CN102799948A - Prediction method for power generating system output power of grid-connected type photovoltaic power station - Google Patents

Abstract

The invention discloses a prediction method for power generating system output power of a grid-connected type photovoltaic power station in the technical field of photovoltaic power generation. The prediction method comprises the following steps of: firstly constructing a sample set: denoising the sample set to obtain the sample set after denoising; then establishing a kernel matching pursuit training model by utilizing the sample set after denoising; afterwards optimizing a coefficient of the kernel matching pursuit training model by utilizing a harmony search algorithm; and finally, predicting the generating power of the photovoltaic power station by utilizing the kernel matching pursuit training model with the optimized coefficient. The prediction method avoids the local optimum phenomenon to enable a prediction model to be accurate, and has very good nonlinear processing capability and high-dimensional data processing capability.

Description

A kind of parallel networking type photovoltaic power station power generation system output power Forecasting Methodology

Technical field

The invention belongs to the photovoltaic power generation technology field, relate in particular to a kind of parallel networking type photovoltaic power station power generation system output power Forecasting Methodology.

Background technology

Solar energy power generating belongs to low, the poor stability of energy density; The energy of regulating power difference; Generated energy receives the influence of weather and region bigger, and when the generated energy of the electricity volume of photovoltaic generation and conventional power plant is in the comparable order of magnitude or becomes can not ignore a part of, parallel network power generation will bring a series of influences to the stablizing of existing power generation mode and electrical network, economy, safe operation and power supply quality; The influence of peak valley of for example loading to electrical network; Variation, the western and eastern's time difference and changes of seasons are to the influence of electrical network round the clock, and meteorological condition changes the influence to electrical network, the influence that remote photovoltaic power delivery is brought etc.

Because photovoltaic generation is the singularity of output power periodically; Can produce the impact of one-period property after being incorporated into the power networks to electrical network; The disturbance of output power might influence the stable of electrical network; Make that the distribution network planning personnel are difficult to accurately predict the load increasing situation more, thereby influence the scheduling of system and the plan of unit output.When therefore doing development plan to the electrical network that contains photovoltaic parallel in system; Be necessary exerting oneself of photovoltaic system predicted, so as to understand large-scale solar photovoltaic grid-connection system the generator operation characteristic and with the matching problem of dispatching of power netwoks, electric load etc.The accurate prediction that photovoltaic generation is exerted oneself helps electric power system dispatching department and in time adjusts operation plan, alleviates the influence of parallel network power generation to electrical network effectively.

The present Forecasting Methodology of photovoltaic generation power mainly contains time series method, expert system approach, neural network method, SVMs etc.The time series method calculated amount is little; Rapid speed; Meteorology change little in prediction effect good, but the factor change that influence generated output is greatly the time, like temperature, intensity of illumination etc.; Then be difficult to reflect exert oneself and these variablees between dynamic, nonlinear relation, so relatively poor to photovoltaic generation prediction effect complicated and changeable.Expert system can be avoided complicated numerical evaluation, but versatility is relatively poor, lacks learning ability.And the model that neural network algorithm is just set up according to the empiric risk minimization principle, speed of convergence is slow, possibly converge to local minimum point, and the knowledge representation difficulty is difficult to make full use of dispatcher's experimental knowledge, and needs the long training time.Also there are some significant disadvantages in SVMs; In training process, need find the solution a quadratic programming problem mainly due to SVMs; This makes for extensive sample set, to have very high time complexity and space complexity; In order to reduce the time and the space complexity of finding the solution quadratic programming, need optimization problem be decomposed into the plurality of sub problem.What find the solution that the algorithms of these subproblems obtains is not optimum solution usually, has therefore reduced the generalization ability of SVMs.

Summary of the invention

In the inferior position aspect precision of prediction and the training speed, the present invention proposes a kind of parallel networking type photovoltaic power station power generation system output power Forecasting Methodology to the Forecasting Methodology of mentioning in the above-mentioned background technology.

Technical scheme of the present invention is, a kind of parallel networking type photovoltaic power station power generation system output power Forecasting Methodology is characterized in that this method may further comprise the steps:

Step 1: make up sample set:

Step 2: to the sample set denoising, obtain the sample set after the denoising based on bent wave conversion;

Step 3: utilize the sample set after the denoising, set up the nuclear matching tracing training pattern;

Step 4: the coefficient of optimizing the nuclear matching tracing training pattern;

Step 5: utilize the nuclear matching tracing training pattern of having optimized coefficient to carry out the photovoltaic power station power generation power prediction.

Said step 1 is specially:

Step 1.1: the generated output of photovoltaic system is classified according to weather pattern;

Step 1.2: find out specific data and add up;

Step 1.3: specific data is carried out sample analysis, form sample set.

Said step 2 is specially:

Step 2.1:, obtain bent wave conversion coefficient to sample set march wave conversion;

Step 2.2: bent wave conversion coefficient is carried out threshold process, obtain revising bent wave system number;

Step 2.3: count the inverse transformation of march ripple to revising bent wave system, obtain the sample set after the denoising.

Said step 3 is specially:

Step 3.1: according to the sample set generating function dictionary after the denoising;

Step 3.2: utilize the function dictionary to expand loss function, obtain the optimal base function;

Step 3.3: try to achieve the nuclear matching tracing training pattern by the optimal base function.

Said step 4 is specially:

Step 4.1: obtain the nuclear matching tracing training pattern by step 3 and obtain the harmony data base;

Step 4.2: generate new harmony, upgrade the harmony data base;

Step 4.3: repeating step 4.2, when reaching the repeatedly setting number of times, the coefficient in the harmony data base is the optimal value of the coefficient of nuclear matching tracing training pattern.

Said nuclear matching tracing training pattern is:

$f_N\left(x\right)=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\ω}}_i{g}_i\left(x\right)$

Wherein:

f _N(x) be the nuclear matching tracing training pattern;

ω _iIt is the coefficient of i basis function;

g _i(x) be i basis function;

N is the quantity of basis function.

It is with bent wave conversion the sample set that obtains to be carried out denoising earlier that the present invention carries out forecast method to photovoltaic power station power generation power; Optimize the coefficient in the KMP model with HS then; Predict that with the KMP model after optimizing the present invention is suitable for the prediction of photovoltaic power station power generation power at last, realized the principle of structural risk minimization; The phenomenon of local optimum has been avoided in the introducing of harmony search, makes forecast model more accurate.

Description of drawings

Fig. 1 is a photovoltaic power station power generation power prediction process flow diagram;

Fig. 2 is for based on the method for the bent wave conversion process flow diagram to the original sample collection denoising that obtains;

Fig. 3 is based on harmony search algorithm optimizes KMP coefficient process flow diagram.

Embodiment

Below in conjunction with accompanying drawing, preferred embodiment is elaborated.Should be emphasized that following explanation only is exemplary, rather than in order to limit scope of the present invention and application thereof.

Nuclear matching tracing (KMP) learning machine is a kind of machine learning method of novelty; Its calculation cost mainly spends in to be sought on basis function and the coefficient thereof; Yet each searching all is stepping type search in all basis function wordbooks; To seek the best atom that is matched with signal structure, so depend primarily on the dictionary scale computing time of KMP and reach the required iterations of fitting precision.Like this,, in a match tracing, a global search to be done, surprising calculated amount will be introduced after the iteration for several times if when dictionary is larger.As a kind of intelligent optimization search strategy; Harmony searching algorithm (HS) has all obtained using widely and having obtained good effect in various fields; Each step in that the nuclear matching tracing iteration is decomposed is adopted the harmony searching algorithm, replaces greedy search method through harmony search choice mechanism, and time-frequency atom dictionary is carried out global optimizing; The harmony storehouse that time-frequency atom dictionary is promptly produced by the weights coefficient of KMP, thus can obtain optimum weights coefficient.Final anticipation function is the weight coefficient that searches among the KMP and the linear combination of basis function, by the anticipation function that obtains the generated output of photovoltaic plant is predicted.

Fig. 1 is a kind of photovoltaic power station power generation power forecasting method process flow diagram based on HS-KMP provided by the invention.Among Fig. 1, method provided by the invention comprises following step:

Step 1: before Collection and Forecast Day with identical 5 days of prediction day weather pattern in intensity of illumination, temperature and the generated output of corresponding photovoltaic system of synchronization, the structure sample set.

Step 11: the generated output of photovoltaic system is divided into four types of fine, cloudy, cloudy, rain according to corresponding weather pattern.

Step 12: statistical dependence data;

According to the influence of the generated output of photovoltaic system, the input data of sample are reduced following several types: the generated output O={o of the photovoltaic system of the preceding 5 day synchronization identical with prediction day weather pattern ₁, o ₂..., o _n, the temperature property T of prediction day, the highest temperature, the lowest temperature, temperature on average T={t ₁, t ₂, t ₃, and weather pattern.

Step 13: data are carried out sample analysis, form sample set.

In the present embodiment; Because the sunshine-duration of annual every day is not quite similar; Therefore photovoltaic time of exerting oneself also is not quite similar, and the sunshine-duration in winter must be lower than summer, for the unified photovoltaic form of exerting oneself; The sampling that the photovoltaic of regulation of the present invention every day goes out force data is a standard with maximum 14 hours sunshine times at 6 (in the morning to point in evenings 19), is that spot patch is neat for the data assignment less than standard.To the force data that goes out in each integral point moment, set up 14 records respectively and predict that photovoltaic system is in corresponding integral point photovoltaic value of exerting oneself constantly.Every record also need be preserved the temperature property and the weather pattern of future position except that needs are preserved load data.

Step 2: data analysis is handled, and to the sample set denoising, formation obtains the sample set after the denoising based on bent wave conversion, and Fig. 2 is the sample set denoising method process flow diagram that utilizes bent wave conversion;

Step 2.1: establish the sample set data and be expressed as { f _{I, j}, i=1,2 ... 5, j=1,2 ..., 14} in the present invention, established M=5 days, and N=14 sampled point arranged every day, and then the load data of noisy is expressed as { g _{I, j}=f _{I, j}+ ξ _{I, j}, i=1,2 ..., 5, j=1,2 ..., 14}, wherein, noise { ξ _IjNormal Distribution N (0, σ ²).Confirm that Qu Bofen separates the direction parameter S on number of plies L and each level n _t,, obtain bent wave conversion coefficient W to sample march wave conversion _g(l, s, i, j) (l=1,2 ... L, s=1,2 ... S _l), wherein, l is that Qu Bofen separates level, and s is a direction parameter, and i is a fate, and j is the sampling time point;

Step 2.2: the bent wave conversion coefficient W that decomposition is obtained _g(i j) carries out threshold process for l, s, obtains bent wave system number ${\hat{W}}_f(l,s,i,j);$

Estimate the noise variance of each subband, calculating formula is:

${\widehat{\mathrm{\σ}}}_{\mathrm{w\ξ}}(l,s)=\frac{\underset{i,j}{\mathrm{Median}}\left(\right|W_g(l,s,i,j)\left|\right)}{0.6745}$

In the formula:

is the estimated value of noise variance;

Median is a median, the 0.6745th, and standard deviation.

Estimate the signal variance of each subband, by:

${\mathrm{\σ}}_{W_g}^2(l,s)={\mathrm{\σ}}_{W_f}^2(l,s)+{\mathrm{\σ}}_{W_{\mathrm{\ξ}}}^2(l,s)$

Can obtain:

${\mathrm{\σ}}_{W_f}(l,s)=\sqrt{\max ({\mathrm{\σ}}_{W_g}^2(l,s)-{\mathrm{\σ}}_{W_{\mathrm{\ξ}}}^2(l,s))}$

Wherein:

is the signal population variance;

is signal variance;

is noise variance.

Calculate the arithmetic mean A on each subband _MWith geometrical mean G _M

$A_M=\frac{1}{{\mathrm{<figure-callout\; id="2"\; label="M"\; filenames="HDA00001796460600011.png,HDA00001796460600021.png"\; state="\{\{state\}\}">M</figure-callout>}}^2}\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{M}{\mathrm{\&Sigma;}}}X(i,j)$

$G_M={\left[\underset{i=1}{\overset{M}{\mathrm{\&Pi;}}}\underset{j=1}{\overset{M}{\mathrm{\&Pi;}}}X(i,j)\right]}^{1/{\mathrm{<figure-callout\; id="2"\; label="M"\; filenames="HDA00001796460600011.png,HDA00001796460600021.png"\; state="\{\{state\}\}">M</figure-callout>}}^2}$

In the formula:

A _MBe arithmetic mean;

G _MBe geometrical mean;

(i j) is the subband matrix to X.

Calculate the threshold value of each subband:

$T_h(l,s)=\frac{{\widehat{\mathrm{\&sigma;}}}_{W_{\mathrm{\&xi;}}}^2}{{\widehat{\mathrm{\&sigma;}}}_{W_f}^2}-|A_M-G_M|$

In the formula:

T _h(l, s) threshold value of subband;

for the signal variance

estimates;

for the signal variance estimate.

Then revising bent wave system number is:

${\hat{W}}_f(l,s,i,j)=\left\{\begin{array}{c}\mathrm{sgn}\left(W_g(l,s,i,j)\right)\left(\right|W_g(l,s,i,j)\left|\right)-T_h(l,s),\left|W_g(l,s,i,j)\right|\&GreaterEqual;T_h(l,s)\\ 0,\left|W_g(l,s,i,j)\right|\&le;T_h(l,s)\end{array}\right.$

The sgn function is meant: work as W _g(l, s, i, j)>0, sgn (W _g(l, s, i, j))=1; Work as W _g(l, s, i, j)=0, sgn (W _g(l, s, i, j))=0; Work as W _g(l, s, i, j)<0, sgn (W _g(l, s, i, j))=-1.

Step 2.3:, obtain the sample set after the denoising to revising bent wave system number

march ripple inverse transformation;

The coefficient

that threshold process is crossed carries out reconstruct, obtains the sample set after the denoising of reconstruct.

Step 3: definition nuclear matching tracing learning machine model and coefficient;

Step 31: given l observation station { x in the sample set after denoising ₁..., x _lAnd kernel function K:R ^d* R ^d→ R utilizes observation station { x ₁..., x _lKernel function value { the y that locates ₁..., y _lGenerating function dictionary: D={g _i=k (, x _i) | i=1 ..., l};

Step 32: utilize the function dictionary to expand loss function, obtain the optimal base function;

Suppose loss function L (y _i, f _n(x _i)), when observed reading is y _iThe time, calculate predicted value f _n(x _i) residual error

${\tilde{R}}_n=(-\frac{\&PartialD;L({\mathrm{<figure-callout\; id="1"\; label="y"\; filenames="HDA00001796460600011.png,HDA00001796460600021.png"\; state="\{\{state\}\}">y</figure-callout>}}_1,f_n\left(x_1\right))}{{\&PartialD;f}_n\left(x_1\right)},...,-\frac{\&PartialD;L(y_l,f_n\left(x_l\right))}{{\&PartialD;f}_n\left(x_l\right)})$

So, by the nuclear matching tracing algorithm, in iteration each time the optimal base function that will seek be:

$g_{i+1}=\arg \underset{g\&Element;D}{\max }\left|\frac{<g_{i+1},>{\tilde{R}}_i}{\left|\right|g_{i+1}\left|\right|}\right|$

The coefficient ω of corresponding this optimal base function _I+1For:

${\mathrm{\&omega;}}_{i+1}=\arg \underset{\mathrm{\&alpha;}\&Element;R}{\min }\underset{k=1}{\overset{l}{\mathrm{\&Sigma;}}}L(y_k,f_i\left(x_k\right)+{\mathrm{\&omega;g}}_{i+1}\left(x_k\right))$

Match nuclear matching tracing algorithm then promptly is to carry out following optimizing process:

${\mathrm{\&omega;}}_{1,...i+1}^{(i+1)}=\arg \underset{\left({\mathrm{\&omega;}}_{1,...,i+1}\right)\&Element;R^{i+1}}{\min }\underset{k=1}{\overset{l}{\mathrm{\&Sigma;}}}L(y_k,\underset{m=1}{\overset{i+1}{\mathrm{\&Sigma;}}}{\mathrm{\&omega;}}_m{g}_m\left(x_k\right))$

The loss function that is adopted in the nuclear matching tracing learning machine can for:

Quadratic loss function:

perhaps revises the tanh loss function:

$L(y,f_n\left(x\right))={(\tanh \tilde{f}\left(x\right)-0.65y)}^2$

Step 33: trying to achieve the nuclear matching tracing training pattern by the optimal base function is:

$f_N\left(x\right)=\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{\&omega;}}_i{g}_i\left(x\right)=\underset{i\&Element;\left\{\mathrm{sp}\right\}}{\mathrm{\&Sigma;}}{\mathrm{\&omega;}}_{i}k\left({x,x}_i\right)$

Step 4: the coefficient that utilizes harmony search algorithm optimizes nuclear matching tracing training pattern.Fig. 3 is the process flow diagram that utilizes HS algorithm optimization KMP coefficient.

Step 41: the nuclear matching tracing training pattern that is obtained by step 3 for

wherein; ω is a coefficient; With ω as harmony, initialization harmony data base.

(1) through nuclear mapping training sample is become basis function dictionary D, confirm to be used for the data set s={ (x of training study machine ₁, y ₁) ..., (x _l, y _l), x ∈ E, y _i∈ 1,1} (i=1,2 ..., l), and kernel function k.

(2) utilize given nuclear matching tracing function y and training data to concentrate optional x _i, obtain all basis function y _(i)(x)=k (x, x _i).

(3) utilize $\underset{{\mathrm{\&omega;}}_i}{\mathrm{Min}}\left|\right|y-{\mathrm{\&omega;}}_i{y}_{\left(i\right)}\left(x\right)\left|\right|$ Criterion is obtained ${\mathrm{\&omega;}}_i=\frac{y_{\left(i\right)}^T\left(x\right)\&CenterDot;y}{{\left|\right|y_{\left(i\right)}\left(x\right)\left|\right|}^2};$

(4) therefrom take out N ω at random _i, adopt real coding, as initial sum sound memory storehouse W, W={ ω ₁, ω ₂..., ω _N, harmony data base W is the N tuple of harmony ω.

The harmony data base can be analogous to the population of genetic algorithm, and harmony data base form is following:

$W=\left[\begin{array}{cc}\begin{array}{c}{\mathrm{\&omega;}}^1\\ {\mathrm{\&omega;}}^2\\ .\\ .\\ .\\ {\mathrm{\&omega;}}^{\mathrm{HMS}}\end{array}\left|\begin{array}{c}f\left({\mathrm{\&omega;}}^1\right)\\ f\left({\mathrm{\&omega;}}^2\right)\\ .\\ .\\ .\\ f\left({\mathrm{\&omega;}}^{\mathrm{HMS}}\right)\end{array}\right.& \end{array}\right]=\left[\begin{array}{cccc}\begin{array}{c}{\mathrm{\&omega;}}_1^1\\ {\mathrm{\&omega;}}_1^2\\ .\\ .\\ .\\ {\mathrm{\&omega;}}_1^{\mathrm{HMS}}\end{array}& \begin{array}{c}...\\ ...\\ \\ \\ \\ ...\end{array}& \begin{array}{c}{\mathrm{\&omega;}}_{\mathrm{<figure-callout\; id="1"\; label="N"\; filenames="HDA00001796460600011.png,HDA00001796460600021.png"\; state="\{\{state\}\}">N</figure-callout>}}^1\\ {\mathrm{\&omega;}}_{\mathrm{<figure-callout\; id="2"\; label="N"\; filenames="HDA00001796460600011.png,HDA00001796460600021.png"\; state="\{\{state\}\}">N</figure-callout>}}^2\\ .\\ .\\ .\\ {\mathrm{\&omega;}}_N^{\mathrm{HMS}}\end{array}& \left|\begin{array}{c}f\left({\mathrm{\&omega;}}^1\right)\\ f\left({\mathrm{\&omega;}}^2\right)\\ .\\ .\\ .\\ f\left({\mathrm{\&omega;}}^{\mathrm{HMS}}\right)\end{array}\right]\end{array}\right.$

Wherein, HMS is the size of harmony data base;

Step 42: generate new harmony, upgrade the harmony data base.If maximum update times is Tmax, initial value t=1.

(1) study harmony data base produces new explanation.First variable ω ' of new explanation ₁There is the value probability of HMCR to select among the HM

Get any one value, any one value of the value probability selection HM outer (and in variable range) of 1-HCMR is arranged.Equally, the generating mode of other variable is following:

${\mathrm{\&omega;}}_i^{\&prime;}=\left\{\begin{array}{c}{\mathrm{\&omega;}}_i\&Element;({\mathrm{\&omega;}}_i^1,{\mathrm{\&omega;}}_i^2,...{\mathrm{\&omega;}}_i^{\mathrm{HMS}}),\mathrm{rand}<\mathrm{HMCR}\\ {\mathrm{\&omega;}}_i\&Element;X_i,\mathrm{otherwise}\end{array}\right.i=\mathrm{1,2},...,N$

Wherein, rand representes [0,1] upward equally distributed random number, and X is that other is not by the weights coefficient set of taking out immediately.

(2.) tone fine setting.If new harmony ω ' _iFrom harmony data base HM, finely tune it.

Concrete operations are following:

${\mathrm{\&omega;}}_i^{\&prime;}=\left\{\begin{array}{c}{\mathrm{\&omega;}}_i+\mathrm{rand}*\mathrm{bw},\mathrm{rand}<\mathrm{PAR}\\ {\mathrm{\&omega;}}_i,\mathrm{otherwise}\end{array}\right.i=\mathrm{1,2},...N$

Wherein, bw is a tone fine setting bandwidth, and PAR is a tone fine setting frequency, and rand representes [0,1] upward equally distributed random number.

(3) new explanation is assessed, if be superior to the poorest one of functional value among the HM, then new explanation is updated among the HM, concrete operations are following:

Iff(ω')＞f(ω ^worst)Then?ω ^worst＝ω′

Step 43: repeating step 42, till reaching Tmax.Coefficient in the harmony data base of this moment is the optimal value of the coefficient of nuclear matching tracing training pattern;

Step 5: utilize the KMP after HS optimizes to carry out the photovoltaic power station power generation power prediction.

Element in the sample set of handling well brought into optimize back nuclear matching tracing training pattern and photovoltaic power station power generation power is predicted got final product.

The calculation cost of traditional nuclear matching tracing (KMP) mainly spends in to be sought on basis function and the coefficient thereof; Yet each searching all is stepping type search in all basis function wordbooks; To seek the best atom that is matched with signal structure, so depend primarily on the dictionary scale computing time of KMP and reach the required iterations of fitting precision.Like this,, in a match tracing, a global search to be done, surprising calculated amount will be introduced after the iteration for several times if when dictionary is larger.Because the limitation that standard nuclear coupling is pursued learning machine self makes it can not handle these problems effectively.The present invention has adopted harmony searching algorithm (HS) that the coefficient of KMP is optimized.Research shows that HS-KMP has stronger popularization ability, and more strong non-linear processing power and higher-dimension processing power are compared than other nuclear machines simultaneously, and the sparse property that it is separated is more excellent.

The above; Be merely the preferable embodiment of the present invention, but protection scope of the present invention is not limited thereto, any technician who is familiar with the present technique field is in the technical scope that the present invention discloses; The variation 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 (6)

1. parallel networking type photovoltaic power station power generation system output power Forecasting Methodology is characterized in that this method may further comprise the steps:

Step 1: make up sample set:

Step 2: to the sample set denoising, obtain the sample set after the denoising based on bent wave conversion;

Step 3: utilize the sample set after the denoising, set up the nuclear matching tracing training pattern;

Step 4: the coefficient of optimizing the nuclear matching tracing training pattern;

Step 5: utilize the nuclear matching tracing training pattern of having optimized coefficient to carry out the photovoltaic power station power generation power prediction.

2. a kind of parallel networking type photovoltaic power station power generation according to claim 1 system output power Forecasting Methodology is characterized in that said step 1 is specially:

Step 1.1: the generated output of photovoltaic system is classified according to weather pattern;

Step 1.2: find out specific data and add up;

Step 1.3: specific data is carried out sample analysis, form sample set.

3. a kind of parallel networking type photovoltaic power station power generation according to claim 1 system output power Forecasting Methodology is characterized in that said step 2 is specially:

Step 2.1:, obtain bent wave conversion coefficient to sample set march wave conversion;

Step 2.2: bent wave conversion coefficient is carried out threshold process, obtain revising bent wave system number;

Step 2.3: count the inverse transformation of march ripple to revising bent wave system, obtain the sample set after the denoising.

4. a kind of parallel networking type photovoltaic power station power generation according to claim 1 system output power Forecasting Methodology is characterized in that said step 3 is specially:

Step 3.1: according to the sample set generating function dictionary after the denoising;

Step 3.2: utilize the function dictionary to expand loss function, obtain the optimal base function;

Step 3.3: try to achieve the nuclear matching tracing training pattern by the optimal base function.

5. a kind of parallel networking type photovoltaic power station power generation according to claim 1 system output power Forecasting Methodology is characterized in that said step 4 is specially:

Step 4.1: obtain the nuclear matching tracing training pattern by step 3 and obtain the harmony data base;

Step 4.2: generate new harmony, upgrade the harmony data base;

Step 4.3: repeating step 4.2, when reaching the repeatedly setting number of times, the coefficient in the harmony data base is the optimal value of the coefficient of nuclear matching tracing training pattern.

6. a kind of parallel networking type photovoltaic power station power generation according to claim 3 system output power Forecasting Methodology is characterized in that said nuclear matching tracing training pattern is:

$f_N\left(x\right)=\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{\&omega;}}_i{g}_i\left(x\right)$

Wherein:

f _N(x) be the nuclear matching tracing training pattern;

ω _iIt is the coefficient of i basis function;

g _i(x) be i basis function;

N is the quantity of basis function.

Images