CN105447509A - Short-term power prediction method for photovoltaic power generation system - Google Patents

Abstract

The invention relates to a short-term power prediction method for a photovoltaic power generation system. The method comprises the steps of re-dividing weather types: according to the factors that whether the variables of the natural environment are continuous and stable or not and the averaged values of the variables, dividing weather types in the generalized point of view; determining a similar day of a to-be-predicted day: determining a generalized weather type corresponding to the to-be-predicted day according to weather factors, selecting a historical day similar to the to-be-predicted day according to the grey correlation method, adopting the photovoltaic power output of the historical day as a model input, and selecting daily weather feature vectors; defining the degree of similarity between the to-be-predicted day and the above particular historical day; establishing a photovoltaic power generation short-term power prediction model based on a support vector machine, and optimizing the parameters of the prediction model according to the genetic algorithm. The above precision method is high in precision, and fast in convergence.

Description

A kind of short term power Forecasting Methodology of photovoltaic generating system

Technical field

The present invention relates to a kind of photovoltaic generating system power forecasting method.

Background technology

In the face of increasingly serious Energy situation and more poverty-stricken environmental protection present situation, China has formulated the relevant policies and development plan of greatly developing regenerative resource, and sun power as a kind of love knot, technical research input that regenerative resource causes energy circle a large amount of.Solar electrical energy generation have safe and reliable, energy quality is high, without exhausted dangerous, pollution-free, maintain the advantages such as simple, be the important directions of Renewable Energy Development.Although photovoltaic generation is widely applied because of its plurality of advantages, but because photovoltaic generating system is by such environmental effects, there is the features such as uncertainty, undulatory property, intermittence, be unfavorable for sacurity dispatching and the energy management of electrical network, add the operation risk of electrical network.Therefore, accurately predicting is carried out to the short term power of photovoltaic generation, the sacurity dispatching of electrical network and energy management tool are of great significance.

Summary of the invention

The object of this invention is to provide a kind of photovoltaic generating system short term power accurately predicting method, for the adverse effect effectively regulating photovoltaic generating system to bring the sacurity dispatching of electrical network and energy management.Technical scheme of the present invention is as follows:

A short term power Forecasting Methodology for photovoltaic generating system, comprises the following steps:

(1), the repartitioning of weather pattern

Historical data is divided according to weather pattern, thus set up each self-corresponding forecast model, intensity of illumination and environment temperature according to affecting the Major Natural factor that photovoltaic system exerts oneself, little for intensity of illumination undulate quantity, that mean value is lower cloudy day, sleet and haze are done merging treatment, according to the size of physical environment amount whether continous-stable, mean values, weather pattern is divided into A, B, C tri-class from broadest scope:

Broad sense weather divides	A	B	C
				Tradition weather pattern	Fine day	Cloudy	Cloudy day, the rainy day

(2) similar day of day to be predicted, is determined

The broad sense weather pattern corresponding to it is determined according to the weather conditions of day to be predicted, and choose the history day similar to day to be predicted according to Grey Correlation Method, exert oneself as mode input amount using the photovoltaic generation of this history day, Meteorological Characteristics vector every day chosen is:

x _i＝[x _i(1),x _i(2),x _i(3),]＝[t _hi,t _li,h _i]\*MERGEFORMAT(1)

Wherein, t _hirepresent the i-th max. daily temperature, t _lirepresent the i-th Daily minimum temperature, h _irepresent weather humidity on the i-thth;

With x ₀represent day to be predicted, then day x to be predicted ₀with i-th history day x _ithe correlation coefficient of a jth characteristic component be:

$\begin{array}{cc}\begin{array}{c}{\mathrm{\δ}}_i\left(j\right)=\left(\underset{i}{\min }\underset{j}{\min }\left|x_0\left(j\right)-x_i\left(j\right)\right|+\mathrm{\ρ}\underset{i}{\max }\underset{j}{\max }\left|x_0\left(j\right)-x_i\left(j\right)\right|\right)/\\ \left(\left|x_0\left(j\right)-x_i\left(j\right)\right|+\mathrm{\ρ}\underset{i}{\max }\underset{j}{\max }\left|x_0\left(j\right)-x_i\left(j\right)\right|\right)\end{array}& \backslash *MERGEFORMAT\end{array}---\left(2\right)$

Wherein, ρ is generally taken as 0.5, get minimum value in absolute difference for two-layer to calculate, namely ground floor is for first to compare ordered series of numbers x by each respectively _ieach point on curve and reference sequence x ₀get minimum value in the absolute difference of each point on curve, then choose minimum value by the middle of these minimum value; for two-layer gets maximum value calculation in absolute difference, namely ground floor is for first to compare ordered series of numbers x by each respectively _ieach point on curve and reference sequence x ₀get maximal value in the absolute difference of each point on curve, then choose maximal value by the middle of these maximal values; | x ₀(j)-x _i(j) | compare ordered series of numbers x for each _ieach point on curve and reference sequence x ₀the absolute difference of each point on curve;

Day x to be predicted ₀with i-th history day x _isimilarity be defined as:

$\begin{array}{cc}{\mathrm{\λ}}_i=\frac{1}{3}\underset{j=1}{\overset{3}{\Σ}}{\mathrm{\δ}}_i\left(j\right)& \backslash *MERGEFORMAT\end{array}---\left(3\right)$

Using the similar day of similarity maximum history day as day to be predicted;

(3) the photovoltaic generation short term power forecast model based on support vector machine, is set up, sample data selects prediction day and similar day thereof, the environmental factor data of maximum temperature, minimum temperature and weather humidity and the generated output data of similar day are comprised with what predict day, as the input quantity of forecast model, to predict that the photovoltaic generation power of day each future position exports the output data of data as forecast model, the parameter of genetic algorithm to forecast model is adopted to be optimized.

The parameter of employing genetic algorithm to forecast model of step (3) is optimized as follows:

1) initial population scale and the iterations of genetic algorithm are set, and crossing-over rate and aberration rate;

2) initialization parameter value to be optimized, namely utilizes real coding mode to carry out initialization to regularization parameter C and nuclear parameter σ;

3) genetic iteration is carried out to population, and calculate the fitness size of each population;

4) using the parameter value after optimizing as the parameter value of support vector machine photovoltaic generation forecast model;

5) whether the output of computational prediction model and the error of actual output meet the demands, if met the demands, then perform step 6), if error does not meet the demands, then return step 3);

6) using the parameter value of the parameter value after this group optimization as forecast model, the power data of the weather data and similar day that input day to be predicted carries out short term power prediction.

This method is by being classified as A, B, C tri-class of broad sense by traditional weather pattern, under the condition of identical weather pattern, the similar day of day to be predicted is found according to Grey Correlation Method, using the input of the photovoltaic generation data of the weather data of day to be predicted and similar day as forecast model, support vector machine method is adopted to carry out the modeling of forecast model, and adopt the parameter of genetic algorithm to forecast model to be optimized, model parameter optimizing is avoided to be absorbed in local optimum, to improve the precision of prediction of forecast model.On basis of the present invention and then effectively can alleviate the adverse effect that photovoltaic generating system brings the energy management of electrical network and dispatching of power netwoks.

Accompanying drawing explanation

Fig. 1 support vector machine structural drawing

The process flow diagram of Fig. 2 Optimized model and forecast model

Predicted power and real power under Fig. 3 category-A meteorological condition

Predict the outcome under Fig. 4 category-A meteorological condition relative error

Predicted power and actual work under Fig. 5 category-B meteorological condition

Predict the outcome under Fig. 6 category-B meteorological condition relative error

Predicted power and real power under Fig. 7 C class meteorological condition

Predict the outcome under Fig. 8 C class meteorological condition relative error

Predicted power and real power under Fig. 7 C class meteorological condition

Predict the outcome under Fig. 8 C class meteorological condition relative error

Embodiment

Below technical scheme of the present invention is described in detail

(1), the repartitioning of weather pattern

The classification forecast model setting up photovoltaic generating system just needs to divide according to weather pattern historical data, thus sets up each self-corresponding forecast model.Weather pattern can be divided into fine day, cloudy, cloudy, rainy day, snow, mist etc. by current weather forecast.Too much category division can cause data set features vector too to disperse, for the training of model brings difficulty.The Major Natural factor of exerting oneself owing to affecting photovoltaic system is intensity of illumination and environment temperature, little for intensity of illumination undulate quantity, that mean value is lower cloudy day, sleet and haze can be done merging treatment, according to the size etc. of physical environment amount whether continous-stable, mean values, weather pattern is divided into A, B, C tri-class from broadest scope.

Broad sense weather divides	A	B	C
				Tradition weather pattern	Fine day	Cloudy	Cloudy day, the rainy day

Concrete operation step is:

1) traditionally weather pattern data carry out broad sense classification to sample data;

2) its broad sense weather pattern is determined according to the weather pattern data of day to be predicted;

3) according to broad sense weather pattern, utilize and day to be predicted of the same type under sample data carry out modeling and forecasting.

(2), the similar day of day to be predicted determined by model

Determine the weather pattern corresponding to it according to the weather conditions of day to be predicted, and choose the history day similar to day to be predicted according to Grey Correlation Method, exert oneself as mode input amount using the photovoltaic generation of this history day.Meteorological Characteristics vector every day chosen is:

x _i＝[x _i(1),x _i(2),x _i(3),]＝[t _hi,t _li,h _i]\*MERGEFORMAT(1)

Wherein, t _hirepresent the i-th max. daily temperature, t _lirepresent the i-th Daily minimum temperature, h _irepresent weather humidity on the i-thth.

With x ₀represent day to be predicted, then day x to be predicted ₀with i-th history day x _ithe correlation coefficient of a jth characteristic component be:

$\begin{array}{cc}\begin{array}{c}{\mathrm{\δ}}_i\left(j\right)=\left(\min \min \left|x_0\left(j\right)-x_i\left(j\right)\right|+\mathrm{\ρ}\max \max \left|x_0\left(j\right)-x_i\left(j\right)\right|\right)/\\ \left(\left|x_0\left(j\right)-x_i\left(j\right)\right|+\mathrm{\ρ}\max \max \left|x_0\left(j\right)-x_i\left(j\right)\right|\right)\end{array}& \backslash *MERGEFORMAT\end{array}---\left(2\right)$

Wherein, ρ is generally taken as 0.5.

Day x to be predicted ₀with i-th history day x _isimilarity be defined as:

$\begin{array}{cc}{\mathrm{\λ}}_i=\frac{1}{3}\underset{j=1}{\overset{3}{\Σ}}{\mathrm{\δ}}_i\left(j\right)& \backslash *MERGEFORMAT\end{array}---\left(3\right)$

Using the similar day of similarity maximum history day as day to be predicted.

Concrete operation step is:

1) determine sample size, be designated as n;

2) sample data is numbered, is numbered 1 to n;

3) calculate the similarity of No. 1 sample data according to formula (2) and (3), and No. 1 sample is designated as similar day;

4) calculate the similarity of No. 2 sample datas, and compare in the similarity of similar day, if the similarity of No. 2 sample datas is large, then No. 2 samples are designated as similar day, if the similarity of No. 2 samples is little, then similar day is constant;

5) the like, until the similarity of n sample all calculated, the similar day finally obtained is the maximum sample of similarity.

(3), the optimization model of photovoltaic generation short term power forecast model and power prediction model are solved

Genetic algorithm is a kind of optimizing algorithm of analog D arwin theory of biological evolution process, it is for starting point with arbitrary colony, by selection, the crossover and mutation operation of randomness, produce the Cenozoic individuality with better adaptive faculty, by the mechanism of the survival of the fittest, colony is moved on the basis of constraint condition towards better search volume, the highest individuality of fitness is searched out, i.e. the optimum solution of optimization problem through constantly multiplying.The present invention carries out optimizing operation by genetic algorithm to the parameter of forecast model.

To the regularization parameter C of forecast model and nuclear parameter σ by selecting, crossover and mutation operates and is optimized.For avoiding gene delection, improving the global convergence of algorithm, utilize individual fitness to carry out selection operation, its specific operation process is: find out individuality that in current group, fitness is the highest and the minimum individuality of fitness; If in current group, the fitness of optimized individual is more taller than the fitness of total best individuality up to now, just using the optimized individual in current group as new total best individuality; The poorest individuality in current group is replaced with best individuality up to now.And selected fitness function is specially:

$\begin{array}{cc}\begin{array}{c}f=\frac{1}{K+1}\\ K=\frac{1}{2}\underset{i=1}{\overset{l}{\Σ}}e{\left(i\right)}^2\\ e\left(i\right)=y\left(i\right)-y_m\left(i\right)\end{array}& \backslash *MERGEFORMAT\end{array}---\left(4\right)$

Wherein, l is learning sample; The real output value that y (i) is model; y _mi desired output that () is model; E (i) is error between the two.

The specific operation process of interlace operation is: first in chromosome Stochastic choice point as point of crossing, then the string after the string before first former generation point of crossing and second former generation point of crossing is combined into a new chromosome, the string after the string before second former generation point of crossing and first former generation point of crossing is combined into a new chromosome.Mutation operation is then the diversity in order to keep species, changes the character value of certain gene in chromosome with certain probability.

For photovoltaic generation short term power forecast model, what the present invention adopted is algorithm of support vector machine.For given sample { (x _i, y _i) (i=1,2 ..., n), wherein n is sample size, x _ifor input vector, y _ifor exporting data accordingly.Support vector machine adopts Nonlinear Mapping r ⁿ→ R ^minput quantity is mapped to high-dimensional feature space by (m>=n), then adopts following linear function to realize regression forecasting:

Wherein for be mapped to high-dimensional feature space, ω is its weight vectors, and b is the side-play amount of position.For realizing the structural risk minimization principle of SVM, risk function is defined as follows:

$L(y-f\left(x\right),x)=\left\{\begin{array}{cc}\begin{array}{c}|y-f(x,\mathrm{\ω})|-\mathrm{\ϵ},|y-f(x,\mathrm{\ω})|>\mathrm{\ϵ}\\ 0,else\end{array}& \backslash *MERGEFORMAT\end{array}\right.---\left(6\right)$

Wherein for ε is amount of bias, be called loss parameter.In order to training parameter b and ω, need minimization as minor function:

$\begin{array}{cc}Rsvm\left(C\right)=\frac{1}{2}\left|\right|\mathrm{\ω}|{|}^2+\frac{C}{n}\underset{i=1}{\overset{n}{\Σ}}L\mathrm{\ϵ}(d_i,y_i)& \backslash *MERGEFORMAT\end{array}---\left(7\right)$

Wherein, ε is loss parameter, for experience error, be called regular item, C is regularization parameter, is also called extensive coefficient, with deciding experience error with the proportion between regular item.

For solving above-mentioned optimization problem, import slack variable ζ and ζ ^*, make it change into:

$\begin{array}{cc}{\mathrm{MinR}}_{svm}\left(\mathrm{\ω},b,{\mathrm{\ζ}}_i,{\mathrm{\ζ}}_i^{*}\right)=\frac{1}{2}\left|\right|\mathrm{\ω}|{|}^2+C\underset{i=1}{\overset{n}{\Σ}}\left({\mathrm{\ζ}}_i,{\mathrm{\ζ}}_i^{*}\right)& \begin{array}{c}s.t.\left\{\begin{array}{c}{y}_i-f\left(x_i\right)\≤\mathrm{\ϵ}+{\mathrm{\ζ}}_i\\ f\left(x_i\right)-y_i\≤\mathrm{\ϵ}+{\mathrm{\ζ}}_i^{*}\\ {\mathrm{\ζ}}_i,{\mathrm{\ζ}}_i^{*}\≥0\end{array}\right.\\ \backslash /MERGEFORMAT\end{array}\end{array}---\left(8\right)$

Introduce Lagrange construction of function equation:

$\begin{array}{c}L\left(\mathrm{\ω},b,\mathrm{\ζ},{\mathrm{\ζ}}^{*},\mathrm{\α},{\mathrm{\α}}^{*},r,r^{*}\right)=\frac{1}{2}\left|\right|\mathrm{\ω}|{|}^2+C\underset{i=1}{\overset{m}{\Σ}}\left({\mathrm{\ζ}}_i,{\mathrm{\ζ}}_i^{*}\right)-\underset{i=1}{\overset{m}{\Σ}}{\mathrm{\α}}_i\[{\mathrm{\ζ}}_i+\mathrm{\ϵ}-y_i+f\left(x_i\right)\]\\ -\underset{i=1}{\overset{m}{\Σ}}{\mathrm{\α}}^{*}\[{\mathrm{\ζ}}_i^{*}+\mathrm{\ϵ}+y_i-f\left(x_i\right)\]-\underset{i=1}{\overset{m}{\Σ}}{\mathrm{r\ζ}}_i-\underset{i=1}{\overset{m}{\Σ}}{r}^{*}{\mathrm{\ζ}}_i^{*}\\ \backslash *MERGEFORMAT\end{array}---\left(9\right)$

Wherein, i=1,2 ..., m; α _i, according to the extremum conditions of L, and in ω generation, is returned formula (4), the expression formula of regression estimates function can be obtained:

Make φ (x _i) φ (x)=k (x, x _i), then formula (9) becomes:

$f(x,{\mathrm{\α}}_i,{\mathrm{\α}}_i^{*})=\underset{i=1}{\overset{m}{\Σ}}({\mathrm{\α}}_i-{\mathrm{\α}}_i^{*})\·k(x,x_i)+b---\left(11\right)$

Wherein k (x, x _i) be then called kernel function.Through type (10) can be avoided calculating weight vector ω, at known Lagrange operator α _i, and kernel function k (x, x _i) condition under become can calculate f (x).

Its idiographic flow as shown in Figure 2, is specially:

1) initial population scale and iterations are set, and crossing-over rate and aberration rate;

2) initialization parameter value to be optimized, namely utilizes real coding mode to carry out initialization to regularization parameter C and nuclear parameter σ (kernel function selected is gaussian kernel function) herein;

3) each population's fitness size is calculated;

4) selection, crossover and mutation operation is performed successively;

5) using the parameter value after optimizing as the parameter value of supporting vector machine model;

6) in sample data, prediction day and similar day is selected, to predict the environmental factor data (comprising maximum temperature, minimum temperature and weather humidity) of day, and the generated output data of similar day (forecast interval of every day is 6:00-19:00, sampling interval is 15min, comprise the power data of this future position of similar day and the data of former and later two adjacent time points) as the input quantity of forecast model, to predict that the photovoltaic generation power of day each future position exports the output data of data as forecast model

7) whether the output of computational prediction model and the error of actual output meet the demands, if met the demands, then perform step 8), if error does not meet the demands, then return step 3)

8) using the parameter value of the parameter value after this group optimization as forecast model, the power data of the weather data and similar day that input day to be predicted carries out short term power prediction.

By the parameter value input prediction model after optimization, after sample training, prediction obtains the generated output of photovoltaic system day to be predicted, by itself and real output with compare without the SVM model predication value of genetic algorithm optimization, comparison diagram under three kinds of meteorological conditions is as shown in Fig. 3, Fig. 5 and Fig. 7, and the relative error comparison that two class forecast models under three kinds of meteorological conditions predict the outcome is as shown in Fig. 4, Fig. 6 and Fig. 8.

As can be seen from Fig. 3 and Fig. 4, under the category-A weather pattern that intensity of illumination and temperature continuous and stable change, GA-SVM forecast model has good prediction effect, global error maintains in 10% scope, the relative error of SVM forecast model, then in 15% scope, knows that improving supporting vector machine model has better precision of prediction.

As can be seen from Fig. 5 and Fig. 6, under the category-B meteorological condition that intensity of illumination and temperature all have certain undulatory property, photovoltaic generation output power also has certain undulatory property, predict the outcome relative to category-A meteorological condition bigger error, but GA-SVM forecast model has better prediction effect than SVM forecast model equally, the global error of GA-SVM forecast model maintains within the scope of 10%-15%, the relative error of SVM forecast model, then within the scope of 15%-20%, knows that improving supporting vector machine model has better precision of prediction.

As can be seen from Fig. 7 and Fig. 8, under intensity of illumination and all lower C class meteorological condition of temperature, photovoltaic generation output power is also lower, predict the outcome all bigger than normal relative to the error of category-A and category-B meteorological condition, and the fluctuation range predicted the outcome as shown in Figure 8 is also larger, relative error totally maintains between 10%-20% scope, but GA-SVM forecast model has better precision of prediction than SVM forecast model equally.

Simultaneously, as can be seen from the comparison diagram of these three relative errors of Fig. 4, Fig. 6 and Fig. 8, when real output is less, namely within these two time periods of 6:00 ~ 8:00 and 17:00 ~ 19:00, the bigger error that predicts the outcome of model, its main cause is because the environmental factors such as haze reduce atmospheric transparency, have impact on intensity of illumination, and affecting less in this time period of 8:00 ~ 17:00, precision of prediction is also relatively high.

Claims (2)

1. a short term power Forecasting Methodology for photovoltaic generating system, comprises the following steps:

(1), the repartitioning of weather pattern

Historical data is divided according to weather pattern, thus set up each self-corresponding forecast model, intensity of illumination and environment temperature according to affecting the Major Natural factor that photovoltaic system exerts oneself, little for intensity of illumination undulate quantity, that mean value is lower cloudy day, sleet and haze are done merging treatment, according to the size of physical environment amount whether continous-stable, mean values, weather pattern is divided into A, B, C tri-class from broadest scope:

Broad sense weather divides A B C Tradition weather pattern Fine day Cloudy Cloudy day, the rainy day

(2) similar day of day to be predicted, is determined

The broad sense weather pattern corresponding to it is determined according to the weather conditions of day to be predicted, and choose the history day similar to day to be predicted according to Grey Correlation Method, exert oneself as mode input amount using the photovoltaic generation of this history day, Meteorological Characteristics vector every day chosen is:

x _i＝[x _i(1),x _i(2),x _i(3),]＝[t _hi,t _li,h _i]\*MERGEFORMAT(1)

Wherein, t _hirepresent the i-th max. daily temperature, t _lirepresent the i-th Daily minimum temperature, h _irepresent weather humidity on the i-thth;

With x ₀represent day to be predicted, then day x to be predicted ₀with i-th history day x _ithe correlation coefficient of a jth characteristic component be:

$\begin{array}{cc}\begin{array}{c}{\mathrm{\δ}}_i\left(j\right)=\left(\underset{i}{\min }\underset{j}{\min }|x_0\left(j\right)-x_i\left(j\right)|+\mathrm{\ρ}\underset{i}{\max }\underset{j}{\max }|x_0\left(j\right)-x_i\left(j\right)|\right)/\\ \left(|x_0\left(j\right)-x_i\left(j\right)|+\mathrm{\ρ}\underset{i}{\max }\underset{j}{\max }|x_0\left(j\right)-x_i\left(j\right)|\right)\end{array}& \backslash *MERGEFORMAT\end{array}---\left(2\right)$

Wherein, ρ is generally taken as 0.5, get minimum value in absolute difference for two-layer to calculate, namely ground floor is for first to compare ordered series of numbers x by each respectively _ieach point on curve and reference sequence x ₀get minimum value in the absolute difference of each point on curve, then choose minimum value by the middle of these minimum value; for two-layer gets maximum value calculation in absolute difference, namely ground floor is for first to compare ordered series of numbers x by each respectively _ieach point on curve and reference sequence x ₀get maximal value in the absolute difference of each point on curve, then choose maximal value by the middle of these maximal values; | x ₀(j)-x _i(j) | compare ordered series of numbers x for each _ieach point on curve and reference sequence x ₀the absolute difference of each point on curve;

Day x to be predicted ₀with i-th history day x _isimilarity be defined as:

$\begin{array}{cc}{\mathrm{\λ}}_i=\frac{1}{3}\underset{j=1}{\overset{3}{\Σ}}{\mathrm{\δ}}_i\left(j\right)& \backslash *MERGEFORMAT\end{array}---\left(3\right)$

Using the similar day of similarity maximum history day as day to be predicted;

(3) the photovoltaic generation short term power forecast model based on support vector machine, is set up, sample data selects prediction day and similar day thereof, the environmental factor data of maximum temperature, minimum temperature and weather humidity and the generated output data of similar day are comprised with what predict day, as the input quantity of forecast model, to predict that the photovoltaic generation power of day each future position exports the output data of data as forecast model, the parameter of genetic algorithm to forecast model is adopted to be optimized.

2. the short term power Forecasting Methodology of photovoltaic generating system according to claim 1, is characterized in that, in step (3), adopts the parameter of genetic algorithm to forecast model to be optimized as follows:

1) initial population scale and the iterations of genetic algorithm are set, and crossing-over rate and aberration rate;

2) initialization parameter value to be optimized, namely utilizes real coding mode to carry out initialization to regularization parameter C and nuclear parameter σ;

3) genetic iteration is carried out to population, and calculate the fitness size of each population;

4) using the parameter value after optimizing as the parameter value of support vector machine photovoltaic generation forecast model;

5) whether the output of computational prediction model and the error of actual output meet the demands, if met the demands, then perform step 6), if error does not meet the demands, then return step 3);

6) using the parameter value of the parameter value after this group optimization as forecast model, the power data of the weather data and similar day that input day to be predicted carries out short term power prediction.