Distributed wind and light power generation prediction method based on weather information reconstruction

Abstract

The invention relates to the technical field of wind power and photovoltaic power generation prediction, in particular to a weather information reconstruction-based distributed wind and light power generation prediction method, which comprises the following steps of: generating a regional radiation thermodynamic diagram by using reflection brightness information of a satellite remote sensing image; a Doppler radar image is utilized to extract radial wind speed and obtain a potential height contour map, a regional wind speed thermodynamic diagram is generated according to contour density, regional average data is combined to carry out proportion correction, and a finite element coupling model is utilized to complete refined joint reconstruction of wind speed and radiation. High-resolution local wind speed and radiation data in the target area are obtained, and accurate data support is provided for follow-up prediction. And screening main weather characteristics influencing wind power and photovoltaic power by using dynamic mutual information and a maximum information coefficient, further performing variational mode decomposition on wind-light cross characteristics, and extracting a multi-scale characteristic mode. The method has high adaptability and wide application potential, and is especially suitable for wind-solar combined power generation power prediction under different weather conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power and photovoltaic power generation prediction, and in particular to a decentralized wind-solar power generation prediction method based on weather information reconstruction. Background Art

[0002] With the rapid development of the new energy industry, renewable energy sources such as wind energy and solar energy have been widely promoted and applied due to their clean and environmental protection characteristics. However, the uncertainty and volatility of wind energy and solar energy resources seriously affect the stability and reliability of the operation of the power system. Therefore, accurately and efficiently predicting the power of wind power and photovoltaic power generation is an important guarantee for the optimal dispatching and safe operation of the power system.

[0003] Currently, most traditional wind-solar power prediction methods directly rely on regional weather observation data or coarse-resolution remote sensing data for prediction, and cannot effectively reflect the subtle changes of wind speed and solar radiation in space, resulting in difficulty in improving the prediction accuracy. In addition, existing technologies also mostly use single indicators or simple correlation analysis methods in feature selection, and it is difficult to accurately capture the dominant features of wind power and photovoltaic power and their cross-influences, further restricting the improvement of prediction performance. Summary of the Invention

[0004] The purpose of the present invention is to solve the deficiencies existing in the prior art, and to propose a decentralized wind-solar power generation prediction method based on weather information reconstruction. This method can effectively extract the main features affecting wind power and photovoltaic power, and further process the high-frequency components in the features through variational mode decomposition (VMD), thereby improving the accuracy of wind-solar power prediction.

[0005] To achieve the above purpose, the present invention adopts the following technical solutions:

[0006] A decentralized wind-solar power generation prediction method based on weather information reconstruction, including wind-solar joint prediction based on weather information reconstruction, feature selection, and VMD decomposition, comprising the following steps:

[0007] S1. Collect the data set and reconstruct the original weather features;

[0008] Collect weather station data, determine the average wind speed V and average radiation value (total radiation G, etc.) of the target area, obtain high-resolution satellite images of the target area, collect Doppler radar images of the target area, use them to invert the radial wind speed information in the area, and obtain the geopotential height contour lines at various points on a certain pressure surface to form a radar wind speed feature map;

[0009] S2. Data loading and preprocessing;

[0010] S3. Feature selection and correlation analysis;

[0011] S4. VMD decomposition of wind-solar cross features;

[0012] S5. Selection of the final feature set;

[0013] S6. Training and prediction model.

[0014] Preferably, in S1, the following steps are included:

[0015] S1.1. Photovoltaic radiation inversion and normalization correction

[0016] Image preprocessing steps: Geometric correction, radiometric correction, and atmospheric correction are performed on the original satellite image, and it is normalized into a reflectance image:

[0017]

[0018] In the formula, I(x,y) is the original image value, and I raw (x,y) is the reflectance after normalization, with a value range of [0,1], and I max and I min are the maximum and minimum pixel values of the image, respectively;

[0019] In the normalized image, with the target point (x,y) as the center, a neighborhood B(x,y) is selected, and the local weighted average brightness is extracted using the Gaussian weight function:

[0020]

[0021] In the formula, σ is the standard deviation parameter. The larger σ is, the "flatter" the weight distribution is, and pixels at a farther distance within the neighborhood can also have a greater impact on the center point brightness; the smaller σ is, the weights are concentrated near the center, and the influence of distant pixels is rapidly attenuated. L(x,y) is the local weighted average brightness, W(x′,y′) is the weight value of the coordinate (x′,y′) with respect to the coordinate (x,y), I(x′,y′) is the brightness value at the coordinate (x′,y′), and exp is the two-dimensional Gaussian distribution kernel, which assigns weights according to the Euclidean distance between pixels. The farther the distance, the smaller the weight. The Gaussian kernel has the characteristics of being smooth and rapidly decaying, and can well retain the local structure and suppress noise;

[0022] Calculate the average brightness within the region:

[0023]

[0024] In the formula, Ω is the region to be inverted, A is the area of the region, is the average brightness of the region;

[0025] Construct a non-linear inversion model to make the local radiation positively correlated with the brightness. To ensure the overall average radiation value of the region Perform normalization correction:

[0026]

[0027] Wherein, G(x,y) is the reconstructed local radiation value, G avg is the regional average radiation value, and γ is the non-linear adjustment parameter;

[0028] S1.2, Wind speed reconstruction and ratio correction method

[0029] From the Doppler radar image, after geometric correction and filtering, extract the radial wind speed information in the region and generate the geopotential height contour map; divide the image into small blocks, count the contour density in each block, and obtain the local heat value index D according to the relationship that the density is proportional to the heat value (x,y) (The normalized value is between 0 and 1), reflecting the wind speed intensity of this region;

[0030] Assume that the average wind speed of the radar sampling region is The heat value of the target small block is D target , the maximum heat value in the region is D max , and the wind speed of the target point is reconstructed proportionally as:

[0031]

[0032] Wherein, V(x,y) is the wind speed after the initial reconstruction of the target point, D target (x,y), D max (x,y) are the heat values of the target block and in the region respectively;

[0033] S1.3, Radiation and wind speed joint finite element coupling model

[0034] To realize the joint reconstruction of wind speed and radiation data, construct the following coupling finite element model:

[0035] Let the local weather parameter vector be:

[0036] u represents the weather parameter to be solved at the point (X,Y), G(x,y) is the radiation value after the initial reconstruction of the target point, and V(x,y) is the wind speed after the initial reconstruction of the target point.

[0037] Construct the coupling model, considering spatial diffusion and data recovery:

[0038]

[0039] Wherein

[0040]

[0041]

[0042] In the formula, K(x, y) is the diffusion coefficient matrix, Q(x, y) is the coupling matrix, and k V (x, y) is the wind speed diffusion coefficient, k G (x, y) is the radiation diffusion coefficient, α V (x, y), α G (x, y) is the regression coefficient, δ(x, y) is the coupling coefficient, reflecting the mutual influence between wind speed and radiation, F(x, y) is the externally applied load vector, including image recognition correction information, u ref (x, y) is the reference data, is the regional average wind speed data, G avg is the regional average radiation value; is the gradient operator, that is is the divergence operator, which maps a vector field to a scalar field, is the gradient of the vector u, understood as taking partial derivatives of the components of V(x, y) and G(x, y) in the X and Y directions respectively, and then multiplying by K(x, y) to obtain the diffusion effect.

[0043] S1.4. Finite element discretization and solution

[0044] Discretize the region Ω into finite elements Ω e (e = 1, 2,..., n):

[0045]

[0046] Within each element, use the shape function N i (x, y) to approximate the reconstructed weather parameter u(x, y):

[0047]

[0048] In the formula, u(x, y) represents the weather parameter to be solved at the point (X, Y) (such as wind speed or radiation value), N i (x, y) is the i-th shape function, describing the weight of the contribution of node i to any point (X, Y) within the element, U i (x, y) is the value of the unknown parameter at the i-th node within the element, and m is the number of nodes within the element;

[0049] Construct the element stiffness matrix:

[0050]

[0051] In the formula, is the matrix element in the i-th row and j-th column of element e, reflecting the mutual influence between node i and node j; is the gradient of the shape function, representing the rate of change of the \(i\)-th shape function in the \(X\) and \(Y\) directions. is the rate of change of the \(j\)-th shape function in the \(X\) and \(Y\) directions;

[0052]

[0053] In the formula, \(F\) i (e) is the load value corresponding to the \(i\)-th node in element \(e\), representing the "force" or "source term" externally acting on this node; \(F(x, y)\) is the body load term, which contains the basic data in the region and the correction information obtained by image inversion, and usually reflects the external input, such as local radiation changes; \(\Gamma\) is the part of the boundary of element \(e\) where the load is applied. is the boundary load, reflecting the boundary conditions;

[0054] Through the assembly matrix \(A\) e (a sparse, 0 / 1 mapping matrix used to map the element node numbers to the global node numbers), the element matrices are assembled into an overall system;

[0055]

[0056] Solve the overall finite element equation: \(KU = F\)

[0057] In the formula, \(K\) is the global stiffness matrix, which is assembled from the local stiffness matrices of each element. \(K\) (e) is the local stiffness matrix of the \(e\)-th element, \(U\) is the global node unknown parameter vector, \(F\) is the global load vector, and \(F\) (e) is the local load vector of the \(e\)-th element, and \(n\) is the number of elements, representing the total number of elements obtained after dividing the entire region into a grid;

[0058] Then the global node vector \(U\) can be obtained, that is, \(U\) i (x, y), and then according to the formula the local data \(u(x, y)\) at any target point can be obtained:

[0059]

[0060] Preferably, in S2, the following steps are included:

[0061] S2.1. Load the wind-solar combined data from the reconstructed dataset and extract relevant weather features (such as wind speed, temperature, humidity, air pressure, photovoltaic irradiance, etc.);

[0062] S2.2. Normalize the feature data and the target variables (wind power and photovoltaic power) to make them have the same dimension to improve the training effect of the model;

[0063] Let the original data sequence be \(x = \{x_1, x_2, \cdots, x\}\), and \(L\) be the total number of samples in the data sequence. Then the normalized data sequence is: L}, where \(x_{max}\) is the maximum value of the data sequence, \(x_{min}\) is the minimum value of the data sequence, and \(\hat{x}\) is the normalized data with a value range of \([0, 1]\).

[0064]

[0065] where \(\hat{x}_i=\frac{x_i - x_{min}}{x_{max}-x_{min}}\) max is the maximum value of the data sequence, \(x\) min is the minimum value of the data sequence, and \(\hat{x}\) is the normalized data with a value range of \([0, 1]\).

[0066] Preferably, in S3, the following steps are included:

[0067] S3.1. Calculate the correlation between each weather feature and the target variable using dynamic mutual information (DMI) and maximum information coefficient (MIC) respectively;

[0068] For the weather feature sequence \(X\) and the power target sequence \(Y\), the dynamic mutual information between them is defined as:

[0069]

[0070] where \(I(X; Y)\) is the dynamic mutual information between sequences \(X\) and \(Y\), \(p(x, y)\) is the joint probability distribution of the feature sequence and the target sequence, and \(p(x)\) and \(p(y)\) are the marginal probability distributions of sequences \(X\) and \(Y\) respectively;

[0071] The maximum information coefficient is defined as:

[0072]

[0073] where \(I\cdot(X, Y, a, b)\) is the maximum mutual information value after dividing the sequences \(X\) and \(Y\) into \(a\times b\) cells by grid;

[0074] When MIC is 0, it means the feature is irrelevant to the target variable, and the closer the value is to 1, the stronger the relationship between the feature and the target variable;

[0075] S3.2. Calculate the influence of each weather feature on photovoltaic power and wind power, and obtain the DMI and MIC scores of each feature;

[0076] S3.3. Calculate the contribution ratio of each feature according to the weighted sum of DMI and MIC scores, and divide the features into three categories: wind power dominant, photovoltaic power dominant, and wind-solar cross features;

[0077] The calculation method for the contribution ratio of weather features to wind power is:

[0078]

[0079] The calculation method for the contribution ratio of weather characteristics to photovoltaic power is as follows:

[0080]

[0081] In the formula, CR i_wind represents the contribution ratio of the i-th characteristic to wind power, and CR i_pv represents the contribution ratio of the i-th characteristic to photovoltaic power. DMI wind,i and MIC wind,i are the dynamic mutual information and maximum information coefficient of the i-th characteristic and wind power; DMI pv,i and MIC pv,i are the dynamic mutual information and maximum information coefficient of the i-th characteristic and photovoltaic power;

[0082] Classify the characteristics according to the contribution ratio: CR i_wind ≥0.7, CR i_pv ≤0.4 indicates wind power dominant characteristics, CR i_wind ≤0.4, CR i_pv ≥0.7 indicates photovoltaic power dominant characteristics, 0.15 ≤ CR i_wind ≤0.7, 0.15 ≤ CR i_pv ≤0.7 for the rest indicates cross characteristics.

[0083] Preferably, in S4, it includes the following steps:

[0084] S4.1. Perform variational mode decomposition (VMD) on the wind-solar cross characteristics, and decompose these characteristics into multiple modal signals to capture different frequency components in the data;

[0085] VMD obtains the characteristic modes by optimizing the following objective function:

[0086]

[0087] In the formula, u k (t) is the k-th modal signal obtained by decomposition, ω k is the center frequency of the k-th mode, f(t) is the original signal, and K is the set number of modes;

[0088] By solving the above optimization problem, multiple modal signals of the original cross characteristic signal are obtained;

[0089] S4.2. Extract more representative characteristic modes through variational mode decomposition, and use principal component analysis for subsequent dimensionality reduction processing of these modes;

[0090] The dimensionality reduction process of PCA is as follows:

[0091] Let the matrix composed of the original VMD characteristic modes be X ∈ R n×m, and its covariance matrix is:

[0092]

[0093] Obtained by eigenvalue decomposition:

[0094] CV = VΛ

[0095] Where C is the covariance matrix used to reflect the linear correlation degree between features, X is the data matrix, each row corresponds to a sample, each column corresponds to a feature, L is the total number of data sequence samples, V is the matrix composed of eigenvectors; Λ is the diagonal matrix composed of corresponding eigenvalues;

[0096] Select the eigenvectors corresponding to the first q largest eigenvalues (the selection of q is related to the dimensionality reduction target. If the original data dimension is m, theoretically up to m principal components can be retained, but in order to compress the dimension, reduce redundancy or noise, usually only the first q are selected), and form the transformation matrix V q ;

[0097] X PCA = XV q

[0098] Where X PCA is the data matrix after dimensionality reduction.

[0099] Preferably, in S5, the following steps are included:

[0100] S5.1. Select the most influential features from the wind power dominant features and the photovoltaic dominant features respectively;

[0101] S5.2. Select the most representative modal signals from the cross features obtained by VMD decomposition to form the final feature set;

[0102] S5.3. The final feature set includes 1 wind power dominant feature, 3 photovoltaic dominant features, 1 wind-solar cross feature and 2 modal signals.

[0103] Preferably, in S6, the following steps are included:

[0104] S6.1. Use the wind-solar historical data set after reconstructing the weather features to train the SABO-LSTM multi-input multi-output regression prediction model;

[0105] S6.2. The model learns the relationship of wind-solar combined power generation through the training data and uses the reconstructed numerical weather forecast information for prediction, and finally obtains the wind-solar combined prediction result.

[0106] By adopting the above technical solutions: through high-resolution remote sensing data such as remote sensing satellite images and radar images, the wind speed and radiation parameters of the local area are finely reconstructed. Specifically, the reflection brightness information in the satellite image is used to realize the fine reconstruction of the radiation heat map, and the radial wind speed characteristics are extracted using the Doppler radar image. The wind speed heat map is generated by the density of potential height contour lines, and the proportion correction and finite element coupling model are combined to further realize the joint fine reconstruction of wind speed and radiation. Based on the dynamic mutual information (DMI) and maximum information coefficient (MIC) methods, the dominant and cross-features in the weather data that are closely related to wind and solar power are deeply screened, and finally the variational mode decomposition (VMD) method is used to further mine multi-scale information, thereby effectively improving the accuracy and robustness of wind power and photovoltaic power forecasting.

[0107] Compared with the prior art, the present invention has the following beneficial effects:

[0108] 1. The present invention introduces a meteorological data reconstruction module. Based on high-resolution satellite and Doppler radar images, the radiation and wind speed heat maps are generated by inverting the reflection brightness and the density of potential height contour lines, and the local wind speed and radiation data are finely reconstructed, thus overcoming the problems of spatial averaging and insufficient resolution of remote sensing data, thereby improving the prediction accuracy.

[0109] 2. The present invention integrates dynamic mutual information (DMI) and maximum information coefficient (MIC) to accurately screen key features that are highly relevant to power prediction, and uses variational mode decomposition (VMD) to mine high-frequency information of wind-solar crossover features to further improve model performance.

[0110] 3. The present invention proposes an intelligent feature selection mechanism, which automatically and efficiently divides the wind power-dominated, photovoltaic-dominated and wind-solar cross features by calculating the contribution ratio of the features to wind power and photovoltaic power, thereby avoiding the subjectivity and complexity of manual selection.

[0111] 4. The present invention has adaptive adjustment capability and can dynamically select appropriate features and model parameters. It is suitable for wind-solar combined power prediction tasks in different regions and under complex weather conditions and has good adaptability and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0112] Figure 1 is a flow chart of the present invention;

[0113] Figure 2 The radiation heat map reconstructed based on the satellite reflection brightness according to the embodiment of the present invention;

[0114] Figure 3 A wind speed heat map generated based on the density of potential height contour lines according to an embodiment of the present invention;

[0115] Figure 4Wind speed field reconstruction diagram based on the finite element coupling model in the embodiment of the present invention;

[0116] Figure 5 Prediction result diagram of the embodiment of the present invention. Specific implementation manner

[0117] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings, so that those skilled in the art can better understand the advantages and features of the present invention, and thus more clearly define the protection scope of the present invention. The embodiments described in the present invention are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the protection scope of the present invention.

[0118] A decentralized wind-solar power generation prediction method based on weather information reconstruction, including wind-solar joint prediction based on weather information reconstruction, feature selection, and VMD decomposition, comprising the following steps:

[0119] S1. Collect a data set and reconstruct the original weather features;

[0120] Collect weather station data, determine the average wind speed V and radiation values (total radiation G, etc.) in the target area, obtain high-resolution satellite images of the target area, collect Doppler radar images of the target area, and use them to invert the radial wind speed information in the area, and obtain the geopotential height contour lines at various places on a certain pressure surface to form a radar wind speed feature map;

[0121] S2. Data loading and preprocessing;

[0122] S3. Feature selection and correlation analysis;

[0123] S4. VMD decomposition of wind-solar cross features;

[0124] S5. Selection of the final feature set;

[0125] S6. Training and prediction model.

[0126] Specifically, in S1, it includes the following steps:

[0127] S1.1. Photovoltaic radiation inversion and normalization correction

[0128] Image preprocessing steps: Perform geometric correction, radiation correction, and atmospheric correction on the original satellite image, and normalize it into a reflectance image:

[0129]

[0130] Where I(x, y) is the original image value, I raw(x, y) is the normalized reflectance, with a value range of [0, 1], and I max and I min are the maximum and minimum pixel values of the image respectively;

[0131] In the normalized image, with the target point (x, y) as the center, select the neighborhood B(x, y), and use the Gaussian weight function to extract the local weighted average brightness:

[0132]

[0133] In the formula, σ is the standard deviation parameter. The larger σ is, the "flatter" the weight distribution is, and pixels at a farther distance within the neighborhood can also have a greater impact on the brightness of the center point; the smaller σ is, the weights are concentrated near the center, and the influence of distant pixels is rapidly attenuated. L(x, y) is the local weighted average brightness, W(x′, y′) is the weight value of the coordinate (x′, y′) with respect to the coordinate (x, y), I(x′, y′) is the brightness value at the coordinate (x′, y′), exp is the two-dimensional Gaussian distribution kernel, which assigns weights according to the Euclidean distance between pixels. The farther the distance, the smaller the weight. The Gaussian kernel has the characteristics of smoothness and rapid attenuation, and can well retain the local structure and suppress noise;

[0134] Calculate the average brightness within the region:

[0135]

[0136] In the formula, Ω is the region to be inverted, A is the area of the region, is the average brightness of the region;

[0137] Construct a non-linear inversion model to make the local radiation positively correlated with the brightness. To ensure the overall average radiation value of the region Perform normalization correction:

[0138]

[0139] In the formula, G(x, y) is the reconstructed local radiation value, G avg is the regional average radiation value, and γ is the non-linear adjustment parameter;

[0140] S1.2, Wind speed reconstruction and ratio correction method

[0141] From the Doppler radar image, after geometric correction and filtering, extract the radial wind speed information within the region and generate a geopotential height contour map; divide the image into small blocks, count the contour density within each block, and obtain the local thermal value index D (x,y) (normalized value between 0 and 1), which reflects the wind speed intensity of the region;

[0142] Assume that the average wind speed in the radar sampling area is The thermal value of the target small block is D target , and the maximum thermal value in the area is D max , and the wind speed of the target point is reconstructed proportionally as:

[0143]

[0144] In the formula, V(x,y) is the wind speed after the initial reconstruction of the target point, D target (x,y), D max (x,y) are the thermal values of the target block and the area respectively;

[0145] S1.3. Radiation and wind speed joint finite element coupling model

[0146] To realize the joint reconstruction of wind speed and radiation data, the following coupling finite element model is constructed:

[0147] Let the local weather parameter vector be:

[0148] u represents the weather parameter to be solved at the point (X,Y), G(x,y) is the radiation value after the initial reconstruction of the target point, and V(x,y) is the wind speed after the initial reconstruction of the target point.

[0149] Construct a coupling model, considering spatial diffusion and data recovery:

[0150]

[0151] Among them

[0152]

[0153]

[0154] In the formula, K(x,y) is the diffusion coefficient matrix, Q(x,y) is the coupling matrix, k V (x,y) is the wind speed diffusion coefficient, k G (x,y) is the radiation diffusion coefficient, α V (x,y), α G (x,y) are the regression coefficients, δ(x,y) is the coupling coefficient, reflecting the mutual influence between wind speed and radiation, F(x,y) is the external load vector, including image recognition correction information, u ref (x,y) is the reference data, is the regional average wind speed data, G avg is the regional average radiation value; is the gradient operator, that is is the divergence operator, which maps a vector field to a scalar field. is the gradient of the vector u, understood as taking partial derivatives of the components of V(x, y) and G(x, y) in the X and Y directions respectively, and then multiplying by K(x, y) to obtain the diffusion effect.

[0155] S1.4, Finite Element Discretization and Solution

[0156] Discretize the region Ω into finite elements Ω e (e = 1, 2, …, n):

[0157]

[0158] Adopt the shape function N i (x, y) to approximate the reconstructed weather parameter u(x, y):

[0159]

[0160] where u(x, y) represents the weather parameter to be solved at the point (X, Y) (such as wind speed or radiation value), N i (x, y) is the i-th shape function, describing the weight of the contribution of node i to any point (X, Y) within the element, U i (x, y) is the value of the unknown parameter at the i-th node within the element, and m is the number of nodes within the element;

[0161] The shape function is generated by the MATLAB PDE ToolboX for two-dimensional quadratic triangular elements. For each element, there are 6 nodes, and the shape function is composed of quadratic Lagrange polynomials, which take the value 1 at their respective nodes and 0 at other nodes (i.e., satisfying the Kronecker delta property).

[0162] Construct the element stiffness matrix:

[0163]

[0164] where is the matrix element in the i-th row and j-th column of element e, reflecting the mutual influence between node i and node j; is the gradient of the shape function, representing the rate of change of the i-th shape function in the X and Y directions, is the rate of change of the j-th shape function in the X and Y directions;

[0165]

[0166] where F i (e)$F_{i}$ is the load value corresponding to the $i$-th node in element $e$, representing the "force" or "source term" externally acting on this node; $F(x, y)$ is the body load term, which contains the basic data within the region and the correction information obtained by image inversion, and usually reflects the external input, such as local radiation changes; $\Gamma$ is the part where the load is applied on the boundary of element $e$, which is the boundary load and reflects the boundary conditions;

[0167] Through the assembly matrix $A$ e (a sparse, 0 / 1 mapping matrix used to map the element node numbers to the global node numbers), the element matrices are assembled into the overall system;

[0168]

[0169] Solve the overall finite element equation: $KU = F$

[0170] where $K$ is the global stiffness matrix, assembled from the local stiffness matrices of each element, $K$ (e) is the local stiffness matrix of the $e$-th element, $U$ is the global node unknown parameter vector, $F$ is the global load vector, and $F$ (e) is the local load vector of the $e$-th element, and $n$ is the number of elements, representing the total number of elements obtained after meshing the entire region;

[0171] Then the global node vector $U$ can be obtained, that is, $U$ i (x, y), and then according to the formula the local data $u(x, y)$ at any target point can be obtained:

[0172]

[0173] The shape functions are two-dimensional quadratic triangular elements generated by the MATLAB PDE ToolboX. For each element, there are 6 nodes, and the shape functions are composed of quadratic Lagrange polynomials, which take the value 1 at their respective nodes and 0 at other nodes (i.e., satisfying the Kronecker delta property).

[0174] Specifically, in S2, it includes the following steps:

[0175] S2.1. Load the combined wind-solar data from the reconstructed dataset and extract relevant weather features (such as wind speed, temperature, humidity, air pressure, photovoltaic irradiance, etc.);

[0176] S2.2. Normalize the feature data and target variables (wind power and photovoltaic power) to make them have the same dimension to improve the training effect of the model;

[0177] Let the original data sequence be $x = \{x_1, x_2,..., x$L}, where \(L\) is the total number of data sequence samples, then the normalized data sequence is:

[0178]

[0179] In the formula, \(x\) max is the maximum value of the data sequence, and \(x\) min is the minimum value of the data sequence. is the normalized data, and its value range is \([0, 1]\).

[0180] Specifically, step S3 includes the following steps:

[0181] S3.1. Calculate the correlation between each weather feature and the target variable using dynamic mutual information (DMI) and maximum information coefficient (MIC) respectively;

[0182] For the weather feature sequence \(X\) and the power target sequence \(Y\), the dynamic mutual information between the two is defined as:

[0183]

[0184] In the formula, \(I(X; Y)\) is the dynamic mutual information between sequences \(X\) and \(Y\), \(p(x, y)\) is the joint probability distribution of the feature sequence and the target sequence, and \(p(x)\) and \(p(y)\) are the marginal probability distributions of sequences \(X\) and \(Y\) respectively;

[0185] The maximum information coefficient is defined as:

[0186]

[0187] In the formula, \(I\cdot(X, Y, a, b)\) is the maximum mutual information value after the sequences \(X\) and \(Y\) are divided into \(a\times b\) cells by grid;

[0188] When MIC is 0, it means that the feature is irrelevant to the target variable, and the closer the value is to 1, the stronger the relationship between the feature and the target variable;

[0189] S3.2. Calculate the influence of each weather feature on photovoltaic power and wind power, and obtain the DMI and MIC scores of each feature;

[0190] S3.3. Calculate the contribution ratio of each feature according to the weighted sum of the DMI and MIC scores, and divide the features into three categories: wind power dominant, photovoltaic power dominant, and wind-solar cross features;

[0191] The calculation method for the contribution ratio of weather features to wind power is:

[0192]

[0193] The calculation method for the contribution ratio of weather features to photovoltaic power is:

[0194]

[0195] wherein, CR i_wind represents the contribution ratio of the i-th feature to the wind power, and CR i_pv represents the contribution ratio of the i-th feature to the photovoltaic power. DMI wind,i , MIC wind,i are the dynamic mutual information and the maximum information coefficient between the i-th feature and the wind power; DMI pv,i , MIC pv,i are the dynamic mutual information and the maximum information coefficient between the i-th feature and the photovoltaic power;

[0196] Classify the features according to the contribution ratio: CR i_wind ≥0.7, CR i_pv ≤0.4 indicates the wind power dominant features, CR i_wind ≤0.4, CR i_pv ≥0.7 indicates the photovoltaic power dominant features, 0.15≤CR i_wind ≤0.7, 0.15≤CR i_pv ≤0.7 and the rest indicate the cross features.

[0197] Specifically, in S4, the following steps are included:

[0198] S4.1. Perform variational mode decomposition (VMD) on the wind-solar cross features to decompose these features into multiple modal signals in order to capture different frequency components in the data;

[0199] VMD obtains the characteristic modes by optimizing the following objective function:

[0200]

[0201] wherein, u k (t) is the k-th modal signal obtained by decomposition, ω k is the central frequency of the k-th mode, f(t) is the original signal, and K is the set number of modes;

[0202] By solving the above optimization problem, multiple modal signals of the original cross feature signal are obtained;

[0203] S4.2. Extract more representative characteristic modes through variational mode decomposition, and perform subsequent dimensionality reduction processing on these modes using principal component analysis;

[0204] The dimensionality reduction process of PCA is as follows:

[0205] Let the matrix formed by the original VMD characteristic modes be X∈R n×m , and its covariance matrix is:

[0206]

[0207] Obtained by eigenvalue decomposition:

[0208] CV = VΛ

[0209] Where C is the covariance matrix used to reflect the linear correlation degree between features, X is the data matrix, each row corresponds to a sample, each column corresponds to a feature, L is the total number of data sequence samples, V is the matrix composed of eigenvectors; Λ is the diagonal matrix composed of corresponding eigenvalues;

[0210] Select the eigenvectors corresponding to the top q largest eigenvalues (the selection of q is related to the dimensionality reduction target. If the original data dimension is m, theoretically up to m principal components can be retained, but in order to compress the dimension, reduce redundancy or noise, usually only the top q are selected), and form the transformation matrix V q ;

[0211] X PCA = XV q

[0212] Where X PCA is the data matrix after dimensionality reduction.

[0213] Preferably, in S5, it includes the following steps:

[0214] S5.1. Select the features with the greatest influence from the wind power dominant features and the photovoltaic dominant features respectively;

[0215] S5.2. Select the most representative modal signals from the cross features obtained by VMD decomposition to form the final feature set;

[0216] The final feature set includes 1 wind power dominant feature, 3 photovoltaic dominant features, 1 wind-solar cross feature and 2 modal signals.

[0217] Specifically, in S6, it includes the following steps:

[0218] S6.1. Use the historical wind-solar power generation dataset after reconstructing the weather features to train the SABO-LSTM multi-input multi-output regression prediction model;

[0219] S6.2. This model learns the relationship of wind-solar combined power generation through training data and uses the reconstructed numerical weather prediction information for prediction, and finally obtains the wind-solar combined prediction result.

[0220] Example:

[0221] The decentralized wind-solar power generation prediction technology based on weather information reconstruction is a wind-solar combined prediction based on feature selection and VMD decomposition, including the following steps:

[0222] Step 1: Collect the required dataset, satellite images, and radar information from a wind-solar joint demonstration project in Xinjiang Uygur Autonomous Region, and reconstruct the original weather features to obtain a radiation thermal map reconstructed based on satellite reflected brightness ( Figure 2 ) and a wind speed thermal map generated based on the density of geopotential height isolines ( Figure 3 ). On this basis, use the radiation and wind speed joint finite element coupling model to obtain the reconstructed weather information and a reconstructed wind speed field map based on the finite element coupling model ( Figure 4 );

[0223] Step 2: Load the wind-solar joint data from the dataset, extract the relevant weather features, and normalize the feature data and target variables (wind power and photovoltaic power) to have the same dimension;

[0224] Step 3: Calculate the correlation between each weather feature and the target variable using dynamic mutual information (DMI) and maximum information coefficient (MIC) respectively, and then calculate the contribution ratio of each feature according to the weighted sum of the DMI and MIC scores, and divide the features into three categories: wind power-dominated, photovoltaic power-dominated, and wind-solar cross features;

[0225] Step 4: For the wind-solar cross features, extract more representative feature modes through variational mode decomposition and perform subsequent dimensionality reduction processing on these modes;

[0226] The final feature set includes 1 wind power-dominated feature, 3 photovoltaic power-dominated features, 1 wind-solar cross feature, and 2 modal signals.

[0227] Table 1 Weighted comprehensive score table of each feature

[0228]

[0229]

[0230] Step 5: Use the reconstructed wind-solar historical dataset to train the SABO-LSTM multi-input multi-output regression prediction model, and make predictions based on the numerical weather prediction information to obtain the wind-solar joint prediction result.

[0231] Reference Figure 5 And Table 2, taking MAE (Mean Absolute Error) as the prediction error evaluation index, it can be seen that whether it is wind power prediction or photovoltaic power prediction, the accuracy of joint prediction is always greater than that of individual prediction.

[0232] Table 2 Comparison table of prediction errors

[0233]

[0234] In summary, the present invention can make the wind power prediction more accurate. First, the weather feature data is reconstructed, and then the correlation between the weather features and the wind power and photovoltaic power is calculated through dynamic mutual information (DMI) and maximum information coefficient (MIC), and the wind power dominant features, photovoltaic power dominant features and cross features are defined; then for the cross features, variational mode decomposition and principal component analysis are used to select the most representative modal signals from the cross features obtained by VMD decomposition to form the final feature set; finally, the final feature set is divided into a training set and a test set, and the SABO-LSTM multi-input multi-output regression prediction model is used for training and prediction, improving the prediction accuracy and having application and promotion value in the field of joint probability prediction of wind, light and storage.

[0235] The descriptions and practices disclosed in the present invention are easy to think and understand for those of ordinary skill in the art, and several improvements and retouches can be made without departing from the principle of the present invention. Therefore, the modifications or improvements made without departing from the spirit of the present invention should also be regarded as within the protection scope of the present invention.

Claims

1. A decentralized wind-solar power generation prediction method based on weather information reconstruction, characterized in that Wind-solar combined prediction based on weather information reconstruction, feature selection, and VMD decomposition includes the following steps: S1. Collect the dataset and reconstruct the original weather features; Collect weather station data and determine the average wind speed in the target area Average radiation value, obtain high-resolution satellite images of the target area, collect Doppler radar images of the target area, use them to retrieve the radial wind speed information within the area, and obtain the geopotential height contour lines at various points on a certain pressure surface to form a radar wind speed characteristic map; S2. Load and preprocess the data; S3. Feature selection and correlation analysis; S4. VMD decomposition of wind-solar cross features; S5. Selection of the final feature set; S6. Train and predict the model.

2. The decentralized wind-solar power generation prediction method based on weather information reconstruction according to claim 1, wherein In S1, it includes the following steps: S1.

1. Photovoltaic radiation inversion and normalization correction Image preprocessing steps: Geometric correction, radiometric correction, and atmospheric correction are performed on the original satellite image, and it is normalized to a reflectance image: Where, I(x, y) is the original image value, and I raw (x, y) is the reflectance after normalization, with a value range of [0, 1], and I max 、I min are the maximum and minimum pixel values of the image respectively; In the normalized image, with the target point (x, y) as the center, the neighborhood B(x, y) is selected, and the local weighted average brightness is extracted using the Gaussian weight function: Where σ is the standard deviation parameter, L(x, y) is the local weighted average brightness, W(x′, y′) is the weight value of the coordinate (x′, y′) with respect to the coordinate (x, y), I(x′, y′) is the brightness value at the coordinate (x′, y′), and exp is the two-dimensional Gaussian distribution kernel; Calculate the average brightness within the region: In the formula, Ω is the area to be inverted, A is the area of the region, is the average brightness of the region; Construct a non-linear inversion model to make the local radiation positively correlated with the brightness, and to ensure the overall average radiation value of the region Perform normalization correction: where G(x, y) is the reconstructed local radiation value, G avg is the regional average radiation value, and γ is the non-linear adjustment parameter; S1.

2. Wind speed reconstruction and ratio correction method From the Doppler radar images, after geometric correction and filtering, the radial wind speed information within the region is extracted, and a geopotential height contour map is generated; the image is divided into small blocks, the contour density within each block is statistically analyzed, and according to the relationship that the density is proportional to the thermal value, the local thermal value index D is obtained (x,y) , which reflects the wind speed intensity in this region; Assume that the average wind speed in the radar sampling area is The thermal value of the target small block is D target , and the maximum thermal value in the area is D max , and the wind speed of the target point is reconstructed proportionally as: Where, V(x,y) is the wind speed after the initial reconstruction of the target point, D target (x,y), D max (x,y) are the thermal values within the target block and the area, respectively; S1.

3. Radiation and wind speed joint finite element coupling model To achieve the joint reconstruction of wind speed and radiation data, the following coupled finite element model is constructed: Let the local weather parameter vector be: u represents the weather parameter to be solved at the point (X, Y), G(x, y) is the radiation value after the initial reconstruction of the target point, and V(x, y) is the wind speed after the initial reconstruction of the target point; Construct the coupling model, considering spatial diffusion and data recovery: Among them where \(K(x,y)\) is the diffusion coefficient matrix, \(Q(x,y)\) is the coupling matrix, \(k\) V (x,y) is the wind speed diffusion coefficient, \(k\) G (x,y) is the radiation diffusion coefficient, \(\alpha\) V (x,y), \(\alpha\) G (x,y) are the regression coefficients, \(\delta(x,y)\) is the coupling coefficient, reflecting the mutual influence between wind speed and radiation, \(F(x,y)\) is the external load vector, including the image recognition correction information, \(u\) ref (x,y) is the reference data, is the regional average wind speed data, \(G\) avg is the regional average radiation value; is the gradient operator, that is \(\nabla\cdot\) is the divergence operator, which maps a vector field to a scalar field, is the gradient of the vector \(u\), understood as taking the partial derivatives of the components of \(V(x,y)\) and \(G(x,y)\) in the \(X\) and \(Y\) directions respectively, and then multiplying by \(K(x,y)\) to obtain the diffusion effect; S1.

4. Finite element discretization and solution Discretize the region Ω into finite elements Ω e (e = 1, 2, …, n): The shape function N is adopted within each unit i (x, y) to approximate the reconstructed weather parameter u(x, y): where \(u(x,y)\) represents the weather parameter to be determined at the point \((X,Y)\), and \(N\) i (x,y) is the \(i\)-th shape function, which describes the weight of the contribution of node \(i\) to any point \((X,Y)\) within the element, and \(U\) i (x,y) is the value of the unknown parameter at the \(i\)-th node within the element, and \(m\) is the number of nodes within the element; Construct the element stiffness matrix: In the formula, is the matrix element in the \(i\)-th row and \(j\)-th column of element \(e\), reflecting the mutual influence between node \(i\) and node \(j\); is the gradient of the shape function, representing the rate of change of the \(i\)-th shape function in the \(X\) and \(Y\) directions, is the rate of change of the \(j\)-th shape function in the \(X\) and \(Y\) directions; where F i (e) is the load value corresponding to the i-th node in element e, representing the force or source term externally acting on this node; F(x, y) is the body load term, which contains the basic data in the region and the correction information obtained by image inversion, reflecting the external input, such as local radiation changes; Γ is the part where the load is applied on the boundary of element e, is the boundary load, reflecting the boundary conditions; Through the assembly matrix A e , assemble the element matrices into the global system; Solve the overall finite element equation: KU = F Where K is the global stiffness matrix, assembled from the element stiffness matrices, and K (e) is the local stiffness matrix of the e-th element, U is the global vector of unknown nodal parameters, F is the global load vector, and F (e) is the local load vector of the e-th element. n is the number of elements, representing the total number of elements obtained after meshing the entire region; The global node vector U, namely U, can be obtained i (x, y), and then according to the formula The local data u(x, y) of any target point can be obtained:

3. A decentralized wind and solar power generation prediction method based on weather information reconstruction according to claim 1, characterized in that, In S2, it includes the following steps: S2.

1. Load the wind-solar combined data from the reconstructed dataset and extract the relevant weather features; S2.

2. Normalize the feature data and the target variable to have the same dimension to improve the training effect of the model; Let the original data sequence be \(x = \{x_1, x_2, \ldots, x\) L \}, and \(L\) be the total number of samples in the data sequence. Then the normalized data sequence is as follows: where x max is the maximum value of the data sequence, and x min is the minimum value of the data sequence, is the normalized data, and its value range is [0, 1].

4. A decentralized wind-solar power generation prediction method based on weather information reconstruction according to claim 1, characterized in that, In S3, it includes the following steps: S3.

1. Calculate the correlation between each weather feature and the target variable using the dynamic mutual information DMI and the maximum information coefficient MIC respectively; For the weather feature sequence X and the power target sequence Y, the dynamic mutual information between the two is defined as: Where I(X; Y) is the dynamic mutual information between the sequences X and Y, p(x, y) is the joint probability distribution of the feature sequence and the target sequence, and p(x) and p(y) are the marginal probability distributions of the sequences X and Y respectively; The maximum information coefficient is defined as: Where I·(X, Y, a, b) is the maximum mutual information value between the sequences X and Y after being divided into a×b cells by grid; When MIC is 0, it means the feature is irrelevant to the target variable, and the closer the value is to 1, the stronger the relationship between the feature and the target variable; S3.

2. Calculate the influence of each weather feature on the photovoltaic power and the wind power, and obtain the DMI and MIC scores of each feature; S3.

3. Calculate the contribution ratio of each feature according to the weighted sum of the DMI and MIC scores, and divide the features into three categories: wind power dominant, photovoltaic power dominant, and wind-solar cross features; The calculation method for the contribution ratio of weather features to wind power is: The calculation method of the contribution ratio of weather characteristics to photovoltaic power is as follows: where CR i_wind represents the contribution ratio of the i-th feature to the wind power, and CR i_pv represents the contribution ratio of the i-th feature to the photovoltaic power. DMI wind,i and MIC wind,i are the dynamic mutual information and the maximum information coefficient between the i-th feature and the wind power; DMI pv,i and MIC pv,i are the dynamic mutual information and the maximum information coefficient between the i-th feature and the photovoltaic power; Classify features according to the contribution ratio: CR i_wind ≥0.7, CR i_pv ≤0.4 indicates the dominant wind power feature, CR i_wind ≤0.4, CR i_pv ≥0.7 indicates the dominant photovoltaic feature, 0.15 ≤ CR i_wind ≤0.7, 0.15 ≤ CR i_pv ≤0.7 The rest indicate cross features.

5. A decentralized wind and solar power generation prediction method based on weather information reconstruction according to claim 1, characterized in that In S4, the following steps are included: S4.

1. Perform variational mode decomposition (VMD) on the wind-solar cross characteristics to decompose these characteristics into multiple mode signals in order to capture different frequency components in the data; VMD obtains the characteristic modes by optimizing the following objective function: where u k (t) is the k-th modal signal obtained by decomposition, ω k is the center frequency of the k-th mode, f(t) is the original signal, and K is the set number of modes; By solving the above optimization problem, multiple mode signals of the original cross-characteristic signal are obtained; S4.

2. Extract more representative characteristic modes through variational mode decomposition, and use principal component analysis to perform subsequent dimensionality reduction processing on these modes; The dimensionality reduction process of PCA is as follows: Let the original VMD eigenmode composition matrix be \(X\in\mathbb{R}\) n×m , and its covariance matrix is: Obtained by eigenvalue decomposition: CV = VΛ In the formula, C is the covariance matrix used to reflect the linear correlation degree between each characteristic, X is the data matrix, each row corresponds to a sample, each column corresponds to a characteristic, L is the total number of data sequence samples, V is the matrix composed of eigenvectors; Λ is the diagonal matrix composed of corresponding eigenvalues; Select the eigenvectors corresponding to the top q largest eigenvalues to form the transformation matrix V q ; X PCA = XV q Where X PCA is the data matrix after dimensionality reduction.

6. The decentralized wind-solar power generation prediction method based on weather information reconstruction according to claim 1, wherein, In S5, the following steps are included: S5.

1. Select the most influential characteristics from the wind power-dominated characteristics and the photovoltaic power-dominated characteristics respectively; S5.

2. Select the most representative mode signals from the cross characteristics obtained by VMD decomposition to form the final feature set; The final feature set includes 1 wind power-dominated characteristic, 3 photovoltaic power-dominated characteristics, 1 wind-solar cross characteristic, and 2 mode signals.

7. A decentralized wind and solar power generation prediction method based on weather information reconstruction according to claim 1, characterized in that, In S6, the following steps are included: S6.

1. Use the historical wind-solar dataset after reconstructing the weather characteristics to train the SABO-LSTM multi-input multi-output regression prediction model; S6.

2. This model learns the relationship between the combined wind-solar power generation through the training data and uses the reconstructed numerical weather forecast information for prediction, and finally obtains the combined wind-solar prediction result.

Images