Photovoltaic power short-term prediction method and system based on MVO optimization mixed kernel PCA dimension reduction and improved FPN network

Abstract

The invention discloses a photovoltaic power short-term prediction method and system based on MVO optimization mixed kernel PCA dimension reduction and an improved FPN network. Optimizing mixed kernel principal component analysis dimensionality reduction on the characteristic data set of the photovoltaic power through a multivariate universe optimization algorithm; an improved feature pyramid network is utilized to input photovoltaic power historical data into a multi-channel convolutional network, features of the historical photovoltaic power data are extracted through multiple convolution kernels, feature fusion is carried out on the features and meteorological and electrical attributes after MVO optimization hybrid kernel PCA dimension reduction, and the features are used as dynamic input features of an FPN network. The dynamic input features combined by the multi-channel convolutional network can continuously adjust the input of the multi-channel network in the prediction process, the dynamic relationship between historical photovoltaic power data and the covariable features after dimension reduction is mined, and finally a prediction result is obtained through a plurality of full connection layers.

Description

Technical Field

[0001] The present invention belongs to the field of photovoltaic power prediction of electric power system, and in particular relates to a photovoltaic power short-term prediction method and system based on MVO optimized hybrid kernel PCA dimensionality reduction and improved FPN network. Background Art

[0002] The rise of the new energy industry has optimized my country's energy structure, and the photovoltaic industry has been widely used in distributed energy. However, with the large-scale integration of distributed energy, mainly photovoltaics, into the power grid, due to its own randomness, volatility and intermittency, there are huge hidden dangers in the stability of the power grid itself. The prediction of photovoltaic power can make scientific decisions in advance and reduce system operating costs. The factors that affect the prediction accuracy are, on the one hand, the characteristic data set of photovoltaic power prediction, and on the other hand, the effective prediction model.

[0003] Factors that affect photovoltaic power include a large number of data sources, such as photovoltaic power and power generation, light intensity, panel temperature, site temperature, wind direction and wind speed, etc. These data have high-dimensional features and complex relationships, which will lead to high direct calculation complexity, large storage requirements and overfitting problems. When faced with high-dimensional feature sets related to photovoltaic power, dimensionality reduction methods can retain key information, reduce noise and delete redundant features, thereby improving the accuracy and efficiency of prediction results.

[0004] In the prior art, the photovoltaic power prediction method using the dimension reduction feature set mainly includes:

[0005] 1. Use principal component analysis (PCA) to reduce the dimension of the relevant features of photovoltaic power, and then use the reduced dimension data to predict photovoltaic power.

[0006] This method has the following problems: This method transforms the original data into a matrix and reduces the dimension by retaining the component with the largest variance. However, PCA assumes that the data is linearly correlated, which is obviously not applicable to most feature sets.

[0007] 2. Use a single kernel function, such as a Gaussian kernel function, to reduce the dimension of the relevant features of photovoltaic power, and then use the reduced-dimensional data to directly predict photovoltaic power through multiple fully connected layers.

[0008] This method has the following problems: a single kernel function is difficult to consider the local and global features between features, and it is difficult to fully integrate the value of mining data using only the fully connected layer. Summary of the invention

[0009] In order to solve the deficiencies in the prior art, the present invention discloses a photovoltaic power short-term prediction method and system based on MVO optimized hybrid kernel PCA dimensionality reduction and improved FPN network, so as to solve the technical problems of high complexity, large storage requirements and insufficient feature fusion of photovoltaic power high-dimensional feature data sets.

[0010] In order to solve the above technical problems, the present invention adopts the following technical solutions.

[0011] The present invention first discloses a photovoltaic power short-term prediction method based on MVO optimized hybrid kernel PCA dimensionality reduction and improved FPN network, which is characterized by comprising the following steps:

[0012] Step 1: discretize all feature quantities of the photovoltaic power high-dimensional feature set by symbolic analysis to obtain a symbol sequence, and divide it into a training set and a test set after normalization;

[0013] Step 2, forming a mutual information matrix from all mutual information values ​​between photovoltaic power characteristic data symbol sequences, and grouping all variables using a heuristic grouping algorithm based on the mutual information matrix;

[0014] Step 3, constructing a hybrid kernel function according to the Gaussian kernel function and the Laplace kernel function, and using the MVO optimization algorithm to determine the optimal parameter combination of the hybrid kernel function;

[0015] Step 4: Substitute the optimal parameter combination into the hybrid kernel function, perform data fusion on the grouped variables through the hybrid kernel function, and obtain features after data dimensionality reduction.

[0016] Step 5 is used to combine the features after dimensionality reduction and the features of photovoltaic power historical data mined by using a multi-channel convolutional network, and use the improved FPN network to perform deep feature mining and fusion to obtain the final prediction result of photovoltaic power.

[0017] The present invention further includes the following preferred embodiments:

[0018] The step 1 further comprises:

[0019] Step 1.1: Normalize the photovoltaic power and its feature set, and transform the time series X(n) = {x1, x2, ..., x n} is mapped to the distribution space from 0 to 1, and we get Y(n)={y1,y2,...,y n};

[0020] Step 1.2: Encode the distribution space into a symbol time sequence containing d symbols:

[0021]

[0022] After symbolization of the time series X(n) and Y(n), symbol sequences S1(m1) and S2(m2) are obtained;

[0023] Step 1.3: Divide the photovoltaic power data set into a training set and a test set in a ratio of 4:1.

[0024] The step 2 further comprises:

[0025] Step 2.1: Get the mutual information from the symbol sequences S1(m1) and S2(m2):

[0026]

[0027] N=m1×m2

[0028] where p ij is the joint probability, i.e., the total number of times k that the combination of the i-th symbol in sequence S1(m1) and the j-th symbol in S2(m2) occurs ij The proportion of the entire sequence combination N, m1 and m2 represent the number of symbolized sequences S1(m1) and S2(m2) respectively.

[0029] Step 2.2: Get the input N data objects, represented as a set of d-dimensional vectors;

[0030] Step 2.3: Initialize the number of classifications, i.e. the final number of dimensionality reduction K and the maximum number of iterations n;

[0031] Step 2.4: For each pair of data objects i and j, calculate the mutual information between them and store the result in the mutual information matrix M;

[0032] Step 2.5: Normalize the mutual information matrix M to obtain the normalized mutual information matrix NM;

[0033] Step 2.6: Initialize an empty grouped list C and an empty ungrouped object list U;

[0034] Step 2.7: Add all data objects to the ungrouped object list U;

[0035] Step 2.8: When the ungrouped object list U is not empty, perform the following steps:

[0036] Select an object u from the ungrouped object list U as a seed object;

[0037] Remove the seed object u from the ungrouped object list U and add it to the grouped list C;

[0038] Iterate the following steps 2.81 to 2.83 until no object can be added to any group or the maximum number of iterations n is met:

[0039] Step 2.81: For each object v in the ungrouped list U, calculate the difference between object v and the grouped list

[0040] The standard measure function for all objects in C is the average of the mutual information within the group, and the result is stored in the object

[0041] The average mutual information of v is MI(v);

[0042] Step 2.82: Select the object v with the largest average mutual information MI(v), and calculate the average mutual information AvgMI(v) between the object v and all the objects in the group list C;

[0043] Step 2.83: If AvgMI(v) is greater than the threshold T, then add object v to the list C and delete object v from the list of ungrouped objects U;

[0044] Step 2.9: If the number of objects in list C is less than K, increase the value of K and return to step 2.9;

[0045] Step 2.10: Output a list of objects in each group.

[0046] The step 3 further comprises:

[0047] Step 3.1: Define the mixing kernel function:

[0048]

[0049] in, is the Gaussian kernel function, is the Laplace kernel function, α, σ1 and σ2 are weight coefficients to be determined in the mixed kernel function;

[0050] Step 3.2: Consider the vector after KPCA dimensionality reduction as the reconstructed vector, and use the reconstruction error as the performance measure of KPCA. Then define the objective function of MVO as:

[0051]

[0052] Among them, X represents the original vector before dimensionality reduction, X r Represents the reconstructed vector, and N represents the number of data samples. By minimizing the objective function, the optimal adjustment coefficient α, kernel function hyperparameters σ1 and σ2 are found to minimize the reconstruction error of MKPCA.

[0053] Step 3.3: For the adjustment coefficient α, the value range is [0, 2], the initial solution is randomly selected within this range, and the selection range of σ1 and σ2 is [0.1, 10]. Each universe represents a set of initial values ​​of the mixing kernel function parameters;

[0054] Step 3.4: Calculate the gravitational force between universes based on the fitness value of each universe;

[0055] Step 3.5: Propagate the calculated gravity to other universes to adjust the position of the universe;

[0056] Step 3.6: According to the results of gravitational propagation and the range of unknown variables, a new solution is generated by fine-tuning and deforming the original solution;

[0057] Step 3.7: For the new solution generated, recalculate its fitness value; if the fitness value of the new solution is better, update the original solution to the new solution, otherwise it remains unchanged;

[0058] Step 3.8: Iterate step 3.4 to step 3.7 until the stopping condition is met, wherein the stopping condition is that the maximum number of iterations is reached or the fitness value converges;

[0059] Step 3.9: Select the hybrid kernel function parameter combination with the best fitness from the last generation of the universe as the optimal solution, that is, the optimal hybrid kernel function parameters.

[0060] The step 4 further comprises:

[0061] Step 4.1: Determine the mixing kernel function K of MKPCA using the normalized data and the optimal parameter value of the mixing kernel function mix (x, y), and calculate the inner product of the kernel function of the normalized data matrix to obtain the kernel matrix W;

[0062] Step 4.2: Calculate the centering matrix W' based on the kernel matrix W:

[0063] W'=W-BW-WB+BWB

[0064] in, l is the sample size;

[0065] Step 4.3: Matrix Perform principal component analysis (PCA) to obtain the eigenvalue λ i and the eigenvector V i ;

[0066] Step 4.4: For the eigenvalue λ i Arrange in descending order and replace the corresponding eigenvector V i Sort by the order of eigenvalues;

[0067] Step 4.5: Take out the eigenvector V corresponding to the maximum eigenvalue m ;

[0068] Step 4.6: Group the matrix Xi The corresponding eigenvector V mi Multiply to get the fused vector:

[0069] F i =X i V mi .

[0071] The step 5 further comprises:

[0072] Step 5.1: Use a multi-channel convolutional network to mine the features of the historical photovoltaic power data and use it as the dynamic input of the subsequent FPN network. The historical photovoltaic power data is convolved once, passed through a residual network channel, and then passed through a convolution layer and a pooling layer. The features of multiple channels are flattened and spliced. The convolution kernel sizes are 3*3, 5*5, and 7*7 respectively, and convolution kernels of different sizes are used to extract features.

[0073] Step 5.2: The FPN network is divided into three parts: bottom-up, top-down and lateral connection through multi-layer sampling. The reduced-dimensional features and the photovoltaic power history data are input into the FPN network together, and divided into multiple different layers according to the bottom-up process. The scale ratio of the feature map between each layer differs by 2.

[0074] Step 5.3: Through the top-down process and lateral connection, the small feature map of the top layer is enlarged to the same size as the feature map of the previous layer by upsampling, and then the features of the previous layer with the same resolution as the current layer after upsampling are fused by adding;

[0075] Step 5.4: The historical PV power data features and the reduced-dimensional meteorological and electrical features are integrated into a feature map as the dynamic input of the FPN network, and then the final PV power prediction result is obtained through multiple fully connected layers.

[0076] Another aspect of the present invention discloses a prediction system based on MVO optimized hybrid kernel PCA dimensionality reduction and improved FPN network, including a mutual information acquisition module, a heuristic grouping module, a kernel function construction optimization module, a data fusion dimensionality reduction module and an FPN network prediction module, characterized in that:

[0077] The mutual information acquisition module is used to discretize all feature quantities of the photovoltaic power high-dimensional feature set by symbolic analysis to obtain a symbol sequence, and divide it into a training set and a test set after normalization;

[0078] The heuristic grouping module is used to form a mutual information matrix from all mutual information values, and to group all variables using a heuristic grouping algorithm based on the mutual information matrix;

[0079] The kernel function construction optimization module is used to construct a hybrid kernel function according to the Gaussian kernel function and the Laplace kernel function, and use the MVO optimization algorithm to determine the optimal parameter combination of the hybrid kernel function;

[0080] The data fusion and dimension reduction module is used to substitute the optimal parameter combination into the hybrid kernel function, perform data fusion on the grouped variables through the hybrid kernel function, and obtain the features after data dimension reduction.

[0081] The FPN network prediction module is used to combine the features after dimensionality reduction and the features of photovoltaic power historical data mined by using a multi-channel convolutional network, and use the FPN network to perform deep feature mining and fusion to obtain the final prediction result of photovoltaic power.

[0082] Accordingly, the present application also discloses a terminal, including a processor and a storage medium;

[0083] The storage medium is used to store instructions;

[0084] The processor is used to operate according to the instructions to execute the steps of the photovoltaic power short-term prediction method based on the aforementioned MVO optimized hybrid kernel PCA dimensionality reduction and improved FPN network.

[0085] Accordingly, the present application also discloses a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the aforementioned photovoltaic power short-term prediction method based on MVO optimized hybrid kernel PCA dimensionality reduction and improved FPN network.

[0086] The beneficial effect of the present invention is that, compared with the prior art, the present invention provides a photovoltaic power short-term prediction method and system based on MVO optimized hybrid kernel PCA dimensionality reduction and improved FPN network.

[0087] In big data mining, mutual information value is used as the criterion for variable grouping, which can characterize the linear correlation and nonlinear correlation between multidimensional variables. Compared with a single kernel, the hybrid kernel function is better at solving complex structure data problems. Mixing the Gaussian kernel function and the Laplace kernel function can comprehensively consider the local and global characteristics of the data. This allows KPCA to more effectively perform feature extraction and dimensionality reduction on photovoltaic power data sets, improving the robustness and generalization ability of the model. The multi-universe optimizer MVO is used to find the global optimal solution of the hybrid kernel function parameters, which has the advantages of global search capability, adaptability, parallel computing capability and robustness. By reducing the dimensionality of the specific feature data after heuristic mutual information grouping and reducing the joint correlation between multiple variables, it can be applied to machine learning to optimize the input information of the prediction model and provide better performance.

[0088] In the photovoltaic power prediction model, an improved FPN network is used. The deep features of the photovoltaic power historical data are first mined through a multi-channel convolutional network, and then these features and the related feature data after dimensionality reduction are used as the dynamic input of the feature pyramid network FPN. The correlation between the data is fully mined to obtain deep dynamic features, and finally the prediction results are obtained through multiple fully connected layers. BRIEF DESCRIPTION OF THE DRAWINGS

[0089] Figure 1 It is a flow chart of the photovoltaic power short-term prediction method based on MVO optimized hybrid kernel PCA dimensionality reduction and improved FPN network in the present invention.

[0090] Figure 2 It is a schematic diagram of the mutual information-based heuristic grouping algorithm in the present invention.

[0091] Figure 3 This is an improved FPN network model diagram.

[0092] Figure 4 It is a multi-channel convolutional network graph.

[0093] Figure 5 It is a dimensionality reduction result diagram for multi-dimensional data in the present invention.

[0094] Figure 6 It is a schematic diagram of the stability results of the heuristic algorithm of the present invention.

[0095] Figure 7 This is the photovoltaic power prediction result diagram. DETAILED DESCRIPTION

[0096] In order to make the purpose, technical solution and advantages of the present invention more clear, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0097] The embodiments described in this application are only some embodiments of the present invention, not all embodiments. Based on the spirit of the present invention, other embodiments obtained by ordinary technicians in this field without creative work are all within the protection scope of the present invention.

[0098] In view of the deficiencies in the prior art, the present invention proposes a photovoltaic power short-term prediction method and system based on MVO optimized hybrid kernel PCA dimensionality reduction and improved FPN network.

[0099] The present invention uses the mutual information matrix as the criterion for grouping feature quantities, and adopts a heuristic algorithm to group all variables. Then, a hybrid kernel PCA is used to fuse multiple feature quantities in each group into a principal component as a higher-level feature. Among them, the parameters of the hybrid kernel use the multi-universe optimizer MVO to select the optimal parameter combination. The result of each symbolized mutual information heuristic grouping is fused into a principal component by MVO-MKPCA and used as the result of the group after dimensionality reduction. Then, using the improved FPN network, the deep features of the photovoltaic power historical data are first mined through a multi-channel convolutional network, and then these features and the related feature data after dimensionality reduction are used as the dynamic input of the FPN network, and then the photovoltaic power prediction results are finally obtained through multiple fully connected layers.

[0100] See also Figure 1 As shown, the present invention discloses a photovoltaic power short-term prediction method based on MVO optimized hybrid kernel PCA dimensionality reduction and improved FPN network, comprising the following steps:

[0101] Step 1: discretize all feature quantities of the photovoltaic power high-dimensional feature set by symbolic analysis to obtain a symbol sequence, and divide it into a training set and a test set after normalization;

[0102] According to a specific embodiment, the step 1 further comprises:

[0103] Step 1.1: Normalize the photovoltaic power and its feature set, and transform the time series X(n) = {x1, x2, ..., x n} is mapped to the distribution space from 0 to 1, and we get Y(n)={y1,y2,...,y n The high-dimensional feature set of photovoltaic power includes meteorological data features such as temperature, humidity, wind speed, irradiance, and its own electrical property data features such as battery conversion efficiency, inverter conversion efficiency, and battery surface cleanliness;

[0104] Step 1.2: Encode the distribution space into a symbol time sequence containing d symbols:

[0105]

[0106] After symbolization, the time series X(n) and Y(n) are obtained into symbol sequences S1(m1) and S2(m2).

[0107] Step 1.3: Divide the photovoltaic power data set into a training set and a test set in a ratio of 4:1.

[0108] Step 2: All mutual information values ​​between photovoltaic power characteristic data symbol sequences are combined into a mutual information matrix, and based on the mutual information matrix, all variables are grouped by a heuristic grouping algorithm.

[0109] The mutual information value between the feature quantities is calculated according to the symbol sequence, and is used as a standard for measuring the similarity between the variables. The feature quantities are grouped using a heuristic grouping algorithm. Figure 2 The figure is a schematic diagram of the mutual information heuristic grouping algorithm. First, the mutual information between all variables is calculated to obtain the mutual information matrix. Then the complementary matrix of the mutual information is solved. The number of groups is set, and the grouping results are obtained through the heuristic grouping algorithm.

[0110] The heuristic grouping algorithm based on mutual information is as follows:

[0111] Step 2.1: Get the mutual information from the symbol sequences S1(m1) and S2(m2):

[0112]

[0113] N=m1×m2 (4)

[0114] where p ij is the joint probability, i.e., the total number of times k that the combination of the i-th symbol in sequence S1(m1) and the j-th symbol in S2(m2) occurs ij The proportion of the entire sequence combination N, m1 and m2 represent the number of symbolized sequences S1(m1) and S2(m2) respectively.

[0115] Step 2.2: Get the input N data objects, represented as a set of d-dimensional vectors;

[0116] Step 2.3: Initialize the number of classifications, i.e. the final number of dimensionality reduction K and the maximum number of iterations n;

[0117] Step 2.4: For each pair of data objects i and j, calculate the mutual information between them and store the result in the mutual information matrix M;

[0118] Step 2.5: Normalize the mutual information matrix M to obtain the normalized mutual information matrix NM;

[0119] Step 2.6: Initialize an empty grouped list C and an empty ungrouped object list U;

[0120] Step 2.7: Add all data objects to the ungrouped object list U;

[0121] Step 2.8: When the ungrouped object list U is not empty, perform the following steps:

[0122] Select an object u from the ungrouped object list U as a seed object;

[0123] Remove the seed object u from the ungrouped object list U and add it to the grouped list C;

[0124] Iterate the following steps 2.81 to 2.83 until no object can be added to any group or the maximum number of iterations n is met:

[0125] Step 2.81: For each object v in the ungrouped list U, calculate the difference between object v and the grouped list

[0126] The standard measure function for all objects in C is the average of the mutual information within the group, and the result is stored in the object

[0127] The average mutual information of v is MI(v);

[0128] Step 2.82: Select the object v with the largest average mutual information MI(v), and calculate the average mutual information AvgMI(v) between the object v and all the objects in the group list C;

[0129] Step 2.83: If AvgMI(v) is greater than the threshold T, then add object v to the list C and delete object v from the list of ungrouped objects U;

[0130] Step 2.9: If the number of objects in list C is less than K, increase the value of K and return to step 2.9;

[0131] Step 2.10: Output a list of objects in each group.

[0132] Step 3: construct a hybrid kernel function based on the Gaussian kernel function and the Laplace kernel function, and use the MVO optimization algorithm to determine the optimal parameter combination of the hybrid kernel function.

[0133] The superiority of MKPCA lies in the selection of kernel function. Different from using a single kernel function, the commonly used Gaussian kernel function and Laplace kernel function are combined and weighted to establish the kernel function. The MVO optimization algorithm is used to select the parameters of the hybrid kernel function, so that the MVO-PCA method is used to fuse data for each group in step 2, and the main component after MKPCA fusion is used as the result of the group after dimensionality reduction, thereby completing the dimensionality reduction. The specific process is:

[0134] Step 3.1: Define the mixing kernel function:

[0135]

[0136] in, is the Gaussian kernel function, is the Laplace kernel function, α is the weight coefficient for adjusting the Gaussian kernel function and the Laplace kernel function. α, σ1 and σ2 are all weight coefficients to be determined in the hybrid kernel function.

[0137] Step 3.2: Consider the vector after KPCA dimensionality reduction as the reconstructed vector, and use the reconstruction error as the performance measure of KPCA. Then define the objective function of MVO as:

[0138]

[0139] Among them, X represents the original vector before dimensionality reduction, X r Represents the reconstructed vector, and N represents the number of data samples. By minimizing the objective function, the optimal adjustment coefficient α, kernel function hyperparameters σ1 and σ2 are found to minimize the reconstruction error of MKPCA, thereby improving the performance of MKPCA.

[0140] Step 3.3: Set the parameter range and initial solution. For the adjustment coefficient α, the value range is [0, 2]. In order to speed up the convergence speed, it is set to [0.25, 1.99] based on experience. The initial solution is randomly selected within this range. The selection range of σ1 and σ2 is [0.1, 10]. Each universe represents a set of initial values ​​of the mixing kernel function parameters.

[0141] Step 3.4: Calculate the gravitational force between universes based on the fitness value of each universe. The calculation of gravitational force usually depends on the distance between universes and the difference in fitness values. Universes with better fitness values ​​will experience greater gravitational force.

[0142] Step 3.5: Propagate the calculated gravitational force to other universes to adjust the position of the universe. This process is similar to the interaction between universes, where the better universe will exert a greater gravitational force on the surrounding universes, thereby attracting them.

[0143] Step 3.6: According to the results of gravitational propagation and the range of unknown variables, a new solution is generated by fine-tuning and deforming the original solution.

[0144] Step 3.7: For the new solution generated, recalculate its fitness value. If the fitness value of the new solution is better, update the original solution to the new solution, otherwise it remains unchanged.

[0145] Step 3.8: Iterate steps 3.4 to 3.7 until the stopping condition is met. The stopping condition can be reaching the maximum number of iterations or the convergence of the fitness value.

[0146] Step 3.9: Select the hybrid kernel function parameter combination with the best fitness from the last generation universe as the optimal solution. According to specific needs, multiple optimal solutions can be set. Extract the best individual, that is, after the iteration, extract the best parameter vector from the best individual, that is, the optimal hybrid kernel function parameter.

[0147] Step 4: Substitute the optimal parameter combination into the hybrid kernel function, perform data fusion on the grouped variables through the hybrid kernel function, and obtain the features after data dimensionality reduction.

[0148] After finding the optimal parameter combination of the hybrid kernel function of MKPCA through the MVO optimization method, the specific parameter values ​​are substituted into the hybrid kernel function, and the optimized MKPCA is used to fuse the heuristic grouped data, and multiple feature quantities in each group are fused into a principal component as a higher-level feature. The specific steps are:

[0149] Step 4.1: Using the normalized data in step 1.1 and the optimal parameter value of the mixed kernel function in step 3, the mixed kernel function K of MKPCA can be determined. mix (x, y), and calculate the inner product of the kernel function of the normalized data matrix to obtain the kernel matrix W.

[0150] Step 4.2: Calculate the centering matrix W' based on the kernel matrix W:

[0151] W'=W-BW-WB+BWB (7)

[0152] in, l is the sample size.

[0153] Step 4.3: Matrix Perform principal component analysis (PCA) to obtain the eigenvalue λ i and the eigenvector V i , the eigenvalue represents the importance of the principal component, and the eigenvector represents the corresponding principal component.

[0154] Step 4.4: For the eigenvalue λ i Arrange in descending order and replace the corresponding eigenvector V i Sort by order of eigenvalues.

[0155] Step 4.5: Take out the eigenvector V corresponding to the maximum eigenvalue m .

[0156] Step 4.6: Group the matrix X i The corresponding eigenvector V mi Multiply to get the final fused vector:

[0157] F i =X i V mi (9)

[0158] Step 5: Combine the reduced-dimensional features with the photovoltaic power historical data features mined using a multi-channel convolutional network, and use the improved dynamic input FPN network to perform deep feature mining and fusion to obtain the final photovoltaic power prediction result.

[0159] The input feature map of the traditional FPN network is fixed, but the multi-channel convolutional network improves and optimizes its input and turns it into a dynamic input feature map, so that feature fusion and mining can be better performed to obtain the final prediction result. The specific architecture of the FPN network combined with the multi-channel convolutional network is as follows: Figure 3 shown.

[0160] The specific steps are:

[0161] Step 5.1: Use a multi-channel convolutional network to mine the features of the historical photovoltaic power data and use it as the dynamic input of the subsequent FPN network. Specifically, the historical photovoltaic power data is convolved once, passed through a residual network channel, and then passed through a convolution layer and a pooling layer. Finally, the features of multiple channels are flattened and spliced. The convolution kernel sizes are 3*3, 5*5, and 7*7 respectively. Convolution kernels of different sizes are used to extract features as much as possible. The multi-channel convolutional network is as follows: Figure 4 As shown;

[0162] Step 5.2: The FPN network is divided into three parts: bottom-up, top-down and horizontal connection through multi-layer sampling. The reduced-dimensional features and the photovoltaic power historical data features are input into the FPN network together, and divided into multiple different layers according to the bottom-up process. The scale ratio of the feature map between each layer differs by 2.

[0163] Step 5.3: Through the top-down process and lateral connection, the small feature map of the top layer is enlarged to the same size as the feature map of the previous layer by upsampling, and then the features of the previous layer with the same resolution as the current layer after upsampling are fused by adding. In order to correct the number of channels, the current layer is generally subjected to a 1*1 convolution operation before fusion.

[0164] Step 5.4: The historical PV power data features and the reduced-dimensional meteorological and electrical features are integrated into a feature map as the dynamic input improvement of the FPN network, and then the final PV power prediction result is obtained through multiple fully connected layers.

[0165] The beneficial effect of the present invention is that, compared with the prior art, the present invention provides a photovoltaic power short-term prediction method and system based on MVO optimized hybrid kernel PCA dimensionality reduction and improved FPN network, which uses mutual information value as the criterion for variable grouping, and can characterize the linear correlation and nonlinear correlation between multidimensional variables. Compared with a single kernel, mixing the Gaussian kernel function and the Laplace kernel function can comprehensively consider the local and global characteristics of the data. The multi-universe optimizer MVO is used to find the global optimal solution of the hybrid kernel function parameters, which has the advantages of global search capability, adaptability, parallel computing capability and robustness. The improved FPN network uses the reduced feature set and the dynamic characteristics of the multi-channel convolution historical photovoltaic data values ​​as fusion input, which can effectively mine data information and improve the prediction accuracy.

[0166] In order to demonstrate that the present invention can effectively mine the high-dimensional feature set of photovoltaic power, five high-dimensional data sets with different characteristics are used to verify as benchmark data. The data sets are from public data sets in the UCI repository applied to different fields of physics, life sciences and computer. The benchmark data table is shown in Table 1:

[0167] Table 1

[0168]

[0169] The number of features and the number of samples in the data set both provide useful feature indicators for the classifier, and these features affect the classification accuracy. Therefore, the accuracy of the classifier serves as a proof of the dimensionality reduction effect of the method of the present invention. In order to prove the generalization ability of the dimensionality reduction method of the present invention, three classifiers, namely random forest (RF), k nearest neighbor (KNN) and support vector machine (SVM), were used in the experiment. Due to the different number of features in different data sets, the percentage of dimensionality reduction is determined by multiplying the number of features in the original data set by different proportions, including 5%, 10%, 15%, 20%, 30%, 40% and 50%. The feature quantities of different dimensionality reduction are divided into training and test data sets in a ratio of 7:3, and these data sets are fed into RF, KNN and SVM classifiers for learning. The probability of correct classification (i.e., accuracy) is used as an intuitive observation result, and this process is repeated 15 times. Figure 4 The classification effects of the three classifiers on the selected data sets under different dimensionality reduction methods are comprehensively demonstrated. Figure 5 Contains 5*3 subgraphs, which contain three types of information. Each row represents the classification performance of the dataset in RF, KNN and SVM classifiers. Each column represents the classification effect of five datasets under the same classifier. The horizontal axis of each subgraph represents the percentage of the final dimension reduction to the total number of features (rounded off), and the vertical axis represents the classification accuracy.

[0170] exist Figure 5 In the figure, the method of the present invention (the line with an upward triangle) is superior and stable compared to other methods, and its performance is even more obvious on the Sonar and Musk datasets. In the third column, due to the large number of samples in SVM, when solving the support vector, the storage and calculation of the matrix will consume a lot of machine memory and computing time. The method of the present invention just makes up for this deficiency. Compared with other dimensionality reduction methods, under the 5% dimension selection, the method of the present invention can still extract key feature information, so that the classification effect is still very ideal and the accuracy is the highest.

[0171] Considering that the heuristic grouping algorithm is prone to fall into the local optimum during the iteration process (that is, the optimization result is related to the selection of the initial center), a box plot of 15 experimental results was drawn using the SVM classifier as a representative to reflect the fluctuation of the results, as shown in the figure. Figure 6 As shown. Figure 6 In the above data, except for the SCADI data set, the results of the other data sets are relatively stable, and the highest accuracy can reach 100%. Figure 5 In the SCADI data set, the comparison methods ICA and FA have poor results and great fluctuations. However, the method of the present invention is still stable with an accuracy rate of more than 87%. Figure 6 It can be seen that there are two obvious outliers when the dimension is reduced by 20% and 50%. Although the present invention does not delete this value when obtaining the average result, the accuracy is 88.019% and 87.067% respectively, which is still higher than other comparison methods. Moreover, in the upper boundary of the experimental results, all the accuracy rates can reach more than 90%. The above analysis shows that the algorithm of the present invention can approach the optimal global classification effect through multiple iterations. For data with difficult-to-distinguish features, multiple experiments can reduce contingency to improve the accuracy of the test set. While reducing the data dimension, the feature information can be effectively retained, and it performs superiorly compared with other methods.

[0172] Figure 7 In order to use the improved FPN network to perform feature fusion for photovoltaic power prediction results, the inputs of the comparison networks GRU, LSTM and BiLSTM are consistent with the inputs of the FPN network, which are all photovoltaic power historical data and feature sets after dimensionality reduction. The photovoltaic power historical data of the previous 16 days are used to predict the photovoltaic power on the 17th day. It can be seen that the use of FPN network to fuse the feature information after dimensionality reduction can better mine the dynamic feature associations of the data and improve the prediction accuracy.

[0173] Table 2 below shows Figure 7 Comparison data table, comparing the FPN model of the present invention with the GRU model, LSTM model and BiLSTM model. Comparison of MAE, RMSE and R in the same time period 2The model of the present invention has the best comprehensive performance. The MAE of the model of the present invention is reduced by 60.6%, 44.1% and 33.9% compared with GRU, LSTM and BiLSTM, and the RMSE is reduced by 75.6%, 53.3% and 42.0% respectively. 2 The index value is 0.9902, indicating that the prediction result is highly consistent with the true value.

[0174] Table 2

[0175]

[0176] The present invention may be a system, method and / or computer program product. The present invention also discloses a photovoltaic power short-term prediction method based on MVO optimized hybrid kernel PCA dimensionality reduction and improved FPN network, and a prediction system based on MVO optimized hybrid kernel PCA dimensionality reduction and improved FPN network, including a mutual information acquisition module, a heuristic grouping module, a kernel function construction optimization module, a data fusion dimensionality reduction module and an FPN network prediction module.

[0177] The mutual information acquisition module is used to calculate the mutual information value between the feature quantities according to the symbol sequence;

[0178] The heuristic grouping module is used to form a mutual information matrix from all mutual information values, and to group all variables using a heuristic grouping algorithm based on the mutual information matrix;

[0179] The kernel function construction optimization module is used to construct a hybrid kernel function according to the Gaussian kernel function and the Laplace kernel function, and use the MVO optimization algorithm to determine the optimal parameter combination of the hybrid kernel function;

[0180] The data fusion and dimension reduction module is used to substitute the optimal parameter combination into the hybrid kernel function, perform data fusion on the grouped variables through the hybrid kernel function, and obtain the features after data dimension reduction.

[0181] The FPN network prediction module is used to combine the features after dimensionality reduction and the features of photovoltaic power historical data mined by using a multi-channel convolutional network, and use the FPN network to perform deep feature mining and fusion to obtain the final prediction result of photovoltaic power.

[0182] Based on the spirit of the present invention, those skilled in the art can easily think that a computer program product can be obtained based on the aforementioned photovoltaic power short-term prediction method based on MVO optimized hybrid kernel PCA dimensionality reduction and improved FPN network. The computer program product may include a computer-readable storage medium, which carries computer-readable program instructions for enabling a processor to implement various aspects of the present disclosure. That is, the present application also includes a terminal, including a processor and a storage medium; the storage medium is used to store instructions; the processor is used to operate according to the instructions to execute the steps of the aforementioned photovoltaic power short-term prediction based on MVO optimized hybrid kernel PCA dimensionality reduction and FPN network and photovoltaic power short-term prediction method of FPN network.

[0183] Computer readable storage medium can be a tangible device that can keep and store the instructions used by the instruction execution device. Computer readable storage medium can be, for example, - but not limited to - electrical storage device, magnetic storage device, optical storage device, electromagnetic storage device, semiconductor storage device or any suitable combination of the above. More specific examples (non-exhaustive list) of computer readable storage medium include: portable computer disk, hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disk read-only memory (CD-ROM), digital versatile disk (DVD), memory stick, floppy disk, mechanical encoding device, for example, punch card or groove protrusion structure with instructions stored thereon and any suitable combination of the above. Computer readable storage medium used here is not interpreted as instantaneous signal itself, such as radio wave or other free propagating electromagnetic wave, electromagnetic wave propagated by waveguide or other transmission medium (for example, light pulse by optical fiber cable) or electrical signal transmitted by wire.

[0184] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to each computing / processing device, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network can include copper transmission cables, optical fiber transmissions, wireless transmissions, routers, firewalls, switches, gateway computers, and / or edge servers. The network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in the computer-readable storage medium in each computing / processing device.

[0185] The computer program instructions for performing the operation of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-related instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages-such as Smalltalk, C++, etc., and conventional procedural programming languages-such as "C" language or similar programming languages. Computer-readable program instructions may be executed completely on a user's computer, partially on a user's computer, as an independent software package, partially on a user's computer, partially on a remote computer, or completely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer via any type of network-including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., using an Internet service provider to connect via the Internet). In some embodiments, an electronic circuit, such as a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA) may be personalized by utilizing the state information of a computer-readable program instruction, and the electronic circuit may execute a computer-readable program instruction, thereby realizing various aspects of the present disclosure.

[0186] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents, and any modifications or equivalent replacements that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A photovoltaic power short-term prediction method based on MVO optimized hybrid kernel PCA dimensionality reduction and improved FPN network, characterized in that: The following steps are involved: Step 1: discretize all feature quantities of the photovoltaic power high-dimensional feature set by symbolic analysis to obtain a symbol sequence, and divide it into a training set and a test set after normalization; Step 2, forming a mutual information matrix from all mutual information values ​​between photovoltaic power characteristic data symbol sequences, and grouping all variables using a heuristic grouping algorithm based on the mutual information matrix; Step 3, constructing a hybrid kernel function according to the Gaussian kernel function and the Laplace kernel function, and using the MVO optimization algorithm to determine the optimal parameter combination of the hybrid kernel function; Step 4, substituting the optimal parameter combination into the hybrid kernel function, performing data fusion on the grouped variables through the hybrid kernel function, and obtaining features after data dimension reduction; Step 5 is used to combine the features after dimensionality reduction and the features of photovoltaic power historical data mined by using a multi-channel convolutional network, and use the improved FPN network to perform deep feature mining and fusion to obtain the final prediction result of photovoltaic power.

2. According to claim 1, a photovoltaic power short-term prediction method based on MVO optimized hybrid kernel PCA dimensionality reduction and improved FPN network is characterized in that: The step 1 further comprises: Step 1.1: Normalize the photovoltaic power and its feature set, and transform the time series X(n) = {x1, x2, ..., x n } is mapped to the distribution space from 0 to 1, and we get Y(n)={y1,y2,...,y n }; Step 1.2: Encode the distribution space into a symbol time sequence containing d symbols: After symbolization of the time series X(n) and Y(n), symbol sequences S1(m1) and S2(m2) are obtained; Step 1.3: Divide the photovoltaic power data set into a training set and a test set in a ratio of 4:

1.

3. According to claim 2, a photovoltaic power short-term prediction method based on MVO optimized hybrid kernel PCA dimensionality reduction and improved FPN network is characterized in that: The step 2 further comprises: Step 2.1: Get the mutual information from the symbol sequences S1(m1) and S2(m2): N=m1×m2 where p ij is the joint probability, i.e., the total number of times k that the combination of the i-th symbol in sequence S1(m1) and the j-th symbol in S2(m2) occurs ij The proportion of the total sequence combination N, m1 and m2 represent the number of symbolizations of sequences S1(m1) and S2(m2) respectively; Step 2.2: Get the input N data objects, represented as a set of d-dimensional vectors; Step 2.3: Initialize the number of classifications, i.e. the final number of dimensionality reduction K and the maximum number of iterations n; Step 2.4: For each pair of data objects i and j, calculate the mutual information between them and store the result in the mutual information matrix M; Step 2.5: Normalize the mutual information matrix M to obtain the normalized mutual information matrix NM; Step 2.6: Initialize an empty grouped list C and an empty ungrouped object list U; Step 2.7: Add all data objects to the ungrouped object list U; Step 2.8: When the ungrouped object list U is not empty, perform the following steps: Select an object u from the ungrouped object list U as a seed object; Remove the seed object u from the ungrouped object list U and add it to the grouped list C; Iterate the following steps 2.81 to 2.83 until no object can be added to any group or the maximum number of iterations n is met: Step 2.81: For each object v in the ungrouped object list U, calculate the standard measure function of object v and all objects in the grouped list C, that is, the average of the mutual information within the group, and store the result in the mutual information average MI(v) of object v; Step 2.82: Select the object v with the largest average mutual information MI(v), and calculate the average mutual information AvgMI(v) between the object v and all the objects in the group list C; Step 2.83: If AvgMI(v) is greater than the threshold T, then add object v to the list C and delete object v from the list of ungrouped objects U; Step 2.9: If the number of objects in list C is less than K, increase the value of K and return to step 2.9; Step 2.10: Output a list of objects in each group.

4. According to claim 3, a photovoltaic power short-term prediction method based on MVO optimized hybrid kernel PCA dimension reduction and improved FPN network is characterized in that: The step 3 further comprises: Step 3.1: Define the mixing kernel function: in, is the Gaussian kernel function, is the Laplace kernel function, α, σ1 and σ2 are weight coefficients to be determined in the mixed kernel function; Step 3.2: Consider the vector after KPCA dimensionality reduction as the reconstructed vector, and use the reconstruction error as the performance measure of KPCA. Then define the objective function of MVO as: Among them, X represents the original vector before dimensionality reduction, X r Represents the reconstructed vector, and N represents the number of data samples. By minimizing the objective function, the optimal adjustment coefficient α, kernel function hyperparameters σ1 and σ2 are found to minimize the reconstruction error of MKPCA. Step 3.3: For the adjustment coefficient α, the value range is [0, 2], the initial solution is randomly selected within this range, and the selection range of σ1 and σ2 is [0.1, 10]. Each universe represents a set of initial values ​​of the mixing kernel function parameters; Step 3.4: Calculate the gravitational force between universes based on the fitness value of each universe; Step 3.5: Propagate the calculated gravity to other universes to adjust the position of the universe; Step 3.6: According to the results of gravitational propagation and the range of unknown variables, a new solution is generated by fine-tuning and deforming the original solution; Step 3.7: For the new solution generated, recalculate its fitness value; if the fitness value of the new solution is better, update the original solution to the new solution, otherwise it remains unchanged; Step 3.8: Iterate step 3.4 to step 3.7 until the stopping condition is met, wherein the stopping condition is that the maximum number of iterations is reached or the fitness value converges; Step 3.9: Select the hybrid kernel function parameter combination with the best fitness from the last generation of the universe as the optimal solution, that is, the optimal hybrid kernel function parameters.

5. According to claim 4, a photovoltaic power short-term prediction method based on MVO optimized hybrid kernel PCA dimensionality reduction and improved FPN network is characterized in that: The step 4 further comprises: Step 4.1: Determine the mixing kernel function K of MKPCA using the normalized data and the optimal parameter value of the mixing kernel function mix (x, y), and calculate the inner product of the kernel function of the normalized data matrix to obtain the kernel matrix W; Step 4.2: Calculate the centering matrix W' based on the kernel matrix W: W'=W-BW-WB+BWB in, l is the sample size; Step 4.3: Matrix Perform principal component analysis (PCA) to obtain the eigenvalue λ i and the eigenvector V i ; Step 4.4: For the eigenvalue λ i Arrange in descending order and replace the corresponding eigenvector V i Sort by the order of eigenvalues; Step 4.5: Take out the eigenvector V corresponding to the maximum eigenvalue m ; Step 4.6: Group the matrix X i The corresponding eigenvector V mi Multiply to get the fused vector: F i =X i V mi 。 6. The photovoltaic power short-term prediction method based on MVO optimized hybrid kernel PCA dimensionality reduction and improved FPN network according to claim 5 is characterized in that: The step 5 further comprises: Step 5.1: Use a multi-channel convolutional network to mine the features of the historical photovoltaic power data and use it as the dynamic input of the subsequent FPN network. The historical photovoltaic power data is convolved once, passed through a residual network channel, and then passed through a convolution layer and a pooling layer. The features of multiple channels are flattened and spliced. The convolution kernel sizes are 3*3, 5*5, and 7*7 respectively, and convolution kernels of different sizes are used to extract features. Step 5.2: The FPN network is divided into three parts: bottom-up, top-down and lateral connection through multi-layer sampling. The reduced-dimensional features and the photovoltaic power history data are input into the FPN network together, and divided into multiple different layers according to the bottom-up process. The scale ratio of the feature map between each layer differs by 2. Step 5.3: Through the top-down process and lateral connection, the small feature map of the top layer is enlarged to the same size as the feature map of the previous layer by upsampling, and then the features of the previous layer with the same resolution as the current layer after upsampling are fused by adding; Step 5.4: The historical PV power data features and the reduced-dimensional meteorological and electrical features are integrated into a feature map as the dynamic input of the FPN network, and then the final PV power prediction result is obtained through multiple fully connected layers.

7. A prediction system based on MVO optimized hybrid kernel PCA dimensionality reduction and improved FPN network, comprising a mutual information acquisition module, a heuristic grouping module, a kernel function construction optimization module, a data fusion dimensionality reduction module and an FPN network prediction module, characterized in that: The mutual information acquisition module is used to discretize all feature quantities of the photovoltaic power high-dimensional feature set by symbolic analysis to obtain a symbol sequence, and divide it into a training set and a test set after normalization; The heuristic grouping module is used to form a mutual information matrix from all mutual information values, and to group all variables using a heuristic grouping algorithm based on the mutual information matrix; The kernel function construction optimization module is used to construct a hybrid kernel function according to the Gaussian kernel function and the Laplace kernel function, and use the MVO optimization algorithm to determine the optimal parameter combination of the hybrid kernel function; The data fusion and dimension reduction module is used to substitute the optimal parameter combination into the hybrid kernel function, perform data fusion on the grouped variables through the hybrid kernel function, and obtain the features after data dimension reduction; The FPN network prediction module is used to combine the features after dimensionality reduction and the features of photovoltaic power historical data mined by using a multi-channel convolutional network, and use the FPN network to perform deep feature mining and fusion to obtain the final prediction result of photovoltaic power.

8. The prediction system based on MVO optimized hybrid kernel PCA dimensionality reduction and improved FPN according to claim 7, characterized in that: The mutual information acquisition module is further used to: The photovoltaic power and its feature set are normalized and the time series X(n) = {x1, x2, ..., x n } is mapped to the distribution space from 0 to 1, and we get Y(n)={y1,y2,...,y n }; Encode the distribution space into a symbol time sequence containing d symbols: After symbolization of the time series X(n) and Y(n), symbol sequences S1(m1) and S2(m2) are obtained; The photovoltaic power dataset is divided into a training set and a test set in a ratio of 4:

1.

9. The prediction system based on MVO optimized hybrid kernel PCA dimensionality reduction and improved FPN network according to claim 8, characterized in that: The heuristic grouping module is further used for: Get the mutual information from the symbol sequences S1(m1) and S2(m2): N=m1×m2 where p ij is the joint probability, i.e., the total number of times k that the combination of the i-th symbol in sequence S1(m1) and the j-th symbol in S2(m2) occurs ij The proportion of the total sequence combination N, m1 and m2 represent the number of symbolizations of sequences S1(m1) and S2(m2) respectively; Input N data objects, represented as a set of d-dimensional vectors; Initialize the number of classifications, i.e. the final number of dimensionality reduction K and the maximum number of iterations n; For each pair of data objects i and j, calculate the mutual information between them and store the result in the mutual information matrix M; Normalize the mutual information matrix M to obtain the normalized mutual information matrix NM; Initialize an empty grouped list C and an empty ungrouped object list U; Add all data objects to the ungrouped object list U; When the ungrouped object list U is not empty, perform the following steps: Select an object u from the ungrouped object list U as a seed object; Remove the seed object u from the ungrouped object list U and add it to the grouped list C; Iterate the following steps until no object can be added to any group or the maximum number of iterations n is met: for each object v in the ungrouped object list U, calculate the standard measure function between object v and all objects in the grouped list C, that is, the average value of the mutual information within the group, and store the result in the mutual information average value MI(v) of object v; select the object v with the largest mutual information average value MI(v), and calculate the average value AvgMI(v) of the mutual information between object v and all objects in the grouped list C; if AvgMI(v) is greater than the threshold T, add object v to the list C and delete object v from the ungrouped object list U; If the number of objects in list C is less than K, increase the value of K and return to the above iteration step; Output a list of objects in each group.

10. The prediction system based on MVO optimized hybrid kernel PCA dimensionality reduction and improved FPN network according to claim 9, characterized in that: The kernel function constructs an optimization module, which is further used to: Define the mixing kernel function: in, is the Gaussian kernel function, is the Laplace kernel function, α, σ1 and σ2 are weight coefficients to be determined in the mixed kernel function; The vector after KPCA dimensionality reduction is regarded as the reconstructed vector, and the reconstruction error is used as the performance measure of KPCA. Then the objective function of MVO is defined as: Among them, X represents the original vector before dimensionality reduction, X r Represents the reconstructed vector, and N represents the number of data samples. By minimizing the objective function, the optimal adjustment coefficient α, kernel function hyperparameters σ1 and σ2 are found to minimize the reconstruction error of MKPCA. For the adjustment coefficient α, the value range is [0, 2], and the initial solution is randomly selected within this range. The selection range of σ1 and σ2 is [0.1, 10]. Each universe represents a set of initial values ​​of the mixing kernel function parameters; According to the fitness value of each universe, calculate the gravitational force between universes; Propagate the calculated gravity to other universes to adjust the positions of the universes; According to the results of gravitational propagation and the range of unknown variables, a new solution is generated by fine-tuning and deforming the original solution; For the new solution generated, recalculate its fitness value; if the fitness value of the new solution is better, update the original solution to the new solution, otherwise it remains unchanged; Iterate the above steps until a stopping condition is met, wherein the stopping condition is that a maximum number of iterations is reached or the fitness value converges; The hybrid kernel function parameter combination with the best fitness is selected from the last generation of the universe as the optimal solution, that is, the optimal hybrid kernel function parameters.

11. The prediction system based on MVO optimized hybrid kernel PCA dimensionality reduction and improved FPN network according to claim 10, characterized in that: The data fusion dimension reduction module is further used for: Using the normalized data and the optimal parameter value of the mixed kernel function, determine the mixed kernel function K of MKPCA mix (x, y), and calculate the inner product of the kernel function of the normalized data matrix to obtain the kernel matrix W; Calculate the centering matrix W' based on the kernel matrix W: W'=W-BW-WB+BWB in, l is the sample size; Pair Matrix Perform principal component analysis (PCA) to obtain the eigenvalue λ i and the eigenvector V i ; For the eigenvalue λ i Arrange in descending order and replace the corresponding eigenvector V i Sort by the order of eigenvalues; Take out the eigenvector V corresponding to the maximum eigenvalue m ; The grouping matrix X i The corresponding eigenvector V mi Multiply to get the fused vector: F i =X i V mi 。 12. The prediction system based on MVO optimized hybrid kernel PCA dimensionality reduction and FPN network according to claim 11, characterized in that: The FPN network prediction module is further used for: A multi-channel convolutional network is used to mine the features of historical photovoltaic power data, and used as the dynamic input of the subsequent FPN network. The historical photovoltaic power data is convolved once, passed through a residual network channel, and then passed through a convolution layer and a pooling layer. The features of multiple channels are flattened and spliced. The convolution kernel sizes are 3*3, 5*5, and 7*7 respectively, and convolution kernels of different sizes are used to extract features. The FPN network is divided into three parts: bottom-up, top-down and lateral connection through multi-layer sampling. The reduced-dimensional features and photovoltaic power history data are input into the FPN network together, and divided into multiple different layers according to the bottom-up process. The scale ratio of the feature map between each layer differs by 2. Through the top-down process and lateral connection, the small feature map of the top layer is enlarged to the same size as the feature map of the previous layer by upsampling, and then the features of the previous layer that have been upsampled and have the same resolution as the current layer are fused by adding them; The historical photovoltaic power data features and the reduced-dimensional meteorological and electrical features are fused into a feature map as the dynamic input of the FPN network, and then the final photovoltaic power prediction result is obtained through multiple fully connected layers.

13. A terminal comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is used to operate according to the instructions to execute the steps of a photovoltaic power short-term prediction method based on MVO optimized hybrid kernel PCA dimensionality reduction and improved FPN network according to any one of claims 1-6.

14. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of a photovoltaic power short-term prediction method based on MVO optimized hybrid kernel PCA dimensionality reduction and improved FPN network as described in any one of claims 1 to 6 are implemented.

Images