Multi-mode short-term photovoltaic power prediction method and device based on satellite cloud picture

Abstract

The invention discloses a multi-mode short-term photovoltaic power prediction method and device based on a satellite cloud atlas, and the method comprises the steps: obtaining real-time meteorological data, historical photovoltaic power data and cloud cluster image data, and carrying out the preprocessing of the data; a GPAformer model is established, and the meteorological data after noise reduction are predicted; the method comprises the following steps: establishing an SA-Convlstm model, capturing a cloud cluster movement track, extracting time change characteristics, predicting cloud cluster image data, introducing a cloud shielding model, and correcting the prediction deviation of SA-ConvLSTM under the condition that a cloud layer is dense or changes rapidly; establishing a photovoltaic power KAN-COGCN prediction model, and performing photovoltaic power prediction by taking the meteorological factors, the cloud cluster motion time sequence characteristics and historical photovoltaic power data obtained by prediction of the GPAformer and the SA-ConvLSTM model as input; using an improved alpha evolutionary algorithm IAE to optimize hyper-parameters of the three models; and performing error correction on a prediction result by establishing an adaptive wavelet RBF neural network to obtain a final prediction result. According to the invention, the precision and effectiveness of photovoltaic power prediction can be improved.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a multi-modal short-term photovoltaic power prediction method and device based on satellite cloud images. Background Art

[0002] As global energy demand grows, solar energy, as a clean, renewable energy source, has become a key component of energy transformation. However, the accuracy of photovoltaic power generation forecasts remains challenging, particularly in the face of frequent cloud changes and meteorological disturbances. Traditional photovoltaic power forecasting methods often rely on single data sources or simple models, failing to fully utilize the spatiotemporal characteristics of meteorological data, historical power data, and cloud image information.

[0003] In recent years, deep learning technology has made some progress in photovoltaic power forecasting, but its ability to handle complex meteorological and cloud dynamics remains limited. Multimodal fusion methods, combining multiple sources of information, such as meteorological data and cloud images, have demonstrated promising prediction results by capturing the dynamics of photovoltaic power through deep learning models. However, optimizing model parameters, improving generalization, and correcting errors remain pressing challenges. Summary of the Invention

[0004] Purpose of the invention: In response to the problems existing in the prior art, the present invention provides a multimodal short-term photovoltaic power prediction method and device based on satellite cloud images, which can improve the accuracy and effectiveness of the photovoltaic power prediction model.

[0005] Technical solution: The multimodal short-term photovoltaic power prediction method based on satellite cloud images described in the present invention includes the following steps:

[0006] (1) Obtain real-time meteorological data, historical photovoltaic power data, and cloud image data, and pre-process the data;

[0007] (2) Savitzky-Golay filtering is used to reduce noise in meteorological data and historical photovoltaic power data, smoothing the time series signals and retaining key trends; cloud image data is reduced in noise by using a cascade filter of guided filtering to retain the overall brightness distribution and the main morphology of the cloud and reduce noise interference;

[0008] (3) Establish a GPAformer prediction model to capture the temporal characteristics of meteorological factors and predict the denoised meteorological data;

[0009] (4) Establishing the SA-ConvLSTM model to capture cloud motion trajectories and extract temporal variation features, predicting cloud image data, and introducing a cloud cover model to correct the prediction bias of SA-ConvLSTM when the cloud layer is dense and changing rapidly;

[0010] (5) Establish a photovoltaic power KAN-COGCN prediction model, and integrate the meteorological factor data and cloud map motion time series characteristics predicted by the two models in steps (3) and (4) with the historical photovoltaic power data as input to realize photovoltaic power prediction;

[0011] (6) Using the improved alpha evolution algorithm IAE to simultaneously optimize the hyperparameters of the GPAformer, SA-Convlstm model, and photovoltaic power KAN-COGCN prediction model, and search for the optimal solution of the three model parameters; the IAE algorithm is based on the alpha evolution algorithm AE, and adopts the Sobol sequence and normal cloud mutation strategy to improve the original population initialization process to obtain the IAE algorithm;

[0012] (7) Establish an adaptive wavelet RBF neural network to correct the error of the prediction results and obtain the final prediction results.

[0013] Furthermore, the implementation process of step (1) is as follows:

[0014] The meteorological data and historical photovoltaic power data were preprocessed using the Newton interpolation method. The data were screened, all abnormal data were reset to zero, and missing data were filled. The cleaned meteorological and photovoltaic data sets were normalized. The cloud image data were processed to a standard size through image enhancement operations including cloud image acquisition, geographic calibration, and target area cropping.

[0015] Furthermore, the implementation process of step (2) is as follows:

[0016] When processing meteorological data and historical photovoltaic power data sets, Savitzky-Golay filtering is used for noise reduction. i As the center, a window array containing 2T+1 sampling points is constructed, and the data in the window is fitted by a P-order polynomial q(n). The signal is smoothed using the least squares method:

[0017]

[0018] Define a function F. When q(n) reaches its minimum value, the fitting effect is optimal. By sliding the window on the data sequence, the fitting value of each point in the original data is calculated in turn. In the process, the noise is removed and smoothing filtering is achieved. The signal f(i) obtained after smoothing the signal is as follows:

[0019]

[0020] Where f(i) is the filtered signal, D(k) is the sampling response of the smoothing filter, y(x) is a set of data in the original signal, x = -N,…,0,…N;

[0021] When denoising cloud image data, a cascade filter consisting of a multi-level median rational mixture filter and a guided filter is used. First, a multi-level median rational mixture filter with a 3×3 window is used to remove salt and pepper noise from the cloud image through center-weighted median filtering and rational operations. Initial denoising is then performed by calculating new pixel values. The specific calculation formula is as follows:

[0022]

[0023] Where G is the filtered value, β1, β2, and β3 are the medians of the pixel values in windows 1, 2, and 3;

[0024] In the guided filtering process, the input image p is filtered using the guided image E to obtain the output image q, which effectively reduces image noise while keeping the image edge clear. The specific formula is as follows:

[0025] q i =a n E i +b n

[0026] Where i∈ω n ,(a n ,b n ) is the linear coefficient which is assumed to be a constant;

[0027] In the window n In order to solve a n and b n , establish a constraint condition; assuming z i is unnecessary noise in the window, then the relationship between the filtered image q and the input image p is:

[0028] q i =p i -z i

[0029] The least squares method is used to construct the loss function, and the calculation formula is as follows:

[0030]

[0031] Among them, ξ is the regularization parameter, which is used to prevent the coefficient a n Too large; by solving the loss function, we get the linear coefficient a n and b n , the expression is as follows:

[0032]

[0033] Where μ nand Indicates that the guidance image E is in the window ω n The mean and variance in |ω| is the window ω n The number of pixels in is the input image p in window ω n The mean of

[0034] For each pixel i in the image, calculate its filtered q in multiple different windows i values, and take these q i The average value of the values is used as the final filtering result of the pixel to obtain the denoised cloud image data:

[0035]

[0036] Furthermore, the implementation process of step (3) is as follows:

[0037] Autoformer uses sequence decomposition and autocorrelation mechanisms to split meteorological and photovoltaic data into trend and cycle parts, while suppressing local disorder. The specific calculation formula is as follows:

[0038] L t =Avg(pad(L))

[0039] L s =LL t

[0040] Where L is a sequence of length X, L∈R X×d ;L t is the short-term fluctuation part in the sequence; L s represents the periodic smoothing portion remaining after removing short-term fluctuations;

[0041] The autocorrelation mechanism calculates the autocorrelation coefficient P(τ) of the sequence by aggregating similar subsequences. It is used to reflect the periodic dependency of the sequence. The calculation formula of P(τ) is as follows:

[0042]

[0043] Where P(τ) represents the sequence L G With delay G steps after L G-τ The similarity between them, Y represents the number of delayed operations;

[0044] Autoformer uses fast Fourier transform FFT to calculate the autocorrelation coefficient P(τ). The calculation formula is as follows:

[0045] S(f)=F(L G )F * (L G )

[0046] P(τ)=F -1 (Sxx(f))

[0047] Where S(f) represents the result of converting the time series to the frequency domain, F represents Fourier transform, and F * represents the Fourier transform conjugate, f represents the frequency, F -1 represents the inverse Fourier transform;

[0048] For a given time series data {y1,y2,…y n}, smoothed value y′ after Gaussian process processing i The calculation formula is as follows:

[0049]

[0050] Where K(j) is the weight function of the Gaussian kernel; σ is the standard deviation of the Gaussian kernel; and U is the size of the smoothing window.

[0051] Furthermore, the implementation process of step (4) is as follows:

[0052] The SA-ConvLSTM model includes the ConvLSTM module and the self-attention memory module SAM;

[0053] The ConvLSTM module uses a convolutional neural network (CNN) to extract spatial features and replaces the element-by-element multiplication operation in the traditional LSTM with a convolution operation to capture the spatiotemporal variation of cloud clusters and complete the construction of the ConvLSTM model. The specific calculation formula is as follows:

[0054]

[0055] Among them, X t is the input data; H t-1 is the hidden state at the previous moment; C t-1 is the state memory of the previous moment; b is the deviation term; y t is the input gate; v t For the forget gate; t is the output gate; * represents the convolution operation; Represents the Hadamard product operation;

[0056] The self-attention memory module SAM is used to store and update historical spatial information, so that the model can remember the relationship between adjacent points for a long time. SAM consists of the historical memory state M t-1 and self-attention calculation unit; SAM is composed of input information X t Perform linear transformation to generate query Q h , key K h Sum Vh ; According to the memory state M of the previous moment t-1 , generate the corresponding key-value pair K m Sum V m ;

[0057] Using input information X t Generated query Q h , respectively with key K h and K m Calculate the similarity and get Z h , Z m , combine these two results into Z, and then generate the output H t And the corresponding gate unit; according to the calculated input gate y′ t , memory update gate g′ t and output gate O′ t , for the self-attention memory state M t and output results To update, the calculation formula is as follows:

[0058]

[0059] Where M t-1 is the self-attention memory state at time t-1; SAM is integrated into the ConvLSTM model to construct the SA-ConvLSTM model;

[0060] A cloud cover model was introduced to extract cloud image data from the latitude and longitude coordinates of the region captured by the all-sky imager tracking and monitoring. This data was used to determine the distribution of clouds within the region and capture the dynamic changes in cloud motion and thickness. Semantic segmentation technology using the DeepLabv3+ network was used to classify satellite cloud images at the pixel level and extract pixel information about the cloud region. The classified cloud images were grayscaled by setting the background to white and the clouds to dark, and grayscale values were extracted to assess the impact of cloud cover on photovoltaic power.

[0061] Calculate the grayscale mean i of the cloud map in the area where the photovoltaic power station is located n , as an indicator to measure the degree of cloud cover, and calculate the correction coefficient η to correct the preliminary predicted photovoltaic power value:

[0062]

[0063] Among them, i n is the grayscale mean of the cloud image, i max is the maximum value of the grayscale value;

[0064] The preliminary prediction results are corrected to obtain the final photovoltaic prediction value. The calculation formula is as follows:

[0065] P=k×η×Ht +m×H r

[0066] Among them, p represents the final prediction result; H t Indicates the preliminary prediction result; H r represents the power value at the previous moment; η represents the correction coefficient; k represents the power prediction correction coefficient; m represents the power prediction correction coefficient.

[0067] Furthermore, the implementation process of step (5) is as follows:

[0068] The model formula of KAN is:

[0069]

[0070] in, is an external function; is an internal function. The model uses a spline function as a learnable activation function, and its shape is changed by adjusting parameters. A KAN layer is regarded as a one-dimensional function matrix. Multiple KAN layers are simply stacked to construct a deep KAN model. The cascade relationship of multiple layers of functions is further expressed in matrix form to enhance the computational efficiency, parameter sharing capability, and parallel computing performance of the photovoltaic power prediction model. The specific expression is as follows:

[0071]

[0072] Among them, the symbol in the upper right corner of the formula Represents the composite operation of a function;

[0073] The CSEConv module of the COGCN model constructs a directed graph L through the KNN algorithm to extract the spatiotemporal characteristics of cloud movement and capture the impact of cloud cover on photovoltaic power. i And its K nearest neighbor points H (i，j) Based on the KNN algorithm, each neighboring point is connected to the center point through a directed edge. The directed graph L constructed using the KNN algorithm is defined as:

[0074] L=(M,E)

[0075]

[0076] I={e i =(e i1 ,e i2 ,...,e ij )|i=1,2,3,...,n}

[0077] Where Q represents n local vertices, and each local vertex is a C-dimensional feature point H i, I represents the set of edges between vertices; after constructing the graph structure L through the KNN algorithm, a multi-layer perceptron w is used, which includes a convolutional layer Conv, a normalization layer BN, and a RELU activation layer. θ Extract edge features u from the graph structure i,j ; Let u i,j Indicates the center point H i The edge feature between its nearest neighbor H(i, j) is defined as follows:

[0078] u ij =w θ (H i ,H i -H (i,j) )

[0079]

[0080] Based on edge feature u i,j , further normalized and activated, the local domain feature Y is finally obtained i :

[0081] Y i =ReLu(BN(u ij )

[0082] The extracted local neighborhood features are further enhanced by the CBAM module, and the key information at different locations is extracted through channel weighting and spatial weighting processing; the CSEConv module finally obtains the local domain features CSE of the cloud data through the multi-layer perceptron and the maximum pooling function. out :

[0083] CSE out =MaxP(MLP(F″))

[0084] Introduce the Offset Attention module to capture context features. For the input feature B in ∈R N×de , generate query Q, key K and value V matrices through linear transformation, and the transformation process is as follows:

[0085] (Q,K,V)=B in ·(W Q ,W K ,W V )

[0086]

[0087] Among them, W Q ,W K , W V is the shared learnable weight, T a is the dimension of Q and K, Te is the dimension of V; the attention weight is calculated using the matrix dot product of the query matrix Q and the key matrix K. The specific calculation formula is as follows:

[0088] E=α i,j =Q·K T

[0089] Then, the obtained weights are normalized to obtain E = α i,j :

[0090]

[0091] The calculated attention weights are weighted and summed with the value matrix V to obtain the self-attention output feature B sa :

[0092] B sa =L·V

[0093] Calculate input feature J in And the self-attention output feature B sa The difference is input to the LBR layer and compared with B in Add them together and process them through the multi-layer perceptron MLP to finally get the output feature B of the OffsetAttention module out :

[0094] B out =OL(B in )=MLP(LBR(B in -B sa )+B in )

[0095] Among them, the LBR layer is a network layer that combines Linear, BatchNorm and ReLU functions.

[0096] Furthermore, the implementation process of step (6) is as follows:

[0097] Set the population size, dimension, number of iterations, and upper and lower limits of the search space of the AE algorithm; introduce the Sobol sequence strategy into the algorithm to enhance the uniform distribution of the population, and combine it with the normal cloud mutation strategy to improve the diversity of the population, thereby completing the population initialization. The improved formula is as follows:

[0098] X i =Sobol(LB,UB,N)+ΔX i

[0099] Among them, X irepresents the initialized i-th individual, Sobol(LB,UB,N) represents the generation of N Sobol sequence samples in the search space (LB,UB), ΔX i ~N(0,σ) represents the normal cloud mutation strategy, and σ controls the amplitude of the mutation;

[0100] By performing a sampling with replacement operation, some individuals are selected to form the evolution matrix, and the evolution operation is performed only on these individuals. The specific expression is as follows:

[0101]

[0102] The search process is implemented through the α operator, which can simultaneously extract and utilize evolutionary information within a single operator:

[0103]

[0104] The adaptive basis vector plays a core role in the solution update process. Its specific expression is as follows:

[0105]

[0106] Among them, diag is a function that generates a D-order square matrix, which is obtained by sampling by multiplying D with the candidate solution; the dimension of matrix B is K×D, ω i:K represents the weight of the solution in the equation;

[0107] Construct two different evolutionary paths. The specific calculation formula is as follows:

[0108]

[0109] Where c a and c b Represents L a and L b The learning rate;

[0110] Random step size αΔr i The algorithm is given the ability to search globally, and the nonlinear decreasing attenuation factor is used to gradually shrink the search range. The specific calculation formula is as follows:

[0111]

[0112] Where s represents the fitness evaluation, and Max represents the maximum fitness evaluation; Δr is a disturbance that gradually weakens with time delay, and is calculated as follows:

[0113] Δr=(ub-lb)·(2R1)·R2-R2)·S

[0114] Here, R1 and R2 represent random real number matrices generated by rand(0,1,[N,D]), while represents a random integer matrix containing only 0 or 1 generated by randi(0,1,[N,D]), weighing the perturbation based on the dimension;

[0115] The boundary constraint method ensures that the algorithm performs an effective search within the search space. In the AE algorithm, this goal is achieved by adopting the distance halving method. The specific calculation formula is as follows:

[0116]

[0117] In the AE algorithm, a greedy selection strategy is introduced:

[0118]

[0119] Based on the individual update formula in the population, the global optimal solution is searched, and the hyperparameters of the three models are gradually optimized to obtain the optimal parameter combination of the GPAformer, SA-convLSTM, and KAN-COGCN models. During the optimization process, the root mean square error (RMSE) is used to evaluate the parameters in the three models, and the accuracy of the model prediction results is measured by minimizing the RMSE value, thereby obtaining the optimal prediction result.

[0120] Furthermore, the hyperparameters of the GPAformer, SA-Convlstm model and photovoltaic power KAN-COGCN prediction model in step (6) include the standard deviation σ of the Gaussian kernel, the smoothing window size k, the size of the pooling window, the time step in the autocorrelation mechanism and the number of heads in the self-attention mechanism, the hidden layer dimension, the activation function type, and the number of delay operations L in the GPAformer model; the convolution kernel size, time step, and number of attention heads of SA-convLSTM; the number of layers in KAN in the KAN-COGCN model, the number of nodes in each layer, the learning rate and the number of training rounds of KAN; the number of CSEConv modules, the number of OffsetAttention modules, the number of nearest neighbors K, and the output dimension of the feature extraction layer in COGCN.

[0121] Furthermore, the implementation process of step (7) is as follows:

[0122] Construct a wavelet neural network based on radial basis function network structure, and assume that the input layer receives the sample sequence X=(x1,x2,…,x n ), at the same time, the wavelet function and radial basis function are established, and the output of the first hidden layer is:

[0123]

[0124] Where F1(X) is the wavelet function, ψ(xi ) is the wavelet basis function, a is the wavelet expansion factor, b is the wavelet translation factor, is the output of the first hidden layer, is the wavelet network scale variable, t i is the wavelet network translation variable, M is the number of neurons in the first hidden layer, R i is a scale variable;

[0125] The output of the second hidden layer is expressed as:

[0126]

[0127] Where, L u is the center point of the radial basis function, σ u is the width variable of the radial basis function, and N is the number of neurons in the second hidden layer;

[0128] The K-means clustering algorithm is introduced to perform cluster analysis on the signal after wavelet transformation of the first hidden layer, and the center point l of the radial basis function of the second hidden layer is quickly determined. u ; Thus completing the output of the wavelet RBF neural network:

[0129]

[0130] Where y is the output of the wavelet RBF network, ω u is the weight between the second hidden layer and the output layer, is the output of the second hidden layer, g is the quantization noise;

[0131] The AGA algorithm is used to correct the network parameters. First, the prediction error e of the network is calculated. The calculation formula is as follows:

[0132]

[0133] Where, is the expected output value of the network, is the actual output of the network;

[0134] Calculate the gradient G e With the squared gradient accumulation R and the learning rate r, the specific calculation formula is as follows:

[0135]

[0136] Where η is the initial learning rate and ε is the minimum value; according to the network's prediction error e, the network parameters are corrected:

[0137]

[0138]

[0139] Where W(u) is the weight of the wavelet neural network, a(k) is the scaling scale of the wavelet function, b(k) is the translation factor of the wavelet function, and γ is the momentum factor, which ranges from 0 to 1. The cycle operation is performed until the output of the wavelet neural network reaches the expected value, and the final photovoltaic power prediction result is obtained.

[0140] A device according to the present invention includes a memory and a processor, wherein:

[0141] a memory for storing computer programs capable of running on the processor;

[0142] The processor is configured to execute the steps of the above-mentioned multi-modal short-term photovoltaic power prediction method based on satellite cloud images when running the computer program.

[0143] Beneficial effects: Compared with the prior art, the present invention has the following beneficial effects:

[0144] 1. This invention introduces Savitzky-Golay filtering to reduce noise in meteorological data and historical photovoltaic power data, effectively removing high-frequency noise in the time series while retaining key trends in the signal. It also uses a cascade filter of guided filtering to reduce noise in cloud image data. By processing cloud images with the cascade filter, random noise in the image can be removed while preserving the brightness distribution and main cloud morphology.

[0145] 2. Establish a GPAformer prediction model. By capturing the temporal characteristics of meteorological data, GPAformer can learn how meteorological factors change over time, thereby effectively predicting future meteorological data. This model has a significant impact on photovoltaic power, and accurately predicting meteorological data provides a solid foundation for photovoltaic power prediction.

[0146] 3. SA-ConvLSTM combines a convolutional neural network (CNN) with a long short-term memory network (LSTM). It can effectively extract the spatiotemporal features of cloud images, capturing the cloud's motion trajectory and temporal changes. By introducing a cloud cover model, it can correct model prediction deviations when the cloud layer is dense or changing rapidly, ensuring that it can still provide relatively accurate prediction results under complex meteorological conditions.

[0147] 4. Establishing the KAN-COGCN model, which can effectively integrate meteorological factors, cloud image features, and historical photovoltaic power data; through the collaborative processing of multimodal data, the model can capture the complex relationships between various input sources and improve the prediction accuracy of photovoltaic power;

[0148] 5. This paper improves the Alpha Evolution Algorithm (AE) by introducing the Sobol sequence and normal cloud mutation strategy to improve the initialization process of the original Alpha Evolution Algorithm (AE). This allows for better exploration of the parameter space and finds the optimal parameter combination for the GPAformer, SA-Convlstm, and KAN-COGCN models. Furthermore, through a global optimization algorithm, the performance of the model is improved, ensuring the robustness and accuracy of the prediction model under different conditions.

[0149] 6. The adaptive wavelet RBF (radial basis function) neural network can correct the errors in the prediction results and adapt to the errors of the model in practical applications. Through wavelet analysis, it can effectively capture the multi-scale characteristics of the signal and dynamically adjust the prediction error of the model, thereby improving the accuracy of the final prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0150] Figure 1 This is a flow chart of the multi-modal short-term photovoltaic power prediction method based on satellite cloud images;

[0151] Figure 2 A schematic diagram of the GPAformer model structure provided by the present invention;

[0152] Figure 3 Schematic diagram of the SAM module structure in the SA-ConvLSTM model provided by the present invention;

[0153] Figure 4 Schematic diagram of the overall structure of the SA-ConvLSTM model provided by the present invention;

[0154] Figure 5 Schematic diagram of the CSEConv module in the KAN-COGCN model provided by the present invention;

[0155] Figure 6 Schematic diagram of the OffsetAttention module in the KAN-COGCN model provided by the present invention;

[0156] Figure 7 Schematic diagram of the flow chart of the improved IAE algorithm. DETAILED DESCRIPTION

[0157] The present invention will be described in further detail below with reference to the accompanying drawings.

[0158] like Figure 1 As shown, the present invention proposes a multi-modal short-term photovoltaic power prediction method based on satellite cloud images, and the specific steps are as follows:

[0159] Step 1: Collect real-time meteorological data and historical power data of photovoltaic power plants; at the same time, obtain cloud image data and preprocess the data set.

[0160] (1.1): Establish a meteorological dataset whose variables include real-time meteorological information such as temperature, humidity, wind speed, wind direction, air pressure, irradiance, and cloud cover; establish a historical power dataset of photovoltaic power stations; divide the meteorological dataset and historical photovoltaic power dataset into training and test sets in a ratio of 4:1; and simultaneously, obtain cloud image data, including cloud thickness, cloud motion trajectory, and density, through all-sky imaging tracking and monitoring.

[0161] (1.2): Data cleaning is performed on the above three data sets respectively. The photovoltaic data and meteorological data are preprocessed using the Newton interpolation method. The data are screened, all abnormal data are reset to zero, and missing data are filled to effectively improve the quality and integrity of the data. The specific calculation formula is as follows:

[0162]

[0163] Where: P represents the specific time point associated with the valid data, P′ is the time corresponding to the missing data, B is the zero-order difference quotient parameter, x1 is the linear data parameter, x2 is the nonlinear data parameter, and α is the missing data feature.

[0164] (1.3): To unify the dimensions and improve the model prediction fit, the cleaned meteorological and photovoltaic data sets are normalized so that the data values are all in the range of [0,1]. The specific calculation formula is as follows:

[0165]

[0166] in, represents the sequence after standardization, k is the original photovoltaic sequence or meteorological sequence sample, k max and k min Respectively represent the maximum and minimum values of the column data.

[0167] (1.4): Satellite cloud images are generated based on the differences in sunlight reflection from clouds of varying thickness. However, variations in the solar altitude can lead to discrepancies in cloud pixel recognition between morning and evening. Therefore, through image enhancement operations such as cloud image acquisition, geocalibration, and target area cropping, the cloud images are processed to a standard size of 224×224×3.

[0168] Step 2: Use Savitzky-Golay filtering to reduce noise in meteorological data and historical photovoltaic power data; use guided filtering cascade filter to reduce noise in cloud image data, including the following steps:

[0169] (2.1): The Savitzky-Golay smoothing filter method uses the principle of linear least squares to smooth the data by selecting a continuous range of data points and fitting them with a preset polynomial. The specific formula of the SG smoothing process is as follows:

[0170]

[0171] Where: is the smoothed data set; x j+i is the data set after collection; a i , a0 is the smoothing coefficient, that is, the original data x of period i j+i The weight in the smoothing window; n is the number of data in the sliding window; b is the window width, where n = 2b + 1.

[0172] (2.2): When processing meteorological data and historical photovoltaic power data sets, Savitzky-Golay filtering is used for noise reduction. i As the center, a window array containing 2T+1 sampling points is constructed, and the data in the window is fitted by a P-order polynomial q(n). The signal is smoothed using the least squares method, effectively removing noise while keeping the shape and width of the original signal unchanged. The specific formula is as follows:

[0173]

[0174] (2.3): Define a function F. When q(n) reaches its minimum value, the fitting effect is optimal. By sliding the window over the data sequence, the fitting value of each point in the original data is calculated in sequence. In the process, the noise is removed and smoothing filtering is achieved. The signal f(i) obtained after smoothing the signal is as follows:

[0175]

[0176] Where: f(i) is the filtered signal, D(k) is the sampling response of the smoothing filter, y(x) is a set of data in the original signal, x = -N,…,0,…N.

[0177] (2.4): When denoising cloud image data, a cascade filter consisting of a multi-level median rational mixture filter and a guided filter is used. First, a 3×3 window multi-level median rational mixture filter is used to remove salt and pepper noise from the cloud image through center-weighted median filtering (CWMF) and rational operations. Initial denoising is then performed by calculating new pixel values. The specific calculation formula is as follows:

[0178]

[0179] Among them, G is the filtered value, β1, β2, and β3 are the medians of the pixel values in windows 1, 2, and 3.

[0180] (2.5): In the guided filtering process, the input image p is filtered using the guided image E to obtain the output image q. This effectively reduces image noise while maintaining clear image edges. The specific formula is as follows:

[0181] q i =a n E i +b n

[0182] Where i∈ω n ,(a n ,b n ) is a linear coefficient that is assumed to be a constant.

[0183] (2.6): In window w n In order to solve a n and b n , a constraint condition needs to be established. Assume that z i is unnecessary noise in the window, then the relationship between the filtered image q and the input image p can be expressed as:

[0184] q i =p i -z i

[0185] (2.7): Combining the formulas in steps (2.5) and (2.6), in order to maintain the accuracy of the local linear model and minimize the deviation between the filtered output image q and the original input image p, thereby reducing the information loss caused by filtering, the least squares method is used to construct the loss function. The calculation formula is as follows:

[0186]

[0187] Among them, ξ is the regularization parameter, which is used to prevent the coefficient a n Too big.

[0188] (2.8): By solving the above loss function, we can get the linear coefficient a n and b n , the expression is as follows:

[0189]

[0190] Where: μ n and Indicates that the guidance image E is in the window ω n The mean and variance in |ω| is the window ω n The number of pixels in is the input image p in window ω n The mean in .

[0191] (2.9): For each pixel i in the image, calculate its q after filtering in multiple different windows i values, and take these q i The average value of the values is used as the final filtering result of the pixel to obtain the cloud image data after noise reduction. The calculation formula is as follows:

[0192]

[0193] Step 3: Create Figure 2 The GPAformer prediction model shown captures the temporal characteristics of meteorological factors and predicts the denoised meteorological data, including the following steps:

[0194] (3.1): The GPAformer model combines the advantages of Gaussian process and Autoformer, using Gaussian process to capture local changes and global trends, and Autoformer's ability to handle long time series dependencies and multi-scale dynamic characteristics to predict meteorological data sets.

[0195] (3.2): Autoformer splits meteorological and photovoltaic data into trend and cycle parts through sequence decomposition and autocorrelation mechanism, while suppressing local disorder. The specific calculation formula is as follows:

[0196] L t =Avg(pad(L))

[0197] L s =LL t

[0198] Where L is a sequence of length X, L∈R X×d ;L t is the short-term fluctuation part in the sequence; L s Represents the periodic smoothing portion remaining after removing short-term fluctuations.

[0199] (3.3): The autocorrelation mechanism calculates the autocorrelation coefficient P(τ) of the sequence by aggregating similar subsequences, which is used to reflect the periodic dependence of the sequence. The calculation formula of P(τ) is as follows:

[0200]

[0201] Where P(τ) represents the sequence L G With delay G steps after L G-τ The similarity between them is , and Y represents the number of delayed operations.

[0202] (3.4): To improve computational efficiency, Autoformer uses Fast Fourier Transform (FFT) to efficiently calculate the autocorrelation coefficient P(τ). The calculation formula is as follows:

[0203] S(f)=F(L G )F * (L G )

[0204] P(τ)=F -1 (Sxx(f))

[0205] Where S(f) represents the result of converting the time series to the frequency domain, F represents Fourier transform, and F * represents the Fourier transform conjugate, f represents the frequency, F -1 Represents the inverse Fourier transform.

[0206] (3.5): Gaussian process uses Gaussian kernel function to perform weighted averaging on time series data, removes high-frequency noise and retains the main trends and change characteristics, effectively achieving smoothing effect. For a given time series data {y1,y2,…y n}, smoothed value y′ after Gaussian process processing i The calculation formula is as follows:

[0207]

[0208] Where K(j) is the weight function of the Gaussian kernel; σ is the standard deviation of the Gaussian kernel; and U is the size of the smoothing window.

[0209] Step 4: Create Figure 4 The SA-ConvLSTM cloud image prediction model shown captures cloud motion trajectories and extracts temporal variation features to predict denoised cloud image data. It includes the following steps:

[0210] (4.1): The denoised cloud image data is fed into the SA-ConvLSTM model for prediction. The SA-ConvLSTM model consists of a ConvLSTM module and a self-attention memory module (SAM).

[0211] (4.2): Construct a ConvLSTM module. This module uses a convolutional neural network (CNN) to extract spatial features and replaces the element-by-element multiplication operation in the traditional LSTM with a convolution operation, thereby effectively capturing the spatiotemporal variation of the cloud clusters and completing the construction of the ConvLSTM model. The specific calculation formula is as follows:

[0212]

[0213] Among them, X t is the input data; Ht-1 is the hidden state at the previous moment; C t-1 is the state memory of the previous moment; b is the deviation term; y t is the input gate; v t For the forget gate; t is the output gate; * represents the convolution operation; Represents the Hadamard product operation.

[0214] (4.3): If Figure 3 The self-attention mechanism (SA) shown in Figure 3 can capture the global information of spatiotemporal data, but it lacks memory capacity and is difficult to model long-term dependencies. Therefore, the self-attention memory module (SAM) is introduced to store and update historical spatial information, so that the model can remember the relationship between adjacent points for a long time. SAM is mainly composed of the historical memory state M t-1 and self-attention calculation unit.

[0215] (4.4): SAM takes input information X t Perform linear transformation to generate query Q h , key K h Sum V h , the specific calculation formula is as follows:

[0216] Q h =W q X t ,K h =W k X t ,V h =W v X t

[0217] (4.5): According to the memory state M of the previous moment t-1 , generate the corresponding key-value pair K m Sum V m , the expression is as follows:

[0218] K m ,V m =f(M t-1 )

[0219] (4.6): Using input information X t Generated query Q h , respectively with key K h and K m Calculate the similarity and get Z h , Z m , combine these two results into Z, and then generate the output H t And the corresponding gate control unit, the calculation formula is:

[0220]

[0221] y t ′=β(W yz *Z+W yh *H t +b y )

[0222] g t ′=tanh(W gz *Z+W gh *H t +b g )

[0223] O′ t =β(W oz *Z+W oh *H t +b o )

[0224] Among them, y′ t represents the input gate, g′ t represents the memory update gate, O′ t Represents the output gate.

[0225] (4.7): According to the calculated input gate y′ t , memory update gate g′ t and output gate O′ t , for the self-attention memory state M t and output results To update, the calculation formula is as follows:

[0226]

[0227] Where M t-1 is the self-attention memory state at time t-1.

[0228] (4.8) A cloud cover model is introduced to extract cloud image data from the latitude and longitude coordinates of the area tracked and monitored by the all-sky imager. This data is used to determine the distribution of clouds within the area and capture the dynamic changes in cloud movement and thickness, thereby reducing the impact of cloud cover on the accuracy of photovoltaic power station power forecasts.

[0229] The DeepLabv3+ network's semantic segmentation technology was used to classify satellite cloud images at the pixel level, extracting pixel information from cloud regions. The classified cloud images were grayscaled by setting the background to white and the clouds to dark, extracting grayscale values to assess the impact of cloud cover on photovoltaic power.

[0230] Calculate the grayscale mean of the cloud map in the area where the photovoltaic power station is located (i n), as an indicator of the degree of cloud cover, and calculate the correction coefficient (η) to correct the preliminary predicted photovoltaic power value:

[0231]

[0232] Among them, i n is the grayscale mean of the cloud image, i max is the maximum grayscale value.

[0233] The preliminary prediction results are corrected to obtain the final photovoltaic prediction value. The calculation formula is as follows:

[0234] P=k×η×H t +m×H r

[0235] Among them, p represents the final prediction result; H t Indicates the preliminary prediction result; H r represents the power value at the previous moment; η represents the correction coefficient; k represents the power prediction correction coefficient; m represents the power prediction correction coefficient.

[0236] The self-attention memory module SAM is integrated into the ConvLSTM model to construct the SA-ConvLSTM model. Compared with the traditional ConvLSTM, the SAM module enhances the hidden state H t The spatial location information of the model improves the model's ability to capture spatiotemporal features, thereby improving prediction accuracy.

[0237] Step 5: Establish the photovoltaic power KAN-COGCN prediction model, and use the data predicted by the two models in steps (3) and (4) as input to achieve the final photovoltaic power prediction:

[0238] (5.1): The KAN model introduces meteorological factors or photovoltaic power generation patterns to construct an adjacency matrix that is more consistent with actual patterns and can handle the complex nonlinear relationship between meteorological factors and historical photovoltaic power. At the same time, by combining spline functions and multi-layer perceptrons (MLPs), KAN achieves high prediction accuracy in low-dimensional space, thereby efficiently handling photovoltaic power prediction tasks. The KAN model formula is:

[0239]

[0240] in, is an external function; It is an internal function. The model uses spline function as a learnable activation function and changes its shape by adjusting parameters.

[0241] (5.2): A KAN layer can be viewed as a one-dimensional function matrix. To construct a deep KAN model, we only need to stack multiple KAN layers. The transfer relationship between the input and output layers can be expressed in matrix form as follows:

[0242]

[0243] in, is the matrix function representing the l-th layer KAN, is the activation function on each edge, and the number of nodes in the KAN layer is determined by the number of input nodes.

[0244] (5.3): The cascade relationship of multi-layer functions is further expressed in matrix form to enhance the computational efficiency, parameter sharing capability, and parallel computing performance of the photovoltaic power prediction model. The specific expression is as follows:

[0245]

[0246] Among them, the symbol in the upper right corner of the formula Represents a composite operation of a function.

[0247] (5.4): The core of the COGCN model is to convert Figure 5 The local features extracted by the CSEConv module are fused with the contextual features captured by the OffsetAttention module. The CSEConv module constructs a directed graph L through the KNN algorithm to extract the spatiotemporal features of cloud movement and capture the impact of cloud cover on photovoltaic power. Specifically, with the center point H i and its K nearest neighbor points H( i,j )(j=1,2,…,k), the KNN algorithm is used to connect each neighboring point with the center point through a directed edge. The directed graph L constructed by the KNN algorithm is defined as:

[0248] L=(M,E)

[0249]

[0250] I={e i =(e i1 ,e i2 ,...,e ij )|i=1,2,3,...,n}

[0251] Where Q represents n local vertices, and each local vertex is a C-dimensional feature point H i , I represents the set of edges between vertices.

[0252] (5.5): After constructing the graph structure L through the KNN algorithm, a multi-layer perceptron w is used, which includes a convolutional layer (Conv), a normalization layer (BN), and a RELU activation layer. θ Extract edge features u from the graph structure i,j . Let u i,j Indicates the center point H i The edge feature between its nearest neighbor H(i, j) is defined as follows:

[0253] u ij =w θ (H i ,H i -H (i,j) )

[0254]

[0255] (5.6): Based on the edge feature u calculated above i,j , further normalized and activated, the local domain feature Y is finally obtained i , specifically defined as follows:

[0256] Y i =ReLu(BN(u ij )

[0257] (5.7): The extracted local neighborhood features are further enhanced by the CBAM module, which extracts key information at different locations through channel weighting and spatial weighting. Specifically, the input feature F∈R C*H*W First, one-dimensional convolution O with the channel weighted module c ∈R C*1*1 Multiply them together to get D′, which is then combined with the two-dimensional convolution O of the spatial weighted module. s ∈R 1*H*W Multiply them together and finally get the output feature D", which is calculated as follows:

[0258]

[0259] Among them, O C (D) and O S The formula for (D′) is as follows:

[0260] O C (D)=γ(MLP(AvgPool(D))+MLP(MaxP(D)))

[0261] O s (D′)=γ(Conv(Concat[Avgpool(D′);MaxP(D′)]))

[0262] Among them, γ represents the sigmoid activation function, MLP represents the multi-layer perceptron, AvgPool and MaxP represent the average pooling and maximum pooling operations respectively, Conv represents the convolution operation, and Concat represents the tensor splicing operation.

[0263] (5.8): The CSEConv module uses a multi-layer perceptron and a maximum pooling function to finally obtain the local domain features CSE of the cloud data. out , the expression is as follows:

[0264] CSE out =MaxP(MLP(F″))

[0265] (5.9): Introduce Figure 6 The Offset Attention module shown in the figure captures the context features. For the input feature F in ∈R N×de , generate query Q, key K and value V matrices through linear transformation, and the transformation process is as follows:

[0266] (Q,K,V)=B in ·(W Q ,W K ,W V )

[0267]

[0268] Among them, W Q ,W K , W V is the shared learnable weight, T a is the dimension of Q and K, T e is the dimension of V.

[0269] (5.10): The attention weight is calculated using the matrix dot product of the query matrix Q and the key matrix K. The specific calculation formula is as follows:

[0270] E=α i,j =Q·K T

[0271] (5.11): Normalize the obtained weights to obtain E = α i,j :

[0272]

[0273] (5.12): Use the calculated attention weights and the value matrix V to perform weighted summation to obtain the self-attention output feature B sa :

[0274] B sa =L·V

[0275] (5.13): Calculate the input feature B in And the self-attention output feature B sa The difference is input to the LBR layer and compared with B in Add them together and process them through the multi-layer perceptron MLP to finally get the output feature B of the Offset Attention module out :

[0276] B out =OL(B in )=MLP(LBR(B in -B sa )+B in )

[0277] Among them, the LBR layer is a network layer that combines Linear, BatchNorm and ReLU functions.

[0278] Step 6: Use the improved alpha evolution algorithm (IAE) to simultaneously optimize the hyperparameters of the GPAformer, SA-Convlstm model, and photovoltaic power KAN-COGCN prediction model to search for the optimal solution of the three model parameters; the IAE algorithm is based on the alpha evolution algorithm AE, and uses the Sobol sequence and normal cloud mutation strategy to improve the original population initialization process to obtain the IAE algorithm.

[0279] Hyperparameters of the GPAformer, SA-convLSTM, and KAN-COGCN models, including the standard deviation of the Gaussian kernel (σ), smoothing window size (k), pooling window size, time step in the autocorrelation mechanism, number of heads in the self-attention mechanism, hidden layer dimension, activation function type, and number of delay operations (L) in the GPAformer model; the convolution kernel size, time step, and number of attention heads of SA-convLSTM; the number of layers in KAN and the number of nodes in each layer, the learning rate and number of training rounds in the KAN-COGCN model; the number of CSEConv modules, the number of OffsetAttention modules, the number of nearest neighbors (K), and the output dimension of the feature extraction layer in COGCN.

[0280] like Figure 7 As shown in Figure 2, the implementation process of the improved Alpha evolution algorithm is as follows:

[0281] (6.1): First, set the population size, dimension, number of iterations, and upper and lower limits of the search space of the AE algorithm.

[0282] (6.2): The Sobol sequence strategy is introduced into the algorithm to enhance the uniform distribution of the population, and the normal cloud mutation strategy is combined to improve the diversity of the population, thereby completing the population initialization. The improved formula is as follows:

[0283] X i =Sobol(LB,UB,N)+ΔX i

[0284] Among them, X i represents the initialized i-th individual, Sobol(LB,UB,N) represents the generation of N Sobol sequence samples in the search space (LB,UB), ΔX i ~N(0,σ) represents the normal cloud mutation strategy, where σ controls the mutation amplitude to improve the diversity of the population.

[0285] (6.3): In the algorithm, all candidate solutions form the candidate matrix, and the evolution matrix is its subset. During evolution, instead of directly updating each solution X in the candidate matrix, some individuals are selected through sampling with replacement to form the evolution matrix, and the evolution operation is performed only on these individuals. The specific expression is as follows:

[0286]

[0287] (6.4): An efficient search operator is key to optimizing algorithms in finding optimal solutions in the solution space. The search process is implemented using the α operator, which can simultaneously extract and utilize evolutionary information within a single operator, significantly improving search efficiency. The specific calculation formula is as follows:

[0288]

[0289] (6.5): The adaptive basis vector plays a core role in the solution update process. It not only determines the initial direction of the search, but also affects the development trend of the entire evolution process. Its specific expression is as follows:

[0290]

[0291] Among them, diag is a function that generates a D-order square matrix, which is obtained by sampling by multiplying D with the candidate solution. The dimension of the matrix B is K×D, which is generated by sampling without replacement. And ω i:K Represents the weight of the solution in the equation.

[0292] (6.6): Construct two different evolutionary paths. The specific calculation formula is as follows:

[0293]

[0294] In the formula, c a and cb Represents L a and L b The learning rate.

[0295] (6.7): Random step size αΔr i The algorithm is given the ability to search globally, and the nonlinear decreasing attenuation factor is used to gradually shrink the search range. The specific calculation formula is as follows:

[0296]

[0297] Here, s represents the fitness evaluation, and Max represents the maximum fitness evaluation (the maximum number of calls to the objective function).

[0298] (6.8): Δr is defined as a disturbance that gradually weakens with time delay and is calculated as follows:

[0299] Δr=(ub-lb)·(2R1·R2-R2)·S

[0300] Here, R1 and R2 represent random real number matrices generated by rand(0,1,[N,D]), while represents a random integer matrix containing only 0 or 1 generated by randi(0,1,[N,D]), weighing the perturbation based on the dimension.

[0301] (6.9) The boundary constraint method ensures that the algorithm can search effectively within the search space. In the AE algorithm, this goal is achieved by adopting the distance halving method. The specific calculation formula is as follows:

[0302]

[0303] (6.10): In the AE algorithm, a greedy selection strategy is introduced. This strategy directly passes the best performing solutions to the next generation without additional computation, thereby improving the efficiency of evolution. Its mathematical model is as follows:

[0304]

[0305] (6.11): We search for the global optimal solution based on the individual update formula in the population and gradually optimize the hyperparameters of the three models to obtain the optimal parameter combination for the GPAformer, SA-convLSTM, and KAN-COGCN models. During the optimization process, we use the root mean square error (RMSE) to evaluate the parameters of the three models. The accuracy of the model prediction results is measured by minimizing the RMSE value, thereby obtaining the optimal prediction results.

[0306] (6.12): Determine whether the maximum number of iterations has been reached. If not, return to step (6.3) until the maximum number of iterations is reached and output the optimal parameter configuration. The three models use the optimal parameter configuration to predict the components and obtain the final prediction results.

[0307] Step 7: Establish an adaptive wavelet RBF neural network to correct the error of the prediction result and obtain the final prediction result.

[0308] Construct a wavelet neural network based on radial basis function network structure, including input layer, first hidden layer, second hidden layer and output layer; suppose the input layer receives sample sequence X=(x1,x2,…,x n ), at the same time, the wavelet function and radial basis function are established, and the output of the first hidden layer is:

[0309]

[0310] Where F1(X) is the wavelet function, ψ(x i ) is the wavelet basis function, a is the wavelet expansion factor, b is the wavelet translation factor, is the output of the first hidden layer, is the wavelet network scale variable, t i is the wavelet network translation variable, M is the number of neurons in the first hidden layer, R i is a scale variable.

[0311] The output of the second hidden layer is expressed as:

[0312]

[0313] Where, L u is the center point of the radial basis function, σ u is the width variable of the radial basis function, and N is the number of neurons in the second hidden layer.

[0314] The K-means clustering algorithm is introduced to perform cluster analysis on the signal after wavelet transformation of the first hidden layer, and the center point l of the radial basis function of the second hidden layer is quickly determined. u . Thus the output of the wavelet RBF neural network is completed:

[0315]

[0316] Where y is the output of the wavelet RBF network, ω u is the weight between the second hidden layer and the output layer, is the output of the second hidden layer, and g is the quantization noise.

[0317] The AGA algorithm is used to correct the network parameters. First, the prediction error e of the network is calculated. The calculation formula is as follows:

[0318]

[0319] Where, is the expected output value of the network, is the actual output of the network.

[0320] Calculate the gradient G e With the squared gradient accumulation R and the learning rate r, the specific calculation formula is as follows:

[0321]

[0322] Where η is the initial learning rate and ε is the minimum value; according to the network's prediction error e, the network parameters are corrected:

[0323]

[0324] Where W(u) is the weight of the wavelet neural network, a(k) is the scaling scale of the wavelet function, b(k) is the translation factor of the wavelet function, and γ is the momentum factor, which ranges from 0 to 1. The cycle operation is performed until the output of the wavelet neural network reaches the expected value, and the final photovoltaic power prediction result is obtained.

[0325] The present invention also proposes a device comprising a memory and a processor, wherein: the memory is used to store a computer program that can be run on the processor; the processor is used to execute the steps of the above-mentioned multi-modal short-term photovoltaic power prediction method based on satellite cloud images when running the computer program.

[0326] The above embodiments are intended only to illustrate the technical concepts and features of the present invention. Their purpose is to enable those skilled in the art to understand the contents of the present invention and implement them accordingly. They are not intended to limit the scope of protection of the present invention. Any equivalent changes or modifications made in accordance with the spirit of the present invention are intended to be covered by the scope of protection of the present invention.

Claims

1. A multimodal short-term photovoltaic power prediction method based on satellite cloud images, characterized in that: The following steps are involved: (1) Obtain real-time meteorological data, historical photovoltaic power data, and cloud image data, and pre-process the data; (2) Savitzky-Golay filtering is used to reduce noise in meteorological data and historical photovoltaic power data, smoothing the time series signals and retaining key trends; cloud image data is reduced in noise by using a cascade filter of guided filtering to retain the overall brightness distribution and the main morphology of the cloud and reduce noise interference; (3) Establish a GPAformer prediction model to capture the temporal characteristics of meteorological factors and predict the denoised meteorological data; (4) Establishing the SA-ConvLSTM model to capture cloud motion trajectories and extract temporal variation features, predicting cloud image data, and introducing a cloud cover model to correct the prediction bias of SA-ConvLSTM when the cloud layer is dense and changing rapidly; (5) Establish a photovoltaic power KAN-COGCN prediction model, and integrate the meteorological factor data and cloud map motion time series characteristics predicted by the two models in steps (3) and (4) with the historical photovoltaic power data as input to realize photovoltaic power prediction; (6) Using the improved alpha evolution algorithm IAE to simultaneously optimize the hyperparameters of the GPAformer, SA-Convlstm model, and photovoltaic power KAN-COGCN prediction model, and search for the optimal solution of the three model parameters; the IAE algorithm is based on the alpha evolution algorithm AE, and adopts the Sobol sequence and normal cloud mutation strategy to improve the original population initialization process to obtain the IAE algorithm; (7) Establish an adaptive wavelet RBF neural network to correct the error of the prediction results and obtain the final prediction results.

2. The multimodal short-term photovoltaic power prediction method based on satellite cloud images according to claim 1 is characterized in that: The implementation process of step (1) is as follows: The meteorological data and historical photovoltaic power data were preprocessed using the Newton interpolation method. The data were screened, all abnormal data were reset to zero, and missing data were filled. The cleaned meteorological and photovoltaic data sets were normalized. The cloud image data were processed to a standard size through image enhancement operations including cloud image acquisition, geographic calibration, and target area cropping.

3. The multimodal short-term photovoltaic power prediction method based on satellite cloud images according to claim 1 is characterized in that: The implementation process of step (2) is as follows: When processing meteorological data and historical photovoltaic power data sets, Savitzky-Golay filtering is used for noise reduction. i As the center, a window array containing 2T+1 sampling points is constructed, and the data in the window is fitted by a P-order polynomial q(n). The signal is smoothed using the least squares method: Define a function F. When q(n) reaches its minimum value, the fitting effect is optimal. By sliding the window on the data sequence, the fitting value of each point in the original data is calculated in turn. In the process, the noise is removed and smoothing filtering is achieved. The signal f(i) obtained after smoothing the signal is as follows: Where f(i) is the filtered signal, D(k) is the sampling response of the smoothing filter, y(x) is a set of data in the original signal, x = -N,…,0,…N; When denoising cloud image data, a cascade filter consisting of a multi-level median rational mixture filter and a guided filter is used. First, a multi-level median rational mixture filter with a 3×3 window is used to remove salt and pepper noise from the cloud image through center-weighted median filtering and rational operations. Initial denoising is then performed by calculating new pixel values. The specific calculation formula is as follows: Where G is the filtered value, β1, β2, and β3 are the medians of the pixel values in windows 1, 2, and 3; In the guided filtering process, the input image p is filtered using the guided image E to obtain the output image q. This effectively reduces image noise while keeping the image edge clear. The specific formula is as follows: q i =a n E i +b n Where i∈ω n ,(a n ,b n ) is the linear coefficient which is assumed to be a constant; In the window n In order to solve a n and b n , establish a constraint condition; assuming z i is unnecessary noise in the window, then the relationship between the filtered image q and the input image p is: q i =p i -z i The least squares method is used to construct the loss function, and the calculation formula is as follows: Among them, ξ is the regularization parameter, which is used to prevent the coefficient a n Too large; by solving the loss function, we get the linear coefficient a n and b n , the expression is as follows: Where μ n and Indicates that the guidance image E is in the window ω n The mean and variance in |ω| is the window ω n The number of pixels in is the input image p in window ω n The mean of For each pixel i in the image, calculate its filtered q in multiple different windows i values, and take these q i The average value of the values is used as the final filtering result of the pixel to obtain the denoised cloud image data:

4. The multimodal short-term photovoltaic power prediction method based on satellite cloud images according to claim 1 is characterized in that: The implementation process of step (3) is as follows: Autoformer uses sequence decomposition and autocorrelation mechanisms to split meteorological and photovoltaic data into trend and cycle parts, while suppressing local disorder. The specific calculation formula is as follows: L t =Avg(pad(L)) L s =LL t Where L is a sequence of length X, L∈R X×d ;L t is the short-term fluctuation part in the sequence; L s represents the periodic smoothing portion remaining after removing short-term fluctuations; The autocorrelation mechanism calculates the autocorrelation coefficient P(τ) of the sequence by aggregating similar subsequences. It is used to reflect the periodic dependency of the sequence. The calculation formula of P(τ) is as follows: Where P(τ) represents the sequence L G With delay G steps after L G-τ The similarity between them, Y represents the number of delayed operations; Autoformer uses fast Fourier transform FFT to calculate the autocorrelation coefficient P(τ). The calculation formula is as follows: S(f)=F(L G )F*(L G ) P(τ)=F -1 (Sxx(f)) Where S(f) represents the result of converting the time series to the frequency domain, F represents Fourier transform, and F * represents the Fourier transform conjugate, f represents the frequency, F -1 represents the inverse Fourier transform; For a given time series data {y1,y2,…y n }, smoothed value y′ after Gaussian process processing i The calculation formula is as follows: Where K(j) is the weight function of the Gaussian kernel; σ is the standard deviation of the Gaussian kernel; and U is the size of the smoothing window.

5. The multimodal short-term photovoltaic power prediction method based on satellite cloud images according to claim 1 is characterized in that: The implementation process of step (4) is as follows: The SA-ConvLSTM model includes the ConvLSTM module and the self-attention memory module SAM; The ConvLSTM module uses a convolutional neural network (CNN) to extract spatial features and replaces the element-by-element multiplication operation in the traditional LSTM with a convolution operation to capture the spatiotemporal variation of cloud clusters and complete the construction of the ConvLSTM model. The specific calculation formula is as follows: Among them, X t is the input data; H t-1 is the hidden state at the previous moment; C t-1 is the state memory of the previous moment; b is the deviation term; y t is the input gate; v t For the forget gate; t is the output gate; * represents the convolution operation; Represents the Hadamard product operation; The self-attention memory module SAM is used to store and update historical spatial information, so that the model can remember the relationship between adjacent points for a long time. SAM consists of the historical memory state M t-1 and self-attention calculation unit; SAM is composed of input information X t Perform linear transformation to generate query Q h , key K h Sum V h ; According to the memory state M of the previous moment t-1 , generate the corresponding key-value pair K m Sum V m ; Using input information X t Generated query Q h , respectively with key K h and K n Calculate the similarity and get Z h , Z m , combine these two results into Z, and then generate the output H t And the corresponding gate unit; according to the calculated input gate y′ t , memory update gate g′ t and output gate O′ t , for the self-attention memory state M t and output results To update, the calculation formula is as follows: Where M t-1 is the self-attention memory state at time t-1; SAM is integrated into the ConvLSTM model to construct the SA-ConvLSTM model; A cloud cover model was introduced to extract cloud image data from the latitude and longitude coordinates of the region captured by the all-sky imager tracking and monitoring. This data was used to determine the distribution of clouds within the region and capture the dynamic changes in cloud motion and thickness. Semantic segmentation technology using the DeepLabv3+ network was used to classify satellite cloud images at the pixel level and extract pixel information about the cloud region. The classified cloud images were grayscaled by setting the background to white and the clouds to dark, and grayscale values were extracted to assess the impact of cloud cover on photovoltaic power. Calculate the grayscale mean i of the cloud map in the area where the photovoltaic power station is located n , as an indicator to measure the degree of cloud cover, and calculate the correction coefficient η to correct the preliminary predicted photovoltaic power value: Among them, i n is the grayscale mean of the cloud image, i max is the maximum value of the grayscale value; The preliminary prediction results are corrected to obtain the final photovoltaic prediction value. The calculation formula is as follows: P=k×η×H t +m×H r Among them, p represents the final prediction result; H t Indicates the preliminary prediction results; H r represents the power value at the previous moment; η represents the correction coefficient; k represents the power prediction correction coefficient; m represents the power prediction correction coefficient.

6. The multi-modal short-term photovoltaic power prediction method based on satellite cloud images according to claim 1, characterized in that: The implementation process of step (5) is as follows: The model formula of KAN is: in, is an external function; is an internal function. The model uses a spline function as a learnable activation function, and its shape is changed by adjusting parameters. A KAN layer is regarded as a one-dimensional function matrix. Multiple KAN layers are simply stacked to construct a deep KAN model. The cascade relationship of multiple layers of functions is further expressed in matrix form to enhance the computational efficiency, parameter sharing capability, and parallel computing performance of the photovoltaic power prediction model. The specific expression is as follows: The symbol "°" in the upper right corner of the formula represents the composite operation of the function; The CSEConv module of the COGCN model constructs a directed graph L through the KNN algorithm to extract the spatiotemporal characteristics of cloud movement and capture the impact of cloud cover on photovoltaic power. i And its K nearest neighbor points H (i，j) Based on the KNN algorithm, each neighboring point is connected to the center point through a directed edge. The directed graph L constructed using the KNN algorithm is defined as: L=(M,E) I={e i =(e i1 ,e i2 ,...,e ij )|i=1,2,3,...,n} Where Q represents n local vertices, and each local vertex is a C-dimensional feature point H i , I represents the set of edges between vertices; after constructing the graph structure L through the KNN algorithm, a multi-layer perceptron w is used, which includes a convolutional layer Conv, a normalization layer BN, and a RELU activation layer. θ Extract edge features u from the graph structure i,j ; Let u i,j Indicates the center point H i The edge feature between its nearest neighbor H(i, j) is defined as follows: u ij =w θ (H i ,H i -H (i,j) ) Based on edge feature u i,j , further normalized and activated, the local domain feature Y is finally obtained i : Y i =ReLu(BN(u ij ) The extracted local neighborhood features are further enhanced by the CBAM module, and the key information at different locations is extracted through channel weighting and spatial weighting processing; the CSEConv module finally obtains the local domain features CSE of the cloud data through the multi-layer perceptron and the maximum pooling function. out : CSE out =MaxP(MLP(F″)) The OffsetAttention module is introduced to capture context features. For the input feature B in ∈R N×de , generate query Q, key K and value V matrices through linear transformation, and the transformation process is as follows: (Q,K,V)=B in ·(W Q ,W K ,W V ) V∈R N×Te Among them, W Q ,W K , W V is the shared learnable weight, T a is the dimension of Q and K, T e is the dimension of V; the attention weight is calculated using the matrix dot product of the query matrix Q and the key matrix K. The specific calculation formula is as follows: E❝α i,j =Q·K T Then, the obtained weights are normalized to obtain E = α i,j : The calculated attention weights are weighted and summed with the value matrix V to obtain the self-attention output feature B sa : B sa =L·V Calculate input feature J in And the self-attention output feature B sa The difference is input to the LBR layer and compared with B in Add them together and process them through the multi-layer perceptron MLP to finally get the output feature B of the OffsetAttention module out : B out =OL(B in )=MLP(LBR(B in -B sa )+B in ) Among them, the LBR layer is a network layer that combines Linear, BatchNorm and ReLU functions.

7. The multimodal short-term photovoltaic power prediction method based on satellite cloud images according to claim 1, characterized in that: The implementation process of step (6) is as follows: Set the population size, dimension, number of iterations, and upper and lower limits of the search space of the AE algorithm; introduce the Sobol sequence strategy into the algorithm to enhance the uniform distribution of the population, and combine it with the normal cloud mutation strategy to improve the diversity of the population, thereby completing the population initialization. The improved formula is as follows: X i =Sob(LB,UB,N)+ΔX i Among them, X i represents the initialized i-th individual, Sobol(LB,UB,N) represents the generation of N Sobol sequence samples in the search space (LB,UB), ΔX i ~N(0,σ) represents the normal cloud mutation strategy, and σ controls the amplitude of the mutation; By performing a sampling with replacement operation, some individuals are selected to form the evolution matrix, and the evolution operation is performed only on these individuals. The specific expression is as follows: The search process is implemented through the α operator, which can simultaneously extract and utilize evolutionary information within a single operator: The adaptive basis vector plays a core role in the solution update process. Its specific expression is as follows: Among them, diag is a function that generates a D-order square matrix, which is obtained by sampling by multiplying D with the candidate solution; the dimension of matrix B is K×D, ω i:K represents the weight of the solution in the equation; Construct two different evolutionary paths. The specific calculation formula is as follows: Where c a and c a Represents L a and L b The learning rate; Random step size αΔr i The algorithm is given the ability to search globally, and the nonlinear decreasing attenuation factor is used to gradually shrink the search range. The specific calculation formula is as follows: Where s represents the fitness evaluation, and Max represents the maximum fitness evaluation; Δr is a disturbance that gradually weakens with time delay, and is calculated as follows: Δr=(ub-lb)·(2R1·R2-R2)·S Here, R1 and R2 represent random real number matrices generated by rand(0,1,[N,D]), while represents a random integer matrix containing only 0 or 1 generated by randi(0,1,[N,D]), weighing the perturbation based on the dimension; The boundary constraint method ensures that the algorithm performs an effective search within the search space. In the AE algorithm, this goal is achieved by adopting the distance halving method. The specific calculation formula is as follows: In the AE algorithm, a greedy selection strategy is introduced: Based on the individual update formula in the population, the global optimal solution is searched, and the hyperparameters of the three models are gradually optimized to obtain the optimal parameter combination of the GPAformer, SA-convLSTM, and KAN-COGCN models. During the optimization process, the root mean square error (RMSE) is used to evaluate the parameters in the three models, and the accuracy of the model prediction results is measured by minimizing the RMSE value, thereby obtaining the optimal prediction result.

8. The multi-modal short-term photovoltaic power prediction method based on satellite cloud images according to claim 1 is characterized in that: The hyperparameters of the GPAformer, SA-ConvLSTM model and photovoltaic power KAN-COGCN prediction model in step (6) include the standard deviation σ of the Gaussian kernel, the smoothing window size k, the size of the pooling window, the time step in the autocorrelation mechanism and the number of heads in the self-attention mechanism, the hidden layer dimension, the activation function type, and the number of delay operations L in the GPAformer model; the convolution kernel size, time step, and number of attention heads of SA-convLSTM; the number of layers in KAN in the KAN-COGCN model, the number of nodes in each layer, the learning rate of KAN, and the number of training rounds; the number of CSEConv modules, the number of Offset Attention modules, the number of nearest neighbors K, and the output dimension of the feature extraction layer in COGCN.

9. The multimodal short-term photovoltaic power prediction method based on satellite cloud images according to claim 1, characterized in that: The implementation process of step (7) is as follows: Construct a wavelet neural network based on radial basis function network structure, and assume that the input layer receives the sample sequence X=(x1,x2,…,x n ), at the same time, the wavelet function and radial basis function are established, and the output of the first hidden layer is: Where F1(X) is the wavelet function, ψ(x i ) is the wavelet basis function, a is the wavelet expansion factor, b is the wavelet translation factor, is the output of the first hidden layer, is the wavelet network scale variable, t i is the wavelet network translation variable, M is the number of neurons in the first hidden layer, R i is a scale variable; The output of the second hidden layer is expressed as: Where, L u is the center point of the radial basis function, σ u is the width variable of the radial basis function, and N is the number of neurons in the second hidden layer; The K-means clustering algorithm is introduced to perform cluster analysis on the signal after wavelet transformation of the first hidden layer, and the center point l of the radial basis function of the second hidden layer is quickly determined. u ; Thus completing the output of the wavelet RBF neural network: Where y is the output of the wavelet RBF network, ω u is the weight between the second hidden layer and the output layer, is the output of the second hidden layer, g is the quantization noise; The AGA algorithm is used to correct the network parameters. First, the prediction error e of the network is calculated. The calculation formula is as follows: Where, is the expected output value of the network, is the actual output of the network; Calculate the gradient G e With the squared gradient accumulation R and the learning rate r, the specific calculation formula is as follows: Where η is the initial learning rate and ε is the minimum value; according to the network's prediction error e, the network parameters are corrected: Where W(u) is the weight of the wavelet neural network, a(k) is the scaling scale of the wavelet function, b(k) is the translation factor of the wavelet function, and γ is the momentum factor, which ranges from 0 to 1. The cycle operation is performed until the output of the wavelet neural network reaches the expected value, and the final photovoltaic power prediction result is obtained.

10. A device, characterized in that: comprising a memory and a processor, wherein: a memory for storing computer programs capable of running on the processor; A processor is configured to execute the steps of a multimodal short-term photovoltaic power prediction method based on satellite cloud images as described in any one of claims 1 to 9 when running the computer program.

Images