A method for adaptive integrated prediction of photovoltaic power generation at the edge of distribution networks

Abstract

A novel adaptive integrated prediction method for photovoltaic power generation at the edge of a distribution network utilizes the Tradaboost algorithm as the primary integration framework and the extreme learning machine (ELM) as the foundational predictor during the initial model training phase. During the integration process, ELM predictors with low performance are removed and those with high performance are correspondingly improved, reducing the final integration scale and the computational overhead and storage resources required for subsequent predictions. During the day-ahead rolling prediction phase, an online sequential ELM algorithm is used to update the parameters of each foundational ELM predictor, combining actual daily photovoltaic output data. This solves the adaptive nature of the edge-side photovoltaic prediction model and provides an effective basis for generating distributed power generation control strategies for distribution networks.

Description

Technical Field

[0001] The present invention relates to a distributed photovoltaic prediction method at the edge of a distribution network, and more particularly to an adaptive integrated photovoltaic prediction method at the edge of a distribution network. Background Art

[0002] Photovoltaic power generation output is affected by multiple factors, including sunlight intensity, ambient temperature, and the conversion efficiency of photovoltaic modules. The integration of large-scale photovoltaics will exacerbate the randomness and uncertainty of the distribution network, and its impact on the safe operation of the distribution network cannot be ignored. Therefore, accurate prediction of photovoltaic power generation is of great significance for building a next-generation power system and increasing the absorption rate of new energy. It is also of great value in establishing a comprehensive security defense system for the power system and achieving risk control.

[0003] The rapid development of new digital technologies, represented by big data and the Internet of Things, has accelerated the digital transformation of distribution systems. The massive amount of historical data accumulated by distribution networks provides data support for accurate day-ahead photovoltaic forecasts. Based on the form of the forecast results, forecasting methods can be divided into point forecasting and probabilistic forecasting. Point forecasting is a deterministic forecast method, resulting in a fixed output value that is intuitive and easy to analyze. Probabilistic forecasting is a type of uncertain forecast method, resulting in a probability density function of the output that only determines the output fluctuation range. However, massive amounts of data increase the computing and communication burden. Edge computing devices, which integrate data acquisition, storage, computing, and control, can shift some distribution network business applications to the edge, where they can perform basic data analysis and processing and provide rapid service response. This is an effective solution to the pressure of massive data communication. Furthermore, in edge computing models, power data can be transmitted to edge devices more accurately and quickly for processing, avoiding potential data distortion and loss during upload to the cloud. This provides real-time, reliable data for photovoltaic forecasting. Therefore, to improve the accuracy of photovoltaic forecasts, it is necessary to establish photovoltaic forecasting methods at the edge of distribution networks.

[0004] Due to resource constraints such as computing and storage, prediction models used at the edge need to be as lightweight as possible. Traditional neural networks are often trained using gradient descent algorithms, which require updating numerous parameters and hyperparameters during weight iteration. This results in a computationally intensive optimization process and the potential for vanishing gradients. Extreme learning machines, as single-hidden-layer feedforward neural networks, offer the advantages of few training parameters, fast learning speed, and strong generalization, effectively meeting the demands of the distribution network edge. Therefore, using a lightweight neural network, extreme learning machines, for distributed photovoltaic forecasting at the edge can significantly improve forecast accuracy, thereby more precisely empowering power systems. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a photovoltaic adaptive integrated prediction method at the edge of the distribution network that can fully utilize the value of historical data and new measured data to ultimately improve the accuracy of day-ahead photovoltaic output prediction in order to overcome the shortcomings of existing technologies.

[0006] The technical solution adopted by the present invention is: a photovoltaic adaptive integrated prediction method at the edge of a distribution network, comprising the following steps:

[0007] 1) Input the historical data set of photovoltaic output in the area under the jurisdiction of the distribution network edge computing device Numerical Weather Forecast Dataset Set the number of initial training samples Γ of the basic extreme learning machine predictor and the candidate set of hidden layer neuron node parameters Regularization parameter candidate set The ensemble size of the basic extreme learning machine predictor is N0 = N1 × N2, and the number of days in the source domain is T. m Target domain days T d , is the parameter of the hidden layer neuron node, is the regularization parameter, N1 and N2 are the number of hidden layer neuron node parameters and regularization parameters respectively, set the current prediction day n = 1, the maximum prediction day T max ;

[0008] 2) Numerical weather forecast dataset based on distance correlation coefficient The correlation analysis between meteorological factors and photovoltaic output is carried out to screen out the key factors affecting photovoltaic output and form a key meteorological factor data set.

[0009] 3) The photovoltaic output historical data set and the key meteorological factors dataset Perform min-max standardization and sample division respectively to obtain the training set

[0010] 4) Based on the training set Train the basic extreme learning machine predictor of ensemble size N0;

[0011] 5) For the N0 basic extreme learning machine predictors trained in step 4), the sparse adaptive boosting Tradaboost algorithm is used to delete non-critical basic extreme learning machine predictors, and all the basic extreme learning machine predictors {G i (x)} and the corresponding weight coefficient {α i The sparse integration extreme learning machine prediction model G(x) is obtained by integration and used as the photovoltaic output prediction model for the n-th prediction day;

[0012] 6) Input the key meteorological factor data of the nth forecast day into the sparse integration extreme learning machine prediction model G(x) of the nth forecast day to obtain the photovoltaic output prediction result of the nth forecast day

[0013] 7) After the end of the nth forecast day, record the actual photovoltaic output data of the nth forecast day The online sequence extreme learning machine algorithm is used to modify and update the parameters of the sparse ensemble extreme learning machine prediction model G(x) of the nth prediction day, and the prediction day n=n+1 is updated to determine whether n is greater than the maximum prediction day T. max If it is not greater than, return to step 6), otherwise end the prediction.

[0014] The present invention provides a method for adaptive integrated prediction of photovoltaic power at the edge of a distribution network. The method aims to solve the problems of strong randomness of photovoltaic output and adaptability of prediction models in distributed photovoltaic prediction at the edge, while reducing the complexity of the prediction model and storage resource usage as much as possible. In the initial training stage of the model, the Tradaboost algorithm is used as the main integration framework, and the extreme learning machine is used as the basic predictor. During the integration process, by deleting extreme learning machine predictors with low prediction performance and correspondingly improving extreme learning machine predictors with high prediction performance, the final integration scale is reduced, and the computational overhead and storage resources of subsequent predictions are reduced. In the day-ahead rolling prediction stage, combined with the actual photovoltaic output data of each day, the online sequence extreme learning machine algorithm is used to update the parameters of each basic extreme learning machine predictor, so as to solve the adaptive problem of the photovoltaic prediction model at the edge, and provide an effective basis for the control strategy of generating distributed power sources for the distribution network. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a flow chart of a method for adaptive integrated prediction of photovoltaic power generation at the edge of a distribution network according to the present invention;

[0016] Figure 2 This is a comparison chart of photovoltaic prediction results of different schemes;

[0017] Figure 3 This is a comparison chart of the daily average forecast error results of different schemes. DETAILED DESCRIPTION

[0018] The following describes in detail a method for adaptive integrated prediction of photovoltaic power generation at the edge of a distribution network according to the present invention in conjunction with the embodiments and drawings.

[0019] The present invention provides a method for adaptive integrated prediction of photovoltaic power at the edge of a distribution network. In the initial integrated prediction model training phase, the method solves the problems of frequent changes in photovoltaic output and data drift at the edge by performing transfer integrated learning of an extreme learning machine. During the integration process, extreme learning machine predictors with low prediction performance are deleted, thereby reducing the final integration scale and the computational overhead and storage resources of subsequent predictions. In the day-ahead rolling prediction phase, the method combines actual daily photovoltaic output data with an online sequence extreme learning machine algorithm to update the parameters of each basic extreme learning machine predictor, fully leveraging the value of historical data and new measured data, and ultimately improving the accuracy of day-ahead photovoltaic output prediction.

[0020] like Figure 1 As shown, a distribution network edge-side photovoltaic adaptive integrated prediction method of the present invention includes the following steps:

[0021] 1) Input the historical data set of photovoltaic output in the area under the jurisdiction of the distribution network edge computing device Numerical Weather Forecast Dataset Set the number of initial training samples Γ of the basic extreme learning machine predictor and the candidate set of hidden layer neuron node parameters Regularization parameter candidate set The ensemble size of the basic extreme learning machine predictor is N0 = N1 × N2, and the number of days in the source domain is T. m Target domain days T d , is the parameter of the hidden layer neuron node, is the regularization parameter, N1 and N2 are the number of hidden layer neuron node parameters and regularization parameters respectively, set the current prediction day n = 1, the maximum prediction day T max ;

[0022] 2) Numerical weather forecast dataset based on distance correlation coefficient The correlation analysis between meteorological factors and photovoltaic output is carried out to screen out the key factors affecting photovoltaic output and form a key meteorological factor data set.

[0023] The numerical weather forecast dataset The correlation analysis between meteorological factors and photovoltaic output is carried out, and the specific method is as follows:

[0024] (2.1) Calculation of numerical weather forecast dataset Correlation coefficient between medium meteorological factor γ and photovoltaic output distance:

[0025]

[0026] Where, X γ Numerical weather prediction dataset The vector of meteorological factors γ in the equation is: S is the historical data set of photovoltaic output The vector composed of the corresponding photovoltaic output data in ; is the distance correlation coefficient between meteorological factor γ and photovoltaic output; where d Cov (X γ , S) represents the variable X γ The distance covariance between and S is defined as:

[0027]

[0028]

[0029]

[0030] Where A k，l , B k，l Indicates the center distance, x k , x l , x p , x q , s k , s l , s p , s q are vectors X γ , elements in S, k, l, p, q are vectors X γ , the index of the element in S, m is the vector X γ and the S dimension,

[0031] d Cov (X γ , X γ ) and d Cov (S, S) represent vector X γ The variance of the distance from S is defined as

[0032]

[0033]

[0034] (2.2) Selecting numerical weather forecast dataset The meteorological factors with a correlation coefficient greater than 0.2 with the photovoltaic output distance are regarded as key factors affecting photovoltaic output and constitute the key meteorological factor data set.

[0035] 3) The photovoltaic output historical data set and the key meteorological factors dataset Perform min-max standardization and sample division respectively to obtain the training set The min-max normalization process and sample division are specifically as follows:

[0036] (3.1) Photovoltaic output historical data set Key meteorological factors dataset Perform min-max normalization processing separately:

[0037]

[0038] Where U g Represents the g-th data in the dataset U, Indicates U g The standardized values, dataset U refers to the photovoltaic output historical dataset Key meteorological factors dataset

[0039] (3.2) The standardized photovoltaic output historical data set Key meteorological factors dataset Perform sample division, take the key meteorological factors of the weather forecast on the r-th forecast day as the input data of the r-th training sample, r = 1, 2, ... Γ, and merge all the key meteorological factors on the r-th forecast day into the input data column vector x(r) of the r-th forecast day, with a dimension of M×1; the label of the r-th training sample is the actual photovoltaic output column vector y(r) on the r-th forecast day, with a dimension of Z×1, where Z is the number of photovoltaic output periods predicted on the day before. The input data column vector x(r) and the actual photovoltaic output column vector y(r) together constitute the training set

[0040] 4) Based on the training set Train an ensemble of N0 basic extreme learning machine predictors; including:

[0041] (4.1) The candidate set of hidden layer neuron node parameters Regularization parameter candidate set Perform permutations and combinations to obtain N1×N2 hyperparameter combinations, and use the hyperparameter combinations as hyperparameters of N0 basic extreme learning machine predictors, where N1 and N2 are the number of neuron node parameters and regularization parameters, respectively;

[0042] (4.2) Assume that the i-th basic extreme learning machine predictor G i The hyperparameter set of (x) is {L i , C i}, randomly generate input weight matrix Hidden layer bias b i ; Among them, the input weight matrix Each element in is generated by a random number uniformly distributed on [-1, 1], and the hidden layer bias is b iEach element in is generated by a random number uniformly distributed on [0, 1]. Dimension L i ×M,b i Dimension L i ×1, M is the dimension of the input data column vector x(r) for the rth forecast day; L i is the neuron node parameter of the i-th basic extreme learning machine predictor, C i is the regularization parameter of the i-th basic extreme learning machine predictor;

[0043] (4.3) Based on the training set Γ samples in the calculation G i The hidden layer output variables of (x) are as follows:

[0044]

[0045] Where x i (r) is the input vector of the i-th basic extreme learning machine predictor on the r-th prediction day, and tanh is the hyperbolic tangent function;

[0046] (4.4) Form the hidden layer output matrix H of Γ training samples i =[h i (1), h i (2), ..., h i (Γ)], with dimension L i ×Γ; the photovoltaic actual output matrix Y = [y(1), y(2), ..., y(Γ)], with a dimension of Z×Γ, completes the i-th basic extreme learning machine predictor G i (x) Training:

[0047]

[0048] Where I is the identity matrix; is the i-th basic extreme learning machine predictor G i (x) the output weight matrix;

[0049] (4.5) Repeat steps (4.2) to (4.4) until the training of N0 basic extreme learning machine predictors is completed.

[0050] 5) For the N0 basic extreme learning machine predictors trained in step 4), the sparse adaptive boosting Tradaboost algorithm is used to delete non-critical basic extreme learning machine predictors, and all the basic extreme learning machine predictors {G i (x)} and the corresponding weight coefficient {α iThe sparse integration extreme learning machine prediction model G(x) is obtained by integration as the photovoltaic output prediction model for the nth prediction day; including:

[0051] (5.1) For the Γ samples used to train the basic extreme learning machine predictor, sort the first T samples in chronological order. m samples are divided into the source domain dataset, and the remaining T d samples are divided into the target domain dataset; let Calculate the training set The initial weight of each sample in

[0052]

[0053] (5.2) For the i The training set Compute the i-th base extreme learning machine predictor G i (x) in the training set The maximum regression error E of the samples on i :

[0054] E i =max(∑|y(t)-G i (x(t))|),t=1,2,...,T m +T d (11)

[0055] Where G i (x(t)) represents the i-th basic extreme learning machine predictor G i (x) is the predicted photovoltaic output value for the t-th sample, y(t) is the actual photovoltaic value for the t-th sample, and the dimensions are both Z×1, |y(t)-G i (x(t))| represents the i-th basic extreme learning machine predictor G i (x) is the prediction error vector for the t-th sample, ∑|y(t)-G i (x(t))| represents the i-th basic extreme learning machine predictor G i (x) prediction error vector for the tth sample |y(t)-G i Sum all elements in (x(t))|, max(∑|y(t)-Gixt means taking the maximum value in yt-Gixt;

[0056] (5.3) Calculate G i (x) in the training set The relative regression error e on each sample in i，t :

[0057]

[0058] (5.4) Calculate G i (x) in the training set The overall regression error e on i and error coefficient β i :

[0059]

[0060]

[0061] Where, Indicates taking and relatively small values ​​of 0.5;

[0062] (5.5) Calculate G i The weight coefficient α of (x) i And the cumulative weight coefficient R:

[0063]

[0064] R=∑α i (16)

[0065] (5.6) If i = 1, set i = i + 1 and return to step (5.2); otherwise, find two basic extreme learning machine predictors with weight coefficient adjustment according to the following formula They are h1th and h2th respectively:

[0066]

[0067]

[0068] Where G h (x(t)) represents G h (x) The photovoltaic output prediction value for the tth sample, G τ (x(t)) represents G τ (x) The photovoltaic output prediction value for the tth sample, G h (x), G τ (x) are the h-th and τ-th basic extreme learning machine predictors, represents the Hadamard product;

[0069] (5.7) Calculate the optimal weight transfer step size p * , and adjust the weight coefficients of the h1th and h2th basic extreme learning machine predictors:

[0070]

[0071]

[0072] Where, and Represent the basic extreme learning machine predictor The weight coefficient of .

[0073] (5.8) If Then use formula (21) to adjust the cumulative weight coefficient R, reset the weight of the basic extreme learning machine predictor whose weight is less than 0 to 0, and delete the basic extreme learning machine predictor:

[0074]

[0075] (5.9) Update the sample weights of the source domain dataset and the target domain dataset:

[0076]

[0077] In the formula Represents the basic extreme learning machine predictor G i (x) for the training set The weight of sample t in .

[0078] (5.10) If i < N0, set i = i + 1 and return to step (5.2); otherwise, calculate the sparse ensemble extreme learning machine prediction model G(x) for the nth prediction day:

[0079]

[0080] 6) Input the key meteorological factor data of the nth forecast day into the sparse integration extreme learning machine prediction model G(x) of the nth forecast day to obtain the photovoltaic output prediction result of the nth forecast day include:

[0081] (6.1) Assume that the sparse ensemble extreme learning machine prediction model G(x) for the nth prediction day is composed of N sp A basic extreme learning machine predictor {G n，i (x)}, the sparse ensemble extreme learning machine prediction model G(x) of the nth prediction day is composed of the i-th basic extreme learning machine predictor G n，i The input weight matrix of (x) is The output weight matrix is Hidden layer bias b i , the weight coefficient is α i , the corresponding hyperparameter group is {L i , C i}, L i is the neuron node parameter of the i-th basic extreme learning machine predictor, C i is the regularization parameter of the i-th basic extreme learning machine predictor; let i = 1;

[0082] (6.2) Calculate G on the nth forecast day n，i The hidden layer output h of (x) n，i (n):

[0083]

[0084] Where x i (n) is the input vector of the nth forecast day, and tanh is the hyperbolic tangent function;

[0085] (6.3) Output G for the nth forecast day n，i The predicted result y of (x) n，i (n):

[0086]

[0087] (6.4) If i<N sp , then let i = i + 1 and return to step (6.2); otherwise, calculate the photovoltaic output forecast result for the nth forecast day

[0088]

[0089] 7) After the end of the nth forecast day, record the actual photovoltaic output data of the nth forecast day The online sequence extreme learning machine algorithm is used to modify and update the parameters of the sparse ensemble extreme learning machine prediction model G(x) of the nth prediction day, and the prediction day n=n+1 is updated to determine whether n is greater than the maximum prediction day T. max , if not greater than, return to step 6), otherwise end the prediction; wherein, the online sequence extreme learning machine algorithm is used to modify and update the parameters of the sparse ensemble extreme learning machine prediction model G(x) of the n-th prediction day, including:

[0090] (7.1) Assume that the sparse ensemble extreme learning machine prediction model G(x) for the nth prediction day is obtained by N sp A basic extreme learning machine predictor {G i (x)}, the i-th basic extreme learning machine predictor G of the sparse ensemble extreme learning machine prediction model for the n-th prediction day n，i The output weight matrix of (x) is The hidden layer output is h n，i (n), let i=1;

[0091] (7.2) If n = 1, calculate the initial correction matrix P according to formula (27) 1，i Otherwise, iteratively calculate the correction matrix according to formula (28):

[0092]

[0093]

[0094] Among them, H i is the i-th basic extreme learning machine predictor G during initial training i (x)’s hidden layer output matrix; P n，i and P n+1，i are the correction matrices for the nth forecast day and the n+1th forecast day respectively;

[0095] (7.3) we get G n+1，i The output weight matrix of (x)

[0096]

[0097] Where, is the i-th basic extreme learning machine predictor G in the sparse ensemble extreme learning machine prediction model for the n-th prediction day n，i (x) the output weight matrix; is the i-th basic extreme learning machine predictor G in the sparse ensemble extreme learning machine prediction model for the n+1th prediction day n+1，i (x) the output weight matrix; Indicates the actual photovoltaic output value on the nth forecast day;

[0098] (7.4) If i<N sp , let i = i + 1, and return to step (7.2); otherwise, obtain the sparse ensemble extreme learning machine prediction model for the n + 1th prediction day.

[0099] Here are some specific examples:

[0100] This example uses the photovoltaic power station at site 33 in the DKASC dataset as the research object. Located in Australia, this station is ground-mounted and consists of 22 photovoltaic panels with a rated power of 250W, for a total rated power of 5.5kW. The training set uses data from January 1, 2017, to November 30, 2017, and the test set uses data from December 1, 2017, to December 7, 2017, with a sampling interval of 1 hour. Based on the characteristics of photovoltaic output, the photovoltaic output is predicted between 7:00 AM and 7:00 PM daily. Input parameters include the following: the number of training samples Γ = 70, the candidate set of hidden layer neuron parameters for the extreme learning machine ({200, 250, 300, 350, 400, 450, 500, 550, 600, 650}), the candidate set of regularization parameters ({0.1, 1, 10, 100}), the ensemble size N0 = 40, and the number of source days M. d =230, target domain days N d=100.

[0101] To verify the effectiveness of the proposed distributed photovoltaic prediction method, simulation experiments were conducted on the public dataset of the DKASC website, and three test scenarios were set up:

[0102] Solution 1: Use the extreme learning machine as the basic predictor and Tradaboost as the ensemble learning framework to obtain the ensemble extreme learning machine model. The model parameters will not be adjusted in subsequent predictions.

[0103] Solution 2: Based on the ensemble of extreme learning machine models in Solution 1, the basic predictors are automatically screened to reduce the final ensemble size, resulting in a sparse ensemble extreme learning machine model. The model parameters are not adjusted for subsequent predictions.

[0104] Option 3: Based on the sparse integration extreme learning machine model in Option 2, the online sequence extreme learning machine algorithm is used to correct and update the model parameters every day to obtain a sparse integration online sequence extreme learning machine model.

[0105] The evaluation indicators used are the common mean absolute percentage error (MAPE) and root mean square error (RMSE), which are calculated as follows:

[0106]

[0107]

[0108] Where N is the total number of time periods for day-ahead forecast, x i is the actual photovoltaic output in the i-th hour, The predicted value for the i-th hour.

[0109] All model construction and training were performed on Matlab 2019. The hardware platform used an Intel Core i5 CPU with a main frequency of 2.3 GHz and 16 GB of memory.

[0110] All prediction models were tested 10 times independently and the average value was taken as the final result. Figure 2 , attached Figure 3 ,The comparison of the prediction error results of different ,schemes is shown in Appendix 1.

[0111] Table 1 Comparison of prediction error results of different schemes

[0112] Evaluation indicators Option 1 Option 2 Option 3 MAPE (%) 3.5622 3.5697 3.2461 RMSE 0.093765 0.093765 0.088901

[0113] Depend on Figure 2 , Figure 3From the analysis results in Table 1, it can be seen that the prediction performance of the sparse ensemble online sequence extreme learning machine model in Scheme 3 is the best. The sparse ensemble extreme learning machine model in Scheme 2 has a slightly lower prediction effect than the ensemble extreme learning machine model in Scheme 1 because some predictors are deleted and the ensemble scale is reduced.

[0114] Table 2 Comparison of calculation time of different models

[0115] Option 1 Option 2 Option 3 Average training time / s 21.573793 22.836276 23.347497 Average prediction time / s 0.106626 0.084884 0.086706

[0116] Table 2 shows a comparison of computational time for different models. The sparse ensemble extreme learning machine model in Option 2 requires longer training time than the ensemble extreme learning machine model in Option 1 due to the process of sparsifying the ensemble predictors. However, due to the simplification of the prediction model, the prediction time of the sparse ensemble extreme learning machine model in Option 2 is shorter than that of the ensemble extreme learning machine model in Option 1. Furthermore, compared to the ensemble extreme learning machine model in Option 1, the sparse ensemble extreme learning machine model in Option 2 consists of only 27 basic extreme learning machine predictors, and its ensemble scale is smaller than that of the ensemble extreme learning machine model in Option 1, effectively alleviating the computing and storage resource pressure on edge computing devices. The sparse ensemble online sequential extreme learning machine model in Option 3 requires daily corrections and updates, so its training time is longer than that of other models. However, the additional time overhead is relatively low, and its prediction accuracy is superior to that of other models, so it still has an advantage overall.

[0117] In summary, the present invention proposes an adaptive integrated prediction method for photovoltaic power generation at the edge of the distribution network, which can effectively predict the distributed photovoltaic power generation at the edge of the distribution network and revise and update the prediction model based on daily real-time data. It has low computational overhead and storage resource usage, and is feasible for deployment on edge devices.

Claims

1. A photovoltaic adaptive integrated prediction method at the edge of a distribution network, characterized in that: It includes the following steps: 1) Input the historical data set of photovoltaic output in the area under the jurisdiction of the distribution network edge computing device Numerical Weather Forecast Dataset Set the number of initial training samples Γ of the basic extreme learning machine predictor and the candidate set of hidden layer neuron node parameters Regularization parameter candidate set The ensemble size of the basic extreme learning machine predictor is N0 = N1 × N2, and the number of days in the source domain is T. m Target domain days T d , is the parameter of the hidden layer neuron node, is the regularization parameter, N1 and N2 are the number of hidden layer neuron node parameters and regularization parameters respectively, set the current prediction day n = 1, the maximum prediction day T max ; 2) Numerical weather forecast dataset based on distance correlation coefficient The correlation analysis between meteorological factors and photovoltaic output is carried out to screen out the key factors affecting photovoltaic output and form a key meteorological factor data set. 3) The photovoltaic output historical data set and the key meteorological factors dataset Perform min-max standardization and sample division respectively to obtain the training set 4) Based on the training set Train the basic extreme learning machine predictor of ensemble size N0; 5) For the N0 basic extreme learning machine predictors trained in step 4), the sparse adaptive boosting Tradaboost algorithm is used to delete non-critical basic extreme learning machine predictors, and all the basic extreme learning machine predictors {G i (x)} and the corresponding weight coefficient {α i The sparse integration extreme learning machine prediction model G(x) is obtained by integration as the photovoltaic output prediction model for the nth prediction day; including: (5.1) For the Γ samples used to train the basic extreme learning machine predictor, the first T m samples are divided into the source domain dataset, and the remaining T d samples are divided into the target domain dataset; let Calculate the training set The initial weight of each sample in (5.2) For the i The training set Compute the i-th base extreme learning machine predictor G i (x) in the training set The maximum regression error E of the samples on i : E i =max(∑|y(t)-G i (x(t))|),t=1,2,…,T m +T d (11) Where G i (x(t)) represents the i-th basic extreme learning machine predictor G i (x) is the predicted photovoltaic output value for the t-th sample, y(t) is the actual photovoltaic value for the t-th sample, and the dimensions are both Z×1, |y(t)-G i (x(t))| represents the i-th basic extreme learning machine predictor G i (x) is the prediction error vector for the t-th sample, ∑|y(t)-G i (x(t))| represents the i-th basic extreme learning machine predictor G i (x) prediction error vector for the tth sample |y(t)-G i Sum all elements of (x(t))|, max(∑|y(t)-G i (x(t))|) means taking ∑|y(t)-G i the maximum value among (x(t))|; (5.3) Calculate G i (x) in the training set The relative regression error e on each sample in i,t : (5.4) Calculate G i (x) in the training set The overall regression error e on i and error coefficient β i : Where, Indicates taking and relatively small values ​​of 0.5; (5.5) Calculate G i The weight coefficient α of (x) i And the cumulative weight coefficient R: R＝∑α i (16) (5.6) If i = 1, set i = i + 1 and return to step (5.2); otherwise, find two basic extreme learning machine predictors with weight coefficient adjustment according to the following formula They are h1th and h2th respectively: Where G h (x(t)) represents G h (x) The photovoltaic output prediction value for the tth sample, G τ (x(t)) represents G τ (x) The photovoltaic output prediction value for the tth sample, G h (x),G τ (x) are the h-th and τ-th basic extreme learning machine predictors, represents the Hadamard product; (5.7) Calculate the optimal weight transfer step size p * , and adjust the weight coefficients of the h1th and h2th basic extreme learning machine predictors: Where, and Represent the basic extreme learning machine predictor The weight coefficient of ; (5.8) If Then use formula (21) to adjust the cumulative weight coefficient R, reset the weight of the basic extreme learning machine predictor whose weight is less than 0 to 0, and delete the basic extreme learning machine predictor: (5.9) Update the sample weights of the source domain dataset and the target domain dataset: In the formula Represents the basic extreme learning machine predictor G i (x) for the training set The weight of sample t in . (5.10) If i < N0, then let i = i + 1, and return to step (5.2); otherwise, calculate the sparse ensemble extreme learning machine prediction model G(x) for the nth prediction day: 6) Input the key meteorological factor data of the nth forecast day into the sparse integration extreme learning machine prediction model G(x) of the nth forecast day to obtain the photovoltaic output prediction result of the nth forecast day 7) After the end of the nth forecast day, record the actual photovoltaic output data of the nth forecast day The online sequence extreme learning machine algorithm is used to modify and update the parameters of the sparse ensemble extreme learning machine prediction model G(x) of the nth prediction day, and the prediction day n=n+1 is updated to determine whether n is greater than the maximum prediction day T. max If it is not greater than, return to step 6), otherwise end the prediction.

2. A distribution network edge-side photovoltaic adaptive integrated prediction method according to claim 1, characterized in that: The numerical weather forecast dataset described in step 2) The correlation analysis between meteorological factors and photovoltaic output is carried out, and the specific method is as follows: (2.1) Calculation of numerical weather forecast dataset Correlation coefficient between medium meteorological factor γ and photovoltaic output distance: Where, X γ Numerical weather prediction dataset The vector of meteorological factors γ in the equation is: S is the historical data set of photovoltaic output The vector composed of the corresponding photovoltaic output data in ; is the distance correlation coefficient between meteorological factor γ and photovoltaic output; where D Cov (X γ ,S) represents the variable X γ The distance covariance between and S is defined as: Where A k,l ,B k,l Indicates the center distance, x k ,x l ,x p ,x q ,s k ,s l ,s p ,s q are vectors X γ , elements in S, k, l, p, q are vectors X γ , the index of the elements in S, m is the vector X γ and the S dimension, d Cov (X γ ,X γ ) and d Cov (S,S) respectively represent vector X γ The variance of the distance from S is defined as (2.2) Selecting numerical weather forecast dataset The meteorological factors with a correlation coefficient greater than 0.2 with the photovoltaic output distance are regarded as key factors affecting photovoltaic output and constitute the key meteorological factor data set.

3. A distribution network edge-side photovoltaic adaptive integrated prediction method according to claim 1, characterized in that: The min-max normalization processing and sample division described in step 3) are specifically as follows: (3.1) Photovoltaic output historical data set Key meteorological factors dataset Perform min-max normalization processing separately: Where U g Represents the g-th data in the dataset U, Indicates U g The standardized values, dataset U refers to the photovoltaic output historical dataset Key meteorological factors dataset (3.2) The standardized photovoltaic output historical data set Key meteorological factors dataset Perform sample division, take the key meteorological factors of the weather forecast on the r-th forecast day as the input data of the r-th training sample, r = 1, 2, ... Γ, and merge all the key meteorological factors on the r-th forecast day into the input data column vector x(r) of the r-th forecast day, with a dimension of M×1; the label of the r-th training sample is the actual photovoltaic output column vector y(r) on the r-th forecast day, with a dimension of Z×1, where Z is the number of photovoltaic output periods predicted on the day before. The input data column vector x(r) and the actual photovoltaic output column vector y(r) together constitute the training set 4. A distribution network edge-side photovoltaic adaptive integrated prediction method according to claim 1, characterized in that: Step 4) includes: (4.1) The candidate set of hidden layer neuron node parameters Regularization parameter candidate set Perform permutations and combinations to obtain N1×N2 hyperparameter combinations, and use the hyperparameter combinations as hyperparameters of N0 basic extreme learning machine predictors, where N1 and N2 are the number of neuron node parameters and regularization parameters, respectively; (4.2) Assume that the i-th basic extreme learning machine predictor G i The hyperparameter set of (x) is {L i ,C i }, randomly generate input weight matrix Hidden layer bias b i ; Among them, the input weight matrix Each element in is generated by a random number uniformly distributed on [-1,1], and the hidden layer bias is b i Each element in is generated by a random number uniformly distributed on [0,1]. Dimension L i ×M,b i Dimension L i ×1, M is the dimension of the input data column vector x(r) for the rth forecast day; L i is the neuron node parameter of the i-th basic extreme learning machine predictor, C i is the regularization parameter of the i-th basic extreme learning machine predictor; (4.3) Based on the training set Γ samples in the calculation G i The hidden layer output variables of (x) are as follows: Where x i (r) is the input vector of the i-th basic extreme learning machine predictor on the r-th prediction day, and tanh is the hyperbolic tangent function; (4.4) Form the hidden layer output matrix H of Γ training samples i =[h i (1),h i (2),…,h i (Γ)], with dimension L i ×Γ; the photovoltaic actual output matrix Y=[y(1),y(2),…,y(Γ)], with a dimension of Z×Γ, completes the i-th basic extreme learning machine predictor G i (x) Training: Where I is the identity matrix; is the i-th basic extreme learning machine predictor G i (x) the output weight matrix; (4.5) Repeat step (4.2) - step (4.4) until the training of N0 basic extreme learning machine predictors is completed.

5. The method for adaptive integrated prediction of photovoltaic power generation at the edge of a distribution network according to claim 1, characterized in that: Step 6) includes: 6) Input the key meteorological factor data of the nth forecast day into the sparse integration extreme learning machine prediction model G(x) of the nth forecast day to obtain the photovoltaic output prediction result of the nth forecast day (6.1) Assume that the sparse ensemble extreme learning machine prediction model x(x) for the nth prediction day is composed of N sp A basic extreme learning machine predictor {G n,i (x)}, the sparse ensemble extreme learning machine prediction model G(x) of the nth prediction day is composed of the i-th basic extreme learning machine predictor G n,i The input weight matrix of (x) is The output weight matrix is Hidden layer bias b i , the weight coefficient is α i , the corresponding hyperparameter group is {L i ,C i }, L i is the neuron node parameter of the i-th basic extreme learning machine predictor, C i is the regularization parameter of the i-th basic extreme learning machine predictor; let i = 1; (6.2) Calculate G on the nth forecast day n,i The hidden layer output h of (x) n,i (n): In the formula, x(n) is the input vector for the nth prediction day, and tanh is the hyperbolic tangent function; (6.3) Output G for the nth forecast day n,i The predicted result y of (x) n,i (n): (6.4) If i <N sp , then let i = i + 1 and return to step (6.2); otherwise, calculate the photovoltaic output forecast result for the nth forecast day 6. A distribution network edge-side photovoltaic adaptive integrated prediction method according to claim 1, characterized in that: The online sequential extreme learning machine algorithm described in step 7) is used to perform parameter correction and update on the sparse ensemble extreme learning machine prediction model G(x) for the nth prediction day, including: (7.1) Assume that the sparse ensemble extreme learning machine prediction model G(x) for the nth prediction day is obtained by N sp A basic extreme learning machine predictor {G i (x)}, the i-th basic extreme learning machine predictor G of the sparse ensemble extreme learning machine prediction model for the n-th prediction day n,i The output weight matrix of (x) is The hidden layer output is h n,i (n), let i=1; (7.2) If n = 1, calculate the initial correction matrix P according to formula (27) 1,i Otherwise, iteratively calculate the correction matrix according to formula (28): P 1,i =(H i T H i ) -1 a (27) Among them, H i is the i-th basic extreme learning machine predictor G during initial training i (x)’s hidden layer output matrix; P n,i and P n+1,i are the correction matrices for the nth forecast day and the n+1th forecast day respectively; (7.3) we get G n+1,i The output weight matrix of (x) Where, is the i-th basic extreme learning machine predictor G in the sparse ensemble extreme learning machine prediction model for the n-th prediction day n,i (x) the output weight matrix; is the i-th basic extreme learning machine predictor G in the sparse ensemble extreme learning machine prediction model for the n+1th prediction day n+1,i (x) the output weight matrix; Indicates the actual photovoltaic output value on the nth forecast day; (7.4) If i <N sp , let i = i + 1, and return to step (7.2); otherwise, obtain the sparse ensemble extreme learning machine prediction model for the n + 1th prediction day.

Images