CN105069521A - Photovoltaic power plant output power prediction method based on weighted FCM clustering algorithm - Google Patents

Abstract

The invention relates to a photovoltaic power plant output power prediction method based on a weighted FCM clustering algorithm. The method provided by the invention comprises the steps that a weather data sample which matches a meteorological data sample to be predicted and the corresponding photovoltaic power plant output power are selected from the existing photovoltaic power plant operation database and are used as a reference sample; through knowledge evaluation, a typical data matrix is selected and is combined with the meteorological data sample to be predicted; after normalization, a final standard sample matrix is formed and is used as an input variable of the algorithm; and after property-weighted FCM clustering algorithm iteration, output power corresponding to the meteorological data sample to be predicted is acquired. According to the invention, the shortcomings of complex meteorological factors, unbalanced influence on the output power, meteorological data randomness and uncertainty and the like are overcome; the method has the advantages of fast prediction and high accuracy; a prediction result provides a data support for rational resource dispatching and scientific overall planning of the power industry; and good economic and social benefits are acquired.

Description

A kind of photovoltaic plant output power predicting method based on weighted FCM clustering algorithm

Technical field

The present invention relates to a kind of photovoltaic plant output power predicting method based on weighted FCM clustering algorithm, belong to electric power project engineering field.

Background technology

The primary energy such as coal, oil, rock gas are petered out, environmental aspect goes from bad to worse, the development and utilization of regenerative resource becomes the emphasis of every country concern gradually, sun power is inexhaustible, nexhaustible, and clean and safe, conversion convenience, photovoltaic power generation technology has worldwide been furtherd investigate and widespread use, in the energy structure in future, is that the renewable energy power generation of representative will occupy important component with photovoltaic generation.The multiple preferential policy of country and standard criterion support with instruct under, a large amount of photovoltaic plant of China has built up and has dropped into utilization, and meanwhile, the photovoltaic plant is more and more in design and construction.The meteorological condition of photovoltaic plant output power and location, site is closely related, but meteorologic factor Existential Space and temporal ambiguity, randomness and instability, thus the output power of photovoltaic plant also has uncertainty, undulatory property and intermittent feature, the be incorporated into the power networks stability and security that can have a strong impact on electric system of large-scale photovoltaic power station.

Photovoltaic plant output power is carried out to the short-term Accurate Prediction in certain hour section, be conducive to power scheduling department according to regional workload demand rational operation plan on the one hand, the proportioning of the generating of overall arrangement conventional energy resources and photovoltaic generation, reduces the adverse effect that photovoltaic generating system produces electrical network to the full extent.On the other hand, the margin capacity of electric system can be reduced, make full use of solar energy resources, improve economic benefit and the social benefit of photovoltaic plant.And the forecasting research of photovoltaic generation at present gets more and more, such as application number is the patent document of 201310301150.1, this patent document discloses a kind of method for predicting output power of power generation in photovoltaic power station based on similar day, the method adopts Fuzzy Cluster Analysis method to carry out cluster analysis, best cluster number is determined by Clustering Effect evaluation index, utilize BP neural network, set up generated energy forecast model, dope daily generation according to this forecast model.But the method only collecting temperature, humidity and day type as the proper vector of forecast model, consider not comprehensive to the influence factor of photovoltaic plant output power, in fact the total spoke amount of the sun, horizontal plane radiation amount, dip plane radiant quantity, atmospheric pressure and wind speed etc. all can affect the output power of photovoltaic plant to a certain extent; And due to temperature, humidity not identical to the influence degree of photovoltaic plant output power with day type, and do not have in the whole forecasting process of this file consider and distinguish temperature, humidity and day type to the Different Effects degree of output power, cause this Forecasting Methodology not accurate enough.Meanwhile, above-mentioned a kind of method for predicting output power of power generation in photovoltaic power station based on similar day, needs to carry out a large amount of learning trainings by BP neural network to similar day sample set, if training degree is inadequate, is also difficult to the accuracy of guaranteed output forecast model.

Summary of the invention

The object of this invention is to provide a kind of photovoltaic plant output power predicting method based on weighted FCM clustering algorithm, with solve existing photovoltaic plant output power rain inaccurate, caused photovoltaic generation and conventional energy resources generate electricity and coordinate difficult problem not in time.

The present invention solves the problems of the technologies described above to provide a kind of photovoltaic plant output power predicting method based on weighted FCM clustering algorithm, and this Forecasting Methodology comprises the following steps:

1) determine the following precise time section needing to carry out power prediction, gather in this time period the weather data of 8 dimension attributes comprising the sun total spoke amount, horizontal plane radiation amount, dip plane radiant quantity, environment temperature, air humidity, atmospheric pressure, wind speed and wind speed in photovoltaic plant region as weather data sample to be measured;

2) from the knowledge base of this photovoltaic plant history run, choose the reference sample similar to weather data sample to be measured, in each reference sample, include 8 Weather property identical with weather data sample to be measured and the photovoltaic plant output power of corresponding meteorological sample;

3) from reference sample, choose the master sample data set with weather data sample matches to be measured, combine with weather data sample to be measured, form typical sample matrix, and formation master sample matrix is normalized to this typical sample matrix;

4) using the input quantity of master sample matrix as attribute weight FCM clustering algorithm, initializing variable according to setting carries out algorithm iteration to obtain corresponding subordinated-degree matrix, judge according to the subordinated-degree matrix obtained the photovoltaic plant output power that meteorological sample to be measured is corresponding, this output power is the predicted power of meteorological sample to be measured.

Described attribute weight FCM clustering algorithm refers to and utilizes Sample Similarity to give corresponding weights to the percentage contribution of attribute each in meteorological sample to cluster result.

The flow process of described attribute weight FCM clustering algorithm is as follows:

A. each attribute weights in master sample matrix are calculated;

B. according to each attribute weights determination cluster objective function obtained and renewal cluster centre;

C. subordinated-degree matrix is upgraded, until subordinated-degree matrix meets setting requirement after upgrading.

The calculating of each attribute weights of described sample adopts the attribute weight algorithm based on Sample Similarity.

Described step 4) in initializing variable comprise cluster classification, this cluster classification be by the power stage scope of transformer station to be predicted according to prediction class of accuracy refinement segmentation, each section is a cluster classification.

Described step 3) method for normalizing can adoption rate standardization, maximum value standardization, mean variance standardization, any one in absolute value variance criterion and regular standardization.

The method also comprises returns in knowledge base by the power prediction result feedback of meteorological sample to be measured and its correspondence, is convenient to accumulate true and reliable historical data for later weather data sample carries out power prediction.

The invention has the beneficial effects as follows: first the present invention chooses the synoptic data sample that matches with weather data sample to be measured and the photovoltaic plant output power corresponding with it as with reference to sample from existing photovoltaic plant runtime database, pass through knowledge evaluation, filter out typical data matrix, then combine with weather data sample to be measured, through normalized, form final master sample matrix, as the input variable of algorithm, then by after the FCM algorithm iteration of attribute weight, this output power corresponding to weather data sample to be measured is obtained.Instant invention overcomes meteorologic factor complexity, unbalanced to the influence degree of output power, and the randomness that has of weather data and the shortcomings such as uncertain, predetermined speed is fast, accuracy is high, predict the outcome as power industry rational management resource, science overall planning provide data supporting, good economic and social benefit can be obtained simultaneously.

Accompanying drawing explanation

Fig. 1 is the process flow diagram of the photovoltaic plant output power predicting method that the present invention is based on weighted FCM clustering algorithm;

Fig. 2 is the process flow diagram of the weighted FCM clustering algorithm adopted in the embodiment of the present invention;

Fig. 3 is sample attribute weighting system process flow diagram;

Fig. 4 is sample data screening system process flow diagram;

Fig. 5 is the system diagram of the knowledge base adopted in the embodiment of the present invention.

Embodiment

Below in conjunction with accompanying drawing, the specific embodiment of the present invention is described further.

Photovoltaic plant output power directly affects by meteorologic factor, therebetween there is nonlinear correspondence relation, the present invention is based on the service data historical knowledge base of synoptic data data and correspondence thereof, first according to the data of weather forecast in the time period to be predicted, under multiple similar weather condition, the weather data of accomplished fact and the photovoltaic plant output power data sample of correspondence thereof is gathered from knowledge base, form reference sample matrix, after data screening system is preferred, forms typical sample matrix; In order to distinguish in sample, each meteorological attribute is on the impact of the final output power of photovoltaic plant and percentage contribution difference, and what utilize Sample Similarity carries out attribute weight, obtains weighted FCM clustering algorithm.Finally, by weather data sample to be predicted and the combination of typical meteorological sample, and by after normalized, form Standard Gases and resemble sample matrix, as the input matrix of weighted FCM clustering algorithm, by continuous iteration, the photovoltaic plant output power value that the meteorological sample data of the subordinated-degree matrix automatic decision exported by algorithm is corresponding, and being shown in man-machine interface, the flow process of the method is as shown in Figure 1.

1. determine the following precise time section needing to carry out predicted power, gather in photovoltaic plant region corresponding to this time period the weather data of multidimensional property comprising solar irradiance, environment temperature, atmospheric pressure, wind speed and direction.

2. according to the weather data collected, from knowledge base system, obtain this photovoltaic plant and be in the historical empirical data that other identical photovoltaic plants of installed capacity in the same area, operation mode run, weather sample data under the similar meteorological condition of collection some and the station output data of correspondence thereof, and they are merged, as multidimensional reference sample matrix.

Due to gather sample data some may there is the shortcomings such as attribute is imperfect, data are fuzzy, data precision is inadequate, data screening system should be passed through, therefrom optimize the master sample data set with weather data sample matches to be measured, and combine with weather data sample to be measured, form typical sample matrix.

Data in typical sample matrix are measured value, the dimension that each dimension attribute is corresponding and order of magnitude difference larger, ultimate attainment point is converged to faster in order to make weighted FCM clustering algorithm, before cluster, it may be necessary suitable standardized method sample data is normalized, each value is made to be positioned at [0,1] in scope, thus form Standard Gases and resemble sample matrix, method for normalizing can adoption rate standardization, maximum value standardization, mean variance standardization, any one in absolute value variance criterion and regular standardization.

3., using the input quantity of master sample matrix as weighted FCM clustering algorithm, the initializing variable according to setting carries out algorithm iteration to obtain corresponding subordinated-degree matrix, judges according to the subordinated-degree matrix obtained the photovoltaic plant output power that meteorological sample to be measured is corresponding.

Cluster is exactly carry out according to the feature of things self and the similarity between them process that divides and sort out, because clustering algorithm can extract the complex data information of multidimensional automatically, heuristic data structure and inherent laws, the application thus obtained in a lot of fields.Traditional cluster analysis all belongs to hard clustering, the principle that it is followed completely " either-or ", and have strict membership between object to be sorted and set, namely degree of membership only has " 1 " and " 0 " two kinds of forms.But there is blooming and the concept of a large amount of " non-this non-that " in real world, clear and definite membership is not had between object and set, traditional cluster algorithm cannot characterize and process these fuzzy messages, and therefore the generation of fuzzy cluster analysis is significant.

Fuzzy cluster analysis has the characteristic of softening point, and degree of membership is expanded to interval [0,1] by it, strict membership is not had between object and set, two is the sizes according to degree of membership, and judgement sample belongs to the degree of each classification, and the degree of membership sum that certain object is under the jurisdiction of all set is 1.Fuzzy cluster analysis well characterizes the fuzzy corresponding relation between object and set as can be seen here, the FCM clustering algorithm of based target function belongs to the category of fuzzy cluster analysis, and this theory of algorithm basis is good, is easily understood, convenient realization, obtains good application in practice.But the calculating of this algorithm objective function is based on traditional Euclidean distance, think that the effect of each characteristic attribute to cluster result of sample is balanced, they are not distinguished, in order to obtain better cluster result, its significance level should be expressed by giving certain weights to each attribute.

The meteorological condition in the region at photovoltaic plant place and the output power in power station closely related, but these weather datas have very strong randomness and instability along with Changes in weather, cause the output power of photovoltaic plant also to have uncertainty and intermittence.Because photovoltaic power station monitoring system platform is known, meteorological sample comprises the total spoke amount of the sun, horizontal plane radiation amount, dip plane radiant quantity, environment temperature, air humidity, atmospheric pressure and wind speed are at interior multidimensional property, and these property values are all in the process of real-time change, sometime in section, the output power that different meteorological sample possibilities is corresponding same, certain output power value also may cause by multiple meteorological sample, thus between meteorological sample data and output power value, there is fuzzy corresponding relation, FCM clustering algorithm can be adopted predict the output power of photovoltaic plant in certain short-term period following.FCM algorithm is according to the degree of membership relation between multiple history output power samples of accomplished fact in the automatic weather data sample more to be measured of the fuzzy membership matrix exported and knowledge base, and the output power that degree of membership is the highest is output power corresponding to meteorological sample to be measured.

The ultimate principle of FCM algorithm is:

Be provided with one group of sample set X=(x ₁, x ₂..., x _n), comprise n sample, each sample has s characteristic attribute, i.e. x _k=(x _k1, x _k2..., x _ks), wherein k=1,2 ..., n, j=1,2 ..., s.N sample is divided into c classification without supervision by FCM algorithm exactly.

The objective function of FCM cluster is:

$J_m(U,P)=\underset{k=1}{\overset{n}{\Σ}}\underset{i=1}{\overset{c}{\Σ}}{\left({\mathrm{\μ}}_{ik}\right)}^m{\left(d_{ik}\right)}^2---\left(1\right)$

The constraint condition of algorithm is:

$\underset{i=1}{\overset{c}{\Σ}}{\mathrm{\μ}}_{ik}=1---\left(2\right)$

In formula (1), P=(p ₁, p ₂..., p _c) be cluster centre matrix; U=[μ _ik] _{c × n}for subordinated-degree matrix, element μ _ik∈ [0,1] represents sample x _kbe under the jurisdiction of the degree of i-th cluster centre; M ∈ [1, ∞) be smoothing parameter, generally get m=2; d _ikfor sample x _kto cluster centre p _ibetween Euclidean distance.Under the condition of formula (2), algorithm upgrades cluster target function value by continuous iteration, and cluster centre matrix and subordinated-degree matrix.When objective function obtains minimum value, or the degree of membership that when the subordinated-degree matrix norm that produces of twice iteration is less than given threshold values, algorithm exports and cluster centre are respectively:

${\mathrm{\μ}}_{ik}=\frac{1}{\underset{h=1}{\overset{c}{\Σ}}{\left(\frac{d_{ik}}{d_{hk}}\right)}^{\frac{2}{m-1}}}---\left(3\right)$

$p_i=\frac{\underset{k=1}{\overset{n}{\Σ}}{\left({\mathrm{\μ}}_{ik}\right)}^{m}x}{\underset{k=1}{\overset{n}{\Σ}}{\left({\mathrm{\μ}}_{ik}\right)}^m}---\left(4\right)$

In formula (3), d _hk(h=1,2 ..., c) and d _ik(i=1,2 ..., c), represent sample x _kdivide and be clipped to cluster centre p _hand p _ibetween Euclidean distance, d _ikwith d _hkratio less, sample x is described _kdistance cluster centre p _inearer, i.e. x _kthe degree being under the jurisdiction of i-th kind of cluster classification is higher.

Photovoltaic plant output power is by the impact of multiple meteorologic factor, the total spoke amount of the sun, horizontal plane radiation amount, dip plane radiant quantity, environment temperature, air humidity, atmospheric pressure, wind speed, wind direction etc. all can affect the power producing characteristics in power station to a certain extent, but below often kind of weather conditions are different to the influence degree of the final output power in power station, compare the main affecting factors such as the total spoke amount of the sun, horizontal plane radiation amount, dip plane radiant quantity, environment temperature, the influence degree of the factors such as air humidity, atmospheric pressure, wind speed, wind direction is just smaller.

FCM algorithm is based on traditional Euclidean distance, and Euclidean distance represents the distance in hyperspace between 2, when the attribute dimension of sample or classification results more time, algorithm well can not distinguish the significance level of every attribute to cluster result.Therefore in order to improve the accuracy to the prediction of photovoltaic plant output power, needing to improve former FCM algorithm, by giving certain weights to each attribute, carrying out the percentage contribution of outstanding different attribute to cluster result.Sample attribute weight w _jcalculating adopt based on the attribute weight algorithm of Sample Similarity, specific algorithm is as follows:

Objective definition function is:

$E\left(w\right)=\frac{2}{n(n-1)}\underset{p<q}{\&Sigma;}\frac{1}{2}\&lsqb;{{\mathrm{\&rho;}}_{pq}}^{\left(wH\right)}(1-{{\mathrm{\&rho;}}_{pq}}^{\left(1H\right)})+{{\mathrm{\&rho;}}_{pq}}^{\left(1H\right)}(1-{{\mathrm{\&rho;}}_{pq}}^{\left(wH\right)})\&rsqb;$

In formula, ρ _pq ^(wH)for Sample Similarity index; ρ _pq ^(1H)for each dimension attribute weight w of sample _jthe index of similarity of sample when being 1; P, q ∈ [1, n] is the sequence number of two different samples.

Determine w _jprocess as shown in Figure 3, be specially:

Step 1: order for sample attribute weight w _jweighted distance when being 1, determines parameter beta according to formula (5) and (6);

$\frac{2}{n(n-1)}\underset{q<p}{\&Sigma;}\frac{1}{1+{{\mathrm{\&beta;d}}_{pq}}^{\left(1H\right)}}=0.5---\left(5\right)$

${d_{pq}}^{\left(wH\right)}=\underset{j=1}{\overset{s}{\&Sigma;}}{w}_j\&lsqb;K(x_{pj},x_{pj})-2K(x_{pj},x_{qj})+K(x_{qj},x_{qj})\&rsqb;---\left(6\right)$

In formula, K (x, y) is gaussian kernel function, makes δ be Gaussian function width, then $K(x,y)=\exp (-\frac{\left|\right|x-y|{|}^2}{2{\mathrm{\&delta;}}^2}).$

Step 2: reduce direction minimization E (w) by gradient, solve according to formula (6), (7) and (8)

$\left\{\begin{array}{c}\frac{\&part;E\left(w\right)}{\&part;w_j}=\frac{1}{n(n-1)}\underset{q<p}{\mathrm{\&Sigma;}}\&lsqb;1-2{{\mathrm{\&rho;}}_{pq}}^{\left(1H\right)}\&rsqb;\frac{\&part;{{\mathrm{\&rho;}}_{pq}}^{\left(wH\right)}}{\&part;{d_{pq}}^{\left(wH\right)}}\frac{\&part;{d_{pq}}^{\left(wH\right)}}{\&part;w_j}\\ \frac{\&part;{{\mathrm{\&rho;}}_{pq}}^{\left(wH\right)}}{\&part;{d_{pq}}^{\left(wH\right)}}=\frac{-\mathrm{\&beta;}}{{\&lsqb;1+{{\mathrm{\&beta;d}}_{pq}}^{\left(wH\right)}\&rsqb;}^2}\\ \frac{\&part;{d_{pq}}^{\left(wH\right)}}{\&part;w_j}=\frac{w_j{(x_{pj}-x_{qj})}^2}{{d_{pq}}^{\left(wH\right)}}\end{array}\right.---\left(7\right)$

${{\mathrm{\&rho;}}_{pq}}^{\left(wH\right)}=\frac{1}{1+{{\mathrm{\&beta;d}}_{pq}}^{\left(wH\right)}}---\left(8\right)$

Step 3: make λ be nonnegative variable, ask parameter η according to formula (9), that is:

$E\&lsqb;w_1-\mathrm{\&eta;}\frac{\&part;E\left(w\right)}{\&part;w_j},...,w_n-\mathrm{\&eta;}\frac{\&part;E\left(w\right)}{\&part;w_n}\&rsqb;=\underset{\mathrm{\&lambda;}>0}{\min }\&lsqb;E(w_1-\mathrm{\&eta;}\frac{\&part;E\left(w\right)}{\&part;w_j},...,w_n-\mathrm{\&eta;}\frac{\&part;E\left(w\right)}{\&part;w_n})\&rsqb;---\left(9\right)$

Step 4: according to formula (10) by w _j+ Δ w _jupdate attribute weight w _j, that is:

${\mathrm{\&Delta;w}}_j=-\mathrm{\&eta;}\frac{\&part;E\left(x\right)}{\&part;w_j}---\left(10\right)$

Determine w _jafter, the cluster objective function of former FCM algorithm just becomes:

$J_m(U,P,w)=\underset{j=1}{\overset{s}{\&Sigma;}}\underset{k=1}{\overset{n}{\&Sigma;}}\underset{i=1}{\overset{c}{\&Sigma;}}{w}_j{\left({\mathrm{\&mu;}}_{ik}\right)}^m{\left(d_{ik}\right)}^2---\left(11\right)$

Under new cluster objective function, title new algorithm is the FCM clustering algorithm of attribute weight, and its flow process as shown in Figure 2.

Be the photovoltaic plant of 1MW below with installed capacity be example, application the present invention predicts following certain short-term period [t ₁, t ₂] in the output power of this photovoltaic plant, concrete implementation step is as follows.

1. gather the data of weather forecast in this time period local, comprise the total spoke amount of the sun (w/m ²), horizontal plane radiation amount (w/m ²), dip plane radiant quantity (w/m ²), environment temperature (DEG C), air humidity (%), atmospheric pressure (Pa), wind speed (m/s) and wind direction (°) 8 meteorological attributes, get the mean value of each attribute in this time period as weather data sample x to be measured _d=(x _d1, x _d2..., x _d8), that is, wherein x _d1, x _d2, x _d3, x _d4, x _d5, x _d6, x _d7and x _d8represent the total spoke amount of the sun, horizontal plane radiation amount, dip plane radiant quantity, environment temperature, air humidity, atmospheric pressure, wind speed and wind speed respectively.

2. from the historical knowledge base that this photovoltaic plant runs, gather the similar reference sample of weather data sample N number of and to be measured, each sample packages is containing two parts, a part is 8 the weather data samples corresponding with weather data sample to be measured, another part is the photovoltaic plant output power sample corresponding with meteorological sample, if N number of reference gas image data sample X _qfor:

$X_Q=(x_1,x_2,...,x_N)=\left(\begin{array}{cccc}{x}_{11}& x_{12},& ...,& x_{18}\\ x_{21}& x_{22},& ...,& x_{28}\\ ...& ...,& ...,& ...\\ x_{N1}& x_{N2},& ...,& x_{N8}\end{array}\right)$

With N number of reference gas image data sample X _qcorresponding output power sample is:

A _G＝(a ₁,a ₂,...,a _N) ^T

By X _qand A _gmerging can obtain complete reference sample matrix X _kfor:

$X_K=\left(\begin{array}{cccc}{x}_{11}& x_{12},& ...,& x_{18},a_1\\ x_{21}& x_{22},& ...,& x_{28},a_2\\ ...& ...,& ...,& ...\\ x_{N1}& x_{N2},& ...,& x_{N8},a_N\end{array}\right)$

3. pair obtain reference sample matrix X _kcarry out screening coupling

Due to X _kin N number of reference gas image data sample some may with weather data sample attribute to be measured type do not mate, the precision of some property value is inadequate, needs through further screening matching treatment, and flow process of its process is as shown in Figure 4.When the precision of attribute type and property value meets the demands, weather data sample is reduced to n by N number of, and corresponding output power sample is also reduced to n, and reference sample is by X _kbecome X _l, X _lfor typical sample matrix.Namely have

$X_L=\left(\begin{array}{cccc}{x}_{11}& x_{12},& ...,& x_{18},a_1\\ x_{21}& x_{22},& ...,& x_{28},a_2\\ ...& ...,& ...,& ...\\ x_{n1}& x_{n2},& ...,& x_{n8},a_n\end{array}\right)$

Now X _lin n weather data sample and n output power sample have following corresponding relation:

x ₁→a ₁,x ₂→a ₂,...,x _n→a _n

Weather data sample x to be measured _dbefore carrying out output power prediction, need by x _din each attribute data as last 1 row arrangement be merged into typical sample matrix X _lweather data matrix in, formed Standard Gases resemble sample matrix, have:

$X_H=\left(\begin{array}{cccc}{x}_{11}& x_{12},& ...,& x_{18}\\ x_{21}& x_{22},& ...,& x_{28}\\ ...& ...,& ...,& ...\\ x_{n1}& x_{n2},& ...,& x_{n8}\\ x_{D1}& x_{D2},& ...,& x_{D8}\end{array}\right)$

Standard Gases resembles sample matrix X _hin numerical value be measured data, the dimension that each dimension attribute is corresponding and order of magnitude difference are comparatively large, in order to make algorithm converge to extreme point faster, need X _hin each numerical value be normalized, select ratio standard method for normalizing herein, by the measured value of certain attribute divided by the summation of sample this attribute observed readings all, formula is:

${x_{kj}}^{,}=\frac{x_{kj}}{\underset{j=1}{\overset{s}{\&Sigma;}}{x}_{kj}}---\left(12\right)$

In formula, 1≤k≤n, 1≤j≤s, n is number of samples, and s is sample attribute dimension.

Master sample matrix X _hafter normalization, each numerical value is defined in scope [0,1], if the normalization matrix obtained is X _h', namely can be used as the input matrix of FMC algorithm.

4. using the normalization matrix that the obtains input matrix as weighted FCM clustering algorithm algorithm, initializing variable according to setting carries out algorithm iteration to obtain corresponding subordinated-degree matrix, judges according to the subordinated-degree matrix obtained the photovoltaic plant output power that meteorological sample to be measured is corresponding.Concrete iterative step is as follows:

Step 1: initialization, determines cluster classification number c (2≤c≤n), setting iteration stopping valve ε and iteration count b=0, initialization subordinated-degree matrix U ⁽⁰⁾.

FCM algorithm initialization needs to determine cluster classification number, and the classification of output power, due to output power sample a ₁, a ₂..., a _nbe certain measured data, output power classification should be translated into.The generating efficiency of 75% is approximately, if in the present embodiment, installed capacity is the photovoltaic plant output power range of 1MW is [700,800] (kW) according to current photovoltaic plant.Precision of prediction as required, setting every 20kW is a step-length, is further refined as by output power range within the scope of 5 performance numbers or power class, i.e. c=5.

The power span of the 1st classification is: [700,720];

The power span of the 2nd classification is: (720,740];

The power span of the 3rd classification is: (740,760];

The power span of the 4th classification is: (760,780];

The power span of the 5th classification is: (780,800].

Suppose a ₁, a ₂..., a _neach value is all within the scope of one of them in above 5 classifications.

Step 2: determine attribute weight w according to formula (10) _j;

Step 3: upgrade cluster centre matrix P according to formula (13) ^(b):

${p_i}^{\left(b\right)}=\frac{\underset{j=1}{\overset{s}{\mathrm{\&Sigma;}}}\underset{k=1}{\overset{n}{\mathrm{\&Sigma;}}}{w}_j{\left({{\mathrm{\&mu;}}_{ik}}^{\left(b\right)}\right)}^m{x}_k}{\underset{j=1}{\overset{s}{\mathrm{\&Sigma;}}}\underset{k=1}{\overset{n}{\mathrm{\&Sigma;}}}{w}_j{\left({{\mathrm{\&mu;}}_{ik}}^{\left(b\right)}\right)}^m}---\left(13\right)$

Step 4: upgrade subordinated-degree matrix U according to formula (14) ^(b+1):

${{\mathrm{\&mu;}}_{ik}}^{(b+1)}={\{\underset{h=1}{\overset{c}{\&Sigma;}}\&lsqb;{\left(\frac{{d_{ik}}^{(b+1)}}{{d_{hk}}^{(b+1)}}\right)}^{\frac{2}{m-1}}\&rsqb;\}}^{-1}---\left(14\right)$

Step 5: judge whether || U ^(b)-U ^(b+1)|| < ε, if so, then algorithm stops, output matrix U and P.Otherwise make b=b+1, forward step 3 to and continue to perform.Wherein || || be the matrix norm that certain is suitable.

After algorithm stops, the degree of membership obtained and the form of cluster centre are respectively:

${\mathrm{\&mu;}}_{ik}=\frac{1}{\underset{k=1}{\overset{c}{\&Sigma;}}{\left(\frac{d_{ik}}{d_{hk}}\right)}^{\frac{2}{m-1}}}---\left(15\right)$

$p_i=\frac{\underset{j=1}{\overset{s}{\mathrm{\&Sigma;}}}\underset{k=1}{\overset{n}{\mathrm{\&Sigma;}}}{w}_j{\left({\mathrm{\&mu;}}_{ik}\right)}^m{x}_k}{\underset{j=1}{\overset{s}{\mathrm{\&Sigma;}}}\underset{k=1}{\overset{n}{\mathrm{\&Sigma;}}}{w}_j{\left({\mathrm{\&mu;}}_{ik}\right)}^m}---\left(16\right)$

Algorithm is after successive ignition, and the subordinated-degree matrix of output is:

U＝[μ _ik] _c×n(17)

$\underset{i=1}{\overset{c}{\&Sigma;}}{\mathrm{\&mu;}}_{ik}=1---\left(18\right)$

In formula (18), c=5, represents the number of all power classes; N represents the number of decent of Standard Gases; μ _ikexpression kth (k=1,2 ..., n) individual Standard Gases decent x _kbe under the jurisdiction of the degree of i-th power class, as follows:

$U={\left({\mathrm{\&mu;}}_{ik}\right)}_{c\&times;n}=\left(\begin{array}{cccc}{u}_{11}& {\mathrm{\&mu;}}_{12},& ...,& {\mathrm{\&mu;}}_{1n}\\ u_{21}& u_{22},& ...,& u_{2n}\\ ...& ...,& ...,& ...\\ u_{c1}& u_{c2},& ...,& u_{cn}\end{array}\right)$

In matrix U, the degree of membership that meteorological sample to be measured is under the jurisdiction of each power class is in last 1 row, if the maximal value in last 1 row and the degree of membership maximal value of decent of certain Standard Gases appear at same a line, illustrate that the performance number corresponding to decent of this Standard Gases is the output power of weather data sample to be measured, so far complete weather data sample x _dpower prediction, predicting the outcome shows by supervisory system man-machine interface.

In order to enrich further and improve knowledge base system, by the weather data sample x in this power prediction _dand the power prediction result feedback of correspondence returns knowledge base system, knowledge base system as shown in Figure 5, is convenient to accumulate true and reliable historical data for later weather data sample carries out power prediction.

In above-described specific embodiment; object of the present invention, technical scheme are further described; be understood that; the foregoing is only specific embodiments of the invention; be not limited to the present invention; within the spirit and principles in the present invention all, any amendment made, equivalent replacement, improvement etc., all should be included within protection scope of the present invention.

Claims (7)

1. based on a photovoltaic plant output power predicting method for weighted FCM clustering algorithm, it is characterized in that, this Forecasting Methodology comprises the following steps:

1) determine the following precise time section needing to carry out power prediction, gather in this time period the weather data of 8 dimension attributes comprising the sun total spoke amount, horizontal plane radiation amount, dip plane radiant quantity, environment temperature, air humidity, atmospheric pressure, wind speed and wind speed in photovoltaic plant region as weather data sample to be measured;

2) from the knowledge base of this photovoltaic plant history run, choose the reference sample similar to weather data sample to be measured, in each reference sample, include 8 Weather property identical with weather data sample to be measured and the photovoltaic plant output power of corresponding meteorological sample;

3) from reference sample, choose the master sample data set with weather data sample matches to be measured, combine with weather data sample to be measured, form typical sample matrix, and formation master sample matrix is normalized to this typical sample matrix;

4) using the input quantity of master sample matrix as attribute weight FCM clustering algorithm, initializing variable according to setting carries out algorithm iteration to obtain corresponding subordinated-degree matrix, judge according to the subordinated-degree matrix obtained the photovoltaic plant output power that meteorological sample to be measured is corresponding, this output power is the predicted power of meteorological sample to be measured.

2. the photovoltaic plant output power predicting method based on weighted FCM clustering algorithm according to claim 1, it is characterized in that, described attribute weight FCM clustering algorithm refers to and utilizes Sample Similarity to give corresponding weights to the percentage contribution of attribute each in meteorological sample to cluster result.

3. the photovoltaic plant output power predicting method based on weighted FCM clustering algorithm according to claim 2, is characterized in that, the flow process of described attribute weight FCM clustering algorithm is as follows:

A. each attribute weights in master sample matrix are calculated;

B. according to each attribute weights determination cluster objective function obtained and renewal cluster centre;

C. subordinated-degree matrix is upgraded, until subordinated-degree matrix meets setting requirement after upgrading.

4. the photovoltaic plant output power predicting method based on weighted FCM clustering algorithm according to claim 3, is characterized in that, the calculating of each attribute weights of described sample adopts the attribute weight algorithm based on Sample Similarity.

5. the photovoltaic plant output power predicting method based on weighted FCM clustering algorithm according to claim 3, it is characterized in that, described step 4) in initializing variable comprise cluster classification, this cluster classification is that each section is a cluster classification by the class of accuracy refinement segmentation of the power stage scope of transformer station to be predicted according to prediction.

6. the photovoltaic plant output power predicting method based on weighted FCM clustering algorithm according to claim 4, it is characterized in that, described step 3) method for normalizing can adoption rate standardization, maximum value standardization, mean variance standardization, any one in absolute value variance criterion and regular standardization.

7. the photovoltaic plant output power predicting method based on weighted FCM clustering algorithm according to any one of claim 1-6, it is characterized in that, the method also comprises returns in knowledge base by the power prediction result feedback of meteorological sample to be measured and its correspondence, is convenient to accumulate true and reliable historical data for later weather data sample carries out power prediction.