Multi-source feature fusion method for abnormal value self-correction and meta-learning confrontation of distribution network cloud cluster shielding features in extreme weather

Abstract

The invention provides a multi-source feature fusion method for abnormal value self-correction and meta-learning confrontation of cloud cluster shielding features of a power distribution network in extreme weather. The method comprises the following steps: step S110, extracting and predicting cloud cluster shielding features; step S120, screening out meteorological variables with relatively high correlation with the solar irradiance; step S130, based on the cloud cluster shielding characteristics in the step S110, generating a cloud cluster shielding space range; step S140, executing abnormal value detection and correction for the space range shielded by the cloud cluster generated in the step S130; and step S150, fusing the corrected cloud cluster shielding features with meteorological variables based on a generative adversarial network algorithm of meta learning. Compared with a traditional method, focusing of a sudden change area under extreme weather is more accurate, and the high precision and efficiency of cloud cluster shielding feature extraction and prediction are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of distribution networks, and in particular to a multi-source feature fusion method for outlier self-correction and meta-learning confrontation of cloud obscuration features in distribution networks under extreme weather conditions. Background Art

[0002] Amid the accelerated development of new power systems, distribution network operations face the dual challenges of sudden changes in cloud obscuration characteristics and the fusion of multi-source features during extreme weather events. While existing forecasting methods combine satellite cloud imagery with numerical meteorological forecasts, significant deficiencies remain in robust handling of outliers and in the in-depth fusion of multi-source features, limiting the reliability of forecasts in extreme weather scenarios.

[0003] In terms of outlier processing, traditional methods have three shortcomings: First, the anomaly detection mechanism based on fixed thresholds is difficult to adapt to the drastic fluctuations in cloud grayscale values in scenarios such as typhoons and severe convection. When the cloud top temperature drops sharply, causing radiation anomalies, the static threshold is prone to misjudging the physically explanatory mutation as sensor noise. Second, the sensitivity of traditional long-short-term memory networks to outliers increases with the nonlinear enhancement of meteorological elements. Field measurements show that when vertical wind shear exceeds 20 m / s, the error propagation speed caused by outliers in cloud trajectory prediction increases by more than three times compared with normal weather. Third, existing data cleaning methods lack physical constraints. Cloud-missing areas that rely solely on statistical interpolation to repair are prone to violate the laws of atmospheric dynamics, causing the corrected cloud diffusion direction to deviate from the measured wind speed field.

[0004] Existing technologies for multi-source feature fusion face two major bottlenecks: First, the spatiotemporal correlation between meteorological variables and cloud features is not effectively modeled. Second, the nonlinear coupling of multimodal data such as humidity and air pressure exhibits strong time-varying properties under extreme weather conditions, making conventional convolutional neural networks inefficient at learning the cross-modal interaction between meteorological variables and cloud features. Furthermore, outlier interference further disrupts the covariance structure of multi-source features, resulting in a decrease in the signal-to-noise ratio of the fused features.

[0005] These technical flaws present dual risks to existing systems during extreme weather conditions: Misjudgments caused by outliers can lead to increased deviations in cloud cover predictions; and inefficient feature fusion increases fitting errors in the typhoon path-cloud motion correlation model, severely impacting the accuracy of power grid transient stability control. Therefore, a multi-source feature fusion system with the ability to self-correct outliers is urgently needed to overcome the major technical bottleneck in ensuring the safe operation of distribution networks during extreme weather conditions. Summary of the Invention

[0006] The purpose of the present invention is to provide a method for processing outliers and fusing multi-source features of cloud shielding features in distribution networks under extreme weather conditions. First, a cloud shielding feature prediction method based on a spatiotemporal convolutional long short-term memory network is used to predict cloud features, and combined with adaptive threshold segmentation and morphological optimization post-processing, the regional range of cloud shielding features is generated. Then, the abnormal conditions in the region are judged. By adopting a digital twin-driven outlier self-correction method, the prediction accuracy and real-time performance are significantly improved, which can be used for weather forecasting and disaster warning in the field of distribution networks under extreme weather conditions. Furthermore, meteorological variables with a high correlation with solar irradiance are selected and combined with the corrected features as input data. Finally, a meta-learning-based generative adversarial network method is used to realize feature fusion and a Bayesian learning method is used to realize uncertainty quantification of the fused features, which solves the problem of multimodal fusion of meteorological variables and cloud shielding features and supports the reliability of distribution network scheduling decisions.

[0007] The technical solution of the present invention includes:

[0008] A multi-source feature fusion method based on outlier self-correction and meta-learning confrontation for cloud obscuration features in distribution networks under extreme weather conditions, comprising:

[0009] Step S110, extracting and predicting cloud obscuration features;

[0010] Step S120, screening out meteorological variables that have a high correlation with solar irradiance;

[0011] Step S130, generating a spatial range of cloud cover based on the cloud cover characteristics in step S110;

[0012] Step S140 , performing outlier detection and correction on the cloud-obscured spatial range generated in step S130 ;

[0013] Step S150: The meta-learning-based generative adversarial network algorithm fuses the corrected cloud obscuration features with meteorological variables.

[0014] Beneficial effects of the present invention:

[0015] (1) The present invention uses a spatiotemporal convolutional long short-term memory network method to extract the spatiotemporal dynamic characteristics of clouds through three-dimensional convolution, combines it with a bidirectional long short-term memory network to model the evolution of clouds, and uses a spatial-channel attention mechanism to focus on the mutation area under extreme weather conditions. Compared with traditional methods, the three-dimensional convolution operation in this method improves the efficiency of extracting local cloud features, the bidirectional long short-term memory network reduces the error of predicted features, and the spatial-channel attention mechanism makes the focus on the mutation area under extreme weather conditions more accurate. This method significantly improves the high accuracy and efficiency of cloud occlusion feature extraction and prediction.

[0016] (2) This method uses the Pearson correlation coefficient to screen meteorological variables (such as cloud cover and wind speed) that are highly correlated with solar irradiance, effectively eliminating redundant interference factors and strengthening the model's ability to identify the key driving forces of cloud obscuration. This method quantifies the linear correlation between meteorological variables and irradiance, accurately identifying the core parameters that affect photovoltaic output, and achieves a data dimension compression rate of over 60% compared to traditional empirical screening methods. This not only improves the sensitivity of cloud obscuration feature predictions in extreme weather conditions, but also reduces the risk of model overfitting and cloud coverage prediction errors in typhoon scenarios.

[0017] (3) The adaptive threshold segmentation and morphological optimization algorithm proposed in this paper work synergistically to significantly improve the accuracy and robustness of cloud-obscured area generation. Specifically, by dynamically adjusting the segmentation threshold and combining the predicted grayscale mean and standard deviation, the algorithm accurately distinguishes between thick cloud cover and thin cloud noise, thus avoiding over-segmentation (e.g., misidentifying fast-moving fragmented clouds as main clouds) or under-segmentation (e.g., missing low-grayscale but large-scale stratocumulus clouds) caused by sudden changes in cloud cover in extreme weather.

[0018] (4) In the morphological optimization algorithm adopted by the present invention, the opening operation is used to eliminate sensor noise, and the closing operation is used to fill the holes formed in the cumulonimbus cloud due to abnormal ice crystal reflection. The connected area analysis is used to filter out long artifacts and scattered small cloud blocks to ensure that the boundaries of the generated cloud cluster obscured area are complete and the physical characteristics are reasonable, ultimately supporting the rapid response and stable control of the power grid to the sharp drop in photovoltaic output under extreme weather conditions.

[0019] (5) The digital twin-driven outlier self-correction method adopted in the present invention significantly improves the reliability and physical consistency of cloud occlusion feature prediction by integrating physical models with data-driven technology. A digital twin is constructed based on the cloud evolution equation, and the residuals of the predicted values and actual observations are compared in real time. Local model parameters are dynamically updated through federated learning (such as optimizing the weight of the influence of wind speed on cloud trajectory), and data enhancement is combined to repair missing values in abnormal areas to avoid misjudgments caused by noise or sensor failure in traditional methods. At the same time, physical constraints are introduced to correct abnormal radiation fluctuations to ensure that the correction results conform to the laws of atmospheric dynamics. This method improves the accuracy of abnormal area detection and the convergence speed of residuals in typhoon scenarios, supports the power grid to quickly adjust the backup capacity when the cloud layer suddenly changes, and ensures the safe and stable operation of the power system in extreme weather.

[0020] (6) The meta-learning-based generative adversarial network method adopted in this invention realizes feature fusion, and the uncertainty quantification of fused features is realized based on the Bayesian learning method. It has two advantages. First, the meta-learning framework enables the generative adversarial network to have cross-scenario dynamic adaptability by dividing tasks in multiple meteorological scenarios. After the generator is quickly fine-tuned in the support set, it can accurately fuse the spatial correlation between meteorological variables and cloud obscuration features. Compared with the traditional adversarial network, the feature reconstruction error in the typhoon scenario is reduced. Secondly, the probabilistic fusion mechanism constructed by the Bayesian learning method decomposes the total uncertainty into model cognitive uncertainty and data noise accidental uncertainty, quantifies the confidence interval, and achieves a confidence interval coverage rate of 95% in severe convective weather, significantly improving the interpretability of power grid dispatching decisions. The synergistic effect of the two not only strengthens the adversarial network's ability to generalize extreme weather mutation patterns, but also reveals the reliable boundaries of fused features through probabilistic modeling, effectively ensuring the safety margin and operational stability of the new power system under extreme weather conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is the operational flow chart of the multi-source feature fusion method based on outlier self-correction and meta-learning adversarial analysis of cloud obscuration features in distribution networks under extreme weather conditions. DETAILED DESCRIPTION

[0022] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments described herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0023] The present invention provides a multi-source feature fusion method for self-correction of outliers and meta-learning confrontation of cloud shading characteristics of distribution networks under extreme weather conditions. The specific implementation process of this method has five steps, namely: extraction and prediction of cloud shading characteristics, screening of meteorological variables with high correlation with solar irradiance, generation of shading feature area range, outlier detection and processing, feature fusion based on meta-learning generative adversarial network method, and uncertainty quantification of fusion features based on Bayesian learning method. Figure 1 Describe the specific implementation process of this method. Figure 1 As shown, the method includes:

[0024] Step S110: extract and predict cloud obstruction features.

[0025] By extracting the grayscale mean of pixel values in satellite cloud images, we can indirectly infer the vertical optical thickness of clouds. This is because pixel values are correlated with cloud reflectivity, and the grayscale mean reflects cloud density, which in turn correlates with cloud optical thickness. This method is often used in meteorology to analyze the physical properties of clouds and help predict weather changes.

[0026] Step S110 may include the following steps:

[0027] Step S110 - 1 , extracting cloud occlusion features.

[0028] Since the movement of clouds in the atmosphere is random, rapid, and violent, the following process can be used to extract the cloud obscuration features above the test station: first, the original satellite cloud image is converted into a grayscale image, and its pixel mean is calculated; then the original satellite cloud image is preprocessed to enhance the texture details in the image and count the pixel values for weather classification, where the cloud obscuration features correspond to the counted pixel values; finally, the gray-level co-occurrence matrix is used to extract the entropy and correlation of the original satellite cloud image, and a correlation analysis is performed in combination with the pixel mean to select the higher impact factors.

[0029] In image space, cloud masking features (thickness and texture features of cloud images) correspond to the values of pixels in the image and their distribution patterns. Pixel values primarily reveal the brightness level of the cloud image location, which can reflect the strength of the cloud reflectivity. Considering the correlation between pixel values and cloud reflectivity, the grayscale mean of the pixels in the original satellite cloud image area can be calculated (by averaging all pixel values in a certain area of the satellite cloud image, the grayscale mean of the area can be obtained. The grayscale mean can reflect the overall brightness level of the area, and thus indirectly reflects the average level of cloud reflectivity). This can be used as a key parameter to quantify the cloud distribution density over the distribution network area, thereby indirectly inferring the vertical optical thickness characteristics of the cloud layer. The grayscale mean of a pixel is calculated as follows:

[0030] ,

[0031] Where, 、 、 The value corresponding to each pixel, represents the grayscale mean of the pixel, is the number of pixels.

[0032] Gray-level co-occurrence matrix is a statistical method used to describe the texture characteristics of an image. It extracts texture information by analyzing the spatial relationship between the grayscale values of pixels in the image. In image analysis, entropy, as one of the important indicators, reflects the degree of discreteness of the grayscale distribution of the image and is positively correlated with the complexity of the image texture. Specifically, the richer the texture, the higher the entropy; conversely, the simpler the texture, the lower the entropy. The calculation formula is as follows:

[0033] ,

[0034] in, is the entropy value, which is used to quantify the complexity of the image texture; The total number of gray levels of the image; is the gray value under the conditions of scanning step length and arrangement angle and The joint occurrence probability of adjacent pixel pairs, is the scanning step between pixel pairs, is the arrangement angle of pixel pairs.

[0035] Correlation is used to represent the regularity of image texture in a specific direction. If the image has texture features in a certain direction, the correlation is large, otherwise, it is small. The calculation formula is:

[0036] ,

[0037] Among them, the pixel distribution characteristics in the horizontal direction are represented by the horizontal pixel grayscale mean and standard deviation Representation, and the vertical direction corresponds to the vertical pixel grayscale mean and standard deviation .

[0038] Step S110-2: Predicting cloud cover characteristics

[0039] A cloud obscuration feature prediction method based on a spatiotemporal convolutional long-short-term memory (LSTM) network was developed. By integrating satellite cloud imagery with meteorological data, a multi-channel spatiotemporal cube input was constructed. The network was then used to predict future cloud cover. Combined with adaptive threshold segmentation and morphological post-processing, this method significantly improved prediction accuracy and real-time performance, and can be used for weather forecasting and disaster warning within distribution networks during extreme weather conditions.

[0040] Step S110-2-1, data preparation and preprocessing

[0041] Collect high-resolution original satellite cloud image sequences above the station to be measured (height H, width W, cloud cluster characteristics C, including grayscale mean, entropy, correlation) and synchronized meteorological data ( Including temperature, humidity, wind speed, air pressure, etc.), align the original satellite cloud images of N consecutive hours (N≥6) with the meteorological data in time and space, build a multi-channel space-time cube, and integrate the space-time cube into the time window The data in ,For example, with a 30-minute time window, each cube contains 6 sets of cloud images (grayscale mean, entropy, ,correlation) and corresponding meteorological parameters. The spatial dimension covers the target area of the ,distribution network. The cloud images are randomly rotated (angle range ±15°), ,translated (maximum offset 10%) and Gaussian noise injected (standard deviation ≤ 5% ,grayscale range) to improve the robustness of the model.,Next, Min-Max normalization is used. ( is the original grayscale value of the satellite cloud image, The grayscale value of the original satellite cloud image is mapped to [0, 1], and the meteorological data is normalized using the Z-score ,in, is the original value of meteorological data, is the standardized meteorological data. Include and , Include and .

[0042] Step S110-2-2, constructing a spatiotemporal convolutional long short-term memory network architecture, including input layer parameters, spatiotemporal convolution modules, and temporal modeling of the long short-term memory network.

[0043] Input layer parameters: Receive multi-channel space-time cube, where the number of channels is the number of cloud features (grayscale mean, entropy, correlation) + the number of meteorological parameters (temperature, humidity, wind speed, air pressure, etc.). If the cloud layer extracts 3 features and the meteorological data contains 4 parameters, then the number of input channels is This process combines physical and meteorological characteristics, enhancing the model's ability to model cloud dynamics.

[0044] Spatiotemporal convolution module: The spatiotemporal convolution module includes three-dimensional convolution and activation function, Represented by 3D convolution and The local cloud motion features extracted by the activation function are as follows:

[0045] ,

[0046] in, Input cubes built for multi-source data; It is a three-dimensional convolution operation; It is an activation function used to enhance nonlinear expression capabilities and retain significant positive changes in cloud characteristics, such as grayscale-enhanced thunderstorms; convolution kernel is a learnable parameter matrix of size space × space × time, which is used to extract the joint features of local cloud movement and diffusion area.

[0047] Long Short-Term Memory Network (LSTM) ) temporal modeling: a bidirectional long short-term memory network unit is introduced, with the time step consistent with the input cube. The local cloud motion features output by the spatiotemporal convolution module are used as input to model the long-term temporal dependency of the cloud. The expression is:

[0048] ,

[0049] in, represents the convolution feature, is the hidden state of the forward long short-term memory network unit, which can be used to capture the diffusion law of the cloud from the past to the present. is the hidden state of the backward long short-term memory network unit, which can be used to infer the possible future evolution of the cloud. is the time step The hidden state of the bidirectional long short-term memory network unit is integrated with the forward and backward time sequence information to enhance the ability to represent complex dynamics. Then suddenly accelerate and move, will encode the effect of wind speed changes on the trajectory, It is represented by the dimension after the forward and backward long short-term memory networks are combined. The cloud prediction accuracy is high under this dimension.

[0050] By combining convolutional features and long short-term memory networks, spatial-channel attention weights are dynamically generated to focus on the cloud mutation area. The formula is as follows:

[0051] ,

[0052] in, Represented as time step and spatial location The attention weight at is in the range of [0,1]. The closer the value is to 1, the more important the position is for the prediction of cloud occlusion features. is the Sigmoid function , which can compress the output value to [0,1]. Its function is to map the attention score to the probability weight and enhance the sensitivity of the model to the key area; is the transpose of the learning weight vector, Map the nonlinearly transformed features into scalar attention scores; is the hyperbolic tangent function , introducing nonlinearity to enhance the model's ability to fit complex cloud dynamics; For The transformation matrix in the hidden state; is the convolution feature transformation matrix, For the time step and spatial location The local cloud motion features under the convexity can be used to extract the local cloud motion features.

[0053] The specific application of this formula is as follows: In extreme weather conditions such as heavy rain, the distribution network A local grayscale increase is detected. Amplify the signal and Combined with the wind speed information in the , a high weight is generated. Close to 1.

[0054] Step S110-2-3, feature prediction

[0055] The spatial-channel attention weights With the time step and spatial location Local cloud motion characteristics under Multiply point by point to generate cloud masking features weighted by spatial-channel attention weights , used to enhance the contribution of key areas; the prediction target is to predict the grayscale mean of the future time step ,entropy , correlation , through linear transformation (fully connected layer), the features weighted by spatial-channel attention weights are mapped into three cloud occlusion feature prediction targets. The formula is as follows:

[0056] ,

[0057] in, It is a linear fully connected layer that reduces the dimensionality of high-dimensional features and maps them into the physical quantity space to support the subsequent generation of cloud-obscured areas.

[0058] Step S120: Filter out meteorological variables with high correlation with solar irradiance

[0059] First, we acquired measured solar irradiance data and meteorological variable data (such as cloud cover, temperature, humidity, and wind speed). Then, we performed time alignment to maintain the same resolution for all variables and ensure consistent timestamps. We then used the Pearson correlation coefficient method to select meteorological variables whose correlation with solar irradiance exceeded a correlation threshold.

[0060] The Pearson correlation coefficient formula is as follows:

[0061] ,

[0062] in, are the observed values of meteorological variables (such as cloud cover, air temperature, etc.); is the observed value of solar irradiance; and are the mean values of meteorological variables and solar irradiance, respectively; is the total number of samples (e.g. 24 hour points for hourly data); is the sample size.

[0063] Calculate the Pearson correlation coefficient value of each meteorological variable and solar irradiance ,The larger the absolute value, the stronger the linear correlation.,A positive correlation indicates that the irradiance increases when the meteorological variable increases.,Finally, the correlation strength is graded.,The above method screened out meteorological variables related to solar irradiance and,graded their correlation.

[0064] Step S130: Based on the cloud shielding features in step S110, the spatial range of cloud shielding is generated. The specific method is as follows:

[0065] Step S130-1: Dynamic threshold segmentation and binary mask generation

[0066] The threshold is adaptively adjusted based on the predicted grayscale mean, and the threshold segmentation formula is as follows:

[0067] ,

[0068] in, To predict the grayscale mean value of the future time step, reflecting the cloud density (such as indicates heavy cloud cover); To predict the standard deviation of the grayscale at future time steps, quantify the uncertainty (such as sensor noise); It is a dynamic coefficient, which can be optimized and determined based on historical data, such as when it is sunny and lightly cloudy. , Thunder Cloud By adaptively adjusting the threshold, it can adapt to different weather conditions. It is an adaptive segmentation threshold, a dynamic threshold used to distinguish between clouds and non-clouds.

[0069] The gray value of the satellite cloud image Comparing with the threshold, a preliminary cloud-massage feature area is generated, which provides a basis for subsequent morphological optimization. The formula is as follows:

[0070] ,

[0071] in, Satellite cloud image at location Gray value of A binary mask is an image containing only two pixel values, with 1 marking cloud-covered areas and 0 marking clear areas, making it easy to quickly locate the obscured areas.

[0072] S130-2, morphology optimization

[0073] The morphological optimization operation includes two steps: opening and closing operations and connected region analysis.

[0074] Step S130-2-1 Opening and closing operation

[0075] The opening operation is used to remove the small noise in the cloud image, and the closing operation is used to fill the small holes in the cumulonimbus cloud caused by occlusion. Both the opening operation and the closing operation include two processes: corrosion and expansion. The opening operation is corrosion followed by expansion. The expression is: , which can remove sensor noise; the closing operation is to expand first and then corrode, and the expression is , which can fill the gaps inside the cloud and the breaks in thin clouds; is a structural element, The image to be processed.

[0076] Step S130-2-2 Connected Region Analysis

[0077] After the above-mentioned opening and closing operations, attribute calculation and filtering operations are performed to convert the original binary mask into a high-confidence cloud-occluded area.

[0078] After completing the opening and closing operations, the connected regions (cloud regions) are marked using the union-find method. The method steps are to first initialize each pixel as an independent set, then traverse the pixels and merge the sets of adjacent foreground pixels. Finally, each set corresponds to a connected region. .

[0079] After completing the connected region marking, the region attributes are calculated. The core attributes of the region include the region area, centroid, aspect ratio, and compactness, which can be calculated at once using the following formula.

[0080] area , aspect ratio , compactness , where the area is defined as the total number of all cloud pixels in the area, is the coordinate corresponding to the connected area; the aspect ratio is the ratio of the width to the height of the region bounding box, which can be used to remove long strip artifacts. is the width, is the height; compactness is used to measure the degree to which the shape of the area is close to a circle. is the perimeter of the regional boundary. In short, the coordinated work of these parameters can provide reliable support for weather forecasting and power grid scheduling.

[0081] Step S140: Detect and correct outliers for the cloud-covered space generated in step S130.

[0082] Step S140-1, abnormal value determination condition

[0083] High-precision cloud-shaded area obtained through behavioral optimization , outliers can be detected by the following method.

[0084] Set an unreasonable area that is too small or too large, that is, set is the minimum value of the reasonable area, is the maximum value of the reasonable area, It could be 50 pixels, Can be Pixels.

[0085] if , it is marked as normal, otherwise it is abnormal;

[0086] The areas with aspect ratio > 5 (long strip artifacts) or compactness < 0.2 (highly irregular) are defined as reasonable areas.

[0087] if ,or , it is marked as abnormal, otherwise it is normal;

[0088] Step S140-2: Self-correction of outliers driven by digital twins

[0089] Step S140-2-1, preprocessing of digital twin model

[0090] Input of the digital twin model: satellite cloud images + meteorological variables obtained through Pearson correlation coefficient analysis;

[0091] Physical model prediction: through cloud evolution equations The actual cloud obscuration characteristics can be obtained ; In this cloud evolution equation, is the wind speed vector, in the cloud evolution equation is the advection term, which means that the wind speed drives the cloud to move horizontally, directly affecting spatial distribution of is the turbulent diffusion coefficient. In the cloud evolution equation is the diffusion term, which represents the blurred diffusion caused by turbulence at the edge of the cloud; is the radiation source term, which represents the thickening or evaporation of clouds caused by solar radiation heating.

[0092] Anomaly matching: abnormal areas of connected area analysis The coordinates are mapped to the prediction map, and the residuals are calculated. The residuals are quantified by calculating the Euclidean distance between the predicted value and the observed value in the abnormal area. If the residual exceeds the threshold, the correction process will be triggered. The formula is as follows:

[0093] ,

[0094] in, is the predicted value of cloud occlusion feature mapped by the features weighted by the spatial-channel attention weight; The cloud obscuration features actually observed; are the coordinates of the abnormal regions marked by connected component analysis.

[0095] Step S140-2-2, closed-loop correction mechanism

[0096] The closed-loop correction mechanism includes three steps: local model retraining, data augmentation correction, and physical constraint correction.

[0097] Local model retraining: in abnormal areas Update model parameters in real time through federated learning of the model , improve the ability to adapt to current extreme weather, and jointly optimize data errors and physical consistency ( constraints) to prevent overfitting noise.

[0098] ,

[0099] in, is the trainable parameter of the neural network model; is the learning rate, which controls the step size of parameter update; is the trainable parameter gradient; is the mean square error loss, which is used to measure the predicted cloud cover characteristics Compared with the actual cloud obscuration characteristics differences; is the regularization coefficient, which is used to balance the weight of data error and physical constraints; is the residual term of the physical equation, ensuring that the cloud evolution equation is still satisfied after the model is updated. .

[0100] Data enhancement correction: After completing the dynamic parameter update, data enhancement correction is performed by translating the historical cloud map to generate Similar synthetic cloud images are used to improve the robustness of the model, and then the missing values in the abnormal area are repaired using the adjacent frame data. The formula is as follows:

[0101] ,

[0102] in, To obscure the features of the clouds of historical moments; is the spatial offset, usually taken as Pixels, simulating cloud advection motion; coefficient , which means taking the average of the interpolation of four adjacent spatiotemporal positions to ensure smoothness.

[0103] Physical constraint correction: Enhance the cloud occlusion characteristics obtained after correction This will cause abnormal cloud top temperature. The radiation-temperature relationship can be used to reset the cloud top temperature to prevent the model from outputting cloud parameters that do not conform to atmospheric laws due to excessive reliance on data.

[0104] ,

[0105] in, is the cloud top temperature, is the surface temperature, is the temperature lapse rate, is the cloud height.

[0106] Through the collaborative work of the three, the reliability of cloud obscuration feature prediction in typhoon, thunderstorm and other scenarios has been significantly improved.

[0107] Step S150: The meta-learning-based generative adversarial network algorithm integrates the corrected cloud obscuration features with meteorological variables.

[0108] The meteorological variables obtained in step S120 that have a high correlation with the solar irradiance This is then fused with optimized cloud obscuration features obtained through a meta-learning-based generative adversarial network. This meta-learning-based generative adversarial network fusion algorithm not only integrates multi-source features but also improves the results of the fused predictions, quantifying their uncertainty through Bayesian learning methods.

[0109] Step S150-1: Extract and normalize corresponding features through Bayesian processing

[0110] Step S150-1-1, corresponding feature extraction

[0111] Extraction of meteorological time series features: According to the meteorological variables obtained in step S120 , perform time alignment to obtain the meteorological time series , use Bayesian LSTM to process meteorological time series and output meteorological characteristics for each time step , through the Gaussian distribution model Get the mean of meteorological hidden features and variance , where N is the abbreviation character used to represent Gaussian distribution.

[0112] Corrected cloud occlusion feature extraction for each spatial position: Corrected cloud occlusion feature obtained in step S140 , use Bayesian CNN to process the corrected cloud occlusion features and output the corrected cloud occlusion features at each spatial position , establish a Gaussian distribution model , and obtain the mean of the corrected cloud obscuration characteristics and variance .

[0113] Step S150-1-2, normalization and alignment

[0114] Meteorological variables Normalization: ;

[0115] Corrected cloud obscuration characteristics Normalization: ;

[0116] After normalization of meteorological variables and corrected cloud obscuration characteristics, bilinear interpolation is used to adjust the spatial resolution of meteorological variables to , aligned with satellite cloud images.

[0117] Step S150-2, Generate Adversarial Network Design

[0118] Step S150-2-1, generator G structure and discriminator D structure

[0119] The meteorological variables obtained by normalizing and aligning the above operations and the corrected cloud obscuration characteristics As the input of the generator G, it then passes through the convolutional layer ( Represents a two-dimensional convolution operation) to extract the meteorological variable conditional features, and combine the output of each layer of the encoder inside the generator G with the meteorological variable conditional features After splicing, the encoder is input, and the final generator G outputs the completed cloud occlusion feature ; The output of the discriminator D is .

[0120] Step S150-2-2, define the loss function

[0121] The adversarial loss function is as follows:

[0122] ,

[0123] in, is the expected value, which represents the statistical average of the data distribution; is the gradient penalty coefficient, which is used to balance the weight of the gradient penalty term; is the gradient norm of the discriminator for the interpolated sample, which is used to enforce the continuity of the discriminator. is the interpolation sample Gradient operation of are interpolated samples.

[0124] The reconstruction loss function is as follows:

[0125] ,

[0126] in, is the number of non-zero elements in the mask matrix O, indicating the total number of pixels in the cloud-occluded area; is the binary mask matrix at position The element at , a value of 1 indicates that the location is obscured by clouds and the reconstruction loss needs to be calculated; For real features at position The pixel value of .

[0127] Finally, the total loss function is obtained: , paving the way for meta-learning training strategies, is the weight coefficient used to control the contribution ratio of reconstruction loss to the total loss.

[0128] Step S150-3: Meta-learning framework implementation

[0129] Because meteorological variables vary across datasets and task assignments in different weather scenarios, direct mixed training can lead to confusion between different meteorological characteristics. Therefore, a meta-learning training strategy is employed to achieve dynamic cross-scenario adaptation, making it suitable for extreme weather scenarios. The implementation of the meta-learning training strategy consists of three steps: initialization, outer loop, and meta-parameter update.

[0130] Initialization: Generator parameters (Generator parameters including neural network weights, conditional fusion parameters and normalization parameters) and discriminator parameters (Discriminator parameters Including the discriminative network weights, conditional encoding parameters and spectral normalization parameters) are randomly initialized;

[0131] Outer loop: Divide the meteorological variable dataset into N tasks according to the meteorological scenario, each task contains the support set and queryset , support set A small number of samples, used to simulate rapid adaptation; query set A large number of samples are used to evaluate the generalization ability. In the outer loop process, the inner loop adaptation and query set evaluation need to be satisfied. The inner loop adaptation is to use the support set Calculate the loss and update the parameters. The specific formula is as follows:

[0132] ,

[0133] in, are the adapted generator parameters, are the discriminator parameters after adaptation, is the parameter update step of the generator G, is the parameter update step of the discriminator D, is the generator parameter Gradient operation of is the discriminator parameter Gradient operation, inner loop learning rate , ; Total loss Gradients with respect to the generator parameters; Total loss Gradients with respect to the discriminator parameters;

[0134] After completing the inner loop adaptation operation, the adapted parameters are calculated in the query set Losses .

[0135] Generator parameters and discriminator parameters update:

[0136] ,

[0137] in, and is the inner loop learning rate, is the first tasks, inner loop learning rate , ; and are the updated generator parameters and discriminator parameters respectively.

[0138] After completing the meta-learning training, we get the updated value of the meta-parameters, and reversely inferring them to get the optimized generator. , the meteorological variable condition characteristics and the corrected cloud obscuration characteristics after normalization and alignment As input, the output is the cloud occlusion feature obtained after the optimization operation .

[0139] Step S150-4: Use attention coupling mechanism to couple multimodal features

[0140] The optimized cloud occlusion characteristics obtained by the above operation Meteorological variables with high correlation with solar irradiance Splicing is performed along the channel dimension C to obtain the splicing feature:

[0141] ,

[0142] The obtained splicing features are passed through the convolutional layer learning method to generate the attention weight matrix , the formula is as follows:

[0143] ,

[0144] in, is the attention weight matrix, and each element corresponds to the feature importance of the position. The closer the value is to 1, the more critical the feature at the corresponding position is to the current task. It is a two-dimensional convolution operation; for Activation function, which maps the output to interval.

[0145] Cloud masking features after dynamic fusion optimization based on attention weight matrix Meteorological variables with high correlation with solar irradiance , fusion features The formula is as follows:

[0146] ,

[0147] in, is the attention weight matrix; is element-wise multiplication.

[0148] Step S150-5: quantify the uncertainty of fusion features using Bayesian learning method

[0149] The fusion features It is mapped to probability distribution parameters (mean and variance) to predict the uncertainty of the model.

[0150] Step S150-5-1, Bayesian prediction layer design

[0151] The input of the prediction layer is the fusion feature obtained by uncertainty-aware attention. ; The output is the mean and variance of the fused features, the formula is as follows:

[0152] ,

[0153] ,

[0154] in and is the Bayesian weight coefficient, is the bias term for the mean prediction, is the bias term for the logarithmic prediction of the variance, It is the fusion feature The predicted mean of It is the fusion feature The logarithm of the prediction variance.

[0155] Step S150-5-2, total uncertainty analysis

[0156] Total uncertainty includes epistemic uncertainty and aleatory uncertainty, and uncertainty is reflected by variance.

[0157] Epistemic uncertainty (model uncertainty): Calculate the variance by Monte Carlo sampling T times :

[0158] ,

[0159] in, It is the symbol of variance, which is used to indicate the degree of dispersion of the model prediction.

[0160] Accidental uncertainty (data noise): directly using the variance of the fused features to obtain: .

[0161] Total uncertainty is the sum of epistemic uncertainty and aleatoric uncertainty:

[0162] .

[0163] After obtaining the total uncertainty error, the uncertainty is quantified by the confidence interval. The confidence interval formula is as follows:

[0164] ,

[0165] in, is the masked feature after fusion The confidence interval of , It is usually a small positive number, such as 0.05, indicating that 95% The confidence that the true parameter value falls within this interval is is the fused masking feature, is the mean of the fusion features, For the standard normal distribution (1- ) quantiles (e.g. =0.05 corresponds to =1.96), is the total uncertainty variance of the fusion feature, which is used to measure the discreteness of the fusion feature.

[0166] After completing the uncertainty quantification of the fusion feature, the region range of the fusion feature is generated through step S130 .

[0167] In summary, the multi-source feature fusion method based on meta-learning generative adversarial network realizes the fusion of cloud cluster features under extreme weather conditions and provides key technical support for the safe dispatch of new power systems.

[0168] In the description provided herein, numerous specific details are described. However, it is understood that embodiments of the present invention may be practiced without these specific details. In some instances, well-known methods, structures, and techniques are not shown in detail so as not to obscure the understanding of this description.

[0169] Although the present invention has been described with respect to a limited number of embodiments, those skilled in the art, having benefit of the foregoing description, will appreciate that other embodiments are contemplated within the scope of the invention thus described. Furthermore, it should be noted that the language used in this specification has been selected primarily for readability and instructional purposes, and not for the purpose of explaining or limiting the subject matter of the present invention.

Claims

1. A multi-source feature fusion method based on outlier self-correction and meta-learning for cloud shading features in distribution networks under extreme weather conditions, characterized by: The method includes: Step S110, extracting and predicting cloud obscuration features; Step S120, screening out meteorological variables that have a high correlation with solar irradiance; Step S130, generating a spatial range of cloud cover based on the cloud cover characteristics in step S110; Step S140 , performing outlier detection and correction on the cloud-obscured spatial range generated in step S130 ; Step S150: The meta-learning-based generative adversarial network algorithm fuses the corrected cloud obscuration features with meteorological variables.

2. The multi-source feature fusion method of outlier self-correction and meta-learning confrontation of cloud obscuration features in distribution networks under extreme weather conditions according to claim 1 is characterized in that: Step S110 includes the following steps: Step S110-1, extract cloud cover features: first, convert the original satellite cloud image into a grayscale image and calculate its pixel mean; then pre-process the original satellite cloud image to enhance the texture details in the image and count the pixel values for weather classification, where the cloud cover feature corresponds to the counted pixel values; finally, use the gray level co-occurrence matrix to extract the entropy and correlation of the original satellite cloud image, combine the pixel mean to perform correlation analysis, select the higher impact factor, entropy value The calculation formula is as follows: ， in, is the entropy value, which is used to quantify the complexity of the image texture; The total number of gray levels of the image; is the gray value under the conditions of scanning step length and arrangement angle and The joint occurrence probability of adjacent pixel pairs, is the scanning step between pixel pairs, is the arrangement angle of the pixel pair; Correlation The calculation formula is: ， Among them, the pixel distribution characteristics in the horizontal direction are represented by the horizontal pixel grayscale mean and standard deviation Representation, and the vertical direction corresponds to the vertical pixel grayscale mean and standard deviation ; Step S110 - 2 , predicting cloud obscuration characteristics.

3. The multi-source feature fusion method of outlier self-correction and meta-learning confrontation of cloud obscuration features in distribution networks under extreme weather conditions according to claim 2 is characterized in that: Step S110-2 includes: Step S110-2-1, data preparation and preprocessing Collect high-resolution original satellite cloud image sequences above the station to be measured , the parameters are height H, width W, cloud cluster characteristics C, including grayscale mean, entropy, correlation, and synchronized meteorological data , Including temperature, humidity, wind speed, and air pressure, align the original satellite cloud images for N consecutive hours with the meteorological data in time and space, build a multi-channel space-time cube, and integrate the time window of the space-time cube The data in , using Min-Max normalization , the original satellite cloud image grayscale value is mapped to [0, 1], and the meteorological data is standardized using Z-score ,in, is the raw meteorological data, is the standardized meteorological data. Include and , Include and ; Step S110-2-2, constructing a spatiotemporal convolutional long short-term memory network architecture, including input layer parameters, spatiotemporal convolution modules, and temporal modeling of the long short-term memory network; Input layer parameters: accept multi-channel space-time cube, where the number of channels is the number of cloud features + the number of meteorological parameters; Spatiotemporal convolution module: The spatiotemporal convolution module includes three-dimensional convolution and activation function, Represented by 3D convolution and The local cloud motion features extracted by the activation function are as follows: ， in, Input cubes built for multi-source data; It is a three-dimensional convolution operation; It is an activation function used to enhance nonlinear expression capabilities and retain significant positive changes in cloud characteristics, such as grayscale-enhanced thunderstorms; convolution kernel is a learnable parameter matrix with the size of space × space × time, which is used to extract the joint features of local cloud movement and diffusion area; Long Short-Term Memory Network Temporal modeling: introduce a bidirectional long short-term memory network unit with the same time step as the input cube, and convert the local cloud motion characteristics output by the spatiotemporal convolution module into As the input of the bidirectional long short-term memory network unit, the long-term temporal dependency of the cloud is modeled as follows: ， in, represents the convolution feature, is the hidden state of the forward long short-term memory network unit, which is used to capture the diffusion pattern of the cloud from the past to the present. is the hidden state of the backward long short-term memory network unit, which is used to infer the possible future evolution of the cloud. is the time step The hidden state of the bidirectional long short-term memory network unit is integrated with the forward and backward time sequence information to enhance the ability to represent complex dynamics. It is represented by the dimension of the combined forward and backward long short-term memory networks. The cloud prediction accuracy is high under this dimension. By combining convolutional features and long short-term memory networks, spatial-channel attention weights are dynamically generated to focus on the cloud mutation area. The formula is as follows: ， in, Represented as time step and spatial location The attention weight at is in the range of [0,1]. The closer the value is to 1, the more important the position is for the prediction of cloud occlusion features. is the Sigmoid function , which can compress the output value to [0,1]; is the transpose of the learning weight vector, Map the nonlinearly transformed features into scalar attention scores; is the hyperbolic tangent function , introducing nonlinearity to enhance the model's ability to fit complex cloud dynamics; For The transformation matrix in the hidden state; is the convolution feature transformation matrix, For the time step and spatial location The local cloud motion features under the ,are used to extract the local cloud motion features; Step S110-2-3, feature prediction: The spatial-channel attention weights With the time step and spatial location Local cloud motion characteristics under Multiply point by point to generate cloud masking features weighted by spatial-channel attention weights , used to enhance the contribution of key areas; the prediction target is to predict the grayscale mean of the future time step ,entropy , correlation , the features weighted by the spatial-channel attention weights are mapped into three cloud masking feature prediction targets through the linear transformation fully connected layer. The formula is as follows: ， in, It is a linear fully connected layer that reduces the dimensionality of high-dimensional features and maps them into the physical quantity space to support the subsequent generation of cloud-obscured areas.

4. The multi-source feature fusion method of outlier self-correction and meta-learning confrontation of cloud obscuration features in distribution networks under extreme weather conditions according to claim 3 is characterized in that: Step S120 includes: first, obtaining solar irradiance measured data and meteorological variable data, then performing time alignment to keep all variables at the same resolution and ensure that the timestamps of the data are consistent, and using the Pearson correlation coefficient method to screen out meteorological variables whose correlation with solar irradiance is greater than a correlation threshold; The Pearson correlation coefficient formula is as follows: ， in, is the observed value of the meteorological variable; is the observed value of solar irradiance; and are the mean values of meteorological variables and solar irradiance, respectively; is the total number of samples; is the sample size; Calculate the Pearson correlation coefficient value of each meteorological variable and solar irradiance , the larger the absolute value, the stronger the linear correlation. Positive correlation means that the irradiance increases when the meteorological variable increases. Finally, the correlation strength is graded.

5. The multi-source feature fusion method of outlier self-correction and meta-learning confrontation of cloud obscuration features in distribution networks under extreme weather conditions according to claim 4 is characterized in that: Step S130 includes: Step S130-1, dynamic threshold segmentation and binary mask generation: The threshold is adaptively adjusted based on the predicted grayscale mean, and the threshold segmentation formula is as follows: ， in, To predict the grayscale mean value of the future time step, reflecting the cloud density; To quantify the uncertainty, we need to predict the grayscale standard deviation at future time steps; is the dynamic coefficient, It is an adaptive segmentation threshold, a dynamic threshold used to distinguish between clouds and non-clouds; The gray value of the satellite cloud image Comparing with the threshold, a preliminary cloud-massage feature area is generated. The formula is as follows: ， in, Satellite cloud image at location Gray value of A binary mask is an image that contains only two pixel values, with 1 marking cloud-covered areas and 0 marking clear areas; S130-2, morphology optimization, including: Step S130-2-1 Opening and closing operation Opening operation is used to remove small noise in the cloud image, and closing operation is used to fill small holes in the cumulonimbus cloud caused by occlusion; Step S130-2-2 Connected Region Analysis After the aforementioned opening and closing operations, attribute calculation and filtering operations are performed to convert the original binary mask into a highly reliable cloud-occluded area. After the opening and closing operations are completed, the connected regions are marked using the union-find method. After the connected regions are marked, the region attributes are calculated. The core attributes of the region include the area, centroid, aspect ratio, and compactness, which are calculated once using the following formula: area , aspect ratio , compactness , where the area is defined as the total number of all cloud pixels in the area, is the coordinate corresponding to the connected area; the aspect ratio is the ratio of the width to the height of the region bounding box, which is used to remove long strip artifacts. is the width, is the height; compactness is used to measure the degree to which the shape of the area is close to a circle. is the perimeter of the region boundary.

6. The multi-source feature fusion method of outlier self-correction and meta-learning confrontation of cloud obscuration features in distribution networks under extreme weather conditions according to claim 5 is characterized in that: Step S140 includes: Step S140-1, abnormal value determination conditions: Set an unreasonable area that is too small or too large, that is, set is the minimum value of the reasonable area, is the maximum value of the reasonable area, if , it is marked as normal, otherwise it is abnormal; Set the area with aspect ratio > 5 or compactness < 0.2 as the reasonable area: if ,or , it is marked as abnormal, otherwise it is normal; Step S140-2: Self-correction of outliers driven by digital twins.

7. The multi-source feature fusion method of outlier self-correction and meta-learning confrontation of cloud obscuration features in distribution networks under extreme weather conditions according to claim 6 is characterized in that: Step S140-2 includes: Step S140-2-1, preprocessing of digital twin model Inputs of the digital twin model: satellite cloud images and meteorological variables obtained through Pearson correlation coefficient analysis; Physical model prediction: through cloud evolution equations , get the actual cloud obscuration characteristics ; In the cloud evolution equation, is the wind speed vector, in the cloud evolution equation is the advection term, which means that the wind speed drives the cloud to move horizontally, directly affecting spatial distribution of is the turbulent diffusion coefficient. In the cloud evolution equation is the diffusion term, which represents the blurred diffusion caused by turbulence at the edge of the cloud; is the radiation source term, which represents the thickening or evaporation of clouds caused by solar radiation heating; Anomaly matching: abnormal areas of connected area analysis The coordinates are mapped to the prediction map, and the residuals are calculated. The residuals are quantified by calculating the Euclidean distance between the predicted value and the observed value in the abnormal area. If the residual exceeds the threshold, the correction process will be triggered. The formula is as follows: ， in, is the predicted value of cloud occlusion feature mapped by the features weighted by the spatial-channel attention weight; The cloud obscuration features actually observed; are the coordinates of the abnormal regions marked by connected component analysis; Step S140-2-2, closed-loop correction mechanism The closed-loop correction mechanism includes three steps: local model retraining, data augmentation correction, and physical constraint correction: Local model retraining: in abnormal areas Update model parameters in real time through federated learning of the model , improve the ability to adapt to current extreme weather, and jointly optimize data errors and physical consistency ( Constraints), to prevent overfitting noise: ， in, is the trainable parameter of the neural network model; is the learning rate, which controls the step size of parameter update; is the parameter gradient; is the mean square error loss, which is used to measure the predicted cloud cover characteristics Compared with the actual cloud obscuration characteristics differences; is the regularization coefficient, which is used to balance the weight of data error and physical constraints; is the residual term of the physical equation, ensuring that the cloud evolution equation is still satisfied after the model is updated. ; Data enhancement correction: After completing the dynamic parameter update, data enhancement correction is performed by translating the historical cloud map to generate Similar synthetic cloud images are used to improve the robustness of the model, and then the missing values in the abnormal area are repaired using the adjacent frame data. The formula is as follows: ， in, To obscure the features of the clouds of historical moments; is the spatial offset, usually taken as Pixels, simulating cloud advection motion; coefficient , which means taking the average of the interpolation of four adjacent spatiotemporal positions to ensure smoothness; Physical constraint correction: Enhance the cloud occlusion characteristics obtained after correction This will cause abnormal cloud top temperature. The radiation-temperature relationship is used to reset the cloud top temperature to prevent the model from outputting cloud parameters that do not conform to atmospheric laws due to excessive reliance on data: ， in, is the cloud top temperature, is the surface temperature, is the temperature lapse rate, is the cloud height.

8. The multi-source feature fusion method of outlier self-correction and meta-learning confrontation of cloud obscuration features in distribution networks under extreme weather conditions according to claim 7 is characterized in that: Step S150 includes: Step S150-1, extracting and normalizing corresponding features through Bayesian processing; Step S150-2, generating an adversarial network design; Step S150-3, meta-learning framework implementation; Step S150-4, using the attention coupling mechanism to perform multimodal feature coupling; Step S150-5: quantify the uncertainty of the fusion feature using a Bayesian learning method.

9. The multi-source feature fusion method of outlier self-correction and meta-learning confrontation of cloud obscuration features in distribution networks under extreme weather conditions according to claim 8 is characterized in that: Step S150-1 includes: Step S150-1-1, corresponding feature extraction: Extraction of meteorological time series features: According to the meteorological variables obtained in step S120 , perform time alignment to obtain the meteorological time series , use Bayesian LSTM to process meteorological time series and output meteorological characteristics for each time step , through the Gaussian distribution model Get the mean of meteorological hidden features and variance , where N is the abbreviation character used to represent Gaussian distribution; Corrected cloud occlusion feature extraction for each spatial position: Corrected cloud occlusion feature obtained in step S140 , use Bayesian CNN to process the corrected cloud occlusion features and output the corrected cloud occlusion features at each spatial position , establish a Gaussian distribution model , and obtain the mean of the corrected cloud obscuration characteristics and variance ; Step S150-1-2, normalization and alignment: Meteorological variables Normalization: ; Corrected cloud obscuration characteristics Normalization: ; After normalization of meteorological variables and corrected cloud obscuration characteristics, bilinear interpolation is used to adjust the spatial resolution of meteorological variables to , aligned with satellite cloud images.

10. The multi-source feature fusion method of outlier self-correction and meta-learning confrontation of cloud obscuration features in distribution networks under extreme weather conditions according to claim 9 is characterized in that: Step S150-2 includes: Step S150-2-1, generator G structure and discriminator D structure: The meteorological variables obtained by normalizing and aligning the above operations and the corrected cloud obscuration characteristics As the input of the generator G, it then passes through the convolutional layer ( Represents a two-dimensional convolution operation) to extract the meteorological variable conditional features, and combine the output of each layer of the encoder inside the generator G with the meteorological variable conditional features After splicing, the encoder is input, and the final generator G outputs the completed cloud occlusion feature ; The output of the discriminator D is ; Step S150-2-2, define the loss function: The adversarial loss function is as follows: ， in, To counter the loss, it is used to balance the ability of generation and discrimination; is the expected value, which represents the statistical average of the data distribution; is the gradient penalty coefficient, which is used to balance the weight of the gradient penalty term; is the gradient norm of the discriminator for the interpolated sample, which is used to enforce the continuity of the discriminator. is the interpolation sample Gradient operation of is the interpolation sample; The reconstruction loss function is as follows: ， in, is the number of non-zero elements in the mask matrix O, indicating the total number of pixels in the cloud-occluded area; is the binary mask matrix at position The element at , a value of 1 indicates that the location is obscured by clouds and the reconstruction loss needs to be calculated; For real features at position Pixel value of Finally, the total loss function is obtained: , paving the way for meta-learning training strategies, is the weight coefficient used to control the contribution ratio of reconstruction loss to the total loss.

11. The multi-source feature fusion method of outlier self-correction and meta-learning confrontation of cloud obscuration features in distribution networks under extreme weather conditions according to claim 10 is characterized in that: The implementation process of the meta-learning training strategy includes three steps: initialization, outer loop, and meta-parameter update: Initialization: Generator parameters and the discriminator parameters Random initialization, generator parameters Including neural network weights, conditional fusion parameters and normalization parameters, discriminator parameters Including discriminant network weights, conditional encoding parameters and spectral normalization parameters; Outer loop: Divide the meteorological variable dataset into N tasks according to the meteorological scenario, each task contains the support set and queryset , support set A small number of samples, used to simulate rapid adaptation; query set A large number of samples are used to evaluate generalization ability; In the outer loop process, the inner loop adaptation and query set evaluation need to be satisfied. The inner loop adaptation is to use the support set Calculate the loss and update the parameters. The specific formula is as follows: ， in, are the adapted generator parameters, are the discriminator parameters after adaptation, is the parameter update step of the generator G, is the parameter update step size of the discriminator D, is the generator parameter Gradient operation of is the discriminator parameter Gradient operation, inner loop learning rate , ; Total loss Gradients with respect to the generator parameters; Total loss Gradients with respect to the discriminator parameters; After completing the inner loop adaptation operation, the adapted parameters are calculated in the query set Losses ; Generator parameters and discriminator parameters update: ， in, and is the inner loop learning rate, is the first tasks, inner loop learning rate , ; and are the updated generator parameters and discriminator parameters respectively; After completing the meta-learning training, the updated value of the meta-parameter is obtained, and the optimized generator is obtained by reverse deduction. , the meteorological variable condition characteristics and the corrected cloud obscuration characteristics after normalization and alignment As input, the output is the cloud occlusion feature obtained after the optimization operation .

12. The multi-source feature fusion method of outlier self-correction and meta-learning confrontation of cloud obscuration features in distribution networks under extreme weather conditions according to claim 11 is characterized in that: Step S150-4 includes: optimizing the cloud masking feature Meteorological variables with high correlation with solar irradiance Splicing is performed along the channel dimension C to obtain the splicing feature: ， The obtained splicing features are passed through the convolutional layer learning method to generate the attention weight matrix , the formula is as follows: ， in, is the attention weight matrix, each element corresponds to the feature importance of the position; It is a two-dimensional convolution operation; for Activation function, which maps the output to interval; Cloud masking features after dynamic fusion optimization based on attention weight matrix Meteorological variables with high correlation with solar irradiance , fusion features The formula is as follows: ， in, is the attention weight matrix; is element-wise multiplication.

13. The multi-source feature fusion method of outlier self-correction and meta-learning confrontation of cloud shading features in distribution networks under extreme weather conditions according to claim 12 is characterized in that: Step S150-5 includes: fusing features Mapped to probability distribution parameters (mean and variance), and then predicting the uncertainty of the model: Step S150-5-1, Bayesian prediction layer design: The input of the prediction layer is the fusion feature obtained by uncertainty-aware attention. ; The output is the mean and variance of the fused features, the formula is as follows: ， ， in and is the Bayesian weight coefficient, is the bias term for the mean prediction, is the bias term for the logarithmic prediction of the variance, It is the fusion feature The predicted mean of It is the fusion feature The logarithm of the prediction variance of ; Step S150-5-2, total uncertainty analysis: Total uncertainty includes epistemic uncertainty and aleatory uncertainty, and uncertainty is reflected by variance; Epistemic uncertainty: Calculate variance by Monte Carlo sampling T times : ， in, is the symbol for variance, which is used to indicate the degree of dispersion of the model prediction; Accidental uncertainty (data noise): directly using the variance of the fused features to obtain: ; Total uncertainty is the sum of epistemic uncertainty and aleatoric uncertainty: ； After obtaining the total uncertainty error, the uncertainty is quantified by the confidence interval. The confidence interval formula is as follows: ， in, is the masked feature after fusion The confidence interval of , It is usually a small positive number, such as 0.05, indicating that there is a 95% confidence that the true parameter value falls within this interval. is the fused masking feature, is the mean of the fusion features, For the standard normal distribution (1- ) quantiles (e.g. =0.05 corresponds to =1.96), is the total uncertainty variance of the fusion feature, which is used to measure the discreteness of the fusion feature; After completing the uncertainty quantification of the fusion feature, the region range of the fusion feature is generated through step S130 .

Images