CN102567636B - High-precision wind electric field power interval forecasting method based on relevance vector machine - Google Patents

Abstract

The invention discloses a high-precision wind electric field power interval forecasting method based on a relevance vector machine, which comprises the steps of collecting data, carrying out normalization on the data, and selecting a training sample of a relevance vector machine forecasting model; optimizing parameters of the relevance vector machine forecasting model to obtain an optimized iteration initial value of kernel function width and the relevance vector machine forecasting model; obtaining a kernel function, and then obtaining relevance vector machine forecasting model parameters after convergence; and finally obtaining a forecasting value and a variance of a wind electric field, so as to obtain a forecasted interval of wind electric field power. The method can improve adaptability of the model, improve forecasting accuracy, reduce training sample size and reduce training time.

Description

Based on the high precision wind power interval prediction method of Method Using Relevance Vector Machine

Technical field

The invention belongs to technical field of wind power, particularly relate to a kind of high precision wind power interval prediction method based on Method Using Relevance Vector Machine.

Background technology

Along with the great-leap-forward development of wind-power electricity generation, the economy of the intrinsic intermittence of wind-power electricity generation, stochastic volatility serious threat Operation of Electric Systems, stability and power supply quality.Wind farm power prediction technology alleviates one of wind farm grid-connected effective way to electric system adverse effect.Wind farm power prediction also contributes to accurately and reliably: (1) reduces spinning reserve capacity, reasonable arrangement turnaround plan, thus reduces operating cost; (2) wind-electricity integration ratio is improved; (3) improve wind-powered electricity generation enterprise competitiveness, surf the Net at a competitive price for wind-powered electricity generation and advantage is provided.

Wind electric field power prediction method conventional both at home and abroad has: artificial neural network (ANN), support vector machine (SVM) etc.Artificial Neural Network can approach any non-linear continuous function in theory, is therefore successfully applied to wind power prediction field.But ANN easily caused problem concerning study in training.Its consequence is: when for given data, and predicated error is very little; When unknown data outside input amendment, predicated error is increased sharply.This is also referred to as limited generalization ability.Based on this understanding, improving one of way of ANN model generalization ability is exactly increase training sample quantity.Demand for a large amount of training sample too increases the application limitation of ANN conversely, as: China's wind energy turbine set is not long for most working time, and historical data is not enough, is difficult to obtain enough training samples and sets up ANN forecast model; More be not easy to divide and month or season set up ANN forecast model.

In order to improve the generalization ability of model and the demand broken away from for a large amount of training sample, SVM technology is applied in wind power prediction field by many scholars, use less training sample to realize predicting comparatively accurately, and effectively avoid the danger being absorbed in Local Minimum.But SVM also embodies some shortcomings part in Practical Project uses:

1. kernel function must meet horse plucked instrument condition, and namely kernel function is confined to continuous symmetric function;

2. predict the outcome as point estimation, any unascertained information can not be provided;

3. the number of support vector along with the increase of number of training linear increase, add computation complexity;

4. must estimate some insensitive parameters, cause the meaningless increase of calculated amount and parameter amount.

In order to overcome above-mentioned shortcoming, calendar year 2001 Michael E.Tipping proposes a kind of probability learning method based on bayesian theory, edge likelihood theory---Method Using Relevance Vector Machine (Relevance Vector Machine, RVM).It is the simple nonlinear pattern recognition model of a kind of structure, not only remains the outstanding predictive ability of SVM and generalization ability, further improves the weak point of SVM.Therefore, the method Successful utilization in multiple fields, as load prediction, failure modes etc., but to be not yet applied in wind farm power prediction.

Summary of the invention

Weak and cannot the deficiencies such as unascertained information be provided for the existing wind power forecasting method generalization ability mentioned in above-mentioned background technology, the present invention proposes a kind of high precision wind power interval prediction method based on Method Using Relevance Vector Machine.

Technical scheme of the present invention is, based on the high precision wind power interval prediction method of Method Using Relevance Vector Machine, it is characterized in that the method comprises the following steps:

Step 1: image data normalization, selects the training sample of Method Using Relevance Vector Machine forecast model according to the absolute error of wind speed in numerical weather forecast;

Step 2: be optimized the parameter of Method Using Relevance Vector Machine forecast model on the basis of step 1, obtains the iterative initial value of optimized kernel function width and Method Using Relevance Vector Machine forecast model;

Step 3: try to achieve kernel function on the basis of step 2, and then try to achieve the Method Using Relevance Vector Machine prediction model parameters after convergence;

Step 4: predicted value and the variance of trying to achieve wind power, obtain the forecast interval of wind power.

The described training sample selecting Method Using Relevance Vector Machine forecast model according to the absolute error of wind speed in numerical weather forecast is specially:

Step 1: be station work sample according to the absolute error of wind speed in numerical weather forecast;

Step 2: each group is tested with Method Using Relevance Vector Machine forecast model respectively, chooses one group of training sample as Method Using Relevance Vector Machine forecast model that wherein precision of prediction is the highest.

Described step 2 is specially:

Step 2.1: using training sample as the input amendment of particle cluster algorithm, tries to achieve the adaptive value of the fitness function of each particle in particle cluster algorithm in each group;

Step 2.2: select the position of the minimum particle of adaptive value in all particles as global extremum;

Step 2.3: whether inspection adaptive value meets iteration requirement, if meet iteration requirement, then global extremum is optimum kernel function width parameter and the iterative initial value of Method Using Relevance Vector Machine model; Otherwise more the speed of new particle and position, recalculate adaptive value.

Described fitness function is:

$f=\frac{\sqrt{{\mathrm{\Σ}}_{i=1}^n{(P_{\mathrm{Mi}}-P_{\mathrm{Pi}})}^2}}{\mathrm{Cap}\×\sqrt{n}}$

Wherein:

F is fitness function;

P _mifor the real power in i moment;

P _pifor the predicted power in i moment;

Cap is wind energy turbine set start total volume;

N is all number of samples.

Described iteration requires to be less than 0.1 for adaptive value.

Described step 3 is specially:

Step 3.1: calculate the kernel function of Method Using Relevance Vector Machine forecast model, the variance of training sample Posterior distrbutionp and the weighted value of Posterior distrbutionp;

Step 3.2: on the basis of step 3.1, utilizes the maximal margin estimation technique to calculate the maximization numerical value of the prior distribution of hyper parameter;

Step 3.3: in inspection current iteration step, whether hyper parameter meets specified requirements, if meet iteration requirement, then hyper parameter is now Method Using Relevance Vector Machine prediction model parameters; Otherwise, upgrade hyper parameter, until meet iteration requirement.

The computing formula of described kernel function is:

$K(X_i,X_j)=\exp (-\frac{{\left|\right|X_i-X_j\left|\right|}^2}{2{\mathrm{\θ}}^2})$

Wherein:

K (X _i, X _j) be kernel function;

X _i, X _jfor the mode input parameter after normalization;

θ is the kernel function width after PSO optimizes.

The computing formula of the variance of described training sample Posterior distrbutionp is:

∑＝(σ ^-2φ ^Tφ+A) ^-1

Wherein:

∑ is the variance of Posterior distrbutionp;

A is hyper parameter diagonal entry, A=diag (α ₀, α ₁... α _n);

σ ^-2for iterative initial value;

φ is basis function vector.

The computing formula of the weighted value of described Posterior distrbutionp is:

w _i＝σ ^-2∑φ ^Tt

Wherein:

W _ifor Posterior distrbutionp weighted value;

T is the target component of training sample.

The predicted value of described wind power and the computing formula of variance be respectively:

y _*＝w ^Tφ

σ ² _*＝σ ² _MP+φ ^T∑φ

Wherein:

Y _*for wind farm power prediction monodrome;

σ ² _*for wind farm power prediction is worth variance.

Advantage of the present invention:

One, present invention employs Method Using Relevance Vector Machine (RVM) method, the precision of prediction of RVM be high, generalization ability is strong, and there is following considerable advantage:

(1) diversity of Selection of kernel function: without the need to meeting horse plucked instrument condition, improve the adaptive faculty of model, can more widely, the power generation characteristics of the better different wind energy turbine set of simulation;

(2) probabilistic forecasting: single power prediction numerical value is not only provided, predict the outcome under also providing certain confidence level contingent fluctuation upper lower limit value;

(3) training process needs less sample: the power Method Using Relevance Vector Machine forecast model can setting up different month on the one hand very easily, more meets numerical weather forecast accuracy rate characteristic distributions, likely improves the precision of Method Using Relevance Vector Machine forecast model; Be applicable to the less newly-built wind energy turbine set of historical data on the other hand;

(4) openness: most associated vector automatically goes to zero in training, make the number of number much smaller than support vector of associated vector, and can not along with the increase of training sample linear increase.Because learning process only pays close attention to those vectors relevant with high-precision forecast result, greatly reduce model complexity, improve the efficiency of model training;

(5) optimum configurations is simple: being different from support vector machine needs to arrange additional parameter, and Method Using Relevance Vector Machine only needs manually to arrange kernel function width, at utmost reduces the impact of subjective factor for model.

Two, the present invention is based on numerical weather forecast (NWP) precision distribution characteristics and propose a kind of training sample screening technique, the method effectively improves precision of prediction, reduces training sample scale, decreases the training time.

Three, the present invention adopts particle group optimizing method (PSO) to carry out optimizing to kernel function width and iterative initial value, improves precision of prediction further.

Accompanying drawing explanation

Fig. 1 is wind power interval prediction system construction drawing;

Fig. 2 is training sample screening process figure;

Fig. 3 is Model Parameter Optimization process flow diagram;

Fig. 4 is Method Using Relevance Vector Machine forecast model training process flow diagram;

Fig. 5 is prediction and uncertainty analysis process flow diagram;

Fig. 6 is that 1 month by month correlation vector machine error prediction model in 2010 is with NWP accuracy class change curve;

Fig. 7 is each moon RSP method, SVM method, GA-ANN method root-mean-square error figure;

Fig. 8 is each moon RSP method, SVM method, GA-ANN method mean absolute error figure;

Fig. 9 is spring (May 24 was example) wind power interval prediction result figure;

Figure 10 is summer (July 29 was example) wind power interval prediction result figure;

Figure 11 is autumn (September 25 was example) wind power interval prediction result figure;

Figure 12 is winter (Dec 26 was example) wind power interval prediction result figure.

Embodiment

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

For the defect of existing wind electric field power prediction method, the present invention proposes a kind of high precision wind power interval prediction method based on Method Using Relevance Vector Machine.The method carries out the screening of training sample according to the feature that numerical weather forecast (NWP) error distributes, particle group optimizing method (PSO) is adopted to carry out optimizing to kernel function width and iterative initial value, and fully utilized the advantage of Method Using Relevance Vector Machine (RVM), the interval Method Using Relevance Vector Machine forecast model of wind power that establish that precision of prediction is high, generalization ability is strong, uncertainty analysis can be provided.The method also has lower to training sample scale requirements, that computation process is easy advantage.

Technical scheme of the present invention is, wind power interval prediction method, is characterized in that the method comprises the following steps:

The present invention quotes Method Using Relevance Vector Machine (RVM) method and carries out wind power interval prediction, and have stronger predictive ability and generalization ability, training sample demand is few.In order to improve the precision of Method Using Relevance Vector Machine forecast model further, the present invention, according to the characteristic distributions of numerical weather forecast (NWP) error, proposes a kind of screening technique of training sample; Particle group optimizing method (PSO) is utilized to be optimized kernel functional parameter and iterative initial value.In addition, it is simple that the present invention also has optimum configurations, the advantages such as working time is short, is convenient to promote in Practical Project adopt.

The technical problem to be solved in the present invention is: for existing wind electric field power prediction method Problems existing, proposes that a kind of precision of prediction is high, generalization ability is strong, provides uncertainty analysis, the wind power interval prediction method that few to training sample quantity demand, operation time is short; Propose a kind of training sample screening technique based on numerical weather forecast (NWP) accuracy characteristic, to improve precision of prediction, reduce the training time; A kind of wind power interval prediction method based on Method Using Relevance Vector Machine (RVM) is provided, and adopts particle group optimizing method (PSO) to carry out optimizing to kernel function width and Method Using Relevance Vector Machine forecast model iterative initial value.

The present invention proposes the high precision wind power interval prediction method based on Method Using Relevance Vector Machine (Relevance vector machine with samples Selection and Particle swarm optimization is called for short RSP method) comprising 4 stages, 18 steps.

Training sample screening stage comprises 3 steps, is respectively: data acquisition and normalization; Packet: by whole sample by after month grouping again according to the absolute error of NWP wind speed by each moon sample packet; Choose each moon to the best NWP accuracy class of accurately predicting, screen the training sample of Method Using Relevance Vector Machine forecast model accordingly, enter the Model Parameter Optimization stage.

The Model Parameter Optimization stage comprises 5 steps, is respectively: population parameter initialization; The calculating of particle fitness function value; The calculating of individual extreme value and global extremum; Upgrade particle rapidity and particle position; After meeting iterated conditional, deposit optimum kernel function width and Method Using Relevance Vector Machine forecast model iterative initial value, enter the Method Using Relevance Vector Machine forecast model training stage.

The Method Using Relevance Vector Machine forecast model training stage comprises 7 steps, is respectively: input training sample and Optimal Parameters; Kernel function calculates; Posterior distrbutionp variance calculates; Posterior distrbutionp weighted value calculates; Hyper parameter α _iand σ ²calculating; Inspection iterated conditional, upgrades hyper parameter α _iand σ ²; Delete invalid vector, after meeting iterated conditional, preservation model parameter, enters prediction and uncertainty analysis stage.

Prediction and uncertainty analysis stage comprise 3 steps, are respectively: calculate wind farm power prediction monodrome and variance; Renormalization process; Wind farm power prediction uncertainty analysis.

Operating process of the present invention as shown in Figure 1, is described in detail below in conjunction with type selecting example:

A. training sample screening stage, as shown in Figure 2

B. the Model Parameter Optimization stage, as shown in Figure 3

C. the Method Using Relevance Vector Machine forecast model training stage, as shown in Figure 4

D. prediction and the uncertainty analysis stage, as shown in Figure 5

Comprise in above-mentioned steps A (training sample screening stage):

A1. from wind farm data collection and supervisor control (SCADA) system and anemometer tower, extract reality based on the data acquisition module of the interval Method Using Relevance Vector Machine forecast model of high precision wind power of Method Using Relevance Vector Machine and survey wind data, Power Output for Wind Power Field, numerical weather forecast data (NWP), input parameter using NWP as model training power Output for Wind Power Field is as model training target parameter note: x _ifor sample parameter, be respectively the wind speed in NWP, wind direction sine value, wind direction cosine value, temperature, pressure, humidity; N is sample size.Then normalized input parameter and corresponding target parameter, normalization formula is:

x＝[x _ori-min(x _ori)]/[[max(x _ori)-min(x _ori)]]

Wherein:

X _orifor original sample parameter;

X is the sample parameter after normalization.

A2. first whole sample was divided into groups according to month, then the absolute error of the precision of each moon NWP according to NWP wind speed is divided into groups.Such as, < 1.2m/s representative is chosen NWP wind speed absolute error and is less than the sample of 1.2m/s as training sample, and namely < 1.2m/s is higher than the NWP accuracy class of < 1.8m/s.The absolute error e of NWP wind speed _nWPexpression formula is as follows:

e＝|x _NWP-x _act|

Wherein:

E is NWP wind speed absolute error;

X _nWPfor the wind speed numerical value in NWP result;

X _actfor the wind speed numerical value of actual measurement.

A3. find best NWP accuracy class and set up training sample.As shown in Fig. 6 and table 1, the precision of prediction in example wind energy turbine set January first raises along with the raising of NWP accuracy class, and when continuing to improve NWP accuracy class, precision of prediction declines from mxm. again; Therefore the training sample in January is the data of NWP wind speed absolute error < 1.5m/s.NWP grade of errors corresponding to different month Method Using Relevance Vector Machine precision of forecasting model mxm. is not identical, that is: be station work sample according to the absolute error of wind speed in numerical weather forecast; Each group is tested with Method Using Relevance Vector Machine forecast model respectively, chooses one group of training sample as Method Using Relevance Vector Machine forecast model that wherein precision of prediction is the highest.As shown in table 2.

Method Using Relevance Vector Machine precision of forecasting model under table 1.1 month different N WP accuracy class

NWP accuracy class	＜1.8m/s	＜1.7m/s	＜1.6m/s	＜1.5m/s	＜1.4m/s	＜1.3m/s	＜1.2m/s
								Method Using Relevance Vector Machine precision of forecasting model	0.14254	0.14226	0.14049	0.13801	0.14477	0.15791	0.15864

The NWP accuracy class that table 2. each month by month correlation vector machine precision of forecasting model mxm. is corresponding

Through the screening to example wind energy turbine set training sample, for each month sets up the training sample meeting NWP precision characteristic distributions, final each month, training sample number was as shown in table 3: visible by time Filtering system, training sample scale reduces, therefore saved computing time, proved that the method has the advantage little to training sample quantity demand simultaneously.

Table 3. each moon whole sample size, screening after training sample quantity

Above-mentioned B-stage comprised in (Model Parameter Optimization stage):

B1. with the input amendment that the training sample of A stage gained is the Model Parameter Optimization stage.This stage adopts particle cluster algorithm (PSO), first the initial value of particle position and speed in each colony of PSO is set, in example wind energy turbine set, number of particles gets 20, particle initial position is between (0,30) random number between, speed initial value is the random number between (-2,2);

B2. calculate the adaptive value of each particle, fitness function f expression formula is:

$f=\frac{\sqrt{{\mathrm{\&Sigma;}}_{i=1}^n{(P_{\mathrm{Mi}}-P_{\mathrm{Pi}})}^2}}{\mathrm{Cap}\&times;\sqrt{n}}$

Wherein:

P _mifor the real power in i moment;

P _pifor the predicted power in i moment;

Cap is wind energy turbine set start total volume;

N is all number of samples.

B3. select particle position that in each colony, particle fitness function value is minimum respectively as the individual extreme value of each colony, select particle position minimum in all particles fitness function value as global extremum;

B4. upgrade particle rapidity and position, expression formula is respectively:

$v_{i,d}^{k+1}=\mathrm{\&omega;}\&CenterDot;v_{i,d}^k+c_1\&CenterDot;\mathrm{rand}\left(\right)({\mathrm{pd}}_{i,d}^k-y_{i,d}^k)+c_2\&CenterDot;\mathrm{rand}\left(\right)({\mathrm{gb}}_{i,d}^k-y_{i,d}^k)$

$y_{i,d}^{k+1}=y_{i,d}^k+v_{i,d}^{k+1}$

Wherein:

C ₁and c ₂for Studying factors;

Rand () is the random number between (0,1);

with be respectively speed and the position of d dimension in particle i kth time iteration;

for the position of the individual extreme value that particle i ties up at d;

for the position of the global extremum that colony ties up at d;

ω is the inertia weight factor.

B5. adaptive value is checked whether to meet iteration requirement, as met then finishing iteration, when the global extremum obtained after secondary iteration is optimum kernel function width parameter and Method Using Relevance Vector Machine forecast model iterative initial value, do not meet and then return B2 until restrain or reach maximum iteration time.

Comprise in the above-mentioned C stage:

C1: kernel function width, Method Using Relevance Vector Machine forecast model iterative initial value (B-stage obtains result) after the best training sample (the A stage obtains result) after input screening and PSO optimize;

C2: the kernel function K (X calculating the interval Method Using Relevance Vector Machine forecast model of high precision wind power based on Method Using Relevance Vector Machine _i, X _j), expression formula is:

$K(X_i,X_j)=\exp (-\frac{{\left|\right|X_i-X_j\left|\right|}^2}{2{\mathrm{\&theta;}}^2})$

Wherein:

K (X _i, X _j) be kernel function;

X _i, X _jfor the mode input parameter after normalization;

θ is the kernel function width after PSO optimizes.

C3: the variance ∑ of calculation training sample Posterior distrbutionp, expression formula is:

∑＝(σ ^-2φ ^Tφ+A) ^-1

Wherein:

∑ is the variance of Posterior distrbutionp;

A=diag (α ₀, α ₁... α _n) be hyper parameter diagonal entry;

σ ^-2for iterative initial value;

φ is basis function vector, φ=(1, K (x, t)) ^t, x is input parameter, and t is target parameter; C4: the weighted value w calculating Posterior distrbutionp _i, expression formula is:

w _i＝σ ^-2∑φ ^Tt

Wherein:

W _ifor Posterior distrbutionp weighted value;

T is the target component of training sample.

C5: utilize the maximal margin estimation technique to calculate about hyper parameter { α _iand σ ²maximized prior distribution, maximal margin estimation technique expression formula is:

L＝-0.5[log|β ^-1I+φA ^-1φ ^T|+t ^T(β ^-1I+φA ^-1φ ^T) ^-1t]

Wherein:

β is iterative initial value σ ^-2inverse;

I is unit matrix;

C6: if do not meet iterated conditional (α between twice iteration _ithe variable quantity of numerical value is less than 10 ^-6) then continue to upgrade to calculate { α _iand σ ²until convergence, parameter more new formula is:

${\mathrm{\&alpha;}}_i^{\mathrm{new}}=\frac{1-{\mathrm{\&alpha;}}_i{\mathrm{\&Sigma;}}_{\mathrm{ii}}}{w_i^2};$ ${\left({\mathrm{\&sigma;}}^2\right)}^{\mathrm{new}}=\frac{{\left|\right|t-\mathrm{\&phi;}{w}_i\left|\right|}^2}{N-\mathrm{\&Sigma;}(1-{\mathrm{\&alpha;}}_i{\mathrm{\&Sigma;}}_{\mathrm{ii}})}$

Wherein:

∑ _iii-th diagonal entry of Posterior distrbutionp variance ∑;

N is sample size;

C7: delete α in iterative process _iw corresponding to → ∞ _i; The w retained _icorresponding vector is " associated vector " of Method Using Relevance Vector Machine forecast model, and after convergence, Method Using Relevance Vector Machine prediction model parameters is: α _mP, σ ² _mP.

The above-mentioned D stage comprises:

D1: substitute into kernel function width that test sample book, B-stage optimization obtains and Method Using Relevance Vector Machine forecast model iterative initial value, Method Using Relevance Vector Machine prediction model parameters α that C stage-training obtains _mP, σ ² _mP, try to achieve the wind farm power prediction value between (0,1) and variance thereof, expression formula is as follows:

y _*＝w ^Tφ

σ ² _*＝σ ² _MP+φ ^T∑φ

Wherein:

Y _*for wind farm power prediction monodrome;

σ ² _*for wind farm power prediction is worth variance.

D2: renormalization process, expression formula is:

x _anti＝x _0-1×[max(x _ori)-min(x _ori)]+min(x _ori)

Wherein:

X _antifor renormalization result;

X _0-1for the parameter of [0,1];

X _orifor original parameter;

D3: given confidence level parameter δ, calculate the contingent fluctuation higher limit of wind farm power prediction monodrome and fluctuation lower limit, i.e. uncertainty analysis, expression formula is as follows:

P _upper＝P _prep×(1+σ _*z _δ/2)

P _lower＝P _prep×(1-σ _*z _δ/2)

Wherein:

P _upperfor the contingent fluctuation higher limit of wind farm power prediction value;

P _lowerfor the contingent fluctuation lower limit of wind farm power prediction value;

P _prepfor wind farm power prediction value;

σ _*for wind farm power prediction is worth mean square deviation;

Z _δ/2for the critical value that standardized normal distribution under confidence level 1-δ is corresponding;

Through wind power interval prediction, not only can obtain the predicted value of a real number, the contingent fluctuation range under certain confidence level can also be tried to achieve.The decision in the face of risk that this probabilistic forecasting mechanism substantially increases model is worth.

Sample calculation analysis

1 data

Check analysis is carried out with the data instance of north of China wind energy turbine set, data comprise average wind power measured value, the mean wind speed of the every 15min extracted from SCADA system, and the same period numerical weather forecast result, the time is (except October) in 2010.Have 80% for training in each moon sample, 20% for test.

Choose 2 kinds of instantly popular intelligent learning methods: neural network (GA-ANN) method as a comparison of support vector machine (SVM), genetic algorithm optimization, from precision of prediction, uncertainty analysis, operation time, model complexity 4 aspect, Forecasting Methodology analyzed respectively and contrast.All methods adopt same input parameter, training sample and test sample book.Note: because GA-ANN is comparatively large to training sample demand, then use whole 11 training samples of RVM as training data, only set up a model.

2 accuracy evaluation indexs

$\mathrm{RMSE}=\frac{\sqrt{{\mathrm{\&Sigma;}}_{i=1}^n{(P_{\mathrm{Mi}}-P_{\mathrm{Pi}})}^2}}{\mathrm{Cap}\&times;\sqrt{n}}$

$\mathrm{MAE}=\frac{{\mathrm{\&Sigma;}}_{i=1}^n|P_{\mathrm{Mi}}-P_{\mathrm{Pi}}|}{\mathrm{Cap}\&times;n}$

In formula:

RMSE is root-mean-square error;

MAE is mean absolute error;

P _mifor the real power in i moment;

P _pifor the predicted power in i moment;

Cap is wind energy turbine set start (or wind power generating set) total volume;

N is all number of samples.

3 test analysis

As shown in table 4 and Fig. 7, Fig. 8, RSP method, SVM method, GA-ANN method each moon root-mean-square error peace all absolute errors.The wind farm power prediction precision of RSP method is far above SVM method and GA-ANN method.Average of the whole year RMSE and MAE value comparatively SVM and GA-ANN reduce (12.565%-9.920%)/12.565%=21.05% and (9.088%-6.654%)/9.088%=26.778% respectively, embody the predictive ability that RSP method is stronger.

Table 4.RSP method, SVM method, GA-ANN method each moon predicated error

For May 24, July 29, September 25 and Dec 26, Fig. 9-Figure 12 depicts contingent fluctuation range under the predicted value of wind power, actual value and 90% confidence level, demonstrates anticipation trend and the uncertainty analysis ability of RSP method.This is one of important feature of the present invention, does not only give the predicted value determined in Output rusults, also provides the uncertainty analysis of this prediction simultaneously, the fluctuation higher limit namely predicted the outcome and lower limit.Can find out, the method has well doped the variation tendency of real power, and when power is undergone mutation, prediction curve also can well be followed the trail of.The bound plots changes of predicted power fluctuation is also consistent with real power curve, and is clipped in by real power curve between fluctuation bound; Through inspection: under 90% fiducial interval, the point of annual power actual value between predicted power fluctuation upper lower limit value has 89.928%.Prove the method can instruct wind energy turbine set operations staff and electric power system dispatching personnel more science carry out decision in the face of risk.

As can be seen from Table 5: the training time of RSP method and computing time are all much smaller than other two kinds of Forecasting Methodologies, and vectorial number involved in calculating is few, has fully demonstrated learning ability and the learning efficiency of RSP wind electric field power prediction method.In test, model running environment is: dominant frequency 2.79GHz internal memory 3.12GB.Prove that RSP method meets the requirement that the scheduling decision carrying the previous day requires and the ultra-short term operational plan of 10min-4h is formulated completely.

Table 5. model computing time and vectorial quantity

The above; be only the present invention's preferably embodiment, but protection scope of the present invention is not limited thereto, is anyly familiar with those skilled in the art in the technical scope that the present invention discloses; the change that can expect easily or replacement, all should be encompassed within 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 (2)

1., based on the high precision wind power interval prediction method of Method Using Relevance Vector Machine, it is characterized in that the method comprises the following steps:

Step 1: image data normalization, selects the training sample of Method Using Relevance Vector Machine forecast model according to the absolute error of wind speed in numerical weather forecast;

Step 2: be optimized the parameter of Method Using Relevance Vector Machine forecast model on the basis of step 1, obtains the iterative initial value of optimized kernel function width and Method Using Relevance Vector Machine forecast model;

Step 3: try to achieve kernel function on the basis of step 2, and then try to achieve the Method Using Relevance Vector Machine prediction model parameters after convergence;

Step 4: predicted value and the variance of trying to achieve wind power, obtain the forecast interval of wind power;

Described step 1 comprises:

A1. from wind farm data collection and supervisor control and anemometer tower, extract reality based on the data acquisition module of the interval Method Using Relevance Vector Machine forecast model of high precision wind power of Method Using Relevance Vector Machine and survey wind data, Power Output for Wind Power Field, numerical weather forecast data-hereinafter referred to as NWP, input parameter using NWP as model training power Output for Wind Power Field is as model training target parameter wherein x _ifor sample parameter, be respectively the wind speed in NWP, wind direction sine value, wind direction cosine value, temperature, pressure, humidity; N is sample size; Then normalized input parameter and corresponding target parameter, normalization formula is:

x＝[x _ori-min(x _ori)]/[[max(x _ori)-min(x _ori)]]

Wherein:

X _orifor original sample parameter;

X is the sample parameter after normalization;

A2. first whole sample was divided into groups according to month, then the absolute error of the precision of each moon NWP according to NWP wind speed is divided into groups; The absolute error e of NWP wind speed _nWPexpression formula is as follows:

e＝|x _NWP-x _act|

Wherein:

E is NWP wind speed absolute error;

X _nWPfor the wind speed numerical value in NWP result;

X _actfor the wind speed numerical value of actual measurement;

A3. find best NWP accuracy class and set up training sample; NWP grade of errors corresponding to different month Method Using Relevance Vector Machine precision of forecasting model mxm. is not identical, that is: be station work sample according to the absolute error of wind speed in numerical weather forecast; Each group is tested with Method Using Relevance Vector Machine forecast model respectively, chooses one group of training sample as Method Using Relevance Vector Machine forecast model that wherein precision of prediction is the highest; Through the screening to example wind energy turbine set training sample, for each month sets up the training sample meeting NWP precision characteristic distributions;

Described step 2 comprises:

B1. with the input amendment of the training sample of step 1 gained for Model Parameter Optimization step; Adopt particle cluster algorithm, first set the initial value of particle position and speed in each colony of particle cluster algorithm, in example wind energy turbine set, number of particles gets 20, particle initial position is between (0,30) random number between, speed initial value is the random number between (-2,2);

B2. calculate the adaptive value of each particle, fitness function f expression formula is:

$f=\frac{\sqrt{{\mathrm{\&Sigma;}}_{i=1}^n{(P_{\mathrm{Mi}}-P_{\mathrm{Pi}})}^2}}{\mathrm{Cap}\&times;\sqrt{n}}$

Wherein:

P _mifor the real power in i moment;

P _pifor the predicted power in i moment;

Cap is wind energy turbine set start total volume;

N is all number of samples;

B3. select particle position that in each colony, particle fitness function value is minimum respectively as the individual extreme value of each colony, select particle position minimum in all particles fitness function value as global extremum;

B4. upgrade particle rapidity and position, expression formula is respectively:

$v_{i,d}^{k+1}=\mathrm{\&omega;}\&CenterDot;v_{i,d}^k+c_1\&CenterDot;\mathrm{rand}\left(\right)(p{b}_{i,d}^k-y_{i,d}^k)+c_2\&CenterDot;\mathrm{rand}\left(\right)(g{b}_{i,d}^k-y_{i,d}^k)$

$y_{i,d}^{k+1}=y_{i,d}^k+v_{i,d}^{k+1}$

Wherein:

C ₁and c ₂for Studying factors;

Rand () is the random number between (0,1);

with be respectively speed and the position of d dimension in particle i kth time iteration;

for the position of the individual extreme value that particle i ties up at d;

for the position of the global extremum that colony ties up at d;

ω is the inertia weight factor;

B5. adaptive value is checked whether to meet iteration requirement, as met then finishing iteration, when the global extremum obtained after secondary iteration is optimum kernel function width parameter and Method Using Relevance Vector Machine forecast model iterative initial value, do not meet and then return B2 until restrain or reach maximum iteration time;

Described step 3 comprises:

C1: input kernel function width, the Method Using Relevance Vector Machine forecast model iterative initial value after the best training sample after screening and particle cluster algorithm optimization;

C2: the kernel function K (X calculating the interval Method Using Relevance Vector Machine forecast model of high precision wind power based on Method Using Relevance Vector Machine _i, X _j), expression formula is:

$K(X_i,X_j)=\exp (-\frac{{\left|\right|X_i-X_j\left|\right|}^2}{2{\mathrm{\&theta;}}^2})$

Wherein:

K (X _i, X _j) be kernel function;

X _i, X _jfor the mode input parameter after normalization;

θ is the kernel function width after particle cluster algorithm is optimized;

C3: the variance ∑ of calculation training sample Posterior distrbutionp, expression formula is:

∑＝(σ ^-2φ ^Tφ+A) ^-1

Wherein:

∑ is the variance of Posterior distrbutionp;

A=diag (α ₀, α ₁... α _n) be hyper parameter diagonal entry;

σ ^-2for iterative initial value;

φ is basis function vector, φ=(1, K (x, t)) ^t, x is input parameter, and t is target parameter;

C4: the weighted value w calculating Posterior distrbutionp _i, expression formula is:

w _i＝σ ^-2∑φ ^Tt

Wherein:

W _ifor Posterior distrbutionp weighted value;

T is the target component of training sample;

C5: utilize the maximal margin estimation technique to calculate about hyper parameter { α _iand σ ²maximized prior distribution, maximal margin estimation technique expression formula is:

L＝-0.5[log|β ^-1I+φA ^-1φ ^T|+t ^T(β ^-1I+φA ^-1φ ^T) ^-1t]

Wherein:

β is iterative initial value σ ^-2inverse;

I is unit matrix;

C6: if do not meet iterated conditional i.e. α between twice iteration _ithe variable quantity of numerical value is less than 10 ^-6then continue to upgrade to calculate { α _iand σ ²until convergence, parameter more new formula is:

${\mathrm{\&alpha;}}_i^{\mathrm{new}}=\frac{1-{\mathrm{\&alpha;}}_i{\mathrm{\&Sigma;}}_{\mathrm{ii}}}{w_i^2};{\left({\mathrm{\&sigma;}}^2\right)}^{\mathrm{new}}=\frac{{\left|\right|t-\mathrm{\&phi;}{w}_i\left|\right|}^2}{N-\mathrm{\&Sigma;}(1-{\mathrm{\&alpha;}}_i{\mathrm{\&Sigma;}}_{\mathrm{ii}})}$

Wherein:

∑ _iii-th diagonal entry of Posterior distrbutionp variance ∑;

N is sample size;

C7: delete α in iterative process _iw corresponding to → ∞ _i; The w retained _icorresponding vector is " associated vector " of Method Using Relevance Vector Machine forecast model, and after convergence, Method Using Relevance Vector Machine prediction model parameters is: α _mP, σ ² _mP;

Described step 4 comprises:

D1: substitute into kernel function width that test sample book, step 2 optimization obtain and Method Using Relevance Vector Machine forecast model iterative initial value, step 3 trains the Method Using Relevance Vector Machine prediction model parameters α obtained _mP, σ ² _mP, try to achieve the wind farm power prediction value between (0,1) and variance thereof, expression formula is as follows:

y _*＝w ^Tφ

σ ² _*＝σ ² _MP+φ ^T∑φ

Wherein:

Y _*for wind farm power prediction monodrome;

σ ² _*for wind farm power prediction is worth variance;

D2: renormalization process, expression formula is:

x _anti＝x _0-1×[max(x _ori)-min(x _ori)]+min(x _ori)

Wherein:

X _antifor renormalization result;

X _0-1for the parameter of [0,1];

X _orifor original parameter;

D3: given confidence level parameter δ, calculate the contingent fluctuation higher limit of wind farm power prediction monodrome and fluctuation lower limit, i.e. uncertainty analysis, expression formula is as follows:

P _upper＝P _prep×(1+σ _*z _δ/2)

P _lower＝P _prep×(1-σ _*z _δ/2)

Wherein:

P _upperfor the contingent fluctuation higher limit of wind farm power prediction value;

P _lowerfor the contingent fluctuation lower limit of wind farm power prediction value;

P _prepfor wind farm power prediction value;

σ _*for wind farm power prediction is worth mean square deviation;

Z _δ/2for the critical value that standardized normal distribution under confidence level 1-δ is corresponding;

Through wind power interval prediction, can not only obtain the predicted value of a real number, can also try to achieve the fluctuation range of the generation under certain confidence level, the decision in the face of risk that this probabilistic forecasting mechanism substantially increases model is worth.

2. the high precision wind power interval prediction method based on Method Using Relevance Vector Machine according to claim 1, is characterized in that described iteration requires to be less than 0.1 for adaptive value.