Photovoltaic power short-term prediction method based on dynamic combination Stacking ensemble learning

Abstract

The invention discloses a photovoltaic power short-term prediction method based on dynamic combination Stacking ensemble learning. The method is used for solving the problems that feature processing is difficult, photovoltaic output is unstable, meteorological data and photovoltaic output are poor in correlation, data and model structures lack diversity, and model generalization is poor. The method comprises the following steps of 1, preprocessing original data; comprising high-precision data interpolation and meteorological data selection; 2, performing feature extraction on the preprocessed meteorological data by using a variational mode decomposition and kernel principal component analysis method; 3, selecting an optimal base learner for the Stacking ensemble learning model; step 4, performing photovoltaic power short-term prediction by using a Stacking ensemble learning model; 5, evaluating a deterministic prediction result; and firstly carrying out reverse normalization on the prediction result, and then evaluating the prediction result.

Description

Technical Field

[0001] The invention relates to a photovoltaic power short-term prediction method based on dynamic combined Stacking ensemble learning, and belongs to the technical field of photovoltaic prediction of power systems. Background Art

[0002] Energy is the foundation and driving force of human social progress and is vital to a country's economic development and security. In recent years, the development and utilization of photovoltaics, as a clean and renewable energy source, has become an important direction for my country's energy structure adjustment. However, the output of photovoltaic power generation is affected by weather conditions and is volatile and uncertain, which poses great challenges to the management and operation of the power grid. Accurately predicting the output power of photovoltaic power generation can effectively reduce the impact of the uncertainty of photovoltaic output on the power grid and promote the high-quality and stable power supply of the power grid.

[0003] In recent years, with the advancement of technology, the scale of photovoltaic power generation projects has continued to expand, showing good development prospects. Existing photovoltaic output forecasts can be divided into three categories: physical methods, direct forecasts, and indirect forecasts. The physical method estimates the output of distributed power generation systems based on satellite cloud image information, but the cost is relatively high. Direct forecasts establish a mapping relationship between historical data and future data through the periodicity and trend of photovoltaic output curves, but do not consider meteorological information such as solar irradiance, humidity, temperature, pressure, and wind speed. Indirect forecasts predict output by predicting meteorological data and using the mapping relationship between meteorological data and photovoltaic output. Although meteorological information is taken into account, meteorological characteristics are not deeply explored. In order to meet the needs of power grid dispatching, photovoltaic power stations are often required to provide forecast results with shorter step sizes, and photovoltaic day-ahead forecasts face further challenges.

[0004] The nonlinear processing capability of machine learning methods is conducive to fitting the changing law of photovoltaic power and has been widely used in photovoltaic forecasting. As the power grid's requirements for the prediction accuracy of photovoltaic power stations continue to increase, single-layer neural network prediction models often cannot meet the prediction needs. Deep learning networks such as LSTM and CNN are gradually being used in photovoltaic forecasting. However, deep learning methods often require a large amount of operating data support in the process of building models, and their convergence time also increases significantly due to the increase in the number of network layers. To address this problem, more and more research is focusing on combined prediction methods.

[0005] Compared with the traditional single model, the combined prediction method can integrate the advantages of different models and methods in time series prediction, and has become a research hotspot in the field of new energy power prediction. Ensemble learning not only deeply explores the deep feature information of time series, but also improves the prediction accuracy and generalization ability. The Stacking method in ensemble learning has obvious advantages. Stacking combines multiple different types of basic models to capture multiple patterns and features in the data, thereby enhancing the generalization ability of the model. By integrating the predictions of multiple models, it reduces the dependence of a single model on a specific data set, making the final prediction result more accurate. In addition, when processing data, Stacking can select appropriate basic models for different features or data distributions, thereby optimizing prediction performance in different scenarios. This flexibility enables it to have better generalization in a variety of tasks.

[0006] Commonly used deep learning algorithms for photovoltaic prediction in power systems generally have problems such as difficulty in processing complex data, shallow mining of meteorological characteristics, neglect of the relationship between multi-source meteorological characteristics and photovoltaic output, lack of diversity in data and model structure, and poor model generalization, resulting in poor model prediction results. Summary of the invention

[0007] The present invention proposes a photovoltaic power short-term prediction method based on dynamic combination Stacking ensemble learning to solve the problems of difficult feature processing, instability of photovoltaic output, poor correlation between meteorological data and photovoltaic output, lack of diversity in data and model structure, and poor model generalization.

[0008] In order to achieve the above object, the present invention adopts the following technical solutions:

[0009] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0010] The present invention discloses a photovoltaic power short-term prediction method based on stacking ensemble learning of dynamic combination.

[0011] The following steps are involved:

[0012] Step 1: Raw data preprocessing, including high-precision data interpolation and meteorological data selection;

[0013] Step 2: Use variational mode decomposition and kernel principal component analysis methods to extract features from the preprocessed meteorological data; use the VMD algorithm to perform multi-level decomposition on the preprocessed photovoltaic meteorological time series data to enrich the diversity of input variables and highlight the local characteristics of environmental series at different time scales;

[0014] Step 3: Select the optimal base learner for the Stacking ensemble learning model;

[0015] Step 4: Use the Stacking ensemble learning model to perform short-term photovoltaic power prediction; use the above optimal sub-training set as the input data of the Stacking prediction model to train the base learner and meta learner respectively;

[0016] Step 5: Evaluate the deterministic prediction results; first denormalize the prediction results, and then evaluate the prediction results.

[0017] As an improvement of the prediction method disclosed in the present invention, the step 1 uses a piecewise interpolation method, whose interpolation function is a first-order polynomial, to convert low-resolution original meteorological data into high-resolution data; the basic principle of the piecewise interpolation method is as follows:

[0018]

[0019] In formula (1), x is the interpolation node; y is the output after interpolation;

[0020] Next, the Spearman correlation analysis is performed on the interpolated multi-source meteorological data and the photovoltaic output to exclude the influencing factors with less correlation and remove a large amount of redundant information in the meteorological data used. The calculation formula of the Spearman correlation coefficient is:

[0021]

[0022] In formula (2), x i and i Represent the observed values ​​rank of X and Y respectively; and denote the average ranks of X and Y respectively; ρ is the Spearman correlation coefficient of X and Y.

[0023] As an improvement of the prediction method disclosed in the present invention, in step 2, the specific steps of VMD decomposition are as follows:

[0024] (1) Perform Hilbert transform on the sub-mode and obtain the single-sided spectrum of the analytical signal:

[0025]

[0026] In formula (3), u k(t) is the kth intrinsic mode function component; δ(t) is the unit impulse function;

[0027] (2) Correct the estimated center frequency of each modal analytical signal and move the modal spectrum to the corresponding baseband;

[0028] (3) The bandwidth is estimated by the square norm of the demodulated signal. In order to minimize the sum of the bandwidth estimates, a variational model with constraints is constructed as follows:

[0029]

[0030] In formula (4), K is the total number of modes; δ(t) is the Dirac function; ω k is the center frequency; f(t) is the original signal;

[0031] (4) Introducing quadratic penalty terms and Lagrange multipliers to unconstrain the problem, thereby simplifying the solution process; by iteratively updating the variables of each sub-problem, the solution to the problem is finally obtained;

[0032] Next, KPCA dimension reduction is performed on the decomposed meteorological characteristic data to realize multi-source meteorological data extraction.

[0033] As an improvement of the prediction method disclosed in the present invention,

[0034] Assume that the original data is X={x1,x2,…,x n}, the main steps of KPCA dimensionality reduction in step 2 are as follows:

[0035] (1) Preprocess the original data and project it into a high-dimensional space through a mapping function φ to obtain a kernel matrix φ(x i )={φ(x1),φ(x2),…,φ(x n )},i=1,2,…,n;

[0036] (2) Assuming that the kernel matrix has been centered, the covariance matrix is:

[0037]

[0038] (3) Introducing the kernel function K * =φ T φ, perform eigenvalue decomposition on C to generate eigenvalue λ and eigenvector η;

[0039] (4) According to the size of the eigenvalue, calculate the contribution rate of each principal component, set the cumulative contribution rate threshold, and take the contribution rate as 85%;

[0040] (5) Map the original data to the selected principal component to obtain the sample data X after dimensionality reductionT ;

[0041] X T =η T [φ(e1,e),φ(e2,e),…,φ(e i ,e)] T (6)

[0042] Normalize the processed feature data samples to the interval [0,1]; the normalization calculation formula is:

[0043]

[0044] In the formula, P′, P, P max and P min Represent normalized data, power value, maximum power value, and minimum power value, respectively.

[0045] As an improvement of the prediction method disclosed in the present invention, in step 3, the stacking ensemble learning model has two layers: the first layer is a model that takes the original training set as input, which is a base model; the second layer is a model that takes the prediction of the base model on the original training set as the training set and the prediction of the base model on the original test set as the test set, which is a meta-model; wherein the outputs of all base learners are received by a meta-learner, and then the final result is output;

[0046] The new data set after features are extracted by the above VMD and KPCA is divided into an initial training set and a test set. The initial training data set consists of the first 90% of the original data set, and the test data set consists of the last 10% of the original data set. Three sub-training sets are extracted from the initial training set. The number of samples in each sub-training data set is 1 / 4 of the initial training set.

[0047] As an improvement of the prediction method disclosed in the present invention, the training steps of the first layer of the Stacking prediction model are as follows: select the best N training algorithms, input the integrated data into the Stacking first layer prediction model; select the training algorithm A1 to train the base learner by cross-validation to obtain D1, D2, ..., D N A total of N groups of predicted values ​​for the validation set, denoted as t1, t2, ..., t N , and N groups of predicted values ​​for the test set; combine the N groups of predicted values ​​for the validation set, that is, T1=[t1,t2,...,t N ], and the predicted values ​​of the N test sets are averaged to obtain a set of predicted values ​​b1. Similarly, for training algorithms A2~A N Repeat the above steps and finally get the predicted value [T1,T2,...,T N] and the predicted values ​​of the test set [b1,b2,...,b N ];

[0048] The training steps of the second layer of the Stacking prediction model are as follows: the prediction values ​​[T1, T2, ..., T N ] and the predicted values ​​of the test set obtained by the first layer [b1,b2,...,b N ] are used as the second-layer training set and test set respectively, and LSTM is used to train the meta-learner to obtain the final prediction results.

[0049] As an improvement of the prediction method disclosed in the present invention, the step 5 adopts four evaluation indicators to evaluate the performance of the model prediction, namely, root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE) and determination coefficient (R 2 ) to evaluate the prediction effect, which are recorded as RMSE, MAE, MAPE and R 2 , the calculation formula is as follows:

[0050]

[0051] The beneficial effects of the present invention are as follows:

[0052] In view of the common defects of the commonly used deep learning algorithms for photovoltaic power prediction in power systems, the present invention proposes a photovoltaic power short-term prediction method based on dynamic combined Stacking ensemble learning. The problem of low data resolution is solved by a high-precision data interpolation method, the problem of ignoring the connection between multi-source meteorological characteristics and photovoltaic output is solved by the correlation analysis between meteorological data and photovoltaic output, the problem of difficult processing of complex data is solved by deeply mining meteorological characteristics through VMD and KPCA methods, the problem of lack of diversity in data and model structure is solved by the selection of Stacking optimal basis learners, and the problem of poor model generalization is solved by Stacking ensemble learning. Therefore, the present invention has the advantages of deeply mining meteorological characteristics, efficiently processing multi-source data, considering the correlation of photovoltaic data, the diversity of data and model structure, good model generalization performance, and high prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 This is a structural diagram of a combined prediction model according to a specific implementation of the present invention. DETAILED DESCRIPTION

[0054] The specific implementation of the present invention is described below in conjunction with the accompanying drawings and embodiments:

[0055] like Figure 1 As shown, Example 1:

[0056] The present invention provides a photovoltaic power short-term prediction method based on dynamic combination Stacking ensemble learning, the model structure is as follows: Figure 1 As shown, the following steps are included:

[0057] Step 1: Raw data preprocessing, including high-precision data interpolation and meteorological data selection.

[0058] The present invention uses a piecewise interpolation method, whose interpolation function is a first-order polynomial, to transform low-resolution original meteorological data into high-resolution data, increase the integrity and continuity of photovoltaic data, and reduce the impact of weather fluctuations on forecast results. The basic principle of the piecewise interpolation method is as follows:

[0059]

[0060] Where x is the interpolation node and y is the interpolated output.

[0061] Next, the Spearman correlation analysis was performed on the interpolated multi-source meteorological data and the photovoltaic output to exclude the influencing factors with less correlation and remove a large amount of redundant information in the meteorological data used. The calculation formula of the Spearman correlation coefficient is:

[0062]

[0063] In the formula, x i and i Represent the observed values ​​rank of X and Y respectively; and denote the average ranks of X and Y respectively; ρ is the Spearman correlation coefficient of X and Y.

[0064] Step 2: Use variational mode decomposition and kernel principal component analysis methods to extract features from the preprocessed meteorological data.

[0065] The VMD algorithm is used to perform multi-level decomposition on the preprocessed photovoltaic meteorological time series data to enrich the diversity of input variables and highlight the local characteristics of environmental sequences at different time scales. The specific steps of VMD decomposition are as follows:

[0066] (1) Perform Hilbert transform on the sub-mode and obtain the single-sided spectrum of the analytical signal:

[0067]

[0068] In the formula, u k (t) is the kth intrinsic mode function component; δ(t) is the unit impulse function.

[0069] (2) Correct the estimated center frequency of each modal analytical signal and move the modal spectrum to the corresponding baseband.

[0070] (3) The bandwidth is estimated by the square norm of the demodulated signal. In order to minimize the sum of the bandwidth estimates, a variational model with constraints is constructed as follows:

[0071]

[0072] Where K is the total number of modes; δ(t) is the Dirac function; ω k is the center frequency; f(t) is the original signal.

[0073] (4) The quadratic penalty term and Lagrange multiplier are introduced to unconstrain the problem, thereby simplifying the solution process. By iteratively updating the variables of each sub-problem, the solution to the problem is finally obtained.

[0074] Next, KPCA dimensionality reduction is performed on the decomposed meteorological characteristic data to realize multi-source meteorological data extraction. Assume that the original data is X = {x1, x2, ..., x n}, the main steps of KPCA are as follows:

[0075] (1) Preprocess the original data and project it into a high-dimensional space through a mapping function φ to obtain a kernel matrix φ(x i )={φ(x1),φ(x2),…,φ(x n )},i=1,2,…,n.

[0076] (2) Assuming that the kernel matrix has been centered, the covariance matrix is:

[0077]

[0078] (3) Introducing the kernel function K * =φ T φ, perform eigenvalue decomposition on C to produce eigenvalue λ and eigenvector η.

[0079] (4) According to the size of the eigenvalue, the contribution rate of each principal component is calculated, and a cumulative contribution rate threshold is set. The contribution rate of the present invention is set to 85%.

[0080] (5) Map the original data to the selected principal component to obtain the sample data X after dimensionality reduction T .

[0081] X T =η T [φ(e1,e),φ(e2,e),…,φ(e i ,e)]T (6)

[0082] Normalize the processed feature data samples to the interval [0,1]. The normalization calculation formula is:

[0083]

[0084] In the formula, P′, P, P max and P min Represent normalized data, power value, maximum power value, and minimum power value, respectively.

[0085] Step 3: Select the optimal base learner for the Stacking ensemble learning model

[0086] The stacking ensemble learning model has two layers: the first layer is a model that takes the original training set as input, called the base model; the second layer is a model that takes the prediction of the base model on the original training set as the training set and the prediction of the base model on the original test set as the test set, called the meta-model. The outputs of all base learners are received by a meta-learner, which then outputs the final result.

[0087] The new dataset after features are extracted by VMD and KPCA is divided into an initial training set and a test set. The initial training set consists of the first 90% of the original dataset, and the test set consists of the last 10% of the original dataset. Three sub-training sets are extracted from the initial training set. The number of samples in each sub-training set is 1 / 4 of the initial training set.

[0088] For the selection of two-layer training algorithms under the Stacking framework, the present invention uses four algorithms, SVM, RF, KNN and BP, to train base learners, selects LSTM to train meta-learners, and forms a prediction model based on Stacking. The four base learners, SVM, RF, KNN and BP, are trained on three sub-training sets to form a total of 12 base models. Each base learner is trained independently. The base learner is used to fit the corresponding sub-training set, and the error generated is the training error. The training error reflects the fitting and generation capabilities of the base learner on a specific sub-training data set. By comparing the training errors, the optimal sub-training set is selected, and the training effect of each base learner on the optimal sub-training set is compared, and the optimal three base learners are selected as the best base learners for Stacking ensemble learning.

[0089] Step 4: Use the Stacking ensemble learning model to perform short-term photovoltaic power forecasting.

[0090] The above optimal sub-training set is used as the input data of the Stacking prediction model to train the base learner and meta learner respectively.

[0091] The training steps of the first layer of the Stacking prediction model are as follows: select the best N training algorithms and input the integrated data into the first layer of the Stacking prediction model. Select the training algorithm A1 to train the base learner through cross-validation to obtain D1, D2, ..., D N A total of N groups of predicted values ​​for the validation set, denoted as t1, t2, ..., t N , and N groups of predicted values ​​for the test set; combine the N groups of predicted values ​​for the validation set, that is, T1=[t1,t2,...,t N ], and the predicted values ​​of the N test sets are averaged to obtain a set of predicted values ​​b1. Similarly, for training algorithms A2~A N Repeat the above steps and finally get the predicted value [T1,T2,...,T N ] and the predicted values ​​of the test set [b1,b2,...,b N ];

[0092] The training steps of the second layer of the Stacking prediction model are as follows: the prediction values ​​[T1, T2, ..., T N ] and the predicted values ​​of the test set obtained by the first layer [b1,b2,...,b N ] are used as the second-layer training set and test set respectively, and LSTM is used to train the meta-learner to obtain the final prediction results.

[0093] Step 5: Evaluate the deterministic prediction results

[0094] The prediction results are first denormalized and then evaluated. The present invention uses four evaluation indicators to evaluate the performance of model prediction, namely, root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE) and coefficient of determination (R 2 ) to evaluate the prediction effect, which are recorded as RMSE, MAE, MAPE and R 2 , the calculation formula is as follows:

[0095]

[0096] Application examples:

[0097] To verify the performance of the present invention, a data set in the 2014 Global Energy Forecasting Competition was used for simulation verification. The data set includes solar energy data and meteorological data from 2012 to 2014 in a certain place. It is a 24-hour all-weather data set with a data resolution of 1 hour, which realizes the prediction of the actual photovoltaic power of the next day. The original 1-hour resolution data is converted into 15-minute resolution data using a high-precision interpolation method. Then, the photovoltaic output and 12 kinds of meteorological data are subjected to Spearman correlation coefficient calculation, and the relative humidity, 2-meter temperature, surface solar radiation, surface thermal radiation and top net solar radiation under 1000 millibar (r) are finally selected as meteorological feature data. Four algorithms, SVM, RF, KNN and BP, are used to train on 3 sub-training sets, and LSTM is used as a meta-learner. The training errors of the 12 models on their respective sub-training data sets are shown in Table 1.

[0098] Table 1. Training errors of base learners on different sub-training sets

[0099]

[0100] As shown in Table 1, for the three sub-training sets, the four base learners show the best prediction performance in sub-training set 3 compared with other sub-training data sets. Therefore, sub-training set 3 is selected as the optimal sub-training set. By comparing the training effect of each basic model on sub-training set 3, RF, KNN and BP are selected as the best basic models for integration.

[0101] Sub-training set 3 was selected for comparative experiments, and the training set and test set were divided according to a ratio of 7:3. In order to ensure the accuracy of the experimental results, each experiment was conducted 5 times, and finally the prediction results ranked in the median of the prediction accuracy were taken for subsequent analysis and comparison. The photovoltaic power short-term prediction result indicators are summarized in Table 2.

[0102] Table 2 Summary of various photovoltaic power short-term prediction indicators

[0103]

[0104] It can be seen from Table 2 that the model proposed in this invention is superior to other models in each evaluation index and has a better prediction effect. Among them, the RMSE value is 0.020856MW, which is reduced by 22%, 42.6%, and 1.58% compared with the other three models respectively; the MAPE value is 0.009686NW, which is reduced by 41.9%, 54.7%, and 9.1% compared with the other three models respectively; the MAPE value is 3.1944%, which is reduced by 1.18%, 3.3%, and 0.4% compared with the other three models respectively; the determination coefficient R 2In summary, the VMD-KPCA-Stacking model of the present invention has a significant increase in prediction accuracy and a significant decrease in prediction error compared to other models, and the prediction results have a higher degree of fit with the true value, indicating that the photovoltaic power short-term prediction model proposed in the present invention has a significant improvement in performance compared to other comparative prediction models.

[0105] The preferred embodiments of the present invention are described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above embodiments. Various changes can be made within the knowledge scope of ordinary technicians in the field without departing from the purpose of the present invention. These changes involve related technologies well known to those skilled in the art, which all fall within the scope of protection of the patent of the present invention.

[0106] Many other changes and modifications may be made without departing from the concept and scope of the present invention.It should be understood that the present invention is not limited to the specific embodiments, and the scope of the present invention is defined by the appended claims.

Claims

1. Photovoltaic power short-term prediction method based on dynamic combination stacking ensemble learning, It is characterized in that The following steps are involved: Step 1: Raw data preprocessing, including high-precision data interpolation and meteorological data selection; Step 2: Use variational mode decomposition and kernel principal component analysis methods to extract features from the preprocessed meteorological data; The VMD algorithm is used to perform multi-level decomposition on the preprocessed photovoltaic meteorological time series data to enrich the diversity of input variables and highlight the local characteristics of environmental series at different time scales. Step 3: Select the optimal base learner for the Stacking ensemble learning model; Step 4: Use the Stacking ensemble learning model to perform short-term photovoltaic power forecasting; The above optimal sub-training set is used as the input data of the Stacking prediction model to train the base learner and meta learner respectively; Step 5: Evaluate the deterministic prediction results; The prediction results are first denormalized and then evaluated.

2. According to claim 1, the photovoltaic power short-term prediction method based on dynamic combination stacking ensemble learning, It is characterized in that The step 1 uses a piecewise interpolation method, whose interpolation function is a first-order polynomial, to transform low-resolution original meteorological data into high-resolution data; the basic principle of the piecewise interpolation method is as follows: In formula (1), x is the interpolation node; y is the output after interpolation; Next, the Spearman correlation analysis is performed on the interpolated multi-source meteorological data and the photovoltaic output to exclude the influencing factors with less correlation and remove a large amount of redundant information in the meteorological data used. The calculation formula of the Spearman correlation coefficient is: In formula (2), x i and i Represent the observed values ​​rank of X and Y respectively; and denote the average ranks of X and Y respectively; ρ is the Spearman correlation coefficient of X and Y.

3. The photovoltaic power short-term prediction method based on dynamic combination Stacking ensemble learning according to claim 1, Features: In step 2, the specific steps of VMD decomposition are as follows: (1) Perform Hilbert transform on the sub-mode and obtain the single-sided spectrum of the analytical signal: In formula (3), u k (t) is the kth intrinsic mode function component; δ(t) is the unit impulse function; (2) Correct the estimated center frequency of each modal analytical signal and move the modal spectrum to the corresponding baseband; (3) The bandwidth is estimated by the square norm of the demodulated signal. In order to minimize the sum of the bandwidth estimates, a variational model with constraints is constructed as follows: In formula (4), K is the total number of modes; δ(t) is the Dirac function; ω k is the center frequency; f(t) is the original signal; (4) Introducing quadratic penalty terms and Lagrange multipliers to unconstrain the problem, thereby simplifying the solution process; by iteratively updating the variables of each subproblem, the solution to the problem is finally obtained; Next, KPCA dimension reduction is performed on the decomposed meteorological characteristic data to realize multi-source meteorological data extraction.

4. The photovoltaic power short-term prediction method based on stacking ensemble learning of adaptive dynamic combination according to claim 3, Features: Assume that the original data is X={x1,x2,…,x n }, the main steps of KPCA dimensionality reduction in step 2 are as follows: (1) Preprocess the original data and project it into a high-dimensional space through a mapping function φ to obtain a kernel matrix φ(x i )={φ(x1),φ(x2),…,φ(x n )},i=1,2,…,n; (2) Assuming that the kernel matrix has been centered, the covariance matrix is: (3) Introducing the kernel function K * =φ T φ, perform eigenvalue decomposition on C to generate eigenvalue λ and eigenvector η; (4) According to the size of the eigenvalue, calculate the contribution rate of each principal component, set the cumulative contribution rate threshold, and take the contribution rate as 85%; (5) Map the original data to the selected principal component to obtain the sample data X after dimensionality reduction T ; X T =η T [φ(e1,e),φ(e2,e),…,φ(e i ,and)] T (6) Normalize the processed feature data samples to the interval [0,1]; The normalized calculation formula is: In the formula, P′, P, P max and P min Represent normalized data, power value, maximum power value, and minimum power value, respectively.

5. The photovoltaic power short-term prediction method based on stacking ensemble learning of adaptive dynamic combination according to claim 1, It is characterized in that In step 3, the stacking ensemble learning model has two layers: the first layer is a model that takes the original training set as input, which is the base model; the second layer is a model that takes the prediction of the base model on the original training set as the training set and the prediction of the base model on the original test set as the test set, which is the meta-model; wherein the outputs of all base learners are received by a meta-learner, which then outputs the final result; The new data set after features are extracted by the above VMD and KPCA is divided into an initial training set and a test set. The initial training data set consists of the first 90% of the original data set, and the test data set consists of the last 10% of the original data set. Three sub-training sets are extracted from the initial training set. The number of samples in each sub-training data set is 1 / 4 of the initial training set.

6. The photovoltaic power short-term prediction method based on dynamic combination Stacking ensemble learning according to claim 5, Features: The training steps of the first layer of the Stacking prediction model are as follows: select the best N training algorithms and input the integrated data into the first layer of the Stacking prediction model; select the training algorithm A1 to train the base learner through cross-validation to obtain D1, D2, ..., D N A total of N groups of predicted values ​​for the validation set, denoted as t1, t2, ..., t N , and N groups of predicted values ​​for the test set; combine the N groups of predicted values ​​for the validation set, that is, T1=[t1,t2,...,t N ], and the predicted values ​​of the N test sets are averaged to obtain a set of predicted values ​​b1. Similarly, for training algorithms A2~A N Repeat the above steps and finally get the predicted value [T1,T2,...,T N ] and the predicted values ​​of the test set [b1,b2,...,b N ]; The training steps of the second layer of the Stacking prediction model are as follows: the prediction values ​​[T1, T2, ..., T N ] and the predicted values ​​of the test set obtained by the first layer [b1,b2,...,b N ] are used as the second-layer training set and test set respectively, and LSTM is used to train the meta-learner to obtain the final prediction results.

7. The photovoltaic power short-term prediction method based on dynamic combination Stacking ensemble learning according to claim 1, Features: The step 5 adopts four evaluation indicators to evaluate the performance of model prediction, namely, root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE) and coefficient of determination (R 2 ) to evaluate the prediction effect, which are recorded as RMSE, MAE, MAPE and R 2 , the calculation formula is as follows:

Images