction method in the described step 5 adopts formula

Carry out weighted optimization, optimizing the back error is e _k=ω _Pke _Pk+ ω _Ake _Ak+ ω _Nke _Nk, variance does

To optimize back variance minimum is objective function, at ω _Pk+ ω _Ak+ ω _NkConstruct Lagrange's equation under=1 the constraint condition: $f={\mathrm{\ω}}_{\mathrm{Pk}}^2\mathrm{Var}\left(e_{\mathrm{Pk}}\right)+{\mathrm{\ω}}_{\mathrm{Ak}}^2\mathrm{Var}\left(e_{\mathrm{Ak}}\right)+{\mathrm{\ω}}_{\mathrm{Nk}}^2\mathrm{Var}\left(e_{\mathrm{Nk}}\right)+\mathrm{\λ}(1-{\mathrm{\ω}}_{\mathrm{Pk}}-{\mathrm{\ω}}_{\mathrm{Ak}}-{\mathrm{\ω}}_{\mathrm{Nk}}),$ Minimum value through asking this equation can be tried to achieve weight coefficient ω _Pk, ω _AkAnd ω _NkFor: $\left\{\begin{array}{c}{\mathrm{\ω}}_{\mathrm{Pk}}=\frac{\mathrm{Var}\left(e_{\mathrm{Ak}}\right)\mathrm{Var}\left(e_{\mathrm{Nk}}\right)}{\mathrm{Var}\left(e_{\mathrm{Pk}}\right)\mathrm{Var}\left(e_{\mathrm{Ak}}\right)+\mathrm{Var}\left(e_{\mathrm{Pk}}\right)\mathrm{Var}\left(e_{\mathrm{Nk}}\right)+\mathrm{Var}\left(e_{\mathrm{Ak}}\right)\mathrm{Var}\left(e_{\mathrm{Nk}}\right)}\\ {\mathrm{\ω}}_{\mathrm{Ak}}=\frac{\mathrm{Var}\left(e_{\mathrm{Pk}}\right)\mathrm{Var}\left(e_{\mathrm{Nk}}\right)}{\mathrm{Var}\left(e_{\mathrm{Pk}}\right)\mathrm{Var}\left(e_{\mathrm{Ak}}\right)+\mathrm{Var}\left(e_{\mathrm{Pk}}\right)\mathrm{Var}\left(e_{\mathrm{Nk}}\right)+\mathrm{Var}\left(e_{\mathrm{Ak}}\right)\mathrm{Var}\left(e_{\mathrm{Nk}}\right)}\\ {\mathrm{\ω}}_{\mathrm{Nk}}=\frac{\mathrm{Var}\left(e_{\mathrm{Pk}}\right)\mathrm{Var}\left(e_{\mathrm{Ak}}\right)}{\mathrm{Var}\left(e_{\mathrm{Pk}}\right)\mathrm{Var}\left(e_{\mathrm{Ak}}\right)+\mathrm{Var}\left(e_{\mathrm{Pk}}\right)\mathrm{Var}\left(e_{\mathrm{Nk}}\right)+\mathrm{Var}\left(e_{\mathrm{Ak}}\right)\mathrm{Var}\left(e_{\mathrm{Nk}}\right)}\end{array}\right.;$

With the weight coefficient ω that tries to achieve _Pk, ω _AkAnd ω _NkThe substitution formula In can obtain 0～1 hour ultrashort phase predicted value of wind speed;

Wherein: f is a Lagrangian function; λ is Lagrangian coefficient; V _k ^*Be leading k step combined prediction value; ω _PkBe the weight coefficient of persistence forecasting model in leading k step combined prediction, ω _AkBe the weight coefficient of ARMA forecast model in leading k step combined prediction, ω _NkBe the weight coefficient of wavelet-neural net forecast model in leading k step combined prediction; e _PkBe the leading k step forecasting wind speed error of persistence forecasting model, e _AkBe the leading k step forecasting wind speed error of ARMA forecast model, e _NkBe the leading k step forecasting wind speed error of wavelet-neural net forecast model, e _kError for leading k step combined prediction; Var (e _Pk) be the variance of the leading k step predicated error of persistence forecasting model, Var (e _Ak) be the variance of the leading k step predicated error of ARMA forecast model, Var (e _Nk) be the variance of the leading k step predicated error of wavelet-neural net forecast model, Var (e _k) be the variance of leading k step combined prediction error; K is the predicted time step-length, k=1,2,3,4.

The ultrashort phase combination forecasting method of 3 wind farm wind velocities according to claim 1 is characterized in that, the combined prediction method in the described step 6 adopts formula ${V_k}^{*}={\mathrm{\ω}}_{\mathrm{Ak}}{V}_{\mathrm{Ak}}^{*}+{\mathrm{\ω}}_{\mathrm{Nk}}{V}_{\mathrm{Nk}}^{*},$ Promptly $\left\{\begin{array}{c}{V_k}^{*}={\mathrm{\ω}}_{\mathrm{Pk}}{V}_{\mathrm{Pk}}^{*}+{\mathrm{\ω}}_{\mathrm{Ak}}{V}_{\mathrm{Ak}}^{*}+{\mathrm{\ω}}_{\mathrm{Nk}}{V}_{\mathrm{Nk}}^{*}\\ {\mathrm{\ω}}_{\mathrm{Pk}}=0\end{array}\right.,$ Below the substitution

Formula can obtain weight coefficient ω _AkAnd ω _NkFor:

$\left\{\begin{array}{c}{\mathrm{\ω}}_{\mathrm{ak}}=\underset{\mathrm{Var}\left(e_{\mathrm{pk}}\right)\→\∞}{\lim }\frac{\mathrm{Var}\left(e_{\mathrm{pk}}\right)\mathrm{Var}\left(e_{\mathrm{nk}}\right)}{\mathrm{Var}\left(e_{\mathrm{pk}}\right)\mathrm{Var}\left(e_{\mathrm{ak}}\right)+\mathrm{Var}\left(e_{\mathrm{pk}}\right)\mathrm{Var}\left(e_{\mathrm{nk}}\right)+\mathrm{Var}\left(e_{\mathrm{ak}}\right)\mathrm{Var}\left(e_{\mathrm{nk}}\right)}=\frac{\mathrm{Var}\left(e_{\mathrm{nk}}\right)}{\mathrm{Var}\left(e_{\mathrm{ak}}\right)+\mathrm{Var}\left(e_{\mathrm{nk}}\right)}\\ {\mathrm{\ω}}_{\mathrm{nk}}=\underset{\mathrm{Var}\left(e_{\mathrm{pk}}\right)\→\∞}{\mathrm{lin}}\frac{\mathrm{Var}\left(e_{\mathrm{pk}}\right)\mathrm{Var}\left(e_{\mathrm{ak}}\right)}{\mathrm{Var}\left(e_{\mathrm{pk}}\right)\mathrm{Var}\left(e_{\mathrm{ak}}\right)+\mathrm{Var}\left(e_{\mathrm{pk}}\right)\mathrm{Var}\left(e_{\mathrm{nk}}\right)+\mathrm{Var}\left(e_{\mathrm{ak}}\right)\mathrm{Var}\left(e_{\mathrm{nk}}\right)}=\frac{\mathrm{Var}\left(e_{\mathrm{ak}}\right)}{\mathrm{Var}\left(e_{\mathrm{ak}}\right)+\mathrm{Var}\left(e_{\mathrm{nk}}\right)}\end{array}\right.;$

With the weight coefficient ω that tries to achieve _AkAnd ω _NkThe substitution formula

In can obtain 1～4 hour ultrashort phase predicted value of wind speed;

Wherein: predicted time step-length k=5,6 ... 16.

Images