CN112561178A - Distributed photovoltaic output prediction method considering microclimate factors - Google Patents

Abstract

The invention discloses a distributed photovoltaic output prediction method considering microclimate factors, and belongs to the technical field of renewable energy prediction. The method comprises the steps of establishing an SSA model, considering a microclimate prediction process of similar days, performing singular spectrum analysis-fuzzy information granulation, and establishing a prediction model based on an optimization extreme learning machine. According to the method, micrometeorological factors are considered on the basis of the traditional SSA method, and the model formed by the method has the characteristic of high prediction precision through modeling, prediction and optimization of ICS-ELM; the extreme learning machines are adopted for all the components to respectively establish a prediction model, and then the extreme learning machines are optimized through improving the cuckoo algorithm, so that the randomness of ELM parameter selection can be reduced, the prediction precision is improved, and the problem of the photovoltaic output characteristic reflected by a subsequence after data decomposition neglected by the conventional distributed photovoltaic power station prediction method is finally solved, thereby mining the implicit information and the internal rule of the photovoltaic output and achieving a good prediction effect.

Description

Distributed photovoltaic output prediction method considering microclimate factors

Technical Field

The invention belongs to the technical field of photovoltaic power generation, and particularly relates to a distributed photovoltaic output prediction method considering microclimate factors.

Background

In recent years, distributed photovoltaic power generation has great advantages in the aspects of energy conservation, emission reduction, flexible arrangement and the like, but the influence of intermittence and fluctuation on a power distribution network cannot be ignored. The problems of poor electric energy quality of a power distribution network and the like caused by the fact that a large number of intermittent photovoltaic power generation networks are connected are solved, the power supply safety and reliability of the whole power distribution network are affected, and the prediction of distributed photovoltaic output is realized on the premise of effective regulation and control of the power distribution network.

At present, three types of methods, such as a physical method, a statistical method, a combination method of the methods and the like, are mainly used for photovoltaic output prediction.

The physical method is that a physical model is established according to factors such as the detailed geographic position of the photovoltaic module, the photoelectric conversion efficiency and the like, and meteorological data is directly used as input for prediction according to the power generation principle of the photovoltaic system. The effectiveness depends on the degree of mastering the internal structure and the following rule of the research object and the precision of the model parameters, and the method has the advantages of multiple involved links, complex process and difficult parameter solving.

The statistical method is established by analyzing historical photovoltaic output data by using a certain statistical method, searching for an internal rule in the data and using the internal rule for prediction, and mainly comprises a time series method, a regression analysis method, a grey prediction method, a meta-heuristic series method and the like.

The nature of the meta-heuristic method with a good development prospect is to simulate the work and rest rules of the living beings, and to train sample data by adopting a certain algorithm to obtain the relation between the prediction condition and the quantity to be predicted. The meta-heuristic method mainly comprises a neural network, a support vector machine, a genetic algorithm, a fuzzy system and the like. The neural network method has strong nonlinear fitting capacity, can map any complex nonlinear relation, is very similar to the characteristics of a photovoltaic power generation system, and is very suitable for short-term prediction of the output of a photovoltaic power station. However, a single neural network cannot adapt to various weather types, the generic prediction effect is not good, and particularly, a gradient descent method is adopted in the traditional BP neural network training, so that the problem of easy falling into a local minimum value and slow convergence speed is solved. When the fuzzy system predicts the photovoltaic output, a large amount of historical data and sufficient expert experience are needed for establishing the fuzzy inference rule.

The combination method utilizes the information provided by different models and exerts respective advantages, and selects a proper mode for combination so as to improve the prediction effect. However, compared with the first two methods, the combined method modeling is more complicated than the single method, and the realization process is more difficult.

In summary, the following problems mainly exist in the existing prediction method of the distributed photovoltaic power station:

1. the modeling process is complex;

2. the influence factors are various and are coupled with each other;

3. a large amount of historical sample data is required;

4. the anti-interference capability is poor;

5. the prediction precision is low;

the common characteristics of the above methods are found when various problems are studied: when the photovoltaic output is predicted, the method utilizes modeling prediction after historical data processing, photovoltaic output characteristics reflected by a subsequence after data decomposition are not considered, some implicit information and internal rules of the photovoltaic output cannot be excavated, and therefore a better prediction effect is difficult to achieve.

Disclosure of Invention

The invention aims to provide a distributed photovoltaic output prediction method considering microclimate factors so as to solve the problems in the background art.

In order to solve the problems, the technical scheme of the invention is as follows:

a distributed photovoltaic output prediction method considering micrometeorological factors comprises the following steps:

step A, establishing an SSA model;

trend sequence, oscillation sequence and noise sequence predicted values can be obtained through SSA decomposition, trend/oscillation sequence prediction models are respectively established, after the predicted values of the trend/oscillation sequences are obtained, superposition is carried out according to the formula (1), and the predicted values of the photovoltaic output trend/oscillation sequences can be obtained;

P＝P_low+P_high (1)

in the formula (1), P_low、P_highRespectively predicting values of a trend sequence and an oscillation sequence;

p represents a predicted value of the photovoltaic output trend/oscillation sequence;

step B, considering a similar day selection process of microclimate, wherein the selection process comprises the following steps;

link (1), similar day based on little meteorological information is solved:

solving similar days by adopting a grey correlation theory based on the prediction type and the microclimate information;

and (2) performing correlation analysis on the photovoltaic output time sequence and different microclimate factors:

determining main microclimate factors influencing photovoltaic output according to a curve with the highest correlation degree by adopting a correlation analysis method, specifically determining 4 different main microclimate factors of temperature, irradiation, wind speed and rainfall;

and (3) determining microclimate sensitivity:

analyzing the sensitivity of the main microclimate factors determined in the step (2) to the photovoltaic output change, and determining the sensitivity corresponding to the microclimate factor value interval with the unit length as microclimate sensitivity;

step C, singular spectrum analysis-fuzzy information granulation, wherein the process comprises the following steps;

link (1), singular spectrum analysis:

the SSA is mainly divided into 4 steps of embedding, singular value decomposition, grouping and diagonal averaging and is used for identifying and extracting principal components of data;

and (2) fuzzy information granulation:

an information granulation method based on a fuzzy set is selected, wherein fuzzy information granulation comprises two modules: window division and information fuzzification;

step D, establishing a prediction model based on the optimized extreme learning machine, wherein the process can be realized by adopting the following modes:

(1) an extreme learning machine;

the extreme learning machine is a single hidden layer feedforward neural network, the deviation and the output weight of a hidden layer do not need to be adjusted, the output weight can be simply calculated after the input weight and the hidden layer deviation are randomly set, and the generalized inverse operation of a hidden layer output matrix is applied;

(2) improving a cuckoo algorithm to optimize an extreme learning machine;

in order to reduce the randomness of ELM parameter selection in the link (1) and improve the prediction precision, ICS is used for optimizing the weight and hidden layer bias between an input layer and a hidden layer of an ELM model.

Further, the process of correcting the oscillation sequence in step B is divided into the following steps:

link (1), similar day based on little meteorological information is solved:

selecting a reference day and a reference value of the oscillation sequence according to a similar day theory, taking a historical day which is closest to the day to be predicted and has similar weather types as an oscillation sequence reference day, and taking the photovoltaic output oscillation sequence of the reference day as the reference value of the oscillation sequence of the day to be predicted;

and (2) performing correlation analysis on the photovoltaic output time sequence and different microclimate factors:

according to the curve characteristic with the highest correlation degree, the determined main microclimate factors are temperature and irradiation;

by a coefficient of correlation alpha between temperature and irradiance and photovoltaic output₁And alpha₂Respectively serving as weight coefficients of main microclimate factors influencing photovoltaic output change;

link (3) determining microclimate sensitivity and applying to oscillation sequence P_highAnd (5) correcting:

according to the sensitivity of the microclimate factors to the photovoltaic output change, the temperature difference and the irradiation difference between the day to be predicted and the reference day, the photovoltaic output oscillation sequence P is subjected to the equation (4)_highAnd (5) correcting:

P_high＝P’_high+α₁ΔP₁+α₂ΔP₂ (4)

in the formula (4), P_high、P’_highThe photovoltaic output oscillation sequence of the day to be predicted and the photovoltaic output oscillation sequence of the reference day are obtained;

ΔP₁and Δ P₂The variation of the photovoltaic output oscillation sequence caused by temperature change and the variation of the photovoltaic output oscillation sequence caused by irradiation change are obtained;

α₁and alpha₂Respectively weighting coefficients of the change of the photovoltaic output oscillation sequence influenced by temperature and irradiation;

trend sequence P of photovoltaic output_lowAnd the above P_highThe same applies to the correction.

Further, in formula (4) described in the link (3), the values of the correction amounts including Δ P1 and Δ P2 refer to the following rules:

(1) the value law of the delta P1 is as follows:

a. when the day temperature to be predicted and the reference day temperature are in the same temperature interval (within the set temperature interval range):

ΔP₁＝S_t(t-t’) (5)

b. when the daily temperature to be predicted and the reference daily temperature are in 2 different temperature intervals, for example, 2 adjacent intervals, then:

in formulae (5) and (6):

t and t' are respectively the daily temperature to be predicted and the reference daily temperature value;

S_tand S'_tThe sensitivities corresponding to the respective intervals of the daily temperature to be predicted and the reference daily temperature;

temperature values for 2 interval common endpoints;

(2) Δ P2 takes the following values:

a. when the irradiation of the day to be predicted and the irradiation of the reference day are in the same irradiation interval (within the range of the set irradiation interval):

ΔP₂＝S_l(l-l′) (7)

b. when the daily irradiation to be predicted and the reference daily irradiation are in 2 different irradiation intervals, for example, 2 adjacent intervals are taken as examples, then:

in formulae (7) and (8):

l and l' are respectively a daily irradiation value to be predicted and a reference daily irradiation value;

S_land S'_lRespectively corresponding sensitivities of the regions of the daily irradiation to be predicted and the reference daily irradiation;

irradiation values for the common end point of the 2 intervals.

Further, in the step (1) of the step B, the correlation analysis of the grey correlation theory is to accurately express the difference of the geometric shapes between the curves by using a numerical value;

for a photovoltaic sequence x₀There are typically n comparison series x associated with it_iI-1, 2, …, n, which can be a variety of factors that influence prediction, then:

in the formula (2), xi_i(k) Is a sequence x₀And x_iGrey correlation coefficient at point k;

minimum difference of order 2;

is the 2-step maximum difference;

rho is the resolution subordination, and is generally 0.5;

synthesizing the correlation coefficients of all points to obtain all curves x₀And x_iDegree of association of R_iIs the average of N correlation coefficients, and represents curve x_iFor reference curve x₀The formula of the correlation degree of (c) is:

according to the formula (2), calculating the correlation degree of the photovoltaic output and each microclimate factor curve in a period of time, analyzing to obtain the correlation degree of the photovoltaic output curve and other comparison curves, sequencing the elements, and taking the curve with the highest correlation degree as a main microclimate factor influencing the magnitude of the photovoltaic output;

then, the process proceeds to step B, link (2).

Further, in the element (1) of the step C, the SSA is configured to use (Y) as a one-dimensional time series Y₁,y₂,…,y_N) Decomposed into L (window length) dimensional vectors according to a given nested spatial dimension: x_i＝(y_i,y_i+1,…,y_i+L-1)；

Wherein the L-dimensional vector includes trend, oscillation, and noise;

from K vectors X_iThe trajectory matrix (i ═ 1,2, …, K ═ N-L +1) can be expressed as follows:

considering the covariance S XX^TThe characteristic value is lambda (lambda)₁,λ₂,…,λ_L) And a feature vector U corresponding to the feature value₁,U₂,…,U_L(ii) a Singular value decomposition of X can be expressed as:

X＝E₁+E₂+…+E_i (10)

in formula (10):

is a fundamental matrix;

vector V₁,V₂,…,V_dIs a main component;

collection

The feature triplet is the ith feature triplet after the singular value decomposition of X;

suppose Z is a matrix of L K size, where the number in the matrix is Z_ijWherein i is more than or equal to 1 and less than or equal to L, j is more than or equal to 1 and less than or equal to K, and L is^*＝min(L,K),K^*Max (L, K), N ═ K + L-1, if L < K, order

Otherwise

The reconstructed time series Z ═ Z₁,z₂,…,z_NThe method comprises the following steps:

original sequence Y₀Sum of 2 sequences Y that can be resolved into SSA₀＝Y₁+Y₂Finally, the reconstruction components with larger singular values are selected for addition, and the oscillation rate Y is filtered₂Noise part, resulting Y₀Is approximately equal to Y₁。

Further, in the step (2), the fuzzy information granulation is performed by taking a given photovoltaic sequence X as a window to perform fuzzification processing according to the Pedrycz method in consideration of the single window problem;

fuzzification is aimed at establishing fuzzy particles P on X, and P can describe fuzzy concept G, so the essence of fuzzification is to determine the membership function of fuzzy concept G, namely A ═ μ G, so fuzzy particles P can replace fuzzy concept G, so P ═ A (X);

the most adopted is the triangular form, and the membership function expression is as follows:

in the formula: x is a time domain variable; a. m and b are 3 parameters of the function and respectively correspond to the minimum value, the average value and the maximum value of the data after fuzzy granulation of each window;

the prediction model can be established by ELM for each sequence obtained by the SSA-FIG model.

Further, in step D, the mode of the extreme learning machine is specifically as follows:

assume a set of N-sampled numbers, e.g., (xi, yi), where i is 1,2, …, N, x_i＝[x_i1,x_i2,…,x_in]∈RⁿAnd y_i＝[y_i1,y_i2,…,y_im]∈R^m；

The ELM output with L hidden layer neurons can be expressed as the following formula (13), the number of ELM hidden layer nodes in all models is set to be 6, and the hidden layer function is Sigmoid:

in formula (13): a is_i＝[a_i1,a_i2,…,a_in]^TRepresenting an input weight;

β_i＝[β_i1,β_i2,…,β_im]^Tis the output weight;

b_iis a deviation; g is an activation function;

equation (13) can be simplified as:

Hβ＝T (14)

h is an ELM algorithm hidden layer output matrix, and the coefficient b of the ELM can be obtained by solving the least square solution of the following linear equation:

its special solution can be expressed as follows:

in the formula H⁺Is the moore penrose (moore penrose) generalized inverse of the hidden layer output matrix H.

Further, the ICS algorithm fitness function in step D is as follows:

in formula (19): p is a radical of_tRepresenting an actual value;

representing a predicted value;

n represents the number of training samples;

the individual with the smallest adaptation value is the global optimum value for conversion into the parameters of the ELM.

Further, the specific steps of ICS optimization ELM in step D are as follows:

step (1): determining the input and output of the ELM according to the preliminarily screened noise sequence;

step (2): initializing populations and parameters of ICS, including maximum number of iterations T_{max gen}The number n of bird nests and the probability P of finding bird eggs, and T is set_{max gen}＝200，n＝25，P is 0.25 as the optimal parameter;

randomly generating n bird nests according to the specified range

Each bird nest is a set of parameters that will optimize the ELM for training;

calculating the adaptive value of each bird nest according to the formula (13), and acquiring the position of the current optimal bird nest and recording as X_best；

And (3): preserving the last generation optimal bird nest position X_bestUpdating the population according to the Levy flight mode

Meanwhile, calculating an adaptive value of the updated bird nest, comparing the adaptive value with the previous generation adaptive value, and updating the position if the adaptive value is better;

and (4): randomly generating a number r belonging to (0,1), if r is larger than P, updating the position of the population by random step length, if r is smaller than P, entering a cross mechanism, updating the population according to a formula (13), and judging the optimal bird nest X_bestIf the adaptive value of the position of the bird nest is better than the previous generation, the optimal bird nest position is reserved if the adaptive value of the position of the bird nest is better than the previous generation, and all the updated bird nest positions are obtained

And (5): judging whether the termination condition is met; if the current iteration number is more than T_maxgenThen the algorithm terminates; otherwise, the step (3) is switched to carry out a new iteration.

The invention has the following beneficial effects:

(1) the invention provides a short-term photovoltaic prediction model of a singular spectrum analysis (SSA-MF) (singular spectral analysis of Micrometeorologic factors) embedded with micrometeorologic factors and an improved Cuckoo algorithm (ICS) optimization Extreme Learning Machine (ELM), which integrates the SSA, correlation analysis, sensitivity analysis and other technologies, establishes an SSA-FIG-ICS-ELM model, and can effectively improve the prediction precision of the model.

(2) The method specifically comprises the steps of decomposing a photovoltaic output time sequence into a trend sequence, an oscillation sequence and a noise sequence by adopting an SSA technology, then obtaining small-scale and high-resolution microclimate data of an area where the photovoltaic output time sequence is located through a statistical downscaling method by using large-scale and low-resolution meteorological data obtained through numerical meteorological forecasting, finally analyzing the sensitivity between the photovoltaic output time sequence and microclimate factors, respectively correcting the trend sequence and the oscillation sequence of a day to be predicted according to the result and a reference value of sensitivity analysis, then effectively mining noise components by utilizing fuzzy information granulation, extracting the minimum value, the average value and the maximum value of each window, optimizing the noise component parameters of a prediction model by adopting an improved Cuckoo algorithm (ICS), and finally superposing the correction result and the prediction result to obtain a photovoltaic output prediction result.

(3) The improved SSA method considers micrometeorological factors on the basis of the traditional SSA method, and can decompose the original sequence without any regularity into sequences with different components by extracting the main information of the original data, so that part of the sequences have certain regularity;

because the noise sequence is reconstructed by a sub-matrix with a small ratio of the characteristic values, the noise sequence is processed by utilizing a fuzzy information granulation technology, and the maximum value, the average value and the minimum value of each window are extracted; after the noise sequence is subjected to fuzzy granulation effective excavation, the model formed by the method has the characteristic of high prediction precision through modeling, prediction and optimization of ICS-ELM;

the extreme learning machines are adopted for all the components to respectively establish a prediction model, and then the extreme learning machines are optimized through improving the cuckoo algorithm, so that the randomness of ELM parameter selection can be reduced, the prediction precision is improved, and the problem of the photovoltaic output characteristic reflected by a subsequence after data decomposition neglected by the conventional distributed photovoltaic power station prediction method is finally solved, thereby mining the implicit information and the internal rule of the photovoltaic output and achieving a good prediction effect.

Drawings

FIG. 1 is a schematic diagram of the concept of an embodiment of the present invention;

fig. 2 is a schematic flow chart of similar day selection in the step B link (1) in the embodiment of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be described clearly and completely with reference to the accompanying drawings of the embodiments of the present invention, and it is obvious that the described embodiments are some, but not all embodiments of the present invention. All other embodiments, which can be obtained by a person skilled in the art without any inventive step based on the embodiments of the present invention, are within the scope of the present invention.

Thus, the following detailed description of the embodiments of the present invention, presented in the figures, is not intended to limit the scope of the invention, as claimed, but is merely representative of selected embodiments of the invention.

Examples

As shown in fig. 1-2, a distributed photovoltaic output prediction method considering microclimate factors includes the following steps:

step A, establishing an SSA model.

Trend sequence, oscillation sequence and noise sequence predicted values can be obtained through SSA decomposition, trend/oscillation sequence prediction models are respectively established, after the predicted values of the trend/oscillation sequences are obtained, superposition is carried out according to the formula (1), and the predicted values of the photovoltaic output trend/oscillation sequences can be obtained;

P＝P_low+P_high (1)

in the formula (1), P_low、P_highRespectively predicting values of a trend sequence and an oscillation sequence;

p represents the predicted value of the photovoltaic output trend/oscillation sequence.

The basic idea of this model is shown in fig. 1. The improved SSA method takes microclimate factors into consideration on the basis of the traditional SSA method.

The noise sequence is reconstructed by a submatrix with a small ratio of the characteristic values, then the noise sequence is processed by utilizing a fuzzy information granulation technology, the maximum value, the average value and the minimum value of each window are extracted, and each component adopts an extreme learning machine to respectively establish a prediction model.

And step B, considering a similar day selection process of microclimate, wherein the selection process comprises the following steps.

Link (1), similar day based on little meteorological information is solved:

wherein, the microclimate data is obtained by adopting Statistical Downscaling (SDSM).

The statistical downscaling method is based on 3 assumptions that (1) a large-scale waiting field and a regional climate element field have a significant statistical relationship; (2) the large-scale climate field can be well simulated by a forecasting mode; (3) the statistical relationships established are valid under varying climatic scenarios.

The principle is based on the idea that the regional climate change situation is conditioned by large-scale climate, namely, the output information of the large-scale and low-resolution numerical weather forecast is converted into regional-scale ground climate change information (such as air temperature and precipitation), so that the limitation of the weather forecast on regional climate prediction is made up.

And selecting similar days based on the prediction model and the microclimate information by adopting a grey correlation theory.

Specifically, the method comprises the following steps: the correlation analysis of grey correlation theory is to express the difference of geometric shapes between curves accurately by using numerical values.

For a photovoltaic sequence x₀There are typically n comparison series x associated with it_iI-1, 2, …, n, which can be a variety of factors that influence prediction, then:

in the formula (2), xi_i(k) Is a sequence x₀And x_iGrey correlation coefficient at point k;

minimum difference of order 2;

is the 2-step maximum difference;

rho is the resolution subordination, and is generally 0.5;

synthesizing the correlation coefficients of all points to obtain all curves x₀And x_iDegree of association of R_iIs the average of N correlation coefficients, and represents curve x_iFor reference curve x₀The formula of the correlation degree of (c) is:

according to the formula (2), calculating the correlation degree of the photovoltaic output and each microclimate factor curve in a period of time, analyzing to obtain the correlation degree of the photovoltaic output curve and other comparison curves, and sequencing the essences;

then, the next link is entered.

And (2) performing correlation analysis on the photovoltaic output time sequence and different microclimate factors:

and determining main microclimate factors influencing photovoltaic output according to the curve with the highest correlation degree by adopting a correlation analysis method, specifically determining 4 different main microclimate factors of temperature, irradiation, wind speed and rainfall.

The main microclimate factors determined, taking into account the characteristics of the oscillating sequence, are temperature and irradiation, according to the curve with the highest degree of correlation.

By a coefficient of correlation alpha between temperature and irradiance and photovoltaic output₁And alpha₂And the weight coefficients are respectively used as the weight coefficients of the main microclimate factors influencing the photovoltaic output change.

And (3) determining microclimate sensitivity:

analyzing the sensitivity of the main microclimate factors determined in the step (2) to the photovoltaic output change, and determining the sensitivity corresponding to the microclimate factor value interval with the unit length as microclimate sensitivity;

the sensitivity of the photovoltaic output to the microclimate factor refers to the variation of the photovoltaic output under the unit microclimate factor variation.

In the link (4), the modeling and prediction process of the oscillation sequence is described as an example below, considering that the steps for modeling and predicting the trend/oscillation sequence are the same.

a. And selecting a reference day and a reference value of the oscillation sequence, taking a historical day which is closest to the day to be predicted and has similar weather types as an oscillation sequence reference day, and taking the photovoltaic output oscillation sequence of the reference day as the reference value of the oscillation sequence of the day to be predicted.

b. By a coefficient of correlation alpha between temperature and irradiance and photovoltaic output₁And alpha₂And the weight coefficients are respectively used as the weight coefficients of the main microclimate factors influencing the photovoltaic output change.

c. According to the sensitivity of the microclimate factors to the photovoltaic output change, the temperature difference and the irradiation difference between the day to be predicted and the reference day, the photovoltaic output oscillation sequence P is subjected to the equation (4)_highAnd (5) correcting:

P_high＝P’_high+α₁ΔP₁+α₂ΔP₂ (4)

in the formula (4), P_high、P’_highThe photovoltaic output oscillation sequence of the day to be predicted and the photovoltaic output oscillation sequence of the reference day are obtained;

ΔP₁and Δ P₂The variation of the photovoltaic output oscillation sequence caused by temperature change and the variation of the photovoltaic output oscillation sequence caused by irradiation change are obtained;

α₁and alpha₂The weight coefficients of the change of the photovoltaic output oscillation sequence influenced by the temperature and the irradiation are respectively.

Trend sequence P of photovoltaic output_lowAnd the above P_highThe same applies to the correction.

As can be seen from equation (4), the corrected quantities include Δ P1 and Δ P2, which take on the following rules:

(1) the value law of the delta P1 is as follows:

a. when the day temperature to be predicted and the reference day temperature are in the same temperature interval (within the set temperature interval range):

ΔP₁＝S_t(t-t’) (5)

b. when the daily temperature to be predicted and the reference daily temperature are in 2 different temperature intervals, for example, 2 adjacent intervals, then:

in formulae (5) and (6):

t and t' are respectively the daily temperature to be predicted and the reference daily temperature value;

S_tand S'_tThe sensitivities corresponding to the respective intervals of the daily temperature to be predicted and the reference daily temperature;

the temperature values for the 2 intervals common endpoint.

(2) Δ P2 takes the following values:

a. when the irradiation of the day to be predicted and the irradiation of the reference day are in the same irradiation interval (within the range of the set irradiation interval):

ΔP₂＝S_l(l-l′) (7)

b. when the daily irradiation to be predicted and the reference daily irradiation are in 2 different irradiation intervals, for example, 2 adjacent intervals are taken as examples, then:

in formulae (7) and (8):

l and l' are respectively a daily irradiation value to be predicted and a reference daily irradiation value;

S_land S'_lRespectively corresponding sensitivities of the regions of the daily irradiation to be predicted and the reference daily irradiation;

irradiation values for the common end point of the 2 intervals.

Step C, singular spectrum analysis-fuzzy information granulation, wherein the process comprises the following steps;

link (1), singular spectrum analysis:

the SSA is mainly divided into 4 steps of embedding, singular value decomposition, grouping and diagonal averaging, and is used to identify and extract the principal component of data.

SSA is used to convert a one-dimensional time series Y ═ Y (Y)₁,y₂,…,y_N) Decomposed into L (window length) dimensional vectors according to a given nested spatial dimension: x_i＝(y_i,y_i+1,…,y_i+L-1)；

Where the L-dimensional vector includes trend, oscillation, and noise.

From K vectors X_iThe trajectory matrix (i ═ 1,2, …, K ═ N-L +1) can be expressed as follows:

considering the covariance S XX^TThe characteristic value is lambda (lambda)₁,λ₂,…,λ_L) And a feature vector U corresponding to the feature value₁,U₂,…,U_L(ii) a Singular value decomposition of X can be expressed as:

X＝E₁+E₂+…+E_i (10)

in formula (10):

is a fundamental matrix;

vector V₁,V₂,…,V_dIs a main component;

collection

The feature triplet is the ith feature triplet after the singular value decomposition of X;

suppose Z is a matrix of L K size, where the number in the matrix is Z_ijWherein i is more than or equal to 1 and less than or equal to L, j is more than or equal to 1 and less than or equal to K, and L is^*＝min(L,K),K^*Max (L, K), N ═ K + L-1, if L < K, order

Otherwise

The reconstructed time series Z ═ Z₁,z₂,…,z_NThe method comprises the following steps:

original sequence Y₀Sum of 2 sequences Y that can be resolved into SSA₀＝Y₁+Y₂Finally, the reconstruction components with larger singular values are selected for addition, and the oscillation rate Y is filtered₂Noise part, resulting Y₀Is approximately equal to Y₁。

And (2) fuzzy information granulation:

an information granulation method based on a fuzzy set is selected, wherein fuzzy information granulation comprises two modules: window division and information fuzzification;

fuzzy information granulation according to the Pedrycz method, given photovoltaic sequences X are treated as a window for fuzzification processing, considering the single window problem.

Fuzzification is aimed at establishing fuzzy particles P on X, and P can describe fuzzy concept G, so the essence of fuzzification is to determine the membership function of fuzzy concept G, namely A ═ μ G, so fuzzy particles P can replace fuzzy concept G, so P ═ A (X);

the most adopted is the triangular form, and the membership function expression is as follows:

in the formula: x is a time domain variable; a. m and b are 3 parameters of the function and respectively correspond to the minimum value, the average value and the maximum value of the data after fuzzy granulation of each window;

the prediction model can be established by ELM for each sequence obtained by the SSA-FIG model.

Step D, establishing a prediction model based on the optimized extreme learning machine, wherein the process can be realized by adopting the following modes:

(1) an extreme learning machine;

the extreme learning machine is a single hidden layer feedforward neural network, the deviation and the output weight of a hidden layer do not need to be adjusted, the output weight can be obtained by simple calculation after the input weight and the hidden layer deviation are randomly set, and the generalized inverse operation of a hidden layer output matrix is applied.

The method comprises the following specific steps:

assume a set of N-sampled numbers, e.g., (x)_i,y_i) Where i is 1,2, …, n, x_i＝[x_i1,x_i2,…,x_in]∈RⁿAnd y_i＝[y_i1,y_i2,…,y_im]∈R^m；

The ELM output with L hidden layer neurons can be expressed as the following formula (13), the number of ELM hidden layer nodes in all models is set to be 6, and the hidden layer function is Sigmoid:

in formula (13): a is_i＝[a_i1,a_i2,…,a_in]^TRepresenting an input weight;

β_i＝[β_i1,β_i2,…,β_im]^Tis the output weight;

b_iis a deviation; g is an activation function;

equation (13) can be simplified as:

Hβ＝T (14)

h is an ELM algorithm hidden layer output matrix, and the coefficient b of the ELM can be obtained by solving the least square solution of the following linear equation:

its special solution can be expressed as follows:

in the formula H⁺Is the moore penrose (moore penrose) generalized inverse of the hidden layer output matrix H.

(2) Improving a cuckoo algorithm to optimize an extreme learning machine;

in order to reduce the randomness of ELM parameter selection in the link (1) and improve the prediction precision, ICS is used for optimizing the weight and hidden layer bias between an input layer and a hidden layer of an ELM model.

The ICS algorithm fitness function is as follows:

in formula (19): p is a radical of_tRepresenting an actual value;

representing a predicted value;

n represents the number of training samples;

the individual with the smallest adaptation value is the global optimum value for conversion into the parameters of the ELM.

ICS optimization ELM comprises the following specific steps:

step (1): determining the input and output of the ELM according to the preliminarily screened noise sequence;

step (2): initializing populations and parameters of ICS, including maximum number of iterations T_{max gen}The number n of bird nests and the probability P of finding bird eggs, and T is set_{max gen}200, n is 25, and P is 0.25 as the optimal parameter;

randomly generating n bird nests according to the specified range

Each bird nest is a set of parameters that will optimize the ELM for training;

calculating the adaptive value of each bird nest according to the formula (13), and acquiring the position of the current optimal bird nest and recording as X_best；

And (3): preserving the last generation optimal bird nest position X_bestUpdating the population according to the Levy flight mode

Meanwhile, calculating an adaptive value of the updated bird nest, comparing the adaptive value with the previous generation adaptive value, and updating the position if the adaptive value is better;

and (4): randomly generating a number r belonging to (0,1), if r is larger than P, updating the position of the population by random step length, if r is smaller than P, entering a cross mechanism, updating the population according to a formula (13), and judging the optimal bird nest X_bestIf the adaptive value of the position of the bird nest is better than the previous generation, the optimal bird nest position is reserved if the adaptive value of the position of the bird nest is better than the previous generation, and all the updated bird nest positions are obtained

And (5): judging whether the termination condition is met; if the current iteration number is more than T_maxgenThen the algorithm terminates; otherwise, the step (3) is switched to carry out a new iteration.

Claims (9)

1. A distributed photovoltaic output prediction method considering micrometeorological factors is characterized by comprising the following steps:

the method comprises the following steps:

step A, establishing an SSA model;

trend sequence, oscillation sequence and noise sequence predicted values can be obtained through SSA decomposition, trend/oscillation sequence prediction models are respectively established, after the predicted values of the trend/oscillation sequences are obtained, superposition is carried out according to the formula (1), and the predicted values of the photovoltaic output trend/oscillation sequences can be obtained;

P＝P_low+P_high (1)

in the formula (1), P_low、P_highRespectively predicting values of a trend sequence and an oscillation sequence;

p represents a predicted value of the photovoltaic output trend/oscillation sequence;

step B, considering a similar day selection process of microclimate, wherein the selection process comprises the following steps;

link (1), similar day based on little meteorological information is solved:

solving similar days by adopting a grey correlation theory based on the prediction type and the microclimate information;

and (2) performing correlation analysis on the photovoltaic output time sequence and different microclimate factors:

determining main microclimate factors influencing photovoltaic output according to a curve with the highest correlation degree by adopting a correlation analysis method, specifically determining 4 different main microclimate factors of temperature, irradiation, wind speed and rainfall;

and (3) determining microclimate sensitivity:

analyzing the sensitivity of the main microclimate factors determined in the step (2) to the photovoltaic output change, and determining the sensitivity corresponding to the microclimate factor value interval with the unit length as microclimate sensitivity;

step C, singular spectrum analysis-fuzzy information granulation, wherein the process comprises the following steps;

link (1), singular spectrum analysis:

the SSA is mainly divided into 4 steps of embedding, singular value decomposition, grouping and diagonal averaging and is used for identifying and extracting principal components of data;

and (2) fuzzy information granulation:

an information granulation method based on a fuzzy set is selected, wherein fuzzy information granulation comprises two modules: window division and information fuzzification;

step D, establishing a prediction model based on the optimized extreme learning machine, wherein the process can be realized by adopting the following modes:

(1) an extreme learning machine;

the extreme learning machine is a single hidden layer feedforward neural network, the deviation and the output weight of a hidden layer do not need to be adjusted, the output weight can be simply calculated after the input weight and the hidden layer deviation are randomly set, and the generalized inverse operation of a hidden layer output matrix is applied;

(2) improving a cuckoo algorithm to optimize an extreme learning machine;

in order to reduce the randomness of ELM parameter selection in the link (1) and improve the prediction precision, ICS is used for optimizing the weight and hidden layer bias between an input layer and a hidden layer of an ELM model.

2. The distributed photovoltaic contribution prediction method considering microclimate factors according to claim 1, wherein: the correction process of the oscillation sequence in the step B is divided into the following steps: link (1), similar day based on little meteorological information is solved:

selecting a reference day and a reference value of the oscillation sequence according to a similar day theory, taking a historical day which is closest to the day to be predicted and has similar weather types as an oscillation sequence reference day, and taking the photovoltaic output oscillation sequence of the reference day as the reference value of the oscillation sequence of the day to be predicted;

and (2) performing correlation analysis on the photovoltaic output time sequence and different microclimate factors:

according to the curve characteristic with the highest correlation degree, the determined main microclimate factors are temperature and irradiation;

by a coefficient of correlation alpha between temperature and irradiance and photovoltaic output₁And alpha₂Respectively serving as weight coefficients of main microclimate factors influencing photovoltaic output change;

link (3) determining microclimate sensitivity and applying to oscillation sequence P_highAnd (5) correcting:

according to the sensitivity of the microclimate factors to the photovoltaic output change, the temperature difference and the irradiation difference between the day to be predicted and the reference day, the photovoltaic output oscillation sequence P is subjected to the equation (4)_highAnd (5) correcting:

P_high＝P'_high+α₁ΔP₁+α₂ΔP₂(4)

in the formula (4), P_high、P'_highThe photovoltaic output oscillation sequence of the day to be predicted and the photovoltaic output oscillation sequence of the reference day are obtained;

ΔP₁and Δ P₂The variation of the photovoltaic output oscillation sequence caused by temperature change and the variation of the photovoltaic output oscillation sequence caused by irradiation change are obtained;

α₁and alpha₂Respectively weighting coefficients of the change of the photovoltaic output oscillation sequence influenced by temperature and irradiation;

trend sequence P of photovoltaic output_lowAnd the above P_highThe same applies to the correction.

3. The distributed photovoltaic contribution prediction method considering microclimate factors according to claim 2, wherein: in formula (4) described in the link (3), the values of the correction amounts including Δ P1 and Δ P2 refer to the following rules:

(1) the value law of the delta P1 is as follows:

a. when the day temperature to be predicted and the reference day temperature are in the same temperature interval (within the set temperature interval range):

ΔP₁＝S_t(t-t') (5)

b. when the daily temperature to be predicted and the reference daily temperature are in 2 different temperature intervals, for example, 2 adjacent intervals, then:

in formulae (5) and (6):

t and t' are respectively the daily temperature to be predicted and the reference daily temperature value;

S_tand S'_tThe sensitivities corresponding to the respective intervals of the daily temperature to be predicted and the reference daily temperature;

is 2 intervalsA temperature value of the common endpoint;

(2) Δ P2 takes the following values:

a. when the irradiation of the day to be predicted and the irradiation of the reference day are in the same irradiation interval (within the range of the set irradiation interval):

ΔP₂＝S_l(l-l′)(7)

b. when the daily irradiation to be predicted and the reference daily irradiation are in 2 different irradiation intervals, for example, 2 adjacent intervals are taken as examples, then:

in formulae (7) and (8):

l and l' are respectively a daily irradiation value to be predicted and a reference daily irradiation value;

S_land S'_lRespectively corresponding sensitivities of the regions of the daily irradiation to be predicted and the reference daily irradiation;

irradiation values for the common end point of the 2 intervals.

4. The distributed photovoltaic contribution prediction method considering microclimate factors according to claim 1, wherein: in the step B, the correlation analysis of the grey correlation theory in the link (1) is to accurately express the difference of the geometric shapes among the curves by using numerical values;

for a photovoltaic sequence x₀There are typically n comparison series x associated with it_iI-1, 2, …, n, which can be a variety of factors that influence prediction, then:

in the formula (2), xi_i(k) Is a sequence x₀And x_iGray correlation at kA coefficient;

minimum difference of order 2;

is the 2-step maximum difference;

rho is the resolution subordination, and is generally 0.5;

synthesizing the correlation coefficients of all points to obtain all curves x₀And x_iDegree of association of R_iIs the average of N correlation coefficients, and represents curve x_iFor reference curve x₀The formula of the correlation degree of (c) is:

according to the formula (2), calculating the correlation degree of the photovoltaic output and each microclimate factor curve in a period of time, analyzing to obtain the correlation degree of the photovoltaic output curve and other comparison curves, sequencing the elements, and taking the curve with the highest correlation degree as a main microclimate factor influencing the magnitude of the photovoltaic output;

then, the process proceeds to step B, link (2).

5. The distributed photovoltaic contribution prediction method considering microclimate factors according to claim 3, wherein: in the step C, in the element (1), the SSA is configured to use (Y) as the one-dimensional time series Y₁,y₂,…,y_N) Decomposed into L (window length) dimensional vectors according to a given nested spatial dimension: x_i＝(y_i,y_i+1,…,y_i+L-1)；

Wherein the L-dimensional vector includes trend, oscillation, and noise;

from K vectors X_iThe trajectory matrix (i ═ 1,2, …, K ═ N-L +1) can be expressed as follows:

considering the covariance S XX^TThe characteristic value is lambda (lambda)₁,λ₂,…,λ_L) And a feature vector U corresponding to the feature value₁,U₂,…,U_L(ii) a Singular value decomposition of X can be expressed as:

X＝E₁+E₂+…+E_i (10)

in formula (10):

is a fundamental matrix;

vector V₁,V₂,…,V_dIs a main component;

collection

The feature triplet is the ith feature triplet after the singular value decomposition of X;

suppose Z is a matrix of L K size, where the number in the matrix is Z_ijWherein i is more than or equal to 1 and less than or equal to L, j is more than or equal to 1 and less than or equal to K, and L is^*＝min(L,K),K^*Max (L, K), N ═ K + L-1, if L < K, order

Otherwise

The reconstructed time series Z ═ Z₁,z₂,…,z_NThe method comprises the following steps:

original sequence Y₀Sum of 2 sequences Y that can be resolved into SSA₀＝Y₁+Y₂Finally, selecting the reconstruction component with larger singular valueAdding and filtering out the oscillation rate Y₂Noise part, resulting Y₀Is approximately equal to Y₁。

6. The distributed photovoltaic contribution prediction method considering microclimate factors according to claim 5, wherein: in the step C, the fuzzy information granulation is carried out by taking a given photovoltaic sequence X as a window to carry out fuzzification treatment according to a Pedracz method and considering the problem of a single window;

fuzzification is aimed at establishing fuzzy particles P on X, and P can describe fuzzy concept G, so the essence of fuzzification is to determine the membership function of fuzzy concept G, namely A ═ μ G, so fuzzy particles P can replace fuzzy concept G, so P ═ A (X);

the most adopted is the triangular form, and the membership function expression is as follows:

in the formula: x is a time domain variable; a. m and b are 3 parameters of the function and respectively correspond to the minimum value, the average value and the maximum value of the data after fuzzy granulation of each window;

the prediction model can be established by ELM for each sequence obtained by the SSA-FIG model.

7. The distributed photovoltaic contribution prediction method considering microclimate factors according to claim 1, wherein: in the step D, the mode of the extreme learning machine is specifically as follows:

assume a set of N-sampled numbers, e.g., (xi, yi), where i is 1,2, …, N, x_i＝[x_i1,x_i2,…,x_in]∈RⁿAnd y_i＝[y_i1,y_i2,…,y_im]∈R^m；

The ELM output with L hidden layer neurons can be expressed as the following formula (13), the number of ELM hidden layer nodes in all models is set to be 6, and the hidden layer function is Sigmoid:

in formula (13): a is_i＝[a_i1,a_i2,…,a_in]^TRepresenting an input weight;

β_i＝[β_i1,β_i2,…,β_im]^Tis the output weight;

b_iis a deviation; g is an activation function;

equation (13) can be simplified as:

Hβ＝T(14)

h is an ELM algorithm hidden layer output matrix, and the coefficient b of the ELM can be obtained by solving the least square solution of the following linear equation:

its special solution can be expressed as follows:

in the formula H⁺Is the moore penrose (moore penrose) generalized inverse of the hidden layer output matrix H.

8. The distributed photovoltaic contribution prediction method considering microclimate factors according to claim 1, wherein: the ICS algorithm fitness function in step D is as follows:

in formula (19): p is a radical of_tRepresenting an actual value;

representing a predicted value;

n represents the number of training samples;

the individual with the smallest adaptation value is the global optimum value for conversion into the parameters of the ELM.

9. The distributed photovoltaic contribution prediction method considering microclimate factors according to claim 8, wherein: the ICS optimization ELM in the step D specifically comprises the following steps: step (1): determining the input and output of the ELM according to the preliminarily screened noise sequence;

step (2): initializing populations and parameters of ICS, including maximum number of iterations T_maxgenThe number n of bird nests and the probability P of finding bird eggs, and T is set_maxgen200, n is 25, and P is 0.25 as the optimal parameter;

randomly generating n bird nests according to the specified range

Each bird nest is a set of parameters that will optimize the ELM for training;

calculating the adaptive value of each bird nest according to the formula (13), and acquiring the position of the current optimal bird nest and recording as X_best；

And (3): preserving the last generation optimal bird nest position X_bestUpdating the population according to the Levy flight mode

Meanwhile, calculating the adaptive value of the updated bird nest, comparing the adaptive value with the previous generation adaptive value, and if the adaptive value is not the same as the previous generation adaptive valueIf the position is more optimal, updating the position;

and (4): randomly generating a number r belonging to (0,1), if r is larger than P, updating the position of the population by random step length, if r is smaller than P, entering a cross mechanism, updating the population according to a formula (13), and judging the optimal bird nest X_bestIf the adaptive value of the position of the bird nest is better than the previous generation, the optimal bird nest position is reserved if the adaptive value of the position of the bird nest is better than the previous generation, and all the updated bird nest positions are obtained

And (5): judging whether the termination condition is met; if the current iteration number is more than T_maxgenThen the algorithm terminates; otherwise, the step (3) is switched to carry out a new iteration.

Images