Uncertainty perception multi-source data fusion photovoltaic power prediction method

By combining the improved FCN-Transformer model and OmniCast-NLLPC model with uncertainty perception technology, the problems of cloud change capture and uncertainty quantification in photovoltaic power prediction are solved, achieving high-precision and robust photovoltaic power prediction, supporting grid optimization scheduling and energy storage applications.

CN121863369APending Publication Date: 2026-04-14HUAIYIN INSTITUTE OF TECHNOLOGY
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
HUAIYIN INSTITUTE OF TECHNOLOGY
Filing Date
2026-01-16
Publication Date
2026-04-14

AI Technical Summary

Technical Problem

Existing photovoltaic power prediction methods struggle to accurately capture rapid cloud changes and lack uncertainty quantification, resulting in insufficient prediction accuracy and robustness.

Method used

An improved FCN-Transformer model is used to extract multi-scale spatiotemporal features of cloud images, and Monte Carlo Dropout technique is combined to calculate uncertainty estimates. Meteorological data are denoised by composite chaotic adaptive variational mode decomposition. The OmniCast-NLLPC model is used for meteorological forecasting, and dynamic uncertainty weighting by a hybrid density network is introduced. A generative game fusion module is introduced during the fusion process to achieve active fusion of multi-source data.

Benefits of technology

It effectively reduces prediction errors, enhances prediction stability, provides uncertainty estimation, and supports grid dispatch and energy storage optimization applications.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121863369A_ABST
    Figure CN121863369A_ABST
Patent Text Reader

Abstract

The invention provides an uncertainty perception multi-source data fusion photovoltaic power prediction method, which comprises the following steps that cloud picture data are processed through an improved FCN-Transform model, a meteorological data prediction module adopts a CCAVMD method to carry out noise reduction and enhancement on photovoltaic power and meteorological data, an OmniCard-NLLPC model is input to process the meteorological data, and the photovoltaic power and the meteorological data are subjected to noise reduction and enhancement. Both the two outputs a prediction result with uncertainty estimation; secondly, an uncertainty-driven generative game fusion module is innovatively provided, two paths of information are dynamically evaluated and fused, and traditional passive weighted average is improved into an active decision process. According to the method, high-precision and high-robustness photovoltaic power prediction with confidence interval evaluation is finally realized.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power prediction technology, and in particular to a method for predicting photovoltaic power by fusing multi-source data with uncertainty perception. Background Technology

[0002] With the acceleration of the global energy transition, the penetration rate of distributed photovoltaic (PV) power generation continues to increase. Its randomness and volatility pose a severe challenge to the safe and stable operation of the power system. Achieving high-precision PV power generation forecasting is a key technology to support optimized grid dispatch and improve the capacity for renewable energy absorption. However, distributed PV power generation is affected by a combination of complex factors. Among these, the instantaneous changes in solar irradiance caused by cloud movement and evolution are the main source of uncertainty. Simultaneously, meteorological factors such as ambient temperature and humidity directly affect the conversion efficiency of PV modules. Traditional forecasting methods largely rely on numerical weather prediction, which has low spatiotemporal resolution and struggles to accurately capture the key details of local microclimates and rapid cloud changes, thus limiting forecast accuracy.

[0003] In recent years, deep learning-based prediction models, such as convolutional neural networks and recurrent neural networks, have made progress in time-series forecasting. However, their ability to capture long-range dependencies in spatiotemporal sequence data such as cloud images is limited, and they generally lack the quantification of uncertainty in prediction results. Furthermore, existing fusion methods typically use simple weighting, failing to fully consider the dynamic changes in the reliability of prediction results from different sources, thus limiting the robustness of the models under complex weather conditions. Therefore, developing an intelligent prediction system capable of deeply fusing multi-source heterogeneous data, accurately extracting spatiotemporal features from cloud images, and possessing uncertainty perception capabilities has become a core problem urgently needing to be solved in the field of distributed photovoltaic power prediction. Summary of the Invention

[0004] Purpose of the Invention: To address the shortcomings of existing methods in predicting accuracy and robustness due to their difficulty in accurately capturing rapid cloud changes and lack of uncertainty quantification, this invention provides an uncertainty-aware multi-source data fusion photovoltaic power prediction method, comprising the following steps:

[0005] Step 1: Obtain spatiotemporal sequence cloud map data, historical photovoltaic power data, and meteorological data on solar irradiance, temperature, and humidity;

[0006] Step 2: Perform physical dimension alignment image preprocessing on the spatiotemporal sequence cloud image data, that is, perform physical correction on the original cloud image based on satellite multispectral channel data and atmospheric radiative transfer model;

[0007] Step 3: Input the preprocessed spatiotemporal sequence cloud map data into the improved symmetric hybrid encoder FCN-Transformer model. The FCN-Transformer model captures the multi-scale spatiotemporal features of the cloud map through the encoder-decoder structure, introduces a probability prediction head at the decoder end, calculates the prediction variance through Monte Carlo Dropout technology, and outputs the surface irradiance attenuation coefficient map for future periods, and simultaneously outputs the predicted uncertainty estimate map representing the confidence level.

[0008] Step 4: For historical photovoltaic power data and meteorological data, composite chaotic adaptive variational mode decomposition (CCAVMD) is used for signal denoising and feature enhancement.

[0009] Step 5: Input the preprocessed historical photovoltaic power data and meteorological data into the OmniCast (Omnivorous Forecasting Model) model based on negative log-likelihood and physical constraints for training to obtain predicted data on solar irradiance, temperature, and humidity; the OmniCast model based on negative log-likelihood and physical constraints is denoted as the OmniCast-NLLPC model.

[0010] During model training, the head of a Mixture Density Network (MDN) is used to replace the linear layer, and a composite loss function with three terms is adopted. The dynamic uncertainty weighting term balances the predictive determinism and flexibility, improving the adaptability to complex scenarios.

[0011] Step 6: Input meteorological features and cloud image features into the uncertainty-driven generative game fusion module. By introducing a generative completer and a game selector, the fusion process is transformed from a passive weighted average into an active, creative, and decision-making game process.

[0012] Step 1 includes: acquiring spatiotemporal sequence cloud map data of the power grid in the coverage area using geostationary meteorological satellite remote sensing technology; synchronously collecting real-time total solar irradiance, ambient temperature and humidity data through solar irradiance sensors, temperature and humidity sensors deployed at and around each distributed site; and retrieving historical photovoltaic power data through the monitoring cloud platform of the distributed photovoltaic cluster.

[0013] Step 2 includes: obtaining the cloud optical thickness by inverting the radiative transfer model based on satellite observation data in the visible, shortwave infrared, and thermal infrared bands. The cloud optical thickness The inversion is based on the physical principle of Beer-Lambert's law, and the radiation attenuation relationship is described by the following formula:

[0014] ,

[0015] in, The radiation intensity received by the satellite sensor. Let be the theoretical incident radiation intensity, and exp be the natural exponential function.

[0016] The inversion process is achieved by querying a pre-established lookup table based on an atmospheric radiative transfer model, which converts the satellite-measured radiative intensity into a retrieval value. The simulated values ​​are matched against those in the lookup table to determine the optimal cloud optical thickness. ; to cloud optical thickness Using satellite observation geometric parameters as input into the atmospheric radiative transfer model, the total solar irradiance reaching the Earth's surface under cloud cover conditions is simulated and calculated. Total solar irradiance reaching the Earth's surface under ideal clear sky conditions ; Generate a map of surface irradiance attenuation coefficients:

[0017] ,

[0018] in, The surface irradiance attenuation coefficient is a dimensionless physical quantity between 0 and 1; the surface irradiance attenuation coefficient map is a preprocessed cloud map after physical dimension alignment.

[0019] In step 3, the improved symmetric hybrid encoder FCN-Transformer model includes an encoder and a decoder. The encoder consists of parallel fully convolutional network branches and a visual Transformer branch, used to extract local spatial features and global spatiotemporal correlation features from the input cloud map in parallel. The decoder receives the fused encoded features and restores the spatial resolution through upsampling. At the end of the decoder, a probabilistic prediction head is connected. The probabilistic prediction head enables Monte Carlo Dropout technology during the inference phase and obtains T sets of prediction results by performing T random forward propagations. The mean of the T sets of prediction results is calculated as the final output map of the surface irradiance attenuation coefficient for future periods, and the variance of the T sets of prediction results is calculated as the synchronously output map representing the uncertainty estimate of the prediction confidence level.

[0020] In step 3, the encoder further includes a feature fusion module, which integrates the local spatial feature map output by the fully convolutional network branch with the global spatiotemporal feature sequence output by the visual Transformer branch;

[0021] The feature fusion module performs the following operations:

[0022] The global spatiotemporal feature sequence is reconstructed into a feature map, aligning the spatial dimensions of the feature map with those of the local spatial feature map. A cross-attention mechanism is then used to adaptively compute and fuse the local spatial features with the reconstructed global spatiotemporal features, generating fused features that are fed into the decoder.

[0023] Step 4 includes:

[0024] Parameter initialization is performed and a multi-objective optimization framework is built. The key parameters to be optimized and the search range in the VMD (Variational Mode Decomposition) algorithm are determined: the number of modes. Penalty parameters In addition, bandwidth control parameters are introduced. As a new optimization variable, the search range is: This extends the traditional two-parameter optimization to three-parameter optimization, enhancing the model's flexibility; among which These are the minimum and maximum values ​​of the number of modes, respectively; These are the minimum and maximum values ​​of the penalty parameter, respectively.

[0025] Set the population size N and the maximum number of iterations for the CASBO algorithm (Complex Adaptive System-Based Optimization Algorithm). Instead of a single Logistic mapping, a novel composite chaotic mapping, the Tent-Logistic hybrid mapping, is adopted. The formula is improved as follows:

[0026] ,

[0027] in, Indicates the current chaotic sequence value; To adjust parameters, Furthermore, it avoids fixed points; this mapping exhibits more complex dynamic behavior and can better prevent population initialization from falling into periodicity;

[0028] The following three strategies are adopted:

[0029] Individuals in G1 (typically 70%) are generated by the aforementioned complex chaotic mapping;

[0030] G2 (typically 20%) individuals are generated through Latin hypercube sampling to ensure uniform coverage of the parameter space;

[0031] The G3 (typically 10%) individuals are generated by introducing directional perturbations based on historical best solution information to accelerate convergence;

[0032] The solutions generated by the three strategies are combined to form the initial population;

[0033] For each individual in the population Perform VMD decomposition to improve the traditional VMD constrained variational problem by introducing a Total Variation (TV) regularization term. The new objective function is:

[0034] ,

[0035] Where e is the natural constant, and st indicates that it is constrained by, It is the total variation of the modal components. It is the regularization coefficient, and Related; The k mode functions obtained from the decomposition; The center frequency; The partial derivative operator with respect to time t; Let be the Dirac delta function, which is infinite at t=0 and zero elsewhere, with an integral of 1; j is the imaginary unit, satisfying . t is a time variable; It is the value of the k-th mode function at time t; This means that the sum of all modal functions must equal the original signal f.

[0036] After decomposition, we discard the single envelope entropy metric and construct a fitness function that integrates multiple metrics: weighted envelope entropy, with the following formula:

[0037] ,

[0038] Among them, weight With the center frequency of the mode Negative correlation gives low-frequency modes greater importance; Let be the original envelope entropy of the i-th mode; Let be the weighted envelope entropy of the i-th mode.

[0039] Calculate the correlation coefficient between each Intrinsic Mode Function (IMF) and the original signal, eliminate invalid modes with excessively low correlation coefficients, and calculate the energy concentration of each IMF in the frequency domain.

[0040] The improved adaptive function Fitness is a weighted sum of multiple objective functions:

[0041] ,

[0042] in, It is the number of valid modes after filtering by the Mode Completeness Index (MCI). These are the weighting coefficients; It is the modal completeness index of the i-th mode; The degree of energy concentration in the i-th mode.

[0043] Based on chaotic adaptive search, the Lévy flight strategy is introduced to enhance global exploration capabilities. For individuals with poor fitness, the position update formula is improved as follows:

[0044] ,

[0045] in, It is the Levi flight stride length, where 's' is the stride length; Let be the exponential parameter of the Lévy distribution. , This represents the dot product; this strategy allows individuals to make long-distance jumps and escape local optima. This represents the new position of the i-th individual after the update in this iteration; This represents the old position of the i-th individual before the update in this iteration; This is the adaptive step size factor.

[0046] The adaptive step size adjustment mechanism has been improved to a dynamic adjustment based on individual ranking and population entropy:

[0047] ,

[0048] in, The search step size is dynamically calculated for the i-th individual; It is the ranking of the i-th individual in the current population according to its fitness. It is the information entropy of the current population solution distribution. It is the initial step size; when the population diversity is high (i.e. (Large), appropriately increase the step size; when the population converges (i.e. When the step size is small, reduce the step size for a more refined search;

[0049] After the CASBO algorithm finishes running, select the parameter corresponding to the individual with the highest fitness value. As the optimal parameters for VMD decomposition; These are the optimal number of modes, the optimal penalty parameter, and the optimal bandwidth control parameter, respectively.

[0050] After performing VMD decomposition on each data sequence using optimal parameters, the criterion for signal denoising and reconstruction is improved to a joint kurtosis-envelope entropy criterion:

[0051] Calculate the envelope entropy and kurtosis value for each IMF component;

[0052] A two-dimensional decision plane is constructed, and fuzzy C-means clustering is used to classify IMFs into three categories: information-dominated, mixed, and noise-dominated.

[0053] Only the information-dominant IMF components are linearly superimposed to reconstruct the denoised sequence; for mixed components, wavelet thresholding is performed for secondary denoising before they are reconstructed.

[0054] Step 5 includes:

[0055] Input data preparation and sequence construction:

[0056] The historical photovoltaic power data, solar irradiance data, temperature data and humidity data after noise reduction and decomposition by the Composite Chaotic Adaptive Variational Mode Decomposition (CCAVMD) method in step 3 are integrated to form a multivariate time series dataset.

[0057] The model input samples are constructed using a sliding time window approach. Each sample contains historical data from consecutive time steps, covering all variables in the multivariate time series dataset. The feature dimension of each time step is equal to the number of variables M.

[0058] The input data of the model has the following shape: ,in It represents the length of the historical sequence; BatchSize represents the number of samples input into the model each time.

[0059] Improved OmniCast model structure:

[0060] The OmniCast model is a multivariate time series forecasting model based on the Transformer architecture. Its core lies in using a self-attention mechanism to capture long-term spatiotemporal dependencies within and between variables. The OmniCast model includes the following components:

[0061] Input embedding and positional encoding: First, a linear projection layer maps the M-dimensional input vector at each time step to a high-dimensional feature space. This forms an input embedding; then, a learnable positional encoding is added to the input embedding to inject relative positional information into each time step in the sequence, thus compensating for the deficiency of the self-attention mechanism itself in not having the ability to perceive order.

[0062] Encoder stacking: An encoder consists of N identical encoder layers stacked together. Each encoder layer contains the following two sub-layers:

[0063] Multi-head self-attention mechanism: The multi-head self-attention mechanism linearly transforms the input sequence into a query, key, and value matrix, and divides the query, key, and value matrix into h sub-matrices; this allows the model to jointly focus on information from different variables and time steps from different representation subspaces in parallel, thereby effectively capturing complex cross-variable dependencies and long-term temporal patterns.

[0064] Feedforward Neural Network: A simple fully connected feedforward network, which typically contains a non-linear activation function to non-linearly transform the output of the self-attention mechanism;

[0065] Each sublayer is surrounded by residual connections and followed by normalized layers to improve the training stability of deep networks.

[0066] Improved probabilistic output layer: Instead of performing point prediction on a single linear layer, the output layer is replaced with a parameterized Mixture Density Network (MDN) head. The MDN head no longer directly predicts future weather values, but instead predicts a function parameter that describes the probability distribution of future weather values.

[0067] For each future time step t and each meteorological variable i, the model outputs a set of parameters. This is used to define an adaptive Gaussian mixture model (GMM), where the probability density function of the GMM is... for:

[0068] ,

[0069] Where K is the number of Gaussian components in the mixture; It is the mixing weight of the k-th Gaussian component; and These are the mean and standard deviation of the k-th Gaussian component, respectively. Let be the probability density function of a Gaussian distribution; The target variable to be predicted; This is the set of parameters for the model.

[0070] The MDN head transmits the encoder output through a linear layer. 3D feature mapping is The output of each dimension corresponds to the three meteorological variables in the future. Parameters of K Gaussian components at each time step ; These are the mixed weights, mean, and standard deviation, respectively.

[0071] Model training and output:

[0072] The training objective has shifted from minimizing the error in point prediction to maximizing the likelihood probability between the predicted distribution and the true data distribution.

[0073] Improved Composite Loss Function: An improved composite loss function based on negative log-likelihood (NLL) and incorporating physical constraints is introduced to measure the difference between the model's predicted future weather sequences and the actual sequences. It consists of the following three parts: core likelihood loss term Physical consistency constraints and dynamic uncertainty weighted terms ;

[0074] Core likelihood loss term This parameter guides the MDN output probability distribution to cover the true observations as closely as possible. The calculation formula is the sum of the negative log-likelihoods of all samples, all future time steps, and all meteorological variables:

[0075] ,

[0076] in, Minimize the true value This is equivalent to making the probability distribution output by the model most likely to produce the observed true data; For future time steps; For training batches;

[0077] Physical consistency constraints To differentiate it from purely data-driven methods, a physical constraint based on prior knowledge is introduced: solar irradiance must be non-negative, and temperature and humidity must have a negative correlation within a certain range. This penalty model produces a predicted distribution that violates physical laws.

[0078] Dynamic uncertainty weighted term Introducing a regularization term reduces uncertainty when the model is confident, while allowing reasonable uncertainty to be maintained in complex scenarios. The calculation formula is:

[0079] ,

[0080] Ultimately, the total loss function is , , The weighted sum; H is the information entropy function; The regularization coefficient is used. This is the conditional probability distribution output by the model.

[0081] Optimization and Inference: The model is trained using the Adam optimizer, and the loss function is minimized through backpropagation. To improve generalization ability, early stopping is introduced as a regularization strategy during training. When the validation set loss no longer decreases within two or more consecutive batch epochs, training is automatically terminated and the optimal weights are restored. After the model training is completed, the inference phase begins. New historical sequences, which have undergone the same preprocessing, are input into the OmniCast-NLLPC model with negative log-likelihood and physical constraints. The output of the forward propagation is the predicted data for solar irradiance, temperature, and humidity for a specific future period.

[0082] In step 6, the uncertainty-driven generative game fusion module is used to perform the following steps:

[0083] Step 6-1, Extraction of cloud map features and uncertainties;

[0084] Extract more expressive features and uncertainty estimates for each data source to provide rich information for subsequent generative games;

[0085] Backbone Network: Inputs the cloud image prediction data into a lightweight CNN (containing convolutional layers, batch normalization layers, and ReLU activation function), and outputs a high-order feature map. ;

[0086] Uncertainty quantification: transforming high-order feature maps The input is a parallel uncertainty estimation module, which includes convolutional and pooling layers, and the final output is an uncertainty estimation result. Uncertainty map with the same spatial size The map Each pixel value represents the reliability of the corresponding spatial location prediction;

[0087] Feature aggregation: high-order feature maps Perform global average pooling to obtain a fixed-dimensional cloud map feature vector. ;

[0088] Uncertainty aggregation: Uncertainty graph Perform global average pooling to obtain a scalar value. This represents the overall uncertainty of the cloud map branch, while preserving the uncertainty map. ;

[0089] Step 6-2, extraction of meteorological characteristics and uncertainties;

[0090] Feature mapping: Meteorological forecast data (solar irradiance G, temperature T, humidity H, etc.) are mapped into feature vectors through a fully connected layer. ;

[0091] Uncertainty quantification: The forecast uncertainty is calculated for each meteorological variable (G, T, H). This can be achieved using the standard deviation of model forecasts or a parallel, lightweight uncertainty estimation network. The individual uncertainties are then aggregated and weighted to obtain the overall uncertainty scalar for the meteorological branch. ;

[0092] Step 6-3, Uncertainty-driven generative game fusion;

[0093] Dynamic and intelligent feature fusion is achieved by using generative completers and game selectors;

[0094] Generative cloud map completer: converts feature vectors and overall uncertainty scalar After concatenation, the input is used as a completer, which is a small fully connected network that learns to generate an inferred cloud map feature. Overall uncertainty scalar As conditional information, the reliability of the information is input into the completer, thereby influencing the generation process and outputting the inferred cloud map features. ;

[0095] Weather completer Symmetrical to the cloud image completer, the meteorological completer Generate a predicted meteorological feature based on cloud image information. ;

[0096] Game selector: Input two original features , Two generative features , and two uncertain scalars , The game selector is a gated network with fully connected layers using the Softmax activation function, which selects weights based on the magnitude of uncertainty. ;

[0097] Rule 1: If <The first threshold is 0.2, meaning the cloud map is reliable, while If the second threshold is 0.5, indicating that the weather is unreliable, the game selector will assign features from the original cloud image. High weight (i.e., >0.4), and will also impart inferred meteorological characteristics generated from reliable cloud images. higher weight (i.e., >0.3); the model trusts the cloud imagery and uses it to correct unreliable meteorological information;

[0098] Rule 2: If <The first threshold is 0.2, meaning the weather is reliable, while If the second threshold is 0.5, indicating that the cloud map is unreliable, then conversely, assign... and High weight;

[0099] Rule 3: If and If both features are ≥ the second threshold of 0.5, then the game selector will tend to rely on the two generated features. and The fusion is because it represents the model's best guess based on cross-modal reasoning;

[0100] The game selector outputs a set of normalized weights. Then, a weighted sum is performed to obtain the final game fusion feature. :

[0101] ;

[0102] Step 6-4, Regression Prediction and End-to-End Training: Use the fused features to make the final prediction and design a loss function to guide the learning of the entire model;

[0103] Integrating game theory features Input a regressor consisting of fully connected layers, and output the photovoltaic power prediction sequence for the next T time points. ;

[0104] The entire network is trained end-to-end, and the loss function is... It is a composite loss, consisting of three parts:

[0105] Main loss To smooth the mean square error between predicted power and actual power, or to smooth the L1 loss:

[0106] ,

[0107] Where MSE represents the mean square error; This represents the actual power. To predict power;

[0108] Compensation for consistency loss Constrained generative completers require that the generated features be as close as possible to the real features:

[0109] ,

[0110] Uncertainty calibration loss To ensure that the uncertainty u of the prediction matches the prediction error, the total loss is calculated as follows:

[0111] ,

[0112] in, and It is a hyperparameter used to balance the importance of different loss terms.

[0113] This improvement, by introducing a generative completer and a game selector, elevates the fusion process from a passive weighted average to an active, creative, and decision-making game process. The model not only assesses information reliability but also uses reliable information to generate and complete unreliable information, and makes optimal decisions through game theory, thereby achieving more robust and accurate forecasts under complex and variable weather conditions.

[0114] The present invention also provides an electronic device, including a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method.

[0115] The present invention also provides a storage medium storing a computer program or instructions that, when the computer program or instructions are run on a computer, execute the steps of the method described.

[0116] The beneficial effects of this invention are:

[0117] (1) This invention extracts multi-scale spatiotemporal features from cloud images and meteorological data by using the improved FCN-Transformer model negative sum log likelihood and the physically constrained OmniCast model OmniCast-NLLPC, and combines them with an uncertainty-aware progressive fusion mechanism to effectively reduce prediction errors and enhance the prediction stability of the system under different weather conditions.

[0118] (2) By introducing Monte Carlo Dropout and mixed density network, the system not only outputs point prediction results, but also provides uncertainty estimation diagrams and probability distributions of irradiance attenuation coefficient and meteorological variables, providing confidence intervals for prediction results and supporting decision credibility assessment.

[0119] (3) The cloud map preprocessing method with physical dimension alignment is adopted, combined with the composite chaotic adaptive variational mode decomposition (CCAVMD) signal denoising technology, which effectively improves the quality and feature discrimination of the input data, and lays a reliable data foundation for subsequent model training and prediction.

[0120] (4) By integrating multimodal information such as cloud images and meteorological data, and outputting power prediction results with confidence intervals, it can support practical engineering applications such as power grid dispatch and energy storage optimization, and has good scalability and deployment flexibility. Attached Figure Description

[0121] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0122] Figure 1 This is a flowchart of the model of the present invention.

[0123] Figure 2 This is a structural diagram of the improved symmetric hybrid encoder FCN-Transformer model in this embodiment of the invention.

[0124] Figure 3 This is a structural diagram of the improved meteorological factor decoupling OmniCast model in this embodiment of the invention. Detailed Implementation

[0125] like Figure 1 , Figure 2 and Figure 3 As shown in the figure, this embodiment of the invention provides a photovoltaic power prediction method based on uncertainty-aware multi-source data fusion, comprising the following steps:

[0126] Step 1: Obtain spatiotemporal sequence cloud map data, historical photovoltaic power data, and meteorological data on solar irradiance, temperature, and humidity;

[0127] Step 2 performs physical dimension alignment image preprocessing on the spatiotemporal sequence cloud image data, that is, physical correction is performed on the original cloud image based on satellite multispectral channel data and atmospheric radiative transfer model.

[0128] Step 3 inputs the preprocessed cloud image data into the improved symmetric hybrid encoder FCN-Transformer model. This model captures the multi-scale spatiotemporal features of the cloud image through an encoder-decoder structure and introduces a probability prediction head at the decoder end. The prediction variance is calculated using Monte Carlo Dropout technology, thereby outputting a map of the surface irradiance attenuation coefficient for future periods, and simultaneously outputting an uncertainty estimate map representing the confidence level of the prediction.

[0129] Step 4: For the historical photovoltaic power data and meteorological data, composite chaotic adaptive variational mode decomposition (CCAVMD) is used for signal denoising and feature enhancement.

[0130] Step 5 inputs the preprocessed historical photovoltaic power data and meteorological data into the OmniCast-NLLPC model based on negative log-likelihood and physical constraints for training, and obtains solar irradiance, temperature and humidity prediction data. During the model training process, the MDN head is used to replace the linear layer, and a composite loss function with three terms is adopted. The dynamic uncertainty weighting term balances the predictive determinism and flexibility, and improves the adaptability to complex scenarios.

[0131] Step 6 inputs meteorological features and cloud image features into the uncertainty-driven generative game fusion module. By introducing a generative completer and a game selector, the fusion process is transformed from a passive weighted average into an active, creative, and decision-making game process.

[0132] The implementation process of step 1 is as follows:

[0133] The system uses geostationary meteorological satellite remote sensing technology to acquire spatiotemporal sequence cloud image data of the power grid in the coverage area; it synchronously collects real-time meteorological data such as total solar irradiance, ambient temperature and humidity through solar irradiance sensors, temperature and humidity sensors deployed at and around various distributed sites; and it retrieves historical photovoltaic power data through the monitoring cloud platform of the distributed photovoltaic cluster.

[0134] The implementation process of step 2 is as follows:

[0135] Based on satellite observation data in the visible, shortwave infrared, and thermal infrared bands, cloud optical thickness was obtained through radiative transfer model inversion. ;

[0136] The optical thickness of the cloud The inversion is based on the physical principle of the Beer-Lambert law, and its radiation attenuation relationship is described by the following formula:

[0137] ,

[0138] in, The intensity of radiation received by the satellite sensor. This is the theoretical incident radiation intensity. For cloud optical thickness;

[0139] The inversion process is achieved by querying a pre-established lookup table based on an atmospheric radiative transfer model, which retrieves the measured satellite radiation intensity. The simulated values ​​are matched against those in the lookup table to determine the optimal cloud optical thickness. The optical thickness of the cloud The atmospheric radiative transfer model is input with satellite observation geometric parameters to simulate and calculate the total solar irradiance reaching the Earth's surface under cloudy conditions and under ideal clear sky conditions, respectively. A surface irradiance attenuation coefficient map is generated by calculating the ratio of these two values. The calculation formula is as follows:

[0140] ,

[0141] in, The surface irradiance attenuation coefficient is a dimensionless physical quantity between 0 and 1; the surface irradiance attenuation coefficient map is a preprocessed cloud map after physical dimension alignment.

[0142] The implementation process of step 3 is as follows:

[0143] The improved symmetric hybrid encoder FCN-Transformer model adopts an encoder-decoder structure. The encoder consists of parallel fully convolutional network branches and a visual Transformer branch, used to extract local spatial features and global spatiotemporal correlation features from the input cloud map in parallel. The decoder receives the fused encoded features and restores spatial resolution through upsampling. At the end of the decoder, a probabilistic prediction head is connected. This prediction head enables Monte Carlo Dropout technology during the inference phase, obtaining T sets of prediction results through T random forward propagations. Based on the T sets of prediction results, their mean is calculated as the final output map of surface irradiance attenuation coefficients for future periods, and their variance is calculated as a synchronously output map representing the uncertainty estimate of the prediction confidence level.

[0144] The encoder further includes a feature fusion module for integrating the local spatial feature map output by the fully convolutional network branch with the global spatiotemporal feature sequence output by the visual Transformer branch; wherein, the feature fusion module performs the following operations: 1) reconstructing the global spatiotemporal feature sequence into a feature map, aligning its spatial dimensions with the local spatial feature map; 2) using a cross-attention mechanism to adaptively calculate and fuse the local spatial features and the reconstructed global spatiotemporal features to generate fused features that are fed into the decoder.

[0145] Furthermore, the implementation process of the Composite Chaotic Adaptive Variational Mode Decomposition (CCAVMD) method in step 4 is as follows:

[0146] Perform parameter initialization and build a multi-objective optimization framework to determine the key parameters to be optimized in the VMD algorithm and their search range: number of modes. Penalty parameters In addition, bandwidth control parameters are introduced. As a new optimization variable, its search range is This extends the traditional two-parameter optimization to three-parameter optimization, enhancing the model's flexibility.

[0147] Set the population size N and the maximum number of iterations for the CASBO algorithm. Instead of a single Logistic mapping, a novel composite chaotic mapping, the Tent-Logistic hybrid mapping, is adopted, and its formula is improved as follows:

[0148] ,

[0149] in, To adjust parameters, Furthermore, it avoids fixed points. This mapping exhibits more complex dynamics and is better able to prevent population initialization from falling into periodicity.

[0150] To further enhance population diversity and quality, a hybrid initialization strategy is adopted:

[0151] (1) 70% of the individuals are generated by the above-mentioned complex chaotic mapping.

[0152] (2) 20% of the individuals were generated by Latin hypercube sampling to ensure that the parameter space was uniformly covered.

[0153] (3) 10% of the individuals are generated by directional perturbation based on historical optimal solution information, which accelerates convergence.

[0154] The solutions generated by the three strategies are combined to form the initial population.

[0155] For each individual in the population Perform VMD decomposition. Improve the traditional VMD constrained variational problem by introducing a Total Variation (TV) regularization term. The new objective function is:

[0156] ,

[0157] ,

[0158] in, It is the total variation of the modal components. It is the regularization coefficient, and Related.

[0159] After decomposition, we discard the single envelope entropy metric and construct a fitness function that integrates multiple metrics—weighted envelope entropy:

[0160] ,

[0161] Among them, weight With the center frequency of the mode Negative correlation gives low-frequency modes greater importance.

[0162] Calculate the correlation coefficient between each IMF and the original signal, eliminate invalid modes with excessively low correlation coefficients, and calculate the energy concentration of each IMF in the frequency domain.

[0163] The improved adaptive function is a weighted sum of multiple objective functions:

[0164] ,

[0165] in, This is the number of effective modes after MCI screening. These are the weighting coefficients.

[0166] Building upon chaotic adaptive search, a Lévy flight strategy is introduced to enhance global exploration capabilities. For individuals with poor fitness, the position update formula is improved as follows:

[0167] ,

[0168] in, It is Levi's flight stride ( ), This represents a dot product. This strategy allows individuals to make long-distance jumps and escape local optima.

[0169] The adaptive step size adjustment mechanism has been improved to a dynamic adjustment based on individual ranking and population entropy:

[0170] ,

[0171] in, It is a ranking of individuals based on their fitness. It is the information entropy of the current population solution distribution. This is the initial step size. When population diversity is high (i.e. (Large), appropriately increase the step size; when the population converges (i.e. When the step size is small, reduce the step size for a more refined search.

[0172] After the CASBO algorithm finishes running, select the parameter corresponding to the individual with the highest fitness value. ) is used as the optimal parameter for VMD decomposition.

[0173] After performing VMD decomposition on each data sequence using optimal parameters, the criterion for signal denoising and reconstruction is improved to a joint kurtosis-envelope entropy criterion:

[0174] (1) Calculate the envelope entropy and kurtosis value for each IMF component;

[0175] (2) Construct a two-dimensional decision plane and use fuzzy C-means clustering to classify IMFs into three categories: information-dominated, mixed, and noise-dominated;

[0176] (3) Only the information-dominant IMF components are linearly superimposed to reconstruct the denoised sequence. For mixed components, wavelet thresholding can be selectively performed for secondary denoising before participating in the reconstruction.

[0177] The implementation process of decoupling meteorological factors from the OmniCast model in step 5 is as follows:

[0178] Input data preparation and sequence construction:

[0179] The historical photovoltaic power data, solar irradiance data, temperature data, and humidity data after noise reduction and decomposition using the CASBO-VMD method in step 3 are integrated to form a multivariate time series dataset.

[0180] The model input samples are constructed using a sliding time window approach. Each sample contains historical data from consecutive time steps, covering all the variables mentioned above. The feature dimension of each time step is equal to the number of variables M.

[0181] The input data of the model has the following shape: ,in M is the length of the historical sequence, and M is the feature dimension.

[0182] Improved OmniCast model structure:

[0183] The OmniCast model is a multivariate time series forecasting model based on the Transformer architecture. Its core lies in utilizing a self-attention mechanism to capture long-term spatiotemporal dependencies within and between variables. The model mainly consists of the following components:

[0184] Input embedding and position encoding:

[0185] First, a linear projection layer is used to map the M-dimensional input vector at each time step to a high-dimensional feature space. This forms an input embedding.

[0186] Subsequently, a learnable positional encoding is added to the input embedding to inject relative positional information into each time step in the sequence, thus compensating for the deficiency of the self-attention mechanism itself in not having the ability to perceive order.

[0187] Encoder stacking: An encoder is composed of N identical encoder layers stacked together.

[0188] Each encoder layer contains two sub-layers:

[0189] Multi-head self-attention mechanism: This mechanism linearly transforms the input sequence into a query, key, and value matrix, and divides it into h sub-matrices. This allows the model to jointly focus on information from different variables and time steps from different representation subspaces in parallel, thereby effectively capturing complex intervariate dependencies and long-term temporal patterns.

[0190] Feedforward Neural Network: A simple fully connected feedforward network, which typically contains a non-linear activation function to non-linearly transform the output of the self-attention mechanism.

[0191] Each sublayer is surrounded by residual connections, and subsequent layers are normalized to improve the training stability of deep networks.

[0192] Improved probabilistic output layer:

[0193] The improvement to conventional methods, which involve point predictions on a single linear layer, replaces the output layer with a parameterized Mixture Density Network (MDN) head. This MDN head no longer directly predicts future weather values; instead, it predicts the parameters of a function that describes the probability distribution of future weather values.

[0194] For each future time step t and each meteorological variable i, the model outputs a set of parameters. This is used to define an adaptive Gaussian mixture model (GMM). The probability density function of this GMM is:

[0195] ,

[0196] Where K is the number of Gaussian components in the mixture; It is the mixing weight of the k-th Gaussian component; and These are the mean and standard deviation of the k-th Gaussian component, respectively.

[0197] The MDN head transmits the encoder output through a linear layer. 3D feature mapping is The output of each dimension corresponds to the three meteorological variables in the future. Parameters of K Gaussian components at each time step .

[0198] Model training and output:

[0199] The training objective has shifted from minimizing the error in point prediction to maximizing the likelihood probability between the predicted distribution and the true data distribution.

[0200] Improved composite loss function:

[0201] A composite loss function based on negative log-likelihood (NLL) and incorporating physical constraints is introduced to measure the difference between the model's predicted future weather sequences and the actual sequences. This loss function... It consists of the following three parts:

[0202] 1) Core likelihood loss term This parameter guides the MDN output probability distribution to cover the true observations as closely as possible. It is calculated as the sum of the negative log-likelihoods of all samples, all future time steps, and all meteorological variables.

[0203] ,

[0204] in, Minimize the true value This is equivalent to ensuring that the probability distribution output by the model is most likely to produce the observed real data.

[0205] 2) Physical consistency constraints To differentiate it from purely data-driven methods, physical constraints based on prior knowledge are introduced. Solar irradiance must be non-negative, and temperature and humidity must exhibit a negative correlation within a certain range. This penalty model produces predicted distributions that violate physical laws.

[0206] 3) Weighted term for dynamic uncertainty A regularization term is introduced that reduces uncertainty when the model is confident, while allowing reasonable uncertainty in complex scenarios. This term is based on the entropy of the prediction distribution:

[0207] ,

[0208] Ultimately, the total loss function is the weighted sum of the three terms mentioned above.

[0209] Optimization and Inference: The model is trained using the Adam optimizer, and the loss function is minimized through backpropagation. To improve generalization ability, early stopping is introduced as a regularization strategy during training. When the validation set loss no longer decreases over several consecutive epochs, training is automatically terminated and the optimal weights are restored. After training, the model enters the inference phase. New, preprocessed historical sequences are input into the OmniCast-NLLPC model, which uses negative log-likelihood and physical constraints. The output of its forward propagation provides the predicted solar irradiance, temperature, and humidity for specific future periods.

[0210] The implementation process of the uncertainty-driven generative game fusion module in step 6 is as follows:

[0211] (1) Extraction of meteorological characteristics and uncertainties;

[0212] Extract more expressive features and uncertainty estimates for each data source to provide rich information for subsequent generative games.

[0213] Backbone Network: Inputs the cloud image prediction data into a lightweight CNN (containing convolutional layers, batch normalization layers, and ReLU activation function), and outputs a high-order feature map. .

[0214] Uncertainty quantization: The above feature map is input into a parallel uncertainty estimation module. This module consists of several convolutional and pooling layers, and finally outputs an uncertainty estimation result. Uncertainty map with the same spatial size Each pixel value in this map represents the reliability of the corresponding spatial location prediction.

[0215] Feature aggregation: For feature maps Perform global average pooling to obtain a fixed-dimensional cloud map feature vector. .

[0216] Uncertainty aggregation: Uncertainty graph Perform global average pooling to obtain a scalar value. This represents the overall uncertainty of the cloud map branch. Meanwhile, it retains... For use in Phase Two.

[0217] (2) Extraction of meteorological characteristics and uncertainties;

[0218] Feature mapping: Meteorological forecast data (solar irradiance G, temperature T, humidity H, etc.) are mapped into feature vectors through a fully connected layer. ;

[0219] Uncertainty quantification: The forecast uncertainty of each meteorological variable (G, T, H) is calculated. This can be achieved using the standard deviation of model forecasts or a parallel, lightweight uncertainty estimation network. These individual uncertainties are then aggregated and weighted to obtain the overall uncertainty scalar of the meteorological branch. .

[0220] (3) Uncertainty-driven generative game fusion;

[0221] By utilizing generative completers and game selectors, dynamic and intelligent feature fusion is achieved.

[0222] Generative cloud image completer: converts meteorological features... and weather uncertainty After concatenation, the input is used as a completer, which is a small fully connected network that learns to generate an inferred cloud map feature. Uncertainty As conditional information, the reliability of the information is input into the completer, thus influencing its generation process. The output is the inferred cloud map features. .

[0223] Weather completer Symmetrical to the cloud image completer, it generates an inferred meteorological feature based on cloud image information. .

[0224] Game-theoretic selector: Given two original features, two generated features, and two uncertainty scalars, the game-theoretic selector is a gated network consisting of fully connected layers using the Softmax activation function. Weights are selected based on the magnitude of the uncertainty. .

[0225] Rule 1: If Very low, meaning the cloud image is reliable, and If the weather is very high, meaning the weather is unreliable, then the game selector will assign features to the original cloud image. High weight It will also impart inferred meteorological features generated from reliable cloud images. higher weight The model trusts cloud images and uses them to correct unreliable weather information.

[0226] Rule 2: If Very low, meaning the weather is reliable, and If the cloud map is too high, meaning it's unreliable, then conversely, assign... and High weight.

[0227] Rule 3: If both uncertainties are high, the game selector will tend to rely on both generating features. and The fusion is because it represents the model's best guess based on cross-modal reasoning.

[0228] The game selector outputs a set of normalized weights. Then, a weighted sum is performed to obtain the final game fusion feature:

[0229] ;

[0230] (4) Regression prediction and end-to-end training;

[0231] The fused features are used for the final prediction, and a loss function is designed to guide the learning of the entire model.

[0232] Integrating game theory features Input a regressor consisting of fully connected layers, and output the photovoltaic power prediction sequence for the next T time points. .

[0233] The entire network is trained end-to-end, and the loss function is... It is a composite loss, consisting of three parts:

[0234] The main loss is the mean square error or smoothing L1 loss between the predicted power and the actual power:

[0235] ,

[0236] The consistency loss is used to constrain the generative completer, requiring that the generated features be as close as possible to the true features.

[0237] ,

[0238] Uncertainty calibration loss ensures that the prediction uncertainty u matches the prediction error. The total loss is calculated as follows:

[0239] ,

[0240] in, and It is a hyperparameter used to balance the importance of different loss terms.

[0241] This improvement, by introducing a generative completer and a game selector, elevates the fusion process from a passive weighted average to an active, creative, and decision-making game process. The model not only assesses information reliability but also uses reliable information to generate and complete unreliable information, and makes optimal decisions through game theory, thereby achieving more robust and accurate forecasts under complex and variable weather conditions.

[0242] A specific embodiment is provided for complete illustration. This embodiment uses the operation of a 10MW distributed photovoltaic power station on a typical day in the summer of 2024 as a scenario. The algorithm program was written in MATLAB to construct the prediction model of this invention and the VMD+OmniCast comparative prediction model. Table 1 shows the comparison of optimization and improvement.

[0243] Table 1 Comparison of Optimization, Improvement and Enhancement

[0244]

[0245] This example is mainly applied to power prediction of distributed photovoltaic power plants.

[0246] Application scenario: A 10MW distributed photovoltaic power station cluster.

[0247] Deployment and implementation plan:

[0248] (1) Hardware deployment:

[0249] The data acquisition layer synchronously collects real-time total solar irradiance, ambient temperature, and humidity data through solar irradiance sensors, temperature and humidity sensors deployed at and around various distributed sites; it also retrieves historical photovoltaic power data through the monitoring cloud platform of the distributed photovoltaic cluster. This includes inverter data acquisition units and edge computing gateways.

[0250] (2) Software deployment:

[0251] Configure CCAVMD signal noise reduction parameters: ;

[0252] Load the pre-trained FCN-Transformer model (input size 256×256×12, output 6 frames of prediction);

[0253] Deploy the OmniCast-NLLPC model (Transformer encoder layers N=6, head number h=8); configure Monte Carlo Dropout sampling times T=50, MDN mixing component is 3;

[0254] Implement an uncertainty-driven game selector with the following threshold configurations: first threshold = 0.2, second threshold = 0.5.

[0255] Deploy a generative completer network (fully connected layer dimension [256,128,64]);

[0256] (3) Experimental run:

[0257] The optical thickness τ of the test cloud increased sharply from 0.05 to 1.12. The irradiance attenuation coefficient K, predicted in real-time by the FCN-Transformer, decreased from 0.98 to 0.45. Uncertainty maps showed low prediction reliability at the cloud edge (u≥0.3). Meanwhile, CCAVMD decomposition (optimal parameters) was performed. The signal-to-noise ratio was increased from 15.6 dB to 21.3 dB. The OmniCast-NLLPC model outputs a solar irradiance probability distribution of ~N(842,28²) W / m² (95% confidence interval [792,892] W / m²). The game theory fusion module dynamically evaluates the uncertainties of the two paths (cloud map branch). Meteorological Branch According to rule 1 ( and Assign weights (Original cloud map features) (Generate meteorological characteristics), and finally output a predicted power value of 5.62MW (95% range [5.35, 5.89] MW). The actual measured power of 5.58MW fell within the prediction range, with a point prediction error of only 0.04MW. Compared with the conventional method (VMD+OmniCast predicted 5.95MW with an error of 0.37MW), the error was reduced by 89.2%. Full-cycle testing showed that the predicted power RMSE decreased to 52.7kW (compared to 86.3kW for the conventional method), the prediction range coverage PICP reached 94.2%, the error for cloudy weather decreased by 41.9%, and the adaptation time for sudden weather changes was shortened from >45 minutes to <20 minutes.

[0258] This invention provides an uncertainty-aware multi-source data fusion method for photovoltaic power prediction. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. A photovoltaic power prediction method based on uncertainty-aware multi-source data fusion, characterized in that, Includes the following steps: Step 1: Obtain spatiotemporal sequence cloud map data, historical photovoltaic power data, and meteorological data on solar irradiance, temperature, and humidity; Step 2: Perform physical dimension alignment image preprocessing on the spatiotemporal sequence cloud image data, that is, perform physical correction on the original cloud image based on satellite multispectral channel data and atmospheric radiative transfer model; Step 3: Input the preprocessed spatiotemporal sequence cloud map data into the improved symmetric hybrid encoder FCN-Transformer model. The FCN-Transformer model captures the multi-scale spatiotemporal features of the cloud map through the encoder-decoder structure, introduces a probability prediction head at the decoder end, calculates the prediction variance through Monte Carlo Dropout technology, and outputs the surface irradiance attenuation coefficient map for future periods, and simultaneously outputs the predicted uncertainty estimate map representing the confidence level. Step 4: For historical photovoltaic power data and meteorological data, composite chaotic adaptive variational mode decomposition (CCAVMD) is used for signal denoising and feature enhancement. Step 5: Input the preprocessed historical photovoltaic power data and meteorological data into the OmniCast model, a global forecasting model based on negative log-likelihood and physical constraints, for training to obtain predicted data on solar irradiance, temperature, and humidity; the OmniCast model based on negative log-likelihood and physical constraints is denoted as the OmniCast-NLLPC model. During model training, a hybrid density network (MDN) head is used to replace the linear layer, and a composite loss function with three terms is adopted. The dynamic uncertainty weighting term balances the predictive determinism and flexibility, improving adaptability to complex scenarios. Step 6: Input meteorological features and cloud image features into the uncertainty-driven generative game fusion module by introducing a generative completer and a game selector.

2. The method according to claim 1, characterized in that, Step 1 includes: acquiring spatiotemporal sequence cloud map data of the power grid in the coverage area using geostationary meteorological satellite remote sensing technology; synchronously collecting real-time total solar irradiance, ambient temperature and humidity data through solar irradiance sensors, temperature and humidity sensors deployed at and around each distributed site; and retrieving historical photovoltaic power data through the monitoring cloud platform of the distributed photovoltaic cluster.

3. The method according to claim 2, characterized in that, Step 2 includes: obtaining the cloud optical thickness by inverting the radiative transfer model based on satellite observation data in the visible, shortwave infrared, and thermal infrared bands. The cloud optical thickness The inversion is based on the physical principle of the Beer-Lambert law, and the radiation attenuation relationship is described by the following formula: , in, The radiation intensity received by the satellite sensor. Let be the theoretical incident radiation intensity, and exp be the natural exponential function. The inversion process is achieved by querying a pre-established lookup table based on an atmospheric radiative transfer model, which converts the satellite-measured radiative intensity into a retrieval value. The simulated values ​​are matched against those in the lookup table to determine the optimal cloud optical thickness. ; to cloud optical thickness Using satellite observation geometric parameters as input into the atmospheric radiative transfer model, the total solar irradiance reaching the Earth's surface under cloud cover conditions is simulated and calculated. Total solar irradiance reaching the Earth's surface under ideal clear sky conditions ; Generate a map of surface irradiance attenuation coefficients: , in, The surface irradiance attenuation coefficient is the surface irradiance attenuation coefficient map, which is a preprocessed cloud map after physical dimension alignment.

4. The method according to claim 3, characterized in that, In step 3, the improved symmetric hybrid encoder FCN-Transformer model includes an encoder and a decoder. The encoder consists of parallel fully convolutional network branches and a visual Transformer branch, used to extract local spatial features and global spatiotemporal correlation features from the input cloud map in parallel. The decoder receives the fused encoded features and restores the spatial resolution through upsampling. At the end of the decoder, a probabilistic prediction head is connected. The probabilistic prediction head enables Monte Carlo Dropout technology during the inference phase and obtains T sets of prediction results by performing T random forward propagations. The mean of the T sets of prediction results is calculated as the final output map of the surface irradiance attenuation coefficient for future periods, and the variance of the T sets of prediction results is calculated as the synchronously output map representing the uncertainty estimate of the prediction confidence level.

5. The method according to claim 4, characterized in that, In step 3, the encoder further includes a feature fusion module, which integrates the local spatial feature map output by the fully convolutional network branch with the global spatiotemporal feature sequence output by the visual Transformer branch; The feature fusion module performs the following operations: The global spatiotemporal feature sequence is reconstructed into a feature map, aligning the spatial dimensions of the feature map with those of the local spatial feature map. A cross-attention mechanism is employed to adaptively compute and fuse the local spatial features with the reconstructed global spatiotemporal features, generating fused features that are fed into the decoder.

6. The method according to claim 5, characterized in that, Step 4 includes: Perform parameter initialization and build a multi-objective optimization framework to determine the key parameters to be optimized and the search range in the VMD algorithm: number of modes. Penalty parameters In addition, bandwidth control parameters are introduced. As a new optimization variable, the search range is: ;in These are the minimum and maximum values ​​of the number of modes, respectively; These are the minimum and maximum values ​​of the penalty parameter, respectively. Set the population size N and the maximum number of iterations for the CASBO algorithm. Using a Tent-Logistic hybrid mapping, the formula is improved to: , in, Indicates the current chaotic sequence value; To adjust parameters, ; The following three strategies are adopted: Individuals of G1 are generated by complex chaotic mapping; Individuals in G2 were generated through Latin hypercube sampling; G3 introduces directional perturbation generation based on historical optimal solution information; The solutions generated by the three strategies are combined to form the initial population; For each individual in the population Performing VMD decomposition improves the traditional VMD constrained variational problem by introducing a total variational regularization term. The new objective function is: , Where e is the natural constant, and st indicates that it is constrained by, It is the total variation of the modal components. It is the regularization coefficient; The k mode functions obtained from the decomposition; The center frequency; The partial derivative operator with respect to time t; Let j be the Dirac function; j is the imaginary unit, satisfying t is a time variable; It is the value of the k-th mode function at time t; This means that the sum of all modal functions must equal the original signal f; After decomposition, a fitness function fused from multiple indicators is constructed: weighted envelope entropy, with the following formula: , Among them, weight With the center frequency of the mode Negative correlation; Let be the original envelope entropy of the i-th mode; Let be the weighted envelope entropy of the i-th mode; Calculate the correlation coefficient between each intrinsic mode function (IMF) and the original signal, and calculate the energy concentration of each IMF in the frequency domain; The improved adaptive function Fitness is a weighted sum of multiple objective functions: , in, It is the number of effective modes after screening by the Modal Completeness Index (MCI). These are the weighting coefficients; It is the modal completeness index of the i-th mode; The energy concentration of the i-th mode; Based on chaotic adaptive search, the Lévy flight strategy is introduced to enhance global exploration capabilities. For individuals with poor fitness, the position update formula is improved as follows: , in, It is the Levi flight stride length, where 's' is the stride length; Let be the exponential parameter of the Lévy distribution. Dot product; This represents the new position of the i-th individual after the update in this iteration; This represents the old position of the i-th individual before the update in this iteration; This is an adaptive step size factor; The adaptive step size adjustment mechanism has been improved to a dynamic adjustment based on individual ranking and population entropy: , in, The search step size is dynamically calculated for the i-th individual; It is the ranking of the i-th individual in the current population according to its fitness. It is the information entropy of the current population solution distribution. It is the initial step size; After the CASBO algorithm finishes running, select the parameter corresponding to the individual with the highest fitness value. As the optimal parameters for VMD decomposition; These are the optimal number of modes, the optimal penalty parameter, and the optimal bandwidth control parameter, respectively. After performing VMD decomposition on each data sequence using optimal parameters, the criterion for signal denoising and reconstruction is improved to a joint kurtosis-envelope entropy criterion: Calculate the envelope entropy and kurtosis value for each IMF component; A two-dimensional decision plane is constructed, and fuzzy C-means clustering is used to classify IMFs into three categories: information-dominated, mixed, and noise-dominated. Only the information-dominant IMF components are linearly superimposed to reconstruct the denoised sequence; for mixed components, wavelet thresholding is performed for secondary denoising before they are reconstructed.

7. The method according to claim 6, characterized in that, Step 5 includes: Input data preparation and sequence construction: The historical photovoltaic power data, solar irradiance data, temperature data and humidity data after noise reduction and decomposition by the Composite Chaotic Adaptive Variational Mode Decomposition (CCAVMD) method in step 3 are integrated to form a multivariate time series dataset. The model input samples are constructed using a sliding time window approach. Each sample contains historical data from consecutive time steps, and the feature dimension of each time step is equal to the number of variables M. The input data of the model has the following shape: ,in It represents the length of the historical sequence; BatchSize represents the number of samples input into the model each time. Improved OmniCast model structure: The OmniCast model is a multivariate time series forecasting model based on the Transformer architecture. The OmniCast model includes the following components: Input embedding and positional encoding: First, a linear projection layer maps the M-dimensional input vector at each time step to a high-dimensional feature space. This forms the input embedding; subsequently, a learnable positional code is added to the input embedding to inject relative positional information into each time step of the sequence. Encoder stacking: An encoder consists of N identical encoder layers stacked together. Each encoder layer contains the following two sub-layers: Multi-head self-attention mechanism: The multi-head self-attention mechanism linearly transforms the input sequence into a query, key, and value matrix, and divides the query, key, and value matrix into h parts; Feedforward neural network: contains a non-linear activation function to perform a non-linear transformation on the output of the self-attention mechanism; Each sublayer is surrounded by residual connections, and subsequent layers are normalized. Improved probabilistic output layer: Instead of performing point predictions on a single linear layer, the output layer is replaced with a parameterized hybrid density network (MDN) head that predicts a function parameter that describes the probability distribution of future weather values. For each future time step t and each meteorological variable i, the model outputs a set of parameters. This is used to define an adaptive Gaussian mixture model (GMM), where the probability density function of the GMM is... for: , Where K is the number of Gaussian components in the mixture; It is the mixing weight of the k-th Gaussian component; and These are the mean and standard deviation of the k-th Gaussian component, respectively. Let be the probability density function of a Gaussian distribution; The target variable to be predicted; The set of parameters for the model; The MDN head transmits the encoder output through a linear layer. 3D feature mapping is The output of each dimension corresponds to the three meteorological variables in the future. Parameters of K Gaussian components at each time step ; These are the mixed weights, mean, and standard deviation, respectively. Model training and output: The training objective has shifted from minimizing the error in point prediction to maximizing the likelihood probability between the predicted distribution and the true data distribution. Improved Composite Loss Function: A composite loss function based on negative log-likelihood (NLL) and incorporating physical constraints is introduced to measure the difference between the model's predicted future weather sequences and the actual sequences. The loss function... It consists of the following three parts: core likelihood loss term Physical consistency constraints and dynamic uncertainty weighted terms ; Core likelihood loss term The calculation formula is: , in, The actual value; For future time steps; For training batches; Dynamic uncertainty weighted term The calculation formula is: , Ultimately, the total loss function is , , The weighted sum; H is the information entropy function; The regularization coefficient is used. This represents the conditional probability distribution output by the model. Optimization and Inference: The model is trained using the Adam optimizer, and the loss function is minimized through backpropagation. Early stopping is introduced as a regularization strategy during training. When the validation set loss no longer decreases within two or more consecutive batch epochs, training is automatically terminated and the optimal weights are restored. After the model training is completed, the inference phase begins. New historical sequences that have undergone the same preprocessing are input into the OmniCast-NLLPC model with negative log-likelihood and physical constraints. The output of forward propagation is the predicted data of solar irradiance, temperature, and humidity for a specific future period.

8. The method according to claim 7, characterized in that, In step 6, the uncertainty-driven generative game fusion module is used to perform the following steps: Step 6-1, Extraction of cloud map features and uncertainties; Backbone Network: Inputs the cloud image prediction data into a lightweight CNN, which contains convolutional layers, batch normalization layers, and ReLU activation functions, and outputs a high-order feature map. ; Uncertainty quantification: transforming high-order feature maps The input is a parallel uncertainty estimation module, which includes convolutional and pooling layers, and the final output is an uncertainty estimation result. Uncertainty map with the same spatial size ; Feature aggregation: high-order feature maps Perform global average pooling to obtain a fixed-dimensional cloud map feature vector. ; Uncertainty aggregation: Uncertainty graph Perform global average pooling to obtain a scalar value. At the same time, the uncertainty map is preserved. ; Step 6-2, extraction of meteorological features and uncertainties; Feature mapping: Weather forecast data is mapped into feature vectors through a fully connected layer. ; Uncertainty quantification: The forecast uncertainty is calculated for each meteorological variable, and then the individual uncertainties are aggregated and weighted to obtain the overall uncertainty scalar of the meteorological branch. ; Step 6-3, Uncertainty-driven generative game fusion; Dynamic and intelligent feature fusion is achieved by using generative completers and game selectors; Generative cloud map completer: converts feature vectors and overall uncertainty scalar After concatenation, the input is used as a completer, which is a small fully connected network that learns to generate an inferred cloud map feature. Overall uncertainty scalar As conditional information, the reliability of the information is input into the completer, thereby influencing the generation process and outputting the inferred cloud map features. ; Weather completer Symmetrical to the cloud image completer, the meteorological completer Generate a predicted meteorological feature based on cloud image information. ; Game selector: Input two original features , Two generative features , and two uncertain scalars , The game selector is a gated network with fully connected layers using the Softmax activation function, which selects weights based on the magnitude of uncertainty. ; Rule 1: If <The first threshold, i.e., the cloud map is reliable, and If the second threshold, i.e. the weather is unreliable, the game selector will assign features from the original cloud image. High weight It also imparts inferred meteorological characteristics generated from reliable cloud images. Higher weight ; Rule 2: If <The first threshold, i.e., weather reliability, and The second threshold, i.e., if the cloud map is unreliable, is then assigned... and High weight; Rule 3: If and If both features are greater than or equal to the second threshold, then the game selector will tend to rely on the two generated features. and The integration; The game selector outputs a set of normalized weights. Then, a weighted sum is performed to obtain the final game fusion feature. : ; Step 6-4, Regression Prediction and End-to-End Training: Use the fused features to make the final prediction and design a loss function to guide the learning of the entire model; Integrating game theory features Input a regressor consisting of fully connected layers, and output the photovoltaic power prediction sequence for the next T time points. ; The entire network is trained end-to-end, and the loss function is... It is a composite loss, consisting of three parts: Main loss To smooth the mean square error between predicted power and actual power, or to smooth the L1 loss: , Where MSE represents the mean square error; This refers to the actual power. To predict power; Compensation for consistency loss The calculation formula is: , Uncertainty calibration loss To ensure that the uncertainty u of the prediction matches the prediction error, the total loss is calculated as follows: , in, and It is a hyperparameter used to balance the importance of different loss terms.

9. An electronic device, characterized in that, It includes a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 8.

10. A storage medium, characterized in that, It stores a computer program or instructions that, when run on a computer, perform the steps of the method as described in any one of claims 1 to 8.