CN105654207A - Wind power prediction method based on wind speed information and wind direction information - Google Patents

Abstract

The invention relates to a wind power prediction method based on the wind speed information and the wind direction information. The method comprises the steps of collecting and arranging the historical wind data of a wind power plant and the actual wind power output data, conducting the normalization pretreatment on the data as a training sample, then conducting the similarity analysis and the clustering analysis, and dividing data objects similar in characteristic attribute as one type, respectively calculating the cluster center for each type of data based on the k-means clustering algorithm, predicting an active power average value on a prediction day based on an auto-regressive model structure, constructing a model to establish different types of wind power prediction models with the combination of the active power average value with the data of wind speed and wind direction of each type jointly as the input and the actually measured active power data on the prediction day as the output, judging the type of normalized daily vectors highest in similarity on the prediction day, and obtaining the predicted wind power on the prediction day based on a wind power prediction model of the same type. The above method is high in prediction accuracy and high in reliability.

Description

Wind power forecasting method based on wind speed, wind direction information

Technical field

The present invention relates to wind energy turbine set generated power forecasting technology, particularly a kind of wind power forecasting method based on wind speed, wind direction information.

Background technology

Wind energy is a kind of typical non-pollution renewable energy, due to its aboundresources, possess the condition of large-scale development, therefore suffer from paying close attention to widely, becoming the principal mode of future source of energy, the development and utilization of wind energy has become the of paramount importance Renewable Energy Development direction of China. Along with gradually stepping up of the accumulative installed capacity of wind-powered electricity generation, the randomness of its generated output and undulatory property are more and more obvious on the impact of grid-connected rear grid balance. The reliability of stable operation and electric power system in order to ensure electrical network, electric power system must be carried out effectively plan and scheduling, it is thus desirable to wind-power electricity generation power is carried out Accurate Prediction, this is directly connected to the equilibrium of supply and demand of electrical network, also directly affects the operation cost of grid-connected system. 2011, National Energy Board has issued " wind park power prediction forecast management Tentative Measures ", force wind energy turbine set that Wind power forecasting system is installed, and since 1 day January in 2012, do not report and submit the wind energy turbine set of wind-power electricity generation power forecast result also not allow to be incorporated into the power networks on request.

When formulating generation schedule and arranging scheduling, need to ensure that electricity generation system has stronger complex tracking ability, can adapt to minute level or the load fluctuation of hour level, thus it requires system has enough spinning reserve capacities, it is thus desirable to the generated output of wind-driven generator and wind energy turbine set to be carried out the prediction of ultra-short term (< 4h) and short-term (< 72h), and wind power prediction can reduce wind power cost under ensureing the steady of electric power system and the premise of safety accurately, it is possible to reach to improve the purpose of wind power value. The importance of Wind power forecasting as can be seen here.

Summary of the invention

The invention aims to the wind power forecasting method based on wind speed, wind direction information providing a kind of reliability high, utilize wind speed, wind direction information realization ultra-short term or short-term wind-electricity power prediction.

The technical scheme is that

A kind of wind power forecasting method based on wind speed, wind direction information, comprises the following steps:

Step one, carry out collecting to the history wind data of wind energy turbine set and corresponding actual wind power output data and arrange, wherein history wind data includes the wind speed of differing heights, the wind direction information that anemometer tower obtains, and actual wind power output data are the actual active power of wind energy turbine set real-time Transmission;

Step 2, the history wind data of described wind energy turbine set and corresponding actual wind power output data are normalized after pretreatment as training sample, then according to the normalization data sequence of the wind speed of different time points, wind direction and corresponding actual active power is carried out similarity analysis and cluster analysis by time sequencing, data object similar for characteristic attribute is divided into same class, adopt K means clustering algorithm after determining final number of categories K, calculate the cluster centre of each class data respectively;

The actual active power of history before step 3, extraction training Day, autoregression model structure prediction is adopted to go out the active power mean value on the training Day same day, it is combined collectively as input with the wind speed of each apoplexy due to endogenous wind, wind direction data, the actual measurement active power data on training Day same day are modeled as output, by the training to each class data sample, set up inhomogeneous wind power prediction model, namely set up K wind power prediction model;

The normalization data of the active power mean value of step 4, the extraction prediction wind speed on same day day, wind direction and prediction, the normalization day vector on same day predicted composition day, calculate this normalization day vectorial Similarity Parameter with each class cluster centre of K apoplexy due to endogenous wind respectively, judge a class the highest with the normalization day vector similarity of prediction day, utilize such wind power prediction model to obtain the pre-power scale of wind-powered electricity generation on prediction same day day.

The above-mentioned wind power forecasting method based on wind speed, wind direction information, the method carrying out similarity analysis in step 2 is the size adopting Pearson product-moment correlation coefficient to carry out quantitative description similarity, and the computing formula of described Pearson product-moment correlation coefficient is as follows:

$R=\frac{E\left(XY\right)-E\left(X\right)E\left(Y\right)}{\sqrt{E\left(X^2\right)-E^2\left(X\right)}\sqrt{E\left(Y^2\right)-E^2\left(Y\right)}}$

Wherein, E represents mathematic expectaion; X and Y represents the normalization day vector that wind speed and direction data on the same day are not constituted respectively; R represents correlation coefficient, and span is-1��1; | R | < 0.4 is low linear correlation; 0.4��| R | < 0.7 is notable linear correlation; 0.7��| R | < 1 is correlated with for highly linear.

The above-mentioned wind power forecasting method based on wind speed, wind direction information, concretely comprising the following steps of described K means clustering algorithm:

A, all data being divided into K initial classes, choosing K sample point is initial cluster center, is designated as z₁(l),z₂(l),��,z_k(l), wherein, initial value l=1;

B, according to Nearest Neighbor Method, all samples are assigned to the K class �� representated by each cluster centre_j(K), in, all kinds of comprised sample number is N_j(l), wherein, criterion function is Euclidean distance, and definition is:

$d={\left(\underset{k=1}{\overset{m}{\&Sigma;}}{\left(x\left(k\right)-y\left(k\right)\right)}^2\right)}^{\frac{1}{2}}$

In formula, d represents not Euclidean distance between normalization day vector X, Y on the same day; The dimension of m vector X and Y; The subvector of x and y vector X, Y;

C, calculate all kinds of mean vector, and using this vector as new cluster centre:In formula, j=1,2 ..., k; I=1,2 ..., N_j(l);

If d is z_j(l+1)��z_jL (), represents that cluster result is not best, then returns b, continue iterative computation;

If z_j(l+1)=z_jL (), iterative process terminates, and cluster result now is exactly optimum cluster result.

The above-mentioned wind power forecasting method based on wind speed, wind direction information, K=5.

The above-mentioned wind power forecasting method based on wind speed, wind direction information, the expression of the autoregression model structure in step 3 is:

Wherein, p_tRepresent the active power mean value of training Day; p_t-1��p_t-2Represent the active power mean value of two days before training Day respectively;For Parameters of Autoregressive Models, { ��_tFor White Noise Model, by adopting sample autocorrelation function to determine rank and method of least square carries out parameter estimation, obtain

The above-mentioned wind power forecasting method based on wind speed, wind direction information, the wind power prediction model in step 3 selects general regression neural network.

The above-mentioned wind power forecasting method based on wind speed, wind direction information, predicts in step 4 that the normalization day vectorial similarity measurement with each class cluster centre of K apoplexy due to endogenous wind on same day day adopts Euclidean distance to measure.

The invention has the beneficial effects as follows:

Owing to the selection of training sample directly influences the precision of forecast model, so the method adopting cluster analysis before modeling, wind speed with training Day differing heights, wind direction information is as clustering target, adopt Euclidean distance as the measurement criterion of similarity, first sample data is carried out cluster analysis, further according to the sample class the most close with training Day same day, in conjunction with history active power as the training sample modeled, shown by simulation result, after the method adopting cluster analysis carries out data prediction, improve the precision of prediction of forecast model greatly, reliability is high, can under the premise ensureing electric power system stable operation, save operating cost, be conducive to improving the economic benefit that wind energy turbine set is run.

Accompanying drawing explanation

Fig. 1 is the theory diagram of the wind power forecasting method of the present invention.

Detailed description of the invention

As it is shown in figure 1, this wind power forecasting method, comprise the following steps:

1, carry out the history wind data of wind energy turbine set and corresponding actual wind power output data collecting arranging, wherein wind data includes the wind speed of differing heights, the wind direction information that anemometer tower obtains, and actual wind power output data are the actual active power of wind energy turbine set real-time Transmission. In the present embodiment, collect and arrange 10 meters of height wind speed that in April, 2014 to May, anemometer tower obtained, 10 meters of height wind directions, 30 meters of height wind speed, 30 meters of height wind directions, 50 meters of height wind speed and 50 meters of height wind directions, and the actual active power of wind energy turbine set in this period, the resolution of data is 15min, and each day is 96 time points.

2, to 10 meters of height wind speed that in April, 2014 to May, anemometer tower obtained, 10 meters of height wind directions, 30 meters of height wind speed, 30 meters of height wind directions, 50 meters of height wind speed and 50 meters of height wind directions, and the pretreatment that in this period, the actual active power of wind energy turbine set is normalized is as training sample, then according to the normalization data sequence of the wind speed of different time points, wind direction and corresponding actual active power is carried out similarity analysis by time sequencing, wherein, selecting Pearson product-moment correlation coefficient to carry out the size of quantitative description similarity, computing formula is as follows:

$R=\frac{E\left(XY\right)-E\left(X\right)E\left(Y\right)}{\sqrt{E\left(X^2\right)-E^2\left(X\right)}\sqrt{E\left(Y^2\right)-E^2\left(Y\right)}}$

In formula, E mathematic expectaion; X, Y represents the normalization day vector that wind speed and direction data on the same day are not constituted, in the present embodiment, described day vector be made up of 10 meters of 96 time points height wind speed, 10 meters of height wind directions, 30 meters of height wind speed, 30 meters of height wind directions, 50 meters of height wind speed and 50 meters of height wind direction data; The value of correlation coefficient is-1��1, and its character is as follows: as R > 0, represents two variable positive correlations, as R < 0, represents that two variablees are negative correlation; When | R |=1, represent that two variablees are fairly linear relevant, be functional relationship; As R=0, represent between two variablees without linear relationship; As 0 < | R | < 1, representing that two variablees exist a degree of linear correlation, and | R | is closer to 1, two variable linearly relations are more close; | R |, closer to 0, represents that the linear correlation of two variablees is more weak; Generally can by three grades of divisions, | R | < 0.4 is low linear correlation; 0.4��| R | < 0.7 is notable linear correlation; 0.7��| R | < 1 is correlated with for highly linear.

After obtaining different wind speed, wind direction seasonal effect in time series correlation coefficient, historical time data is carried out cluster analysis, data object similar for characteristic attribute is divided into same class.In the present embodiment, select K means clustering algorithm the most classical in Dynamic Clustering Algorithm, it is determined that final number of categories K, and calculate the cluster centre of each class data respectively. In the present embodiment, K=5, calculate the cluster centre of 5 class data respectively. Concretely comprise the following steps:

E, all data being divided into K initial classes, choosing K sample point is initial cluster center, is designated as z₁(l),z₂(l),��,z_k(l), wherein, initial value l=1;

F, according to Nearest Neighbor Method, all samples are assigned to the K class �� representated by each cluster centre_j(K), in, all kinds of comprised sample number is N_j(l), wherein, criterion function is Euclidean distance, and definition is:

$d={\left(\underset{k=1}{\overset{m}{\&Sigma;}}{\left(x\left(k\right)-y\left(k\right)\right)}^2\right)}^{\frac{1}{2}}$

In formula, d represents not Euclidean distance between normalization day vector X, Y on the same day; The dimension of m vector X and Y; The subvector of x and y vector X, Y;

G, calculate all kinds of mean vector, and using this vector as new cluster centre:In formula, j=1,2 ..., k; I=1,2 ..., N_j(l);

If h is z_j(l+1)��z_jL (), represents that cluster result is not best, then returns b, continue iterative computation;

If i is z_j(l+1)=z_jL (), iterative process terminates, and cluster result now is exactly optimum cluster result.

3, extracting the actual active power of history before training Day, adopt autoregression model structure prediction to go out the active power mean value on the training Day same day, in the present embodiment, the expression of autoregression model structure is:

Wherein, p_tRepresent the active power mean value of training Day; p_t-1��p_t-2Represent the active power mean value of two days before training Day respectively;For Parameters of Autoregressive Models, { ��_tFor White Noise Model, by adopting sample autocorrelation function to determine rank and method of least square carries out parameter estimation, obtain

Then the active power mean value on training Day same day is combined collectively as input with the wind speed of each apoplexy due to endogenous wind, wind direction data, the actual measurement active power data on training Day same day are modeled as output, by the training to each class data sample, set up inhomogeneous wind power prediction model, namely set up K wind power prediction model. in the present embodiment, wind power prediction model selects general regression neural network (GRNN), this model is by input layer, hidden layer and output layer are constituted, hidden layer is radial direction basic unit, Gaussian transformation function is adopted to control hidden layer output, thus playing the activation suppressing output unit, in the input space, Gaussian function is symmetrical in acceptance region, the impact exported by network is exponentially decayed by input neuron with the distance between input vector, in GRNN, each trained vector has a corresponding radially base neuron in hidden layer, hidden layer neuron stores each trained vector, when network inputs a new vector, distance between new vector and each unit weight vector of hidden layer can be calculated by below equation:

Dist=| X-W¹|

In formula, X is input vector; W¹For hidden layer unit weight vector; Dist is the spacing of input vector and weight vector. The Gaussian function output of hidden layer calculates as follows:

$a^1=\exp (-{\left(\left|\right|dist\left|\right|\&times;b^1\right)}^2)$

$b^1=\frac{0.8326}{s}$

In formula, s is window width, if dist=s, then distance n after adjusting¹=| | dist | | �� b¹=0.8326, Gaussian function output valve is 0.5, and being equivalent to correlation coefficient is 0.5; If dist is much larger than s, Gaussian function exports close to 0. along with a¹Increase, hidden layer output be gradually reduced, variable s plays a part window, say, that s size plays output layer neuronal activation effect, and s is more big, b¹More little, hidden layer neuron and input vector are apart from reduced, and therefore, the neuron number being activated in window increases, otherwise, s is more little, and b is more big, hidden layer neuron is exaggerated with input vector distance, and hidden layer output reduces, and the neuron number being activated in window reduces.

GRNN output layer is linear layer, and output calculates as follows:

a²=n²=W²a¹+b²

In formula, W²It is the 2nd layer of weights.

Therefore, the feature of general regression neural network (GRNN) is that the artificial parameter regulated is few, only one of which threshold value s. The study of network all relies on data sample. This feature determines GRNN to avoid artificial subjective hypothesis on the impact predicted the outcome to greatest extent. To sum up, for inhomogeneous training sample, in the present embodiment, the input number of nodes of GRNN is 7, and output node number is 1, and the window width parameter s value of network is 0.21.

4, prediction day is extracted (in the present embodiment, prediction is set as-10 days on the 1st June in 2014 day) wind speed on the same day, wind direction and prediction the normalization data of active power mean value, the normalization day vector on same day predicted composition day, calculate this normalization day vectorial Similarity Parameter with the 5 each class cluster centres of apoplexy due to endogenous wind respectively, judge a class the highest with the normalization day vector similarity of prediction day, and utilize such wind power prediction model to obtain the pre-power scale of wind-powered electricity generation on prediction same day day. In the present embodiment, the normalization day vector on prediction same day day adopts Euclidean distance algorithm with the calculating of the Similarity Parameter of each class cluster centre of K apoplexy due to endogenous wind, prediction same day day day vector to the Euclidean distance respectively 0.32,0.43,0.83,0.12,0.48 of each class, nearest with the cluster centre of prediction day vector distance the 3rd class, namely the highest with the 3rd class similarity, therefore selecting the forecast model being training sample with the 3rd class, can obtain its forecast error NMAE (standard absolute value mean error) is 10.47%. Obtain owing to the parameter of forecast model is both for similar sample training, therefore there is higher precision of prediction.

Claims (7)

1., based on a wind power forecasting method for wind speed, wind direction information, comprise the following steps:

Step one, carry out collecting to the history wind data of wind energy turbine set and corresponding actual wind power output data and arrange, wherein history wind data includes the wind speed of differing heights, the wind direction information that anemometer tower obtains, and actual wind power output data are the actual active power of wind energy turbine set real-time Transmission;

Step 2, the history wind data of described wind energy turbine set and corresponding actual wind power output data are normalized after pretreatment as training sample, then according to the normalization data sequence of the wind speed of different time points, wind direction and corresponding actual active power is carried out similarity analysis and cluster analysis by time sequencing, data object similar for characteristic attribute is divided into same class, adopt K means clustering algorithm after determining final number of categories K, calculate the cluster centre of each class data respectively;

The actual active power of history before step 3, extraction training Day, autoregression model structure prediction is adopted to go out the active power mean value on the training Day same day, it is combined collectively as input with the wind speed of each apoplexy due to endogenous wind, wind direction data, the actual measurement active power data on training Day same day are modeled as output, by the training to each class data sample, set up inhomogeneous wind power prediction model, namely set up K wind power prediction model;

The normalization data of the active power mean value of step 4, the extraction prediction wind speed on same day day, wind direction and prediction, the normalization day vector on same day predicted composition day, calculate this normalization day vectorial Similarity Parameter with each class cluster centre of K apoplexy due to endogenous wind respectively, judge a class the highest with the normalization day vector similarity of prediction day, and utilize such wind power prediction model to obtain the pre-power scale of wind-powered electricity generation on prediction same day day.

2. the wind power forecasting method based on wind speed, wind direction information according to claim 1, it is characterized in that: the method carrying out similarity analysis in step 2 is the size adopting Pearson product-moment correlation coefficient to carry out quantitative description similarity, and the computing formula of described Pearson product-moment correlation coefficient is as follows:

$R=\frac{E\left(XY\right)-E\left(X\right)E\left(Y\right)}{\sqrt{E\left(X^2\right)-E^2\left(X\right)}\sqrt{E\left(Y^2\right)-E^2\left(Y\right)}}$

Wherein, E represents mathematic expectaion;X and Y represents the normalization day vector that wind speed and direction data on the same day are not constituted respectively; R represents correlation coefficient, and span is-1��1; | R | < 0.4 is low linear correlation; 0.4��| R | < 0.7 is notable linear correlation; 0.7��| R | < 1 is correlated with for highly linear.

3. the wind power forecasting method based on wind speed, wind direction information according to claim 1, it is characterised in that: concretely comprising the following steps of described K means clustering algorithm:

A, all data being divided into K initial classes, choosing K sample point is initial cluster center, is designated as z₁(l),z₂(l),��,z_k(l), wherein, initial value l=1;

B, according to Nearest Neighbor Method, all samples are assigned to the K class �� representated by each cluster centre_j(K), in, all kinds of comprised sample number is N_j(l), wherein, criterion function is Euclidean distance, and definition is:

$d={\left(\underset{k=1}{\overset{m}{\&Sigma;}}{\left(x\left(k\right)-y\left(k\right)\right)}^2\right)}^{\frac{1}{2}}$

In formula, d represents not Euclidean distance between normalization day vector X, Y on the same day; The dimension of m vector X and Y; The subvector of x and y vector X, Y;

C, calculate all kinds of mean vector, and using this vector as new cluster centre:In formula, j=1,2 ..., k; I=1,2 ..., N_j(l);

If d is z_j(l+1)��z_jL (), represents that cluster result is not best, then returns b, continue iterative computation;

If z_j(l+1)=z_jL (), iterative process terminates, and cluster result now is exactly optimum cluster result.

4. the wind power forecasting method based on wind speed, wind direction information according to claim 1 or 3, it is characterised in that: K=5.

5. the wind power forecasting method based on wind speed, wind direction information according to claim 1, it is characterised in that: the expression of the autoregression model structure in step 3 is:

Wherein, p_tRepresent the active power mean value of training Day; p_t-1��p_t-2Represent the active power mean value of two days before training Day respectively;For Parameters of Autoregressive Models, { ��_tFor White Noise Model, by adopting sample autocorrelation function to determine rank and method of least square carries out parameter estimation, obtain

6. the wind power forecasting method based on wind speed, wind direction information according to claim 1, it is characterised in that: the wind power prediction model in step 3 selects general regression neural network.

7. the wind power forecasting method based on wind speed, wind direction information according to claim 1, it is characterised in that: step 4 being predicted, the normalization day vector on same day day adopts Euclidean distance algorithm with the calculating of the Similarity Parameter of each class cluster centre of K apoplexy due to endogenous wind.