CN103559561B - A kind of ultra-short term prediction method of photovoltaic plant irradiance - Google Patents

Abstract

The invention discloses a kind of ultra-short term prediction method of photovoltaic plant irradiance, comprise step: from historical data base, extract irradiance data, after rejecting night hours segment data, calculate corresponding extraterrestrial theoretical irradiance, carry out data exception inspection with this, and adopt the difference rule of three normalization data of extraterrestrial irradiance theoretical value and actual irradiance; Training sample set is extracted by the input and output dimension of model; Adopt ANFIS to carry out modeling to irradiance time series, wherein determine regular number and the initial parameter of ANFIS model with subtractive clustering, and adopt back-propagation algorithm and least square method Optimization of Fuzzy model parameter; Input prediction sample, calculates predicted value; Add predicted value and form new sample set, circulation realizes multi-step prediction; Predicted value is carried out renormalization process.The present invention only needs to utilize history irradiance time series to realize the ultra-short term prediction of irradiance, and precision of prediction is good, and easy to implement.

Description

Ultra-short-term prediction method for irradiance of photovoltaic power station

Technical Field

The invention relates to an ultra-short-term prediction method of irradiance of a photovoltaic power station.

Background

Large-scale photovoltaic grid-connected power generation is an effective way of utilizing solar energy, but the solar energy is intermittent and fluctuating energy, and large-scale access can bring serious challenges to the safe and stable operation of a power system and the guarantee of electric energy quality. If the power generation power of the photovoltaic power station can be accurately predicted, a power grid dispatching department can be helped to make a reasonable operation mode in time and accurately adjust a dispatching plan, the influence of grid connection of the photovoltaic power station on the whole power grid can be effectively reduced, and the safety and stability of a power system are guaranteed.

The meteorological factors influencing the power generation power of the photovoltaic power station are various, such as irradiation intensity, ambient temperature, humidity and the like. Wherein the influence of irradiance is greatest. The irradiance of the earth surface is influenced by various meteorological factors such as cloud cover, temperature, air pressure and the like, so that the randomness is high, and the difficulty in realizing accurate prediction is high. For ultra-short term prediction of ground irradiance, the following methods are generally used: (1) autoregressive moving average (ARMA) model. The method utilizes historical time series data of irradiance, determines a mathematical model capable of describing the irradiance time series through model identification, parameter estimation and model inspection, and further achieves the purpose of prediction, but because ARMA is a linear model, the prediction precision is limited; (2) a prediction method using sky cloud picture images is provided. The method generally requires a sky imager to be installed on the photovoltaic power station site, and is expensive in cost and complex in prediction algorithm, so that the implementation is difficult. Therefore, how to realize the ultra-short-term prediction of the irradiance of the photovoltaic power station with high precision becomes a problem which is addressed by the applicant.

Disclosure of Invention

The invention aims to provide an ultra-short-term prediction method of irradiance of a photovoltaic power station, which can realize ultra-short-term prediction of irradiance by only using a historical irradiance time sequence, has good prediction precision and easy implementation, and lays a foundation for realizing ultra-short-term prediction of generating power of the photovoltaic power station.

The technical scheme for realizing the purpose is as follows:

an ultra-short-term prediction method for irradiance of a photovoltaic power station comprises the following steps:

step S1, extracting actual irradiance data of the photovoltaic power station in a given time period from the historical database, eliminating invalid irradiation data in night time periods, performing abnormal inspection on the extracted actual irradiance data according to theoretical extraterrestrial irradiance in the same time and region, and performing normalization processing on the irradiance data after the abnormal inspection is completed;

step S2, determining a training sample set according to the input and output dimensions of the prediction model;

step S3, modeling an irradiance time sequence by using ANFIS, inputting a model with 5-dimension and 1-dimension output, determining the rule number and initial parameters of the ANFIS by using subtractive clustering, and optimizing fuzzy model parameters by using a back propagation algorithm and a least square method;

step S4, inputting a prediction sample, and calculating to obtain a prediction value;

step S5, judging whether the multi-step prediction is finished, if yes, entering step S7; if not, go to step S6;

step S6, adding the predicted value to form a new sample set, and returning to step S2;

in step S7, the predicted value is subjected to inverse normalization processing.

The ultra-short term prediction method for irradiance of the photovoltaic power station, wherein the step S1 includes:

step S11, extracting an irradiance time sequence value v (t) of a photovoltaic power station location h days before the day to be predicted from a historical database as an original sample set, wherein h is a positive integer;

step S12, calculating theoretical extraterrestrial irradiance of the same time and region;

step S13, according to the formula: v (t) is more than or equal to 0 and less than or equal to E (t), whether each irradiance time sequence value v (t) is normal or not is checked, and E (t) is the theoretical extraterrestrial irradiance of the area at the time t; if abnormal irradiance data is found, the process goes to step S14; if not, go to step S15;

step S14, rejecting all irradiance data of a data day containing abnormal irradiance data;

step S15, according to the formula:

$x\left(t\right)=\left\{\begin{array}{cc}\frac{E\left(t\right)-v\left(t\right)}{E\left(t\right)}& (E\left(t\right)\≠0)\\ 0& (E\left(t\right)=0)\end{array}\right.$

normalization is carried out, wherein x (t) is normalized data, and x (t) is more than or equal to 0 and less than or equal to 1.

The method for ultra-short-term prediction of irradiance of the photovoltaic power station includes step S12, which uses the following formula to calculate the theoretical extraterrestrial irradiance of the same region at the same time:

wherein E is the extraterrestrial irradiance value; e_scIs the solar constant; (r)₀/r)²Is a day-to-earth distance correction factor; solar declination angle; e_tIs time difference, theta is day angle, N is integral day, △ N is correction value of integral day, Y is year, INT is integer part of number in brackets, tau is solar time angle of time point required by said point, S_dIs the location of the point;is the geographic latitude of the point; γ is the geographic longitude of the point; s is the number of hours in the point-local standard, and F is the number of minutes in the point-local standard.

The method for ultra-short-term prediction of irradiance of the photovoltaic power station, wherein the step S2 refers to:

time series { x) of normalized N +5 consecutive irradiances₁,x₂,x₃,...,x_N+5Decomposed into N +1 5-dimensional vectors V₁,...,V_N+1N is a positive integer, to obtain:

(x₁,x₂,x₃,x₄,x₅)=V₁

(x₂,x₃,x₄,x₅,x₆)=V₂

(x_N,x_N+1,x_N+2,x_N+3,x_N+4)=V_N

(x_N+1,x_N+2,x_N+3,x_N+4,x_N+5)=V_N+1

and further pairing the next moment radiance value of the last one-dimensional data of the previous N vectors with the vectors to form N training sample pairs: { (V)₁,x₆),(V₂,x₇),...,(V_N,x_N+5)}。

The ultra-short term prediction method for irradiance of the photovoltaic power station, wherein the step S3 includes:

step S31, modeling the irradiance time series using ANFIS to obtain:

$x\left(t\right)=\frac{\left(\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\right[\underset{j=1}{\overset{5}{\mathrm{\Σ}}}{\mathrm{\λ}}_j^{i}x(t-j)+{\mathrm{\ξ}}_i]\·\exp [-\left(\underset{j=1}{\overset{5}{\mathrm{\Σ}}}\frac{{(x(t-j)-c_{\mathrm{ij}})}^2}{2{\mathrm{\σ}}_{\mathrm{ij}}^2}\right)\left]\right)}{\left(\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\exp \right[-\left(\underset{j=1}{\overset{5}{\mathrm{\Σ}}}\frac{{(x(t-j)-c_{\mathrm{ij}})}^2}{2{\mathrm{\σ}}_{\mathrm{ij}}^2}\right)\left]\right)}$

wherein i =1,2, ·, n; j =1,2,. 5; x (t-j) is an input quantity; x (t) is the output quantity;ξ_in is the number of rules for the back-end parameter; c. C_ij，σ_ijIs frontPiece parameters;

step S32, carrying out subtractive clustering analysis on the training sample set, specifically:

according to the formula:calculating to obtain data point density D_p，

Wherein p, q = (1, 2.. multidot., m), Y is a sample pair, m is the number of the sample pair,_ais the effective neighborhood radius of the cluster center;

selecting the highest value of the density indexObtaining a first cluster centerReconstructing the density function:

${D^{\′}}_p=D_p-D_{c_1}\exp [-\frac{{\left|\right|Y_p-y_{c_1}\left|\right|}^2}{{({\mathrm{\δ}}_b/2)}^2}]$

wherein,_b=1.25_acalculating the density index of all data points by using the new density function, and determining the next clustering centerA new density function is constructed again, and the process is repeated until the density function is satisfiedThe highest value of the density index of the p-th clustering center is obtained;

thereby obtaining the optimal fuzzy rule number n and the initial model antecedent parameter c_ijAnd σ_ij；

Step S33, model parameters are optimized by adopting a hybrid learning method, namely, parameters are identified by adopting a least square method for the back part, and parameters are optimized by adopting a back propagation algorithm for the front part:

identifying back-part parameters by using least square methodAnd ξ_iConverting the formula obtained in the step S31 into a matrix of X = phi · theta, phi is m × 2n, and theta is a back piece parameter vector of 2n × 1, wherein X is an output vector of m × 1;

let the error index function beTo expect an output, according to the least squares principle, to minimize J (θ), it is necessary to have:thereby obtaining optimized model back-part parametersAnd ξ_i；

Fix back part parametersAnd ξ_iAdjusting the front-part parameter c by using a back propagation algorithm_ijAnd σ_ijThe correction algorithm is as follows:

$\left\{\begin{array}{c}{c}_{\mathrm{ij}}(r+1)=c_{\mathrm{ij}}\left(r\right)-{\mathrm{\α}}_c\frac{\∂E}{{\∂c}_{\mathrm{ij}}}\\ {\mathrm{\σ}}_{\mathrm{ij}}(r+1)={\mathrm{\σ}}_{\mathrm{ij}}\left(r\right)-{\mathrm{\α}}_{\mathrm{\σ}}\frac{\∂E}{{\∂\mathrm{\σ}}_{\mathrm{ij}}}\end{array}\right.$

wherein, α_cAnd α_σIs the learning rate; c. C_ij(r+1)、σ_ij(r+1)、c_ij(r)、σ_ijAnd (r) respectively representing the central parameter and the width parameter of the membership function of the front part in the step r +1 and the step r in the correction algorithm.

The ultra-short-term prediction method of irradiance of the photovoltaic power station is characterized in that,

said $\left\{\begin{array}{c}{c}_{\mathrm{ij}}(r+1)=c_{\mathrm{ij}}\left(r\right)-{\mathrm{\α}}_c\frac{\∂E}{{\∂c}_{\mathrm{ij}}}\\ {\mathrm{\σ}}_{\mathrm{ij}}(r+1)={\mathrm{\σ}}_{\mathrm{ij}}\left(r\right)-{\mathrm{\α}}_{\mathrm{\σ}}\frac{\∂E}{{\∂\mathrm{\σ}}_{\mathrm{ij}}}\end{array}\right.$ In (1),

learning rate α_cAnd α_σThe initial value of the learning method is 0.01, the training cycle number of the mixed learning is 25, and r is more than or equal to 0 and less than or equal to 25.

The method for predicting irradiance of a photovoltaic power station in an ultra-short period includes, in step S4:

judging the number n of the obtained optimal fuzzy rules, and if the number of the rules is one, adopting a continuous method for prediction;

if the number of rules is more than one, then the sample V will be predicted_N+1Inputting the data into a model to obtain a predicted value x_N+6: if x is not less than 0_N+6If the predicted value is less than or equal to 1, the predicted value is effective, otherwise, the continuous prediction method is still adopted.

The method for predicting irradiance of a photovoltaic power station in an ultra-short period includes, in step S7:

according to the formula: p (t) = e (t) = x (t) × e (t)) and (t)) the obtained predicted value is subjected to inverse normalization processing, wherein x (t) is the predicted value, p (t) is data after inverse normalization, and e (t) is a theoretical external illuminance value at the time t.

The invention has the beneficial effects that: the method utilizes historical irradiance data acquired on site by the photovoltaic power station, adopts a difference value proportion method of an extraterrestrial irradiance theoretical value and actual irradiance to normalize the original irradiance data, establishes a prediction model by an adaptive fuzzy neural reasoning system, realizes ultra-short-term prediction in a multi-step circulation mode, can effectively complete the ultra-short-term prediction of the irradiance, has good prediction precision and is easy to implement, thereby providing a basis and guarantee for accurately predicting the ultra-short-term power generation power of the photovoltaic power station.

Drawings

FIG. 1 is a flow chart of a method of ultra-short term prediction of irradiance of a photovoltaic power plant of the present invention;

FIG. 2 is a diagram of the ANFIS model architecture of the present invention;

FIG. 3 is a membership function for input variable 1;

FIG. 4 is a membership function for input variable 2;

FIG. 5 is a membership function of input variable 3;

FIG. 6 is a membership function of input variable 4;

FIG. 7 is a membership function for input variable 5;

FIG. 8 is a plot of predicted versus actual values;

fig. 9 is a prediction error curve.

Detailed Description

The invention will be further explained with reference to the drawings.

Referring to fig. 1, the ultra-short term prediction method of irradiance of a photovoltaic power station of the present invention includes the following steps:

step S1, extracting actual irradiance data of the location of the photovoltaic power station in a given time period from the historical database, carrying out abnormal inspection on the extracted actual irradiance data according to theoretical extraterrestrial irradiance of the same region at the same time, and carrying out normalization processing on the irradiance data after the abnormal inspection is finished; in the embodiment, a meteorological observation station of a photovoltaic power station is used for collecting field irradiance data, the collection frequency is 30 seconds, and then the average value of irradiance for 15 minutes is counted and stored in a historical database; step S1 specifically includes:

step S11, extracting an irradiance time sequence value v (t) (irradiance actual observation data at t moment) of a photovoltaic power station location h days before a day to be predicted from a historical database as an original sample set, wherein h is a positive integer, and in the embodiment, h is 30 days; because the irradiance data is zero at night, taking the average value of 48 15 minutes in the middle of each day, namely the irradiance data from 6 points earlier to 6 points later, forming a continuous time sequence;

step S12, on the premise of neglecting the minor influences of ground axis precession, polar motion and the like, calculating the theoretical extraterrestrial irradiance in the same time and region by using the following formula:

wherein E is the extraterrestrial irradiance value; e_scIs the solar constant, which takes the value of 1367W/m 2; (r)₀/r)²Is a day-to-earth distance correction factor; solar declination angle; e_tIs time difference, unit minute (min), theta is a day angle, N is a product day (the sequence number of the date in one year, the product day of 1 month and 1 day is 0, the product day of 12 months and 31 days in an average year is 364, and the product day of a leap year is 365), △ N is the correction value of the product day, Y is the year, INT is the integer part of the number in brackets, tau is the solar time angle of the moment sought by the point, and S is the time difference_dIs the location of the point;is the geographic latitude of the point; γ is the geographic longitude of the point; s is the hours of the local standard (i.e. Beijing time) at the point, and F is the local standard at the pointNumber of minutes on time. And inputting the longitude and latitude and the time of the area to obtain the theoretical extraterrestrial irradiance of the area at the moment.

Step S13, according to the formula: v (t) is more than or equal to 0 and less than or equal to E (t), whether each irradiance time sequence value v (t) is normal or not is checked, and E (t) is the theoretical extraterrestrial irradiance of the area at the time t; if abnormal irradiance data is found, the process goes to step S14; if not, go to step S15;

step S14, rejecting all irradiance data of a data day containing abnormal irradiance data;

step S15, according to the formula:

$x\left(t\right)=\left\{\begin{array}{cc}\frac{E\left(t\right)-v\left(t\right)}{E\left(t\right)}& (E\left(t\right)\≠0)\\ 0& (E\left(t\right)=0)\end{array}\right.$

and normalizing the irradiance data after abnormal treatment, wherein x (t) is normalized data, and x (t) is more than or equal to 0 and less than or equal to 1.

Step S2, determining a training sample set according to the input and output dimensions of the prediction model, namely:

time series { x) of normalized N +5 consecutive irradiances₁,x₂,x₃,...,x_N+5Decomposed into N +1 5-dimensional vectors V₁,...,V_N+1N is a positive integer, to obtain:

(x₁,x₂,x₃,x₄,x₅)=V₁

(x₂,x₃,x₄,x₅,x₆)=V₂

(x_N,x_N+1,x_N+2,x_N+3,x_N+4)=V_N

(x_N+1,x_N+2,x_N+3,x_N+4,x_N+5)=V_N+1

and further pairing the next moment radiance value of the last one-dimensional data of the previous N vectors with the vectors to form N training sample pairs: { (V)₁,x₆),(V₂,x₇),...,(V_N,x_N+5)}；V_N+1I.e. the prediction input of the model, x_N+5Next time value x of_N+6I.e. the value we need to predict.

Step S3, modeling an irradiance time sequence by using ANFIS, inputting a model with 5-dimension and 1-dimension output, determining the rule number and initial parameters of the ANFIS by using subtractive clustering, and optimizing fuzzy model parameters by using a back propagation algorithm and a least square method; step S3 includes:

step S31, modeling irradiance time series using ANFIS as follows:

constructing an ANFIS model framework of an irradiance time sequence:

$\begin{array}{cc}\mathrm{if}& x(t-1)\mathrm{is}{A}_1^i,x(t-2)\mathrm{is}{A}_2^i,x(t-3)\mathrm{is}{A}_3^i,x(t-4)\mathrm{is}{A}_4^i,x(t-5)\mathrm{is}{A}_5^i,\\ \mathrm{Then}& x_i\left(t\right)={\mathrm{\λ}}_1^{i}x(t-1)+{\mathrm{\λ}}_2^{i}x(t-2)+{\mathrm{\λ}}_3^{i}x(t-3)+{\mathrm{\λ}}_4^{i}x(t-4)+{\mathrm{\λ}}_5^{i}x(t-5)+{\mathrm{\ξ}}_i\end{array}$

wherein i =1,2, ·, n; j =1,2,. 5; x (t-1), x (t-2), x (t-3), x (t-4) and x (t-5) are input quantities,ξ_in is the number of rules for the back-end parameter;is a fuzzy set of input quantities x (t-j);

the fuzzy set is expressed by a Gaussian membership function:wherein,representing a degree of membership; front part parameter c_ijAnd σ_ijRespectively representing the center and the width of the membership function;

the fuzzy inference rule is obtained using the 5-layer ANFIS network architecture shown in fig. 2:

wherein,x (t) is the output quantity;

through substitution, obtain

$x\left(t\right)=\frac{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{x}_i\left(t\right)\underset{j=1}{\overset{5}{\mathrm{\Π}}}{\mathrm{\μ}}_{A_j^i}}{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\underset{j=1}{\overset{5}{\mathrm{\Π}}}{\mathrm{\μ}}_{A_j^i}}=\frac{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{x}_i\left(t\right)\exp (-\left(\underset{j=1}{\overset{5}{\mathrm{\Σ}}}\frac{{(x-(t-j)-c_{\mathrm{ij}})}^2}{2{\mathrm{\σ}}_{\mathrm{ij}}^2}\right))}{\left(\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\exp \right[-\left(\underset{j=1}{\overset{5}{\mathrm{\Σ}}}\frac{{(x(t-j)-c_{\mathrm{ij}})}^2}{2{\mathrm{\σ}}_{\mathrm{ij}}^2}\right)\left]\right)}$

Then x is divided according to fuzzy inference rule_i(t) substituting the above formula has the following expression:

$x\left(t\right)=\frac{\left(\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\right[\underset{j=1}{\overset{5}{\mathrm{\Σ}}}{\mathrm{\λ}}_j^{i}x(t-j)+{\mathrm{\ξ}}_i]\·\exp [-\left(\underset{j=1}{\overset{5}{\mathrm{\Σ}}}\frac{{(x(t-j)-c_{\mathrm{ij}})}^2}{2{\mathrm{\σ}}_{\mathrm{ij}}^2}\right)\left]\right)}{\left(\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\exp \right[-\left(\underset{j=1}{\overset{5}{\mathrm{\Σ}}}\frac{{(x(t-j)-c_{\mathrm{ij}})}^2}{2{\mathrm{\σ}}_{\mathrm{ij}}^2}\right)\left]\right)};$

in step S32, subtractive clustering is to use each data point as a possible cluster center and calculate the probability of using the point as a cluster center according to the data point density around each data point, which overcomes the disadvantage that the calculated amount of other clustering methods increases exponentially with the dimension of the problem. Carrying out subtraction clustering analysis on a training sample set, specifically comprising the following steps:

according to the formula:calculating to obtain data point density D_p，

Wherein p, q = (1, 2.. multidot., m), Y is a sample pair, m is the number of the sample pair,_athe radius of the effective neighborhood of the cluster center is a positive number, which is set to 0.5 in this embodiment;

selecting the highest value of the density indexObtaining a first cluster centerReconstructing the density function:

${D^{\′}}_p=D_p-D_{c_1}\exp [-\frac{{\left|\right|Y_p-y_{c_1}\left|\right|}^2}{{({\mathrm{\δ}}_b/2)}^2}]$

wherein,_b=1.25_acalculating the density index of all data points by using the new density function, and determining the next clustering centerA new density function is constructed again, and the process is repeated until the density function is satisfied The highest value of the density index of the p-th clustering center is obtained;

thereby obtaining the optimal fuzzy rule number n and the initial model antecedent parameter c_ijAnd σ_ij；

Step S33, identifying the parameters of the back piece by adopting a least square methodAnd ξ_iConverting the formula obtained in the step S31 into a matrix of X = phi theta, phi is m × 2n, and theta is a back piece parameter vector of 2n × 1, wherein X is an output vector of m × 1;

let the error index function beTo expect an output, according to the least squares principle, to minimize J (θ), it is necessary to have:thereby obtaining optimized model back-part parametersAnd ξ_i；

Fix back part parametersAnd ξ_iAdjusting the front-part parameter c by using a back propagation algorithm_ijAnd σ_ijTaking into account the error indicator function，x_i(t) is the current output at time t,is the desired output, the correction algorithm is:

$\left\{\begin{array}{c}{c}_{\mathrm{ij}}(r+1)=c_{\mathrm{ij}}\left(r\right)-{\mathrm{\α}}_c\frac{\∂E}{{\∂c}_{\mathrm{ij}}}\\ {\mathrm{\σ}}_{\mathrm{ij}}(r+1)={\mathrm{\σ}}_{\mathrm{ij}}\left(r\right)-{\mathrm{\α}}_{\mathrm{\σ}}\frac{\∂E}{{\∂\mathrm{\σ}}_{\mathrm{ij}}}\end{array}\right.$

wherein the learning rate α_cAnd α_σThe initial value of (2) is 0.01, and the training cycle number of the mixed learning is 25; c. C_ij(r+1)、σ_ij(r+1)、c_ij(r)、σ_ij(r) respectively representing the central parameter and the width parameter of the membership function of the front part in the (r + 1) th step and the r step in the correction algorithm, wherein r is more than or equal to 0 and less than or equal to 25.

Step S4, inputting a prediction sample, and calculating to obtain a prediction value; the method specifically comprises the following steps:

judging the number n of the obtained optimal fuzzy rules, if the number of the rules is one, adopting a continuous method for prediction, namely using the irradiance value at the current moment as a predicted value at the next moment;

if the number of rules is more than one, then the sample V will be predicted_N+1Inputting the data into a model to obtain a predicted value x_N+6: if x is not less than 0_N+6If the predicted value is less than or equal to 1, the predicted value is effective, otherwise, the continuous prediction method is still adopted. To this end, the single step prediction is completed.

Step S5, judging whether the multi-step prediction is finished, if yes, entering step S7; if not, go to step S6;

step S6, adding the predicted value to form a new sample set, and returning to step S2; repeating and circulating the steps to obtain a plurality of predicted values;

step S7, performing inverse normalization processing on the predicted values, which means:

according to the formula: p (t) = e (t) = x (t) × e (t) and performing inverse normalization processing on the obtained predicted value, and reducing the predicted value of the model to a true value; wherein x (t) is a predicted value, P (t) is data after inverse normalization, and E (t) is a theoretical extraterrestrial irradiance value at the time t.

The following is described in a specific case:

irradiance data of a certain photovoltaic power station in the Shanghai area is selected as an experimental verification object. The average value of irradiance of 15 minutes of 30 days from 18 days to 17 days of 5 months in 2013 is collected as an initial modeling data sample, 2880 in total, and 1440 in total are taken as effective data from 6 points earlier to 6 points later in each day. Calculating theoretical irradiance of the same time period according to the step S12, detecting that bad data exists in two days according to the step S13, and removing the data of the two days to obtain 1344 effective values of irradiance for 28 days. And taking the data of the previous 20 days as an original training sample set to establish an ANFIS model, predicting the irradiance value of 8 days later, predicting once per hour, and predicting 4 average values of 15 minutes each time.

The first 960 data are first normalized and then divided into 955 training sample pairs, each pair of 5-dimensional input vector and 1-dimensional output vector. And determining the number of the fuzzy rules as 5 by adopting subtractive clustering, and then obtaining the parameters of the front part and the rear part of the model by adopting a back propagation method and a least square method. Membership functions for the five inputs of the ANFIS model are shown in fig. 3 to 7. In the figure, in1ct 1-in 1ct4 represent four graphs of membership function of the first-dimension input variable; in2ct 1-in 2ct4 represent four membership function graphs of the second-dimension input variable; in3ct 1-in 3ct4 represent four membership function graphs of the third-dimensional input variable; in4ct 1-in 4ct4 represent four membership function graphs of the fourth-dimensional input variable; in5ct 1-in 5ct4 represent four graphs of membership function for the fifth dimension input variable.

The parameters of the fuzzy rule front part Gaussian membership function are shown in table 1, and the parameters of the back part linear function are shown in table 2:

TABLE 1

Fuzzy rule	Back-piece linearity parameter (λ)1，λ2，λ3，λ4，λ5，ξ）
		R1	[-0.082 -0.063 0.123 -0.169 1.244 -0.060]
R2	[-0.171 0.174 -0.291 0.797 0.588 -0.028]
		R3	[2.334 -0.783 -1.247 0.982 -0.504 0.165]
R4	[0.029 -0.089 0.155 0.026 0.974 -0.036]
		R5	[-0.076 0.277 -0.104 -0.055 1.104 -0.089]

TABLE 2

Month 5, No. 10, No. 6: the average value of the actual irradiance of 5 minutes before 00 (namely, 5 months and 9 days, between 17: 45 and 18: 00) is used as a prediction sample input model, the average value of the next 15 minutes (namely, 5 months and 10 days, between 6: 00 and 6: 15) can be obtained, the prediction value is added into the prediction sample, and the process is repeated for 4 times to obtain the prediction value of 1 hour in the future (namely, 5 months and 10 days, between 6: 00 and 7: 00). The prediction is executed once every hour, the irradiance prediction value of the future 8 days can be obtained after the cycle is 96 times, 384 points are total, the Root Mean Square Error (RMSE) is calculated to be 59.7, and the standard root mean square error (NRMSE) is 0.4259. Wherein, RMSE and NRMSE are respectively calculated according to the following formulas:

$\mathrm{RMSE}={\left[\frac{1}{N^{\′}}\underset{i=1}{\overset{N^{\′}}{\mathrm{\Σ}}}{({\hat{P}}_i-P_i)}^2\right]}^{1/2}$

$\mathrm{NRMSE}=\frac{{\left(\frac{1}{N^{\′}}\underset{i=1}{\overset{N^{\′}}{\mathrm{\Σ}}}{({\hat{P}}_i-P_i)}^2\right)}^{1/2}}{\frac{1}{N^{\′}}\underset{i=1}{\overset{N^{\′}}{\mathrm{\Σ}}}{P}_i}$

in the formula, P_iIn order to predict the irradiance, a target irradiance is determined,and N' is the number of predicted points for the actually measured irradiance.

The comparison curve of the predicted value and the actual value is shown in fig. 8, and the prediction error curve is shown in fig. 9, so that the prediction model has good prediction accuracy.

The above embodiments are provided only for illustrating the present invention and not for limiting the present invention, and those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and therefore all equivalent technical solutions should also fall within the scope of the present invention, and should be defined by the claims.

Claims (7)

1. An ultra-short-term prediction method for irradiance of a photovoltaic power station is characterized by comprising the following steps:

step S1, extracting actual irradiance data of the photovoltaic power station in a given time period from the historical database, eliminating invalid irradiation data in night time periods, performing abnormal inspection on the extracted actual irradiance data according to theoretical extraterrestrial irradiance in the same time and region, and performing normalization processing on the irradiance data after the abnormal inspection is completed;

step S2, determining a training sample set according to the input and output dimensions of the prediction model;

step S3, modeling an irradiance time sequence by using ANFIS, inputting a model with 5-dimension and 1-dimension output, determining the rule number and initial parameters of the ANFIS by using subtractive clustering, and optimizing fuzzy model parameters by using a back propagation algorithm and a least square method;

step S4, inputting a prediction sample, and calculating to obtain a prediction value;

step S5, judging whether the multi-step prediction is finished, if yes, entering step S7; if not, go to step S6;

step S6, adding the predicted value to form a new sample set, and returning to step S2;

step S7, the predicted value is processed with inverse normalization,

the step S1 includes:

step S11, extracting an irradiance time sequence value v (t) of a photovoltaic power station location h days before the day to be predicted from a historical database as an original sample set, wherein h is a positive integer;

step S12, calculating theoretical extraterrestrial irradiance of the same time and region;

step S13, according to the formula: v (t) is more than or equal to 0 and less than or equal to E (t), whether each irradiance time sequence value v (t) is normal or not is checked, and E (t) is the theoretical extraterrestrial irradiance of the area at the time t; if abnormal irradiance data is found, the process goes to step S14; if not, go to step S15;

step S14, rejecting all irradiance data of a data day containing abnormal irradiance data;

step S15, according to the formula:

$x\left(t\right)=\left\{\begin{array}{cc}\frac{E\left(t\right)-v\left(t\right)}{E\left(t\right)}& \left(E\right(t)\≠0)\\ 0& \left(E\right(t)=0)\end{array}\right.$

normalization is carried out, wherein x (t) is normalized data, and x (t) is more than or equal to 0 and less than or equal to 1.

2. The ultra-short term prediction method of photovoltaic power station irradiance as claimed in claim 1, wherein the step S12 is implemented by calculating the theoretical extraterrestrial irradiance at the same time and in the same area according to the following formula:

wherein E is the extraterrestrial irradiance value; e_scIs the solar constant; (r)₀/r)²Is a day-to-earth distance correction factor; solar declination angle; e_tIs time difference, theta is day angle, N is integral day, △ N is correction value of integral day, Y is year, INT is integer part of number in brackets, tau is solar time angle of time point required by said point, S_dIs the location of the point;is the geographic latitude of the point; γ is the geographic longitude of the point; s is the number of hours in the point-local standard, and F is the number of minutes in the point-local standard.

3. The ultra-short term prediction method of photovoltaic power station irradiance as claimed in claim 1, wherein the step S2 refers to:

time series { x) of normalized N +5 consecutive irradiances₁,x₂,x₃,…,x_N+5Decomposed into N +1 5-dimensional vectors V₁,…,V_N+1N is a positive integer, to obtain:

(x₁,x₂,x₃,x₄,x₅)＝V₁

(x₂,x₃,x₄,x₅,x₆)＝V₂

……

(x_N,x_N+1,x_N+2,x_N+3,x_N+4)＝V_N

(x_N+1,x_N+2,x_N+3,x_N+4,x_N+5)＝V_N+1

and further pairing the next moment radiance value of the last one-dimensional data of the previous N vectors with the vectors to form N training sample pairs: { (V)₁,x₆),(V₂,x₇),…,(V_N,x_N+5)}。

4. The ultra-short term prediction method of photovoltaic power station irradiance as recited in claim 1, wherein the step S3 comprises:

step S31, modeling the irradiance time series using ANFIS to obtain:

$x\left(t\right)=\frac{(\underset{i=1}{\overset{n}{\Σ}}\[\underset{j=1}{\overset{5}{\Σ}}{\mathrm{\λ}}_j^{i}x(t-j)+{\mathrm{\ξ}}_i\]\·\exp \[-(\underset{j=1}{\overset{5}{\Σ}}\frac{{\left(x\right(t-j)-c_{ij})}^2}{2{\mathrm{\σ}}_{ij}^2})\])}{(\underset{i=1}{\overset{n}{\Σ}}\exp \[-(\underset{j=1}{\overset{5}{\Σ}}\frac{{\left(x\right(t-j)-c_{ij})}^2}{2{\mathrm{\σ}}_{ij}^2})\])}$

wherein i is 1,2, …, n; j ═ 1,2, …, 5; x (t-j) is an input quantity; x (t) is the output quantity;ξ_in is the number of rules for the back-end parameter; c. C_ij，σ_ijIs a front part parameter;

step S32, carrying out subtractive clustering analysis on the training sample set, specifically:

according to the formula:calculating to obtain data point density D_p，

Wherein p, q is (1,2, …, m), Y is a sample pair, m is the number of sample pairs,_ais the effective neighborhood radius of the cluster center;

selecting the highest value of the density indexObtaining a first cluster centerReconstructing the density function:

${D^{\′}}_p=D_p-D_{c_1}\exp \[-\frac{\left|\right|Y_p-y_{c_1}|{|}^2}{{({\mathrm{\δ}}_b/2)}^2}\]$

wherein,_b＝1.25_acalculating the density index of all data points by using the new density function, and determining the next clustering centerA new density function is constructed again, and the process is repeated until the density function is satisfied The highest value of the density index of the p-th clustering center is obtained;

thereby obtaining the optimal fuzzy rule number n and the initial model antecedent parameter c_ijAnd σ_ij；

Step S33, model parameters are optimized by adopting a hybrid learning method, namely, parameters are identified by adopting a least square method for the back part, and parameters are optimized by adopting a back propagation algorithm for the front part:

identifying back-part parameters by using least square methodAnd ξ_iConverting the formula obtained in the step S31 into a matrix with X being phi & theta, phi being m × 2n and theta being 2n × 1 rear piece parameter vector, and X being an output vector of m × 1;

let the error index function be To expect an output, according to the least squares principle, to minimize J (θ), it is necessary to have:thereby obtaining optimized model back-part parametersAnd ξ_i；

Fix back part parametersAnd ξ_iAdjusting the front-part parameter c by using a back propagation algorithm_ijAnd σ_ijThe correction algorithm is as follows:

$\left\{\begin{array}{c}{c}_{ij}(r+1)=c_{ij}\left(r\right)-{\mathrm{\α}}_c\frac{\∂E}{\∂c_{ij}}\\ {\mathrm{\σ}}_{ij}(r+1)={\mathrm{\σ}}_{ij}\left(r\right)-{\mathrm{\α}}_{\mathrm{\σ}}\frac{\∂E}{\∂{\mathrm{\σ}}_{ij}}\end{array}\right.$

wherein, α_cAnd α_σIs the learning rate; c. C_ij(r+1)、σ_ij(r+1)、c_ij(r)、σ_ijAnd (r) respectively representing the central parameter and the width parameter of the membership function of the front part in the step r +1 and the step r in the correction algorithm.

5. The ultra-short term prediction method of photovoltaic power station irradiance as recited in claim 4,

said $\left\{\begin{array}{c}{c}_{ij}(r+1)=c_{ij}\left(r\right)-{\mathrm{\α}}_c\frac{\∂E}{\∂c_{ij}}\\ {\mathrm{\σ}}_{ij}(r+1)={\mathrm{\σ}}_{ij}\left(r\right)-{\mathrm{\α}}_{\mathrm{\σ}}\frac{\∂E}{\∂{\mathrm{\σ}}_{ij}}\end{array}\right.$ In (1),

learning rate α_cAnd α_σThe initial value of the learning method is 0.01, the training cycle number of the mixed learning is 25, and r is more than or equal to 0 and less than or equal to 25.

6. The ultra-short term prediction method of photovoltaic power station irradiance as claimed in claim 1, wherein the step S4 specifically refers to:

judging the number n of the obtained optimal fuzzy rules, and if the number of the rules is one, adopting a continuous method for prediction;

if the number of rules is more than one, then the sample V will be predicted_N+1Inputting the data into a model to obtain a predicted value x_N+6: if x is not less than 0_N+6If the predicted value is less than or equal to 1, the predicted value is effective, otherwise, the continuous prediction method is still adopted.

7. The ultra-short term prediction method of photovoltaic power station irradiance as claimed in claim 1, wherein the step S7 specifically refers to:

according to the formula: and (p), (t) -x (t) e (t) and (t) performing inverse normalization processing on the obtained predicted value, wherein x (t) is the predicted value, p (t) is data after inverse normalization, and e (t) is a theoretical external illumination value at the time t.