Regional cloud field collaborative prediction and correction method based on multi-site sky imaging network

Abstract

The invention discloses a regional cloud field collaborative prediction and correction method based on a multi-site sky imaging network, and the method comprises the following steps: constructing and deploying the multi-site sky imaging network, and selecting and deploying sparsely distributed network sites; collecting and preprocessing multi-source observation data; carrying out graph structure modeling and node feature extraction; space-time diagram neural network prediction: constructing a model integrating space-time convolution and a diagram neural network for prediction; performing multi-source observation data fusion and dynamic correction, fusing the multi-source observation data, and performing dynamic correction and virtual completion when the data is abnormal or missing; intelligently generating a missing cloud picture by adopting an adversarial network; and outputting a regional cloud field prediction result and early warning information, outputting a future short-time scale regional cloud field prediction result, and automatically generating the early warning information. According to the method, the authenticity and accuracy of regional cloud field collaborative prediction can be improved, so that support is provided for scheduling decision-making of wind and light power generation, and the problem that the scheduling capability of a wind and light power generation station cannot be well improved by using a cloud picture prediction technology at present is solved.

Description

Technical Field

[0001] This invention relates to a cloud image prediction method, specifically a regional cloud field collaborative prediction and correction method based on a multi-site sky imaging network. It belongs to the fields of meteorological forecasting and renewable energy dispatching technology. Background Technology

[0002] As wind and solar power account for an increasing proportion of the power structure, the uncertainty caused by cloud fluctuations is having a growing impact on power grid operation. Therefore, in actual operation, how to utilize cloud image prediction technology to improve the dispatching capabilities of wind and solar power plants has become an important aspect of the stable operation of the new power grid.

[0003] Most existing cloud image prediction methods rely on satellite remote sensing or single ground station monitoring. Satellite remote sensing prediction technology provides large-scale macroscopic cloud system monitoring through geostationary or polar orbit satellites, focusing on the movement trend of large-scale weather systems. Ground station prediction technology, on the other hand, uses equipment such as all-sky imagers deployed at specific locations to provide high temporal resolution local cloud condition monitoring over the station. However, both have the following problems: (1) Satellite remote sensing technology has a low image update frequency and cannot capture the minute-level changes in cloud formation and dissipation details. Ground station prediction can only observe the sky within a limited range and cannot perceive the overall movement path of clouds and regional coordinated evolution. (2) The correlation between data is poor and there is a lack of coordination. Once a single point of equipment fails or the data is abnormal, the entire prediction method fails. (3) There is a lack of effective modeling capabilities for the dynamic propagation of clouds at the regional scale, which makes the prediction results unable to meet the scheduling needs of high-proportion new energy power grids in terms of timeliness, accuracy and regional coordination.

[0004] Based on the above, how to break through the bottlenecks of existing cloud monitoring and prediction technologies, develop low-latency dynamic tracking of cloud clusters and collaborative prediction of regional cloud fields, and provide guidance for the stable operation of wind and solar power generation systems are technical problems that urgently need to be solved by those skilled in the art.

[0005] Relevant patent document: CN120405804A discloses a DNI prediction method based on all-sky imaging and clear-sky DNI fitting, including the following steps: selecting clear-sky DNI data from historical DNI data and fitting a clear-sky DNI model, then using the fitted clear-sky DNI model to calculate the clear-sky DNI at each time point; inputting the real-time all-sky imaging and the all-sky imaging of the previous minute into a trained dual-image convolutional neural network model to predict the clear-sky index one minute later; multiplying the predicted clear-sky index one minute later with the corresponding clear-sky DNI to obtain the predicted DNI value one minute later.

[0006] The above technologies do not solve the problem of how to improve the authenticity and accuracy of regional cloud field collaborative prediction, and thus improve the scheduling capability of wind and solar power plants. Summary of the Invention

[0007] The purpose of this invention is to provide a regional cloud field collaborative prediction and correction method based on a multi-site sky imaging network. This method can improve the realism and accuracy of regional cloud field collaborative prediction, thereby providing support for wind and solar power scheduling decisions and solving the current problem that cloud image prediction technology cannot be effectively used to improve the scheduling capabilities of wind and solar power plants.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0009] A regional cloud field collaborative prediction and correction method based on a multi-site sky imaging network, the technical solution of which includes the following steps:

[0010] S1: Construction and deployment of a multi-site sky imaging network. This step involves selecting and deploying sparsely distributed network sites.

[0011] S2: Multi-source observation data acquisition and preprocessing. This step acquires sky imaging images from multiple stations, and simultaneously obtains multi-source observation data and performs data preprocessing.

[0012] S3: Graph structure modeling and node feature extraction. This step treats the sky imaging network sites as graph nodes and extracts node features.

[0013] S4: Spatiotemporal Graph Neural Network Prediction. This step involves constructing a model that integrates spatiotemporal convolutional neural networks (ST-CNN) and graph neural networks (GNN) for prediction.

[0014] S5: Multi-source observation data fusion and dynamic correction. This step fuses multi-source observation data and performs dynamic correction and virtual completion when data is abnormal or missing.

[0015] S6: Uses adversarial networks to intelligently generate missing cloud maps;

[0016] S7: Output regional cloud field prediction results and early warning information. This step outputs the future short-term regional cloud field prediction results and automatically generates early warning information.

[0017] In the above technical solution, a preferred technical solution may be that, in step S1, the deployment of network sites is prioritized, with the sky imager site being prioritized, and the site layout is dynamically optimized based on cloud hotspots, with different site networks forming a low-power long-distance communication network; step S1 specifically includes the following steps:

[0018] S11: Develop a site selection strategy.

[0019] The distribution of stations must consider geographical location, regional meteorological environment, and the dynamic evolution characteristics of cloud fields. Station placement should prioritize areas with complex climate conditions and frequent cloud movement. The optimal station layout is determined by comprehensively analyzing factors such as historical cloud movement patterns, wind speed, and air pressure changes within the region. The station deployment algorithm uses the following formula to determine station priority:

[0020]

[0021] Among them, P i For the deployment priority of the i-th site, d ij v is the distance between station i and station j. ij For wind speed differences, Δt is the time interval for data collection, α, β, and γ are the weighting coefficients for distance, wind speed, and time interval, respectively, and N is the number of all candidate stations;

[0022] This formula allows for the prioritization of candidate sites, prioritizing deployment in areas with more drastic cloud field changes and greater wind speed variations.

[0023] S12: Perform inter-site communication and data transmission.

[0024] Design a low-power, long-distance communication protocol that can adjust the data transmission frequency according to different weather conditions. During periods of rapid cloud cover changes or high communication demand, the protocol should accelerate data updates, while reducing the frequency when cloud cover is stable or data changes are minimal.

[0025] Inter-site communication also needs to consider transmission latency and bandwidth limitations, requiring the use of appropriate communication scheduling strategies. This strategy is adaptively adjusted using the following optimization formula:

[0026]

[0027] Among them, C ij d represents the communication quality between station i and station j. ij L represents the distance between stations. ij SNR is the length of the data packet. ij It is the signal-to-noise ratio, and α and β are weighting coefficients.

[0028] S13: Perform site layout and dynamic optimization.

[0029] The site layout is optimized based on real-time changes in the cloud field. This optimization mechanism, based on real-time cloud field change data and combined with meteorological conditions such as wind speed and temperature, determines the current direction and speed of cloud movement and adjusts the working frequency and data collection interval between sites. This optimization process is achieved through the following dynamic adjustment formula:

[0030]

[0031] Wherein, ΔP i This is the optimization adjustment amount for site i, Δd ij and Δv ij These represent the changes in distance from other stations and the changes in wind speed, respectively, with γ and δ being dynamically adjusted weighting coefficients.

[0032] In the above technical solution, a preferred technical solution may also be that, in step S2, the multi-source observation data includes data collected by sky imagers, weather stations, radar, etc.; the data preprocessing method is to perform noise filtering, spatiotemporal alignment, and illumination condition correction on the raw data after acquiring the multi-source observation data. Step S2 specifically includes the following steps:

[0033] S21: Data Acquisition.

[0034] Each station is equipped with a sky imager that captures panoramic images of the sky through image acquisition equipment. Each station also collects meteorological information such as wind speed, temperature, humidity, and air pressure.

[0035] The data acquisition frequency for each station is adaptively adjusted based on meteorological conditions and the rate of cloud cover change. The data acquisition frequency is determined using the following formula:

[0036] f 采集 =f 基准 ·(1+α·ΔC) (4)

[0037] Among them, f 采集 f is the data collection frequency of the site. 基准 The base acquisition frequency is α, which is an adjustment coefficient, and ΔC is the rate of change of the current cloud field.

[0038] S22: Data preprocessing.

[0039] In the image data acquired by the sky imager, an image denoising model based on deep convolutional neural networks (CNN) is used;

[0040] In meteorological data, the Kalman filter is used to smooth continuous meteorological data. The Kalman filter formula is as follows:

[0041]

[0042] P k =(IK k H)·P k-1 (6)

[0043] Among them, P k For the estimated state, K k For Kalman gain, Pk-1 Here, H is the observation matrix, represents the estimation error covariance, and I is the identity matrix;

[0044] S23: Perform spatiotemporal alignment and standardization on the data.

[0045] Data collected from different sites needs to be spatiotemporally aligned and standardized.

[0046] Spatiotemporal alignment is achieved through interpolation. Assuming that stations i and j have different observations at the same time point, linear interpolation is used to calculate the spatiotemporal data of the two stations:

[0047]

[0048] Among them, z 对齐 For the aligned data, z i and z j The observations for stations i and j are respectively, t i and t j For the time point of data collection,

[0049] Standardization is the process of unifying the dimensions of data collected from different sites. The standardization process is achieved using the mean-variance standardization formula:

[0050]

[0051] Where z is the original data, μ is the mean of the data, σ is the standard deviation of the data, and z′ is the standardized data.

[0052] S24: Merge the data.

[0053] Data from different sensors is fused. The goal of fusion is to integrate cloud image data from a sky imager with environmental data from a weather station. Data fusion techniques include weighted averaging and Kalman filtering. For data from multiple sensors, a weighted averaging method is used for fusion.

[0054]

[0055] Among them, z 融合 For the merged data, w i The weights for each data source, z i Here, N represents the observations from each sensor, and N is the number of data sources.

[0056] After the above preprocessing and fusion steps, the resulting data will become the input data for subsequent graph neural network (GNN) modeling.

[0057] In the above technical solution, a preferred technical solution may also be that, in step S3, the graph structure modeling includes converting the multi-source observation data of the station into node features in a graph network, and modeling it using a graph neural network (GNN); step S3 specifically includes the following steps:

[0058] S31: Node construction and feature extraction.

[0059] The system treats each station as a node in a graph network. The features of each node include local cloud image information and other meteorological features such as wind speed, temperature, and humidity. These features provide local observation data for the station and input data for subsequent graph neural network modeling.

[0060] Cloud image data consists of panoramic images acquired by sky imagers. After image processing and preprocessing, these images are transformed into features usable by the GNN model. The color, brightness, and corresponding cloud classification information of each pixel in the image, including cumulus, stratus, and cirrus clouds, can all serve as input features for the nodes. These image features are extracted using a deep convolutional neural network (CNN). Assuming the image features of a certain site are F... i Then the node features are represented as:

[0061] F i =CNN(I i (10)

[0062] Among them, I i This represents the sky image acquired by station i. CNN is a convolutional neural network used to extract cloud information from the image. Through the convolutional network, the features of each region in the image are encoded into a vector representation, which serves as the feature of the node.

[0063] S32: Construction of graph structures.

[0064] Once the node features are extracted, the next step is to construct the graph structure. The core idea of ​​graph neural networks is to transmit information through the connections between nodes (i.e., edges). Each node in the graph represents an observation station, and the weights of the edges are dynamically calculated based on the spatial relationships between stations and meteorological conditions.

[0065] Edge connections between stations represent the physical or meteorological correlations between them. The weights of the edges reflect the similarity between stations and their impact on cloud field evolution. Assume the distance between station i and station j is d. ij And the wind speed difference between the two is v ij Then the edge weight w ij Represented as:

[0066]

[0067] Where α and β are the weights of the adjustment factor, l scale and v scale These are the scale factors for distance and wind speed, respectively, ensuring that the calculated weights have appropriate units and ranges. This formula combines spatial distance and meteorological factors, so that the edge weights can truly reflect the relationship between stations.

[0068] S33: Spatiotemporal Feature Modeling and Information Transmission.

[0069] After the graph structure is built, the graph neural network will begin to model the cloud field. The features of each node interact with other nodes through the message passing mechanism of the graph neural network. Through graph convolution operations, each node will update its features based on the information of its neighboring nodes and learn more comprehensive cloud field features.

[0070] Graph Convolutional Networks (GCNs) update node features using the following formula:

[0071]

[0072] in, This represents the feature of node i in the (k+1)th layer. Let w be the set of neighboring nodes of node i. ij It is the edge weight. σ is the feature of node j in the k-th layer, b is the bias term, and σ is the activation function.

[0073] In the above technical solution, a preferred technical solution may also include step S4 further comprising:

[0074] After constructing a model that integrates spatiotemporal convolutional neural networks (ST-CNN) and graph neural networks (GNN) for prediction, cross-site information is shared through a message passing mechanism to achieve real-time early warning.

[0075] The aforementioned real-time early warning system provides advance warnings to downstream sites by predicting the movement of strong convective cloud clusters or large-scale cloud clusters.

[0076] In the above technical solution, a preferred technical solution may also include the following steps in step S4:

[0077] S41: Construction of spatiotemporal graph convolutional networks.

[0078] Spatiotemporal Graph Convolutional Networks (ST-GCN) are models that combine graph neural networks and spatiotemporal convolution, enabling them to capture information in both spatial and temporal dimensions. In graph neural networks, graph convolution operations pass information through the adjacency relationships of nodes. In ST-GCN, the node feature update of each layer is represented by the following formula:

[0079]

[0080] in, w is the feature of node i at layer k+1. ij It is the edge weight. It is the feature of neighbor node j at the k-th layer. Let σ be the set of neighbors of node i, σ be the activation function, and b be the bias term. This formula means that the features of the current node are updated by aggregating the information of neighboring nodes.

[0081] In spatiotemporal graph neural networks, a temporal convolution operation is also introduced. Spatiotemporal convolutional networks capture the evolution of clouds over time by introducing temporal convolution on top of graph convolution. The spatiotemporal convolution update formula is:

[0082]

[0083] Where T is the time dimension, λ t It is the weight of the time step, representing the contribution of each time step to the node features.

[0084] S42: Joint learning of graph convolution and time dependence.

[0085] Design a spatiotemporally weighted graph convolution mechanism such that each time step has a different influence on the update of the current node. For a node i, the update process of the node considers the information of the neighboring nodes and dynamically adjusts its weights according to the characteristics of historical time steps. This process is represented by the following formula:

[0086]

[0087] Where, λ t The adjustment is dynamically based on the degree of influence of each time step on the node features.

[0088] S43: Perform spatiotemporal prediction and cloud field evolution.

[0089] The output prediction result of the spatiotemporal graph neural network is expressed by the following formula:

[0090]

[0091] in, This is the cloud field prediction result at a future time, where X is the input feature of the node, A is the adjacency matrix of the graph, T is the time step, and f is the value of f. ST-GCN It is the prediction function of the spatiotemporal graph neural network model;

[0092] S44: Establish an early warning mechanism and make real-time adjustments.

[0093] When the system predicts strong convective cloud clusters or large-scale cloud cluster movement, it can provide early warnings to downstream sites.

[0094] In the above technical solution, a preferred technical solution may also be that, in step S5, the multi-source observation data fusion and dynamic correction includes using a weighted average, Kalman filter, or Bayesian inference method to correct inconsistencies between data; step S5 specifically includes the following steps:

[0095] S51: Employ a multi-source data fusion method.

[0096] By combining observational data from different sensors and leveraging the complementarity of different data sources to compensate for their respective shortcomings, the input data of the cloud field prediction system includes cloud image data captured by the sky imager, meteorological parameters such as wind speed, temperature, and humidity provided by the weather station, as well as radar data. The fusion of these data is achieved using the weighted average method, Kalman filtering method, or Bayesian inference method.

[0097] First, the data from various data sources are preprocessed and standardized to ensure that different data types have the same dimensions and scale. Then, Kalman filtering is used for data fusion. Kalman filtering is a recursive estimation method that updates the state estimate based on the previous prediction and the current observation, and calculates an optimal estimate. The Kalman filtering update formula is as follows:

[0098]

[0099] P k =(IK k H)·P k-1 (20);

[0100] in, The estimated state at the current moment, z k For the current observation data, K k Let H be the Kalman gain, H be the observation matrix, and P be the Kalman gain. k To estimate the error covariance, where I is the identity matrix, the system can achieve optimal fusion of various types of observation data through Kalman filtering, providing a unified state estimate.

[0101] Different weighting methods are used for fusing different types of data sources. For example, image data from a sky imager and temperature and humidity data from a weather station are fused using a weighted average method.

[0102]

[0103] Among them, z 融合 It is the merged data, w i The weights for each data source, z i For each sensor's observations, N is the number of data sources, and w is the weight. iIt can be dynamically adjusted based on the importance or precision of the data source.

[0104] S52: Establish a dynamic correction mechanism.

[0105] The correction mechanism includes two aspects: one is to fill in missing data, and the other is to correct abnormal data. In the processing of missing data, interpolation or a method based on data from neighboring sites is used to complete the missing data.

[0106] For outlier correction, a statistical model-based approach is used for identification and processing. For example, consider the cloud map data z of a certain site. i If an abnormal deviation occurs at a certain moment, the system will use historical data models or observation data from neighboring stations to make corrections. By calculating the similarity between the cloud map of the current station and its neighboring stations, the system will correct the data based on the characteristics of neighboring stations when the deviation exceeds a certain threshold.

[0107]

[0108] Where, z′ i For the corrected data, z i For the original data, w ij It is the edge weight between site i and its neighboring site j. Let be the set of neighboring nodes of node i, and α be the correction coefficient.

[0109] S53: Establish an adaptive correction and feedback mechanism.

[0110] The core idea of ​​adaptive correction is to dynamically adjust the correction strategies and weights of each site based on real-time cloud field prediction errors. The feedback mechanism relies on the system's real-time monitoring of prediction errors. When the system's output cloud field prediction results deviate significantly from the actual observation results, the model will automatically feed back this error information and adjust the data fusion and correction strategies, assuming the prediction results... The error between the actual result y and the error δ is used by the system to adjust the weights of the data source using the following formula:

[0111]

[0112] in, As the weight for the next time step, γ represents the weight at the current time step, γ is the learning rate, and δ is the current prediction error.

[0113] S54: Perform multi-dimensional data collaborative optimization.

[0114] The system collaboratively optimizes multi-dimensional data, incorporating auxiliary information such as cloud images, meteorological data, radar data, wind speed, radiation, and terrain. During this multi-dimensional data optimization process, the system uses the following weighted model to combine data from different dimensions:

[0115]

[0116] Among them, Z 优化 For the optimized data, Z i For the i-th data source, w i Weights for each data source.

[0117] In the above technical solution, a preferred technical solution may also include step S6 further comprising:

[0118] After generating missing cloud maps using adversarial networks, the spatiotemporal consistency of the generated cloud maps is optimized.

[0119] Step S6 further includes the following steps:

[0120] S61: Identify and locate missing cloud map data.

[0121] First, the system identifies and locates missing parts in the input data. Using sky imagers and weather sensors, the system determines whether a cloud image at a given moment is missing or abnormal. If the cloud image data at a particular moment is significantly inconsistent with data from preceding and following time periods, the system marks that moment as "missing."

[0122] The identification of missing data is based on differences in image features. Suppose the cloud image data at a certain moment is I. t Compare the differences with the cloud map data at adjacent time points:

[0123] ΔI t =||I t -I t-1 ||2 (25);

[0124] Among them, I t and I t-1 These are the cloud map data for the current time and the previous time, respectively, ΔI t To represent the difference between images, by calculating the image difference, if ΔI t If the threshold is exceeded, the data at that moment is considered to be missing or abnormal.

[0125] S62: Supplement missing data based on Generative Adversarial Networks (GANs).

[0126] After identifying missing cloud map data, the system intelligently fills in the missing parts using a Generative Adversarial Network (GAN). In a GAN, the generator G aims to generate a new cloud map based on existing data, while the discriminator D determines whether the generated cloud map is a true cloud map. The loss function of the GAN is expressed by the following formula:

[0127]

[0128] Where, p data For the distribution of real cloud map data, p model For the distribution of data generated by the generator, I fake For the generated fake data, G(Ifake) is the cloud map generated by the generator, D is the discriminator, and the output is the probability of whether the image is real;

[0129] Through repeated training, the generator G learns to generate high-quality cloud maps based on the existing input data, while the discriminator DD ensures that the generated cloud maps are similar to the real cloud maps.

[0130] S63: Perform spatiotemporal consistency optimization.

[0131] When generating missing cloud map data, it is essential to consider the temporal and spatial consistency of the generated image. Specifically, it is crucial to ensure that the generated cloud map reflects the cloud field evolution at consecutive time points.

[0132] The system is designed with a spatiotemporal consistency optimization module. During the cloud map generation process, this module uses a graph neural network (GNN) to model the spatiotemporal dependency, ensuring that the generated cloud map is consistent with the cloud map at adjacent time points. In the spatiotemporal consistency optimization, the system uses cloud map information from previous and subsequent time points to extract features and transmits spatiotemporal information through the graph neural network, ensuring that the generated missing cloud map is consistent with the observation data of neighboring stations in space and synchronized with the dynamic evolution of the cloud field in time.

[0133] This optimization process is controlled by the following spatiotemporal consistency loss function:

[0134]

[0135] Among them, I i and I j These represent the pixels at positions i and j in the cloud map, respectively. For the neighborhood region, I spatial For spatial consistency loss, For temporal consistency loss; these loss functions ensure that the generated missing cloud map is consistent with the surrounding temporal and spatial clouds. Figure 1 To maintain the spatiotemporal continuity of the cloud field.

[0136] In the above technical solution, a preferred technical solution may also be that, in step S7, the output of the regional cloud field prediction result includes the output of cloud coverage rate, cloud cluster location and morphology information at different sites and different time points; step S7 specifically includes the following steps:

[0137] S71: Generate the cloud field prediction results for the generated region.

[0138] Post-processing of cloud field data includes, but is not limited to:

[0139] Spatial interpolation: Interpolation methods include bilinear interpolation, kriging interpolation, etc.

[0140] Time smoothing: The system uses a time smoothing algorithm to smooth the prediction results;

[0141] Assume the cloud field prediction data output by the spatiotemporal graph neural network is At a certain time t, after interpolation and smoothing, the final cloud field prediction result is Yt, which is expressed as:

[0142]

[0143] in, The original prediction results of the ST-GCN model are shown. Interp(·) is the interpolation operation, and Smooth(·) is the time smoothing operation. After interpolation and smoothing, Y is obtained. t This is the processed, spatially and temporally continuous cloud field prediction result.

[0144] S72: Establish a mechanism for predicting and triggering early warnings of cloud movement paths.

[0145] In predicting the movement path of a cloud cluster, the system uses cloud field data from the previous moment, combined with factors such as wind speed and airflow direction, to predict the next movement trend of the cloud cluster. Assume that the position of a cloud cluster at time t is P. t Based on wind speed and airflow information, the system can predict the position P of the cloud cluster at the next time t+1. t+1 It is calculated using the following formula:

[0146] P t+1 =P t +v t ·Δ t (31);

[0147] Among them, P t v represents the current position of the cloud cluster. t The current cloud movement speed (which can be calculated from wind speed) is represented by Δt, where Δt is the time interval.

[0148] Based on the cloud's movement path and predicted speed, the system generates warnings for different regions. If the system detects that a certain cloud will affect the sunlight conditions of a wind and solar power base in a short period, it will issue a warning a certain time before the cloud arrives. The warning mechanism is triggered by the following conditions:

[0149]

[0150] Among them, Time to Impact (P) t ,P target ) is the cloud cluster from its current position P. t To target area P target The estimated travel time, T waring It is the advance warning time (such as 30 minutes, 1 hour, etc.). When the warning triggering conditions are met, the system will output a warning signal, indicating that the wind and solar base may be affected by clouds, and suggest corresponding dispatching measures.

[0151] S73: Establish a scheduling and response mechanism for wind and solar power bases.

[0152] When the system issues an early warning signal, the wind and solar power base will schedule and respond according to the cloud field prediction results. The scheduling system will adjust the operating status of photovoltaic and wind power generation equipment in advance based on the expected movement path and prediction time of the cloud cluster. The scheduling system makes prediction and scheduling decisions using the following formula:

[0153] P adjusted =P forecasted ·(1-f cloud coverage (P target ,t)) (33);

[0154] Among them, P adjusted This is the adjusted power generation capacity, P. forecasted It is the predicted power generation capacity, f cloud coverage (P target ,t) is at the target position P target Above, the cloud coverage function changes with time t;

[0155] S74: Results visualization and report generation.

[0156] The system also provides a visual interface for cloud field prediction. Through the graphical interface, users can view cloud field distribution, cloud movement paths, expected affected areas, and early warning information at different time points. Simultaneously, the system automatically generates reports based on the prediction and early warning results for managers to further analyze and make decisions.

[0157] This invention selects and deploys sparsely distributed sky imager sites to acquire cloud field information over a large area, overcoming the limitations of a single site's observation field of view. After constructing multiple sky imager sites to collect data and performing preprocessing to ensure data accuracy and reliability, a spatiotemporal convolutional prediction model is fused through graph structure modeling and node feature extraction to perform regional cloud field collaborative prediction. Based on the above prediction model, a multi-source data fusion and dynamic correction method is employed to supplement the missing cloud map generated by the adversarial network. This allows for virtual completion based on multi-source data when data from a certain site is abnormal or missing, ensuring the continuity of input data and improving the realism and accuracy of regional cloud field collaborative prediction. This facilitates the use of cloud field collaborative prediction results to better guide the stable operation of wind and solar power generation systems. This invention can accurately reflect the dynamic evolution process of cloud fields, adapt to the prediction needs under complex meteorological conditions, and provide support for wind and solar power generation scheduling decisions.

[0158] In summary, this invention provides a regional cloud field collaborative prediction and correction method based on a multi-site sky imaging network. It can improve the realism and accuracy of regional cloud field collaborative prediction, thereby providing support for wind and solar power scheduling decisions and solving the current problem that cloud image prediction technology cannot be effectively used to improve the scheduling capabilities of wind and solar power plants. Attached Figure Description

[0159] Figure 1 This is a flowchart of the regional cloud field collaborative prediction and correction method based on a multi-site sky imaging network according to the present invention.

[0160] Figure 2 This is a flowchart illustrating the construction and deployment of the multi-site sky imaging network in this invention.

[0161] Figure 3 This is a flowchart of the multi-source observation data acquisition and preprocessing process in this invention.

[0162] Figure 4 This is a flowchart of the graph structure modeling and node feature extraction in this invention.

[0163] Figure 5 This is a flowchart of the spatiotemporal graph neural network prediction in this invention.

[0164] Figure 6 This is a flowchart of the multi-source observation data fusion and dynamic correction in this invention.

[0165] Figure 7 This is a flowchart illustrating the intelligent generation of missing cloud maps using adversarial networks in this invention.

[0166] Figure 8 This is a flowchart illustrating the output of regional cloud field prediction results and early warning information in this invention. Detailed Implementation

[0167] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. Based on these embodiments, all other embodiments obtained by those skilled in the art without creative effort are within the scope of this invention.

[0168] Example 1: As Figures 1 to 8 As shown, the regional cloud field collaborative prediction and correction method based on a multi-site sky imaging network of the present invention includes the following steps:

[0169] S1: Construction and deployment of a multi-site sky imaging network. This step involves selecting and deploying sparsely distributed network sites.

[0170] In this invention, the construction and deployment of a multi-site sky imaging network is the foundation of the entire system. By deploying multiple sky imager sites with collaborative functions, this invention can acquire cloud field information over a large area, overcoming the limitations of a single site's observation field of view. To achieve this goal, it is first necessary to scientifically select the locations of the deployment sites and consider the dynamic evolution characteristics of the cloud field.

[0171] See Figure 2 , Figure 2 A flowchart of the construction and deployment of a multi-site sky imaging network provided by an embodiment of the present invention is given. In step S1, the deployment sites are prioritized, with the sky imager site being prioritized. The site layout is dynamically optimized based on cloud hotspots, and different site networks constitute a low-power long-distance communication network. The method for constructing and deploying a multi-site sky imaging network includes the following steps:

[0172] S11: Develop a site selection strategy.

[0173] The first step in multi-site deployment is the scientific selection of sites. Site distribution must consider not only geographical location but also the regional meteorological environment and the dynamic evolution of cloud fields. Within a specific region, cloud movement speed, cloud thickness, and cloud morphology can vary significantly. Therefore, site placement should prioritize areas with more complex climate conditions and frequent cloud movement. Furthermore, site deployment must be supported by historical meteorological data. By analyzing historical cloud movement patterns, wind speed, and air pressure changes within the region, the optimal site layout can be comprehensively determined.

[0174] Specifically, the site deployment algorithm determines site priority using the following formula:

[0175]

[0176] Among them, P i For the deployment priority of the i-th site, d ij v is the distance between station i and station j. ij For wind speed differences, Δt is the time interval for data collection, α, β, and γ are the weighting coefficients for distance, wind speed, and time interval, respectively, and N is the number of all candidate stations.

[0177] This formula allows for the prioritization of candidate sites, prioritizing deployment in areas with more drastic cloud field changes and greater wind speed variations.

[0178] S12: Perform inter-site communication and data transmission.

[0179] Once the site layout is determined, the next step is to ensure efficient data exchange between these sites. To achieve real-time communication between multiple sites, a low-power, long-distance communication protocol was designed. This protocol can adjust the data transmission frequency according to different weather conditions, accelerating data updates during periods of rapid cloud changes or high communication demand, and reducing the frequency when cloud conditions are stable or data changes are minimal, thereby reducing energy consumption and improving system efficiency.

[0180] Inter-site communication also needs to consider transmission latency and bandwidth limitations, thus requiring a suitable communication scheduling strategy. This strategy can be adaptively adjusted using the following optimization formula:

[0181]

[0182] Among them, C ij d represents the communication quality between station i and station j. ij L represents the distance between stations. ij SNR is the length of the data packet. ij This represents the signal-to-noise ratio, with α and β being weighting coefficients. Real-time monitoring and adjustment of inter-site communication quality ensures the stability and reliability of data transmission.

[0183] S13: Perform site layout and dynamic optimization.

[0184] In addition to the initial deployment plan, this plan also incorporates a dynamic optimization mechanism to optimize the site layout based on real-time changes in the cloud field. This optimization mechanism, based on real-time cloud field change data and combined with meteorological conditions such as wind speed and temperature, determines the current direction and speed of cloud movement and adjusts the operating frequency and data acquisition interval between sites. If cloud formations in a certain area become more intense, the system dynamically increases the density of observation sites in that area to acquire more cloud data. This optimization process can be achieved through the following dynamic adjustment formula:

[0185]

[0186] Wherein, ΔP i This is the optimization adjustment amount for site i, Δd ij and Δv ij These represent the changes in distance from other stations and the changes in wind speed, respectively, with γ and δ being dynamically adjusted weighting coefficients.

[0187] This dynamic optimization mechanism can intelligently adjust the deployment of stations based on rapid changes in cloud formations or fluctuations in meteorological factors, ensuring the efficiency and accuracy of cloud field observation data.

[0188] S2: Multi-source observation data acquisition and preprocessing. This step acquires sky imaging images from multiple stations and simultaneously obtains and preprocesses multi-source observation data. In step S2, the multi-source observation data includes data collected by sky imagers, weather stations, radar, etc. The data preprocessing method involves noise filtering, spatiotemporal alignment, and illumination condition correction of the raw data after acquisition.

[0189] Building upon the multi-site sky imaging network, the next step is the acquisition and preprocessing of multi-source data. The goal of this process is to ensure the consistency of data acquired from all stations and to provide high-quality input features for subsequent model training and prediction. Since the observational data comes from multiple sensors and devices, such as sky imagers, weather stations, and radar, effective fusion and preprocessing of this data are essential to ensure its accuracy and reliability. See also... Figure 3 , Figure 3 A flowchart of multi-source observation data acquisition and preprocessing provided by an embodiment of the present invention is given. The method for multi-source observation data acquisition and preprocessing includes the following steps:

[0190] S21: Data Acquisition.

[0191] Data acquisition is the first step in the entire system, involving the collection of cloud field information and related meteorological data from different stations. Each station's sky imager captures panoramic sky images through image acquisition equipment; these images form the basis for subsequent cloud field prediction models. In addition, each station collects meteorological information such as wind speed, temperature, humidity, and air pressure, which provides additional physical contextual information for subsequent cloud field evolution.

[0192] During data acquisition, it is crucial to ensure synchronization between sensors to avoid data inconsistencies caused by time errors. Clock synchronization of all devices via Network Time Protocol (NTP) ensures that data collected from each site has a consistent time stamp, which is essential for subsequent time-series analysis.

[0193] The data acquisition frequency at each station is adaptively adjusted based on meteorological conditions and the rate of cloud change. For example, areas with rapidly changing cloud cover require a higher acquisition frequency to capture the dynamic evolution of cloud formations in real time. The data acquisition frequency can be determined using the following formula:

[0194] f 采集 =f 基准 ·(1+α·ΔC) (4)

[0195] Among them, f 采集 f is the data collection frequency of the site. 基准 The base acquisition frequency is given by α, an adjustment coefficient is given by ΔC, and ΔC represents the rate of change of the current cloud field. Using this formula, the system can adjust the data acquisition frequency according to the drastic changes in the cloud field.

[0196] S22: Data preprocessing.

[0197] Data preprocessing is the process of cleaning and correcting various types of raw observation data. Its purpose is to remove noise, fill in missing values, and convert data from different sources into a unified format for subsequent processing. Image data acquired by sky imagers often contains noise due to equipment defects, lighting issues, or the complexity of cloud features. To reduce the impact of noise, image denoising techniques are employed, such as image denoising models based on deep convolutional neural networks (CNNs).

[0198] In meteorological data, due to sensor accuracy and environmental interference, missing or outlier values ​​may appear. To address these issues, interpolation methods are used to fill in missing data, or data from adjacent time periods is used for smoothing. For example, a Kalman filter can be used to smooth continuous meteorological data and reduce the impact of noise. The Kalman filter formula is as follows:

[0199]

[0200] P k =(IK k H)·P k-1 (6)

[0201] Among them, P k For the estimated state, K k For Kalman gain, P k-1 Let be the observed values, H be the observation matrix, be the estimation error covariance, and I be the identity matrix. Kalman filtering can effectively reduce noise in meteorological data and improve data accuracy.

[0202] S23: Perform spatiotemporal alignment and standardization on the data.

[0203] Data collected from different stations not only differ in time but also exhibit spatial variations due to geographical location. To ensure that data from different stations can be processed within a unified framework, spatiotemporal alignment and standardization are necessary. The purpose of spatiotemporal alignment is to ensure that observational data from different stations accurately match the same moment and reflect meteorological differences between different geographical locations.

[0204] Spatiotemporal alignment can be achieved through interpolation. Assuming that stations i and j have different observations at the same time point, linear interpolation can be used to calculate the spatiotemporal data of the two stations:

[0205]

[0206] Among them, z 对齐 For the aligned data, z i and z j The observations for stations i and j are respectively, t i and t j This refers to the time point in time when the data was collected.

[0207] Standardization is the process of unifying the dimensions of data collected from different sites. Its purpose is to ensure that various types of data have the same scale, enabling effective data fusion during subsequent model training. The standardization process can be achieved using the mean-variance standardization formula:

[0208]

[0209] Where z represents the original data, μ represents the mean of the data, σ represents the standard deviation of the data, and z′ represents the standardized data. This standardization method unifies meteorological data from different stations to the same units, making subsequent model processing more efficient.

[0210] S24: Merge the data.

[0211] To comprehensively capture the evolution of cloud fields and the influence of meteorological conditions, data from different sensors needs to be fused. The goal of fusion is to integrate cloud image data from sky imagers with environmental data from weather stations, providing richer and more multi-dimensional information for subsequent cloud field modeling. Data fusion techniques can employ methods such as weighted averaging and Kalman filtering. For data from multiple sensors, a weighted averaging method can be used for fusion.

[0212]

[0213] Among them, z 融合 For the merged data, w i The weights for each data source, z i Here, N represents the observations from each sensor, and N is the number of data sources.

[0214] Finally, after the above preprocessing and fusion steps, the obtained data will become the input data for subsequent graph neural network (GNN) modeling, ensuring that the model can be trained and predicted using high-quality data.

[0215] S3: Graph structure modeling and node feature extraction. This step treats the sky imaging network sites as graph nodes and extracts node features.

[0216] In step S3, the graph structure modeling includes converting the multi-source observation data of the station into node features in the graph network and modeling it using a graph neural network (GNN).

[0217] Building upon the multi-site sky imaging network and multi-source data fusion, the next step is graph structure modeling and node feature extraction. The goal of this process is to transform the observation data from each site into node features within a graph network, and then model the cloud field using a graph neural network (GNN). By representing the sites and their relationships as a graph structure, spatial interactions and spatiotemporal variations can be captured in cloud field modeling, thereby improving the accuracy and timeliness of predictions. See also Figure 4 , Figure 4 A flowchart of graph structure modeling and node feature extraction provided by an embodiment of the present invention is given. The method for graph structure modeling and node feature extraction includes the following steps:

[0218] S31: Node construction and feature extraction.

[0219] First, the system treats each station as a node in a graph network. Each node's features include local cloud imagery information and other meteorological features such as wind speed, temperature, and humidity. These features provide local observation data for the station and serve as input data for subsequent graph neural network modeling.

[0220] Cloud image data consists of panoramic images acquired by sky imagers. After image processing and preprocessing, these images are transformed into features usable by GNN models. For example, the color, brightness, and corresponding cloud classification information (such as cumulus, stratus, cirrus, etc.) of each pixel in the image can all serve as input features for nodes. These image features are typically extracted using deep convolutional neural networks (CNNs). Suppose the image features of a certain site are F... i Then the node features can be represented as:

[0221] F i =CNN(I i (10)

[0222] Among them, I iThis represents the sky image acquired at site i. CNN stands for Convolutional Neural Network, used to extract cloud information from the image. Through the convolutional network, the features of each region in the image are encoded into a vector representation, which serves as the feature of the node.

[0223] Besides image features, meteorological data such as wind speed, temperature, and humidity can also be used as node features input to the GNN. For example, assuming the wind speed is v... i The temperature is t i Then the meteorological characteristics of station i are:

[0224] X i =[v i ,t i ,h i ,p i (11);

[0225] Among them, h i For humidity, p i This refers to air pressure. These meteorological features, together with cloud image features, form the comprehensive feature vector F of the node. i .

[0226] S32: Construction of graph structures.

[0227] Once the node features are extracted, the next step is to construct the graph structure. The core idea of ​​graph neural networks is to transmit information through the connections (edges) between nodes. In this invention, each node in the graph represents an observation station, and the weights of the edges are dynamically calculated based on the spatial relationships between stations and meteorological conditions.

[0228] Edge connections between stations represent the physical or meteorological correlations between them. The weights of the edges reflect the similarity between stations and their impact on cloud field evolution. Assume the distance between station i and station j is d. ij And the wind speed difference between the two is v ij Then the edge weight w ij It can be represented as:

[0229]

[0230] Where α and β are the weights of the adjustment factor, l scale and v scale These are the scale factors for distance and wind speed, respectively, ensuring that the calculated weights have appropriate units and ranges. This formula combines spatial distance and meteorological factors, allowing the edge weights to accurately reflect the relationships between stations.

[0231] Besides distance and wind speed, other meteorological factors (such as air pressure and temperature differences) can also serve as supplementary features for edge weights. For example, air pressure difference Δp ij It can also be used to adjust the weight of edges:

[0232]

[0233] In this way, the edge weights not only take into account spatial and meteorological factors, but can also be dynamically adjusted under different environmental conditions, providing more accurate weights for subsequent message transmission mechanisms.

[0234] S33: Spatiotemporal Feature Modeling and Information Transmission.

[0235] After the graph structure is built, the graph neural network will begin modeling the cloud field. Each node's features interact with other nodes through the graph neural network's message passing mechanism. Through graph convolution operations, each node updates its features based on information from its neighbors, thereby learning more comprehensive cloud field features.

[0236] Graph Convolutional Networks (GCNs) update node features using the following formula:

[0237]

[0238] in, This represents the feature of node i in the (k+1)th layer. Let w be the set of neighboring nodes of node i. ij It is the edge weight. Let be the feature of node j at layer k, b be the bias term, and σ be the activation function. By iteratively updating node features, GNN can capture the spatial correlation and dynamic evolution of cloud fields in multi-level graph structures.

[0239] Furthermore, to enhance the modeling capability of spatiotemporal features, a spatiotemporal convolutional network (ST-CNN) can be incorporated, enabling the dynamic evolution of cloud fields to be modeled not only spatially but also temporally, capturing the movement trends of cloud clusters. Through multi-scale spatiotemporal convolution, changes in cloud fields at different time scales can be modeled, providing more refined features for subsequent prediction and early warning.

[0240] S4: Spatiotemporal Graph Neural Network Prediction. This step involves constructing a model that integrates spatiotemporal convolutional neural networks (ST-CNN) and graph neural networks (GNN) for prediction.

[0241] Step S4 also includes: after constructing a model that integrates spatiotemporal convolutional neural networks (ST-CNN) and graph neural networks (GNN) for prediction, sharing cross-site information through a message passing mechanism to achieve real-time early warning; the real-time early warning is given to downstream sites in advance by predicting the movement of strong convective cloud clusters or large-scale cloud clusters.

[0242] After constructing the graph structure and extracting node features, the next crucial step is to perform spatiotemporal prediction of cloud fields using a Graph Neural Network (GNN). The goal of a spatiotemporal graph neural network is to capture the spatial distribution and temporal evolution of cloud fields based on observation data from multiple sites. It learns the dependencies between nodes through a message-passing mechanism and generates future cloud field predictions. This process considers not only the spatial distribution of clouds but also their dynamic changes over time, enabling a comprehensive prediction of the cloud field's evolution. See also... Figure 5 , Figure 5 A flowchart of spatiotemporal graph neural network prediction provided by an embodiment of the present invention is given. The method of spatiotemporal graph neural network prediction includes the following steps:

[0243] S41: Construction of spatiotemporal graph convolutional networks.

[0244] Spatiotemporal Graph Convolutional Networks (ST-GCNs) are models that combine graph neural networks and spatiotemporal convolution, enabling them to capture information in both spatial and temporal dimensions. In graph neural networks, graph convolution operations transmit information through the adjacency relationships of nodes. However, in spatiotemporal modeling, it is necessary to consider not only the spatial dependencies between nodes but also the dynamics of cloud formations that change over time. Therefore, ST-GCNs extend traditional graph convolutional networks by incorporating spatiotemporal convolutional layers, enabling them to process time-series data and transmit information in the spatial dimension.

[0245] In graph convolutional networks, the node feature update of each layer can be represented by the following formula:

[0246]

[0247] in, w is the feature of node i at layer k+1. ij It is the edge weight. It is the feature of neighbor node j at the k-th layer. Let be the set of neighbors of node i, σ be the activation function, and b be the bias term. This formula represents updating the features of the current node by aggregating information from neighboring nodes.

[0248] In spatiotemporal graph neural networks, in addition to traditional graph convolution operations, temporal convolution operations are introduced. It is assumed that the features of each node depend not only on its spatial neighbors but also on its own feature information at different points in time. Spatiotemporal convolutional networks capture the evolution of clouds over time by introducing temporal convolution on top of graph convolution. The spatiotemporal convolution update formula is:

[0249]

[0250] Where T is the time dimension, λ tThese are the weights for each time step, representing the contribution of each time step to the node's features. In this way, the spatiotemporal graph neural network can simultaneously consider the spatial neighbor information and the dynamic information of historical moments of a node in each layer of graph convolution, thereby more comprehensively modeling the evolution of the cloud field.

[0251] S42: Joint learning of graph convolution and time dependence.

[0252] The key to spatiotemporal graph neural networks lies in balancing graph convolution with temporal dependency learning. To effectively capture the dynamic evolution of cloud fields, each layer of the network must not only model spatial relationships based on graph convolution but also incorporate temporal factors to reflect the temporal characteristics of cloud movement and deformation. By introducing a weighting mechanism based on time steps, the network can automatically focus on key moments of cloud changes during the learning process, thereby enhancing its predictive ability for future cloud field evolution.

[0253] To effectively model spatiotemporal dependencies, a spatiotemporally weighted graph convolution mechanism can be designed, where each time step has a different influence on the update of the current node. For example, for a node i, its update process not only considers the information of its neighbors but also dynamically adjusts its weights based on the features of historical moments. This process can be represented by the following formula:

[0254]

[0255] Where, λ t The weighting mechanism dynamically adjusts the model based on the degree of influence of each time step on the node features. This mechanism enables the model to automatically learn the time steps that have a greater impact on cloud evolution at certain times, thereby improving the accuracy of spatiotemporal modeling.

[0256] S43: Perform spatiotemporal prediction and cloud field evolution.

[0257] The ultimate goal of a spatiotemporal graph neural network is to generate predictions of cloud fields at future times by updating the features of the current nodes. The prediction process relies not only on the current cloud image data but also on cloud field information from the previous time step and its neighboring sites. After passing through multiple graph convolutional layers and spatiotemporal convolutional layers, the network outputs predictions of the cloud field at future times. These predictions typically include cloud coverage, cloud movement paths, and cloud morphological changes for a future period.

[0258] To ensure the accuracy of the predictions, the network output can be post-processed to transform it into usable cloud field information. For example, interpolation can be applied to the predictions to fill in any missing regions, or data smoothing can be used to reduce the volatility of the predictions.

[0259] The output prediction result of the spatiotemporal graph neural network can be expressed by the following formula:

[0260]

[0261] in, This is the cloud field prediction result at a future time, where X is the input feature of the node, A is the adjacency matrix of the graph, T is the time step, and f is the value of f. ST-GCN It is the prediction function of the spatiotemporal graph neural network model.

[0262] S44: Establish an early warning mechanism and make real-time adjustments.

[0263] Building upon cloud field prediction, this invention also includes a real-time early warning mechanism. When the system predicts strong convective cloud clusters or large-scale cloud cluster movement, it can issue early warnings to downstream sites. Through the prediction results of the spatiotemporal graph neural network, the system can determine whether a cloud cluster may affect wind and solar power generation in the downstream area at a certain moment and take measures in advance, such as adjusting the operating status of equipment and carrying out preventive scheduling.

[0264] Specifically, when the forecast results indicate the presence of large-scale cloud clusters or high-intensity convective weather at a certain point in time, the system can generate an early warning signal based on the confidence level of the forecast results and issue an early warning to the wind and solar bases in the relevant areas to help them make corresponding preparations.

[0265] S5: Multi-source observation data fusion and dynamic correction. This step fuses multi-source observation data and performs dynamic correction and virtual completion when data is abnormal or missing.

[0266] In step S5, the multi-source observation data fusion and dynamic correction includes using weighted average, Kalman filtering, or Bayesian inference methods to correct inconsistencies between data.

[0267] Building upon the spatiotemporal graph neural network prediction model, the next crucial step is multi-source observation data fusion and dynamic correction. This process aims to further improve the accuracy and reliability of cloud field prediction by effectively fusing data from different sensors and correcting for potential data anomalies or missing data in real time. Multi-source observation data fusion enables the system to integrate real-time information from multiple data sources, such as sky imagers, weather stations, and radar, and corrects inconsistencies between data through intelligent algorithms, ultimately improving prediction accuracy.

[0268] See Figure 6 , Figure 6 A flowchart of the multi-source observation data fusion and dynamic correction method provided in this embodiment of the invention is given. The method of multi-source observation data fusion and dynamic correction includes the following steps:

[0269] S51: Employ a multi-source data fusion method.

[0270] The purpose of multi-source data fusion is to combine observational data from different sensors, leveraging the complementarity of different data sources to compensate for their respective deficiencies. The input data for a cloud field prediction system includes cloud imagery data captured by a sky imager, meteorological parameters such as wind speed, temperature, and humidity provided by weather stations, and radar data. The fusion of these data can be achieved using weighted averaging, Kalman filtering, or Bayesian inference methods.

[0271] In this invention, data from various data sources are first preprocessed and standardized to ensure that different data types have the same dimensions and scale. Then, Kalman filtering is used for data fusion. Kalman filtering is a recursive estimation method that updates the state estimate based on the previous prediction and the current observation, calculating an optimal estimate. The Kalman filtering update formula is as follows:

[0272]

[0273] P k =(IK k H)·P k-1 (20);

[0274] in, The estimated state at the current moment, z k For the current observation data, K k Let H be the Kalman gain, H be the observation matrix, and P be the Kalman gain. k To estimate the error covariance, I is the identity matrix. Through Kalman filtering, the system can achieve optimal fusion of various observation data, providing a unified state estimate.

[0275] Different weighting methods can be used to fuse different types of data sources. For example, image data from a sky imager and temperature and humidity data from a weather station can be fused using a weighted average method.

[0276]

[0277] Among them, z 融合 It is the merged data, w i The weights for each data source, z i For each sensor's observations, N is the number of data sources. Weight w i It can be dynamically adjusted based on the importance or precision of the data source.

[0278] S52: Establish a dynamic correction mechanism.

[0279] Due to weather conditions and equipment errors, data from various stations and sensors may contain inconsistencies, missing values, or outliers. To ensure data quality, the system needs to perform dynamic correction in real time. The goal of the dynamic correction mechanism is to detect and correct outliers in the data in real time, ensuring the consistency of data from each station throughout the entire prediction model.

[0280] The correction mechanism includes two aspects: imputing missing data and correcting outlier data. For handling missing data, interpolation or imputation methods based on neighboring station data can be used. For example, if sky image data from a certain station is missing, the system can use observation data from surrounding stations and supplement it through the message passing mechanism of a graph neural network to generate predicted values ​​for the missing data.

[0281] Outlier correction can be achieved through identification and processing based on statistical models. Suppose a cloud map data z for a certain site... i If an abnormal deviation occurs at a certain moment, the system can correct it using historical data models or observation data from neighboring stations. For example, by calculating the similarity of the cloud map between the current station and neighboring stations, when the deviation exceeds a certain threshold, the system will correct the data based on the characteristics of neighboring stations.

[0282]

[0283] Where, z′ i For the corrected data, z i For the original data, w ij It is the edge weight between site i and its neighboring site j. Let be the set of neighboring nodes of node i, and α be the correction coefficient. In this way, the system can dynamically correct based on data from neighboring sites, ensuring the consistency and accuracy of the cloud map data.

[0284] S53: Establish an adaptive correction and feedback mechanism.

[0285] To further improve the accuracy of data correction, the system also incorporates an adaptive correction and feedback mechanism. The core idea of ​​adaptive correction is to dynamically adjust the correction strategies and weights for each site based on real-time cloud field prediction errors. For example, if the system detects a large prediction error at a particular site, it may indicate a significant bias in the site's observation data. In this case, the system can automatically increase the correction weight for that site, enhancing its influence on the overall model.

[0286] The feedback mechanism relies on the system's real-time monitoring of prediction errors. When the system's cloud field prediction deviates significantly from the actual observations, the model will automatically feed back this error information and adjust the data fusion and correction strategies. For example, assuming the prediction result... The error between the actual result y and the error δ can be adjusted by the system using the following formula:

[0287]

[0288] in, As the weight for the next time step, Let γ be the weight at the current moment, γ be the learning rate, and δ be the current prediction error. In this way, the system can adaptively adjust the weights of each data source according to the feedback mechanism, thereby improving the accuracy of data fusion and correction.

[0289] S54: Perform multi-dimensional data collaborative optimization.

[0290] Finally, to further improve the accuracy of cloud field prediction, the system collaboratively optimizes multi-dimensional data. In addition to cloud images, meteorological data, and radar data, more auxiliary information such as wind speed, radiation, and topography can be incorporated. This multi-dimensional data collaborative optimization method combines the advantages of all data sources, comprehensively improving the accuracy of cloud field prediction by optimizing model parameters.

[0291] In the process of collaborative optimization of multi-dimensional data, the system can combine data from different dimensions using the following weighted model:

[0292]

[0293] Among them, Z 优化 For the optimized data, Z i For the i-th data source, w i The weights for each data source are assigned. By dynamically adjusting the weights of each data source, the system can effectively improve the model's performance under different weather conditions.

[0294] S6: Uses adversarial networks to intelligently generate missing cloud maps.

[0295] Step S6 also includes: optimizing the spatiotemporal consistency of the generated cloud map after generating the missing cloud map using an adversarial network.

[0296] In cloud field prediction systems, data may be missing or incomplete due to equipment malfunctions, severe weather, or other external factors. To ensure system stability and accuracy, an effective mechanism for generating and supplementing missing cloud images must be designed. The goal of this step is to use intelligent algorithms to generate missing cloud image data using existing cloud image data and relevant meteorological information, ensuring the prediction model can continue to function normally and improving prediction reliability. See also Figure 7 , Figure 7 A flowchart of the intelligent generation of missing cloud maps provided in this embodiment of the invention is given. The method for intelligent generation of missing cloud maps includes the following steps:

[0297] S61: Identify and locate missing cloud map data.

[0298] First, the system needs to identify and locate missing portions of the input data. In sky imagers and weather sensors, cloud image data is typically provided as image sequences, and the system must be able to determine whether a cloud image at a given moment is missing or exhibits anomalies. This can be achieved by comparing cloud image data from adjacent moments. For example, if the cloud image data at a particular moment is significantly inconsistent with data from preceding and following time periods, the system can mark that moment as "missing."

[0299] The identification of missing data can be based on differences in image features. For example, suppose the cloud image data at a certain moment is I. t Compare the differences with the cloud map data at adjacent time points:

[0300] ΔI t =||I t -I t-1 ||2 (25);

[0301] Among them, I t and I t-1 These are the cloud map data for the current time and the previous time, respectively, ΔI t This represents the difference between images. By calculating the image difference, if ΔI... t If the threshold is exceeded, the data at that moment is considered to be missing or abnormal.

[0302] S62: Supplement missing data based on Generative Adversarial Networks (GANs).

[0303] After identifying missing cloud image data, the system intelligently fills in the missing parts using a Generative Adversarial Network (GAN). A GAN is a deep learning method that generates high-quality samples through an adversarial process between a generator and a discriminator. For cloud image data completion, the generator aims to generate the missing parts based on existing cloud image data and meteorological information, while the discriminator aims to determine the realism of the generated cloud image.

[0304] In GANs, the generator G aims to generate new cloud maps based on existing data, while the discriminator D determines whether the generated cloud map is a true cloud map. The loss function of a generative adversarial network can be expressed by the following formula:

[0305]

[0306] Where, p data For the distribution of real cloud map data, p model For the distribution of data generated by the generator, I fakeFor the generated fake data, G(Ifake) is the cloud map generated by the generator, and D is the discriminator, which outputs the probability of whether the image is real or not.

[0307] Through repeated training, the generator G learns to generate high-quality cloud maps based on the existing input data, while the discriminator DD ensures that the generated cloud maps are as similar as possible to the real cloud maps.

[0308] S63: Perform spatiotemporal consistency optimization.

[0309] When generating missing cloud image data, in addition to the realism of the image itself, the consistency of the generated image with time and space must also be considered. Cloud fields have obvious spatiotemporal dependencies; the cloud image at a certain moment depends not only on the current meteorological conditions but also on the cloud image information of the preceding and following moments. Therefore, when generating missing data, it is essential to ensure that the generated cloud image is consistent with the cloud field evolution of the preceding and following moments.

[0310] To address this, the system incorporates a spatiotemporal consistency optimization module. This module utilizes a graph neural network (GNN) to model spatiotemporal dependencies during cloud image generation, ensuring that the generated cloud image remains consistent with those at adjacent time points. In the spatiotemporal consistency optimization process, the system extracts features from cloud image information from consecutive time points and transmits spatiotemporal information via the graph neural network, ensuring that the generated missing cloud image is spatially consistent with observation data from neighboring stations and temporally synchronized with the dynamic evolution of the cloud field.

[0311] This optimization process can be controlled using the following spatiotemporal consistency loss function:

[0312]

[0313]

[0314] Among them, I i and I j These represent the pixels at positions i and j in the cloud map, respectively. For the neighborhood region, L spatial For spatial consistency loss, For temporal consistency loss. These loss functions ensure that the generated missing cloud map is consistent with the surrounding temporal and spatial clouds. Figure 1 To maintain the spatiotemporal continuity of the cloud field.

[0315] S64: Perform supplementary optimization based on meteorological data.

[0316] Besides Generative Adversarial Networks (GANs) and spatiotemporal consistency optimization, meteorological data (such as wind speed, temperature, and humidity) can also serve as supplementary information for generating missing cloud maps. For example, clouds may move faster when wind speeds are high, so the generated cloud map should take wind speed into account. Similarly, changes in temperature and humidity affect cloud formation and dissipation, so these meteorological factors should also be taken into account.

[0317] In this step, the missing parts are further optimized by combining meteorological data with the generated cloud map. Assume the cloud map generated by the cloud map generator G is... The corresponding meteorological data is X. t (Such as wind speed, humidity, etc.), the supplemented cloud map can be represented as:

[0318]

[0319] in, To combine the optimized missing cloud map with meteorological data, X t This provides the meteorological data for the current moment. In this way, the system can optimize the generation of cloud maps by incorporating meteorological factors, thereby improving the accuracy and relevance of the supplementary cloud maps.

[0320] S7: Output Regional Cloud Field Prediction Results and Early Warning Information. This step outputs the short-term regional cloud field prediction results and automatically generates early warning information. In step S7, the output regional cloud field prediction results include cloud coverage, cloud cluster location, and morphology information at different sites and time points. Based on the cloud field prediction and supplementation mechanism, the ultimate goal is to output the regional cloud field prediction results and provide timely weather change warnings through the early warning system. This is especially important when large-scale cloud clusters or severe convective weather are about to affect downstream areas, providing effective scheduling basis for wind and solar power bases. The key to this step is to output cloud field information that can be used for practical applications based on the cloud field prediction data generated by the Spatiotemporal Graph Neural Network (ST-GCN) through appropriate post-processing, and to help wind and solar power bases respond to potential weather risks in a timely manner through the early warning mechanism.

[0321] See Figure 8 , Figure 8 A flowchart of the output area cloud field prediction results and early warning information provided in the embodiments of the present invention is given.

[0322] S71: Generate the cloud field prediction results for the generated region.

[0323] After the spatiotemporal graph neural network predicts the cloud field, the generated cloud field data is typically a multidimensional tensor containing information such as cloud coverage, cloud cluster location, and morphology at different sites and time points. These prediction results can be directly used to generate cloud field distribution maps for a certain future time range, enabling wind and solar power generation bases to assess weather impacts and schedule equipment.

[0324] To transform these predictions into an easily understandable and operational form, the system first needs to post-process the cloud field data. This processing includes, but is not limited to:

[0325] Spatial interpolation: Since some sites may not cover the entire area, spatial interpolation is needed to fill in the blank areas within the prediction results. Commonly used interpolation methods include bilinear interpolation and kriging interpolation.

[0326] Time smoothing: In order to reduce the volatility of predictions, especially the large errors that occur on short time scales, the system can use time smoothing algorithms to smooth the prediction results and ensure the stability of cloud field predictions.

[0327] Assume the cloud field prediction data output by the spatiotemporal graph neural network is At a certain time t, after interpolation and smoothing, the final cloud field prediction result is Y. t , can be represented as:

[0328]

[0329] in, This is the original prediction result of the ST-GCN model. Interp(·) is the interpolation operation, and Smooth(·) is the time smoothing operation. After interpolation and smoothing, the resulting Y... t This is the processed, spatially and temporally continuous cloud field prediction result.

[0330] S72: Establish a mechanism for predicting and triggering early warnings of cloud movement paths.

[0331] The movement path and speed of cloud clusters are key factors affecting wind and solar power generation, making accurate prediction and early warning of cloud clusters crucial. Based on cloud field prediction, the system analyzes factors such as the historical trajectory, movement speed, and wind field of cloud clusters to predict their future movement paths and generate early warning signals.

[0332] In predicting the movement path of cloud clusters, the system uses cloud field data from the previous moment, combined with factors such as wind speed and airflow direction, to predict the next movement trend of the cloud cluster. For example, suppose a cloud cluster is located at point P at time t. t Based on wind speed and airflow information, the system can predict the position P of the cloud cluster at the next time t+1. t+1It is calculated using the following formula:

[0333] P t+1 =P t +v t ·Δt (31);

[0334] Among them, P t v represents the current position of the cloud cluster. t Δt represents the current cloud movement speed (which can be calculated from wind speed), and Δt represents the time interval.

[0335] Based on the cloud's movement path and predicted speed, the system can generate early warnings for different regions. For example, if the system detects that a certain cloud will affect the sunlight conditions of a wind and solar power base in the short term, it will issue an early warning a certain time before the cloud arrives. The early warning mechanism is triggered by the following conditions:

[0336]

[0337] Among them, Time to Impact (P) t P target ) is the cloud cluster from its current position P. t To target area P target The estimated travel time, T warning This refers to the advance warning time (e.g., 30 minutes, 1 hour, etc.). When the warning trigger conditions are met, the system will output a warning signal, indicating that the wind and solar base may be affected by clouds, and suggest corresponding dispatching measures.

[0338] S73: Establish a scheduling and response mechanism for wind and solar power bases.

[0339] When the system issues an early warning signal, the wind and solar power base can schedule and respond according to the cloud field forecast results. The scheduling system will adjust the operating status of photovoltaic and wind power generation equipment in advance based on the expected movement path and predicted time of the cloud cluster. For example, when a cloud cluster is about to cover the sunlit area of ​​a wind and solar power base, the photovoltaic system can reduce its output power in advance to avoid overload due to insufficient sunlight; while the wind power generation system can make appropriate power adjustments or equipment protection based on wind speed and the arrival time of the cloud cluster.

[0340] The scheduling system can make predictions and scheduling decisions using the following formula:

[0341] P adjusted =P forecasted ·(1-f cloud coverage (P target ,t)) (33);

[0342] Among them, P adjusted This is the adjusted power generation capacity, P.forecasted It is the predicted power generation capacity, f cloud coverage (P target ,t) is at the target position R target The cloud cover rate function changes with time t. This formula adjusts the power generation based on changes in cloud cover rate, thereby improving the stability and reliability of the system.

[0343] S74: Results visualization and report generation.

[0344] To facilitate viewing of cloud field forecast results and early warning information by managers and operators, the system also provides a visual interface for cloud field forecasting. Through this graphical interface, users can view cloud field distribution, cloud movement paths, expected affected areas, and early warning information at different time points. Simultaneously, the system automatically generates reports based on the forecast and early warning results for managers to further analyze and make decisions.

[0345] Visualization tools allow managers to obtain real-time information on the current cloud field status, understand cloud movement, and anticipate changes in sunlight and wind speed over the next few hours. This information helps wind and solar power bases optimize their operational strategies and reduce losses caused by cloud changes.

[0346] It should be noted that, in this application, the terms "comprising" and "including" do not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the stated elements. Furthermore, the parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of corresponding technical solutions in the prior art have not been described in detail to avoid excessive elaboration.

[0347] In summary, this invention provides a regional cloud field collaborative prediction and correction method based on a multi-site sky imaging network. It can improve the realism and accuracy of regional cloud field collaborative prediction, thereby providing support for wind and solar power scheduling decisions and solving the current problem that cloud image prediction technology cannot be effectively used to improve the scheduling capabilities of wind and solar power plants.

Claims

1. A regional cloud field cooperative prediction and correction method based on a multi-site sky imaging network, characterized in that, It comprises the following steps: S1: construction and deployment of multi-site sky imaging network, this step selects and deploys a sparsely distributed network site; S2: multi-source observation data acquisition and preprocessing, this step collects multi-site sky imaging images, simultaneously acquires multi-source observation data and performs data preprocessing; S3: graph structure modeling and node feature extraction, this step regards the sky imaging network site as a graph node for node feature extraction; S4: spatio-temporal graph neural network prediction, this step constructs a model combining spatio-temporal convolution (ST-CNN) and graph neural network (GNN) for prediction; S5: multi-source observation data fusion and dynamic correction, this step fuses multi-source observation data, and performs dynamic correction and virtual completion when data is abnormal or missing; S6: intelligent generation of missing cloud images using adversarial networks; S7: output of regional cloud field prediction results and warning information, this step outputs future short-time scale regional cloud field prediction results and automatically generates warning information.

2. The regional cloud field cooperative prediction and correction method based on multi-site sky imaging network according to claim 1, characterized in that, In step S1, the deployment of network sites determines the priority, determines that the sky imager site is the priority, dynamically optimizes the site layout according to the cloud field hot spot, and different site networks constitute a low-power long-distance communication network; step S1 comprises the following steps: S11: formulate the site selection strategy, The distribution of the site should consider the geographical location, the regional meteorological environment and the dynamic evolution characteristics of the cloud field, the layout of the site should give priority to those regions with complex climate conditions and frequent cloud movement, and the best site layout is comprehensively judged by analyzing the historical movement mode of the cloud in the region, the wind speed and the change factors of air pressure; the site deployment algorithm determines the priority of the site through the following formula: where P i is the deployment priority of the i-th site, d ij is the distance between site i and site j, v ij is the wind speed difference, Δt is the time interval of data collection, and α, β, γ are the weighting coefficients of distance, wind speed, and time interval, respectively, and N is the number of all candidate sites. Through the formula, the priority of each candidate site is sorted, and the site is preferentially deployed in the region where the cloud field changes more dramatically and the wind speed changes more greatly; S12: communication and data transmission between sites, A low-power and long-distance communication protocol is designed, which can adjust the data transmission frequency according to different meteorological conditions, speed up the data update speed in the period of rapid change of cloud layer or high communication demand, and reduce the frequency in the case of stable cloud layer or small data change, The communication between sites also needs to consider the transmission delay and bandwidth limitation, and needs to use a suitable communication scheduling strategy, which is adaptively adjusted through the following optimization formula: where C ij represents the communication quality between site i and site j, d ij is the distance between sites, L ij is the length of the data packet, SNR ij is the signal-to-noise ratio, and a and β are weight coefficients. S13: site layout and dynamic optimization, According to the real-time change of the cloud field, the site layout is optimized, this optimization mechanism is based on the real-time cloud field change data, combines the wind speed, temperature and meteorological conditions, judges the movement direction and speed of the current cloud cluster, and adjusts the working frequency and data collection interval between sites, this optimization process is realized through the following dynamic adjustment formula: where ΔP i is the optimal adjustment amount for site i, Δd ij and Δv ij are the distance change and wind speed change with other sites, respectively, and γ and δ are the weight coefficients of dynamic adjustment.

3. The regional cloud field collaborative prediction and correction method based on multi-site sky-imaging network according to claim 1, characterized in that, In step S2, the multi-source observation data includes the data collected by the sky imager, the weather station and the radar; the method of data preprocessing is to filter the noise of the original data, align the time and space, and correct the illumination conditions after acquiring the multi-source observation data.

4. The regional cloud field collaborative prediction and correction method based on multi-site sky-imaging network according to claim 1, characterized in that, Step S2 comprises the following steps: S21: data acquisition, The sky imager installed at each site captures the sky panorama image through the image acquisition device, and each site also collects wind speed, temperature, humidity and pressure meteorological information; The data acquisition frequency of each station is adaptively adjusted according to the weather conditions and the speed of the change of the cloud layer, and the data acquisition frequency is determined by the following formula: f 采集 = f 基准 • (1 + a-AC) (4) Wherein, f 采集 is the data collection frequency of the station, f 基准 is the basic collection frequency, and a is an adjustment coefficient and AC is the speed of the current cloud field change. S22: data preprocessing, In the image data obtained by the sky imager, an image denoising model based on a deep convolutional neural network (CNN) is used; In the meteorological data, Kalman filter is used to smooth the continuous meteorological data, S23: spatio-temporal alignment and standardization of data, The data collected by different stations need to be spatio-temporal aligned and standardized, Spatio-temporal alignment is realized by interpolation method, Standardization is a process of unifying the dimension of the data collected by different stations, and the standardization process is realized by the mean-variance standardization formula: S24: data fusion, The data from different sensors are fused, and the goal of fusion is to integrate the cloud image data of the sky imager and the environmental data of the weather station. The data fusion technology adopts weighted average method and Kalman filter method. For the data of multiple sensors, the weighted average method is used for fusion. After the above preprocessing and fusion steps, the obtained data will become the input data for the subsequent graph neural network (GNN) modeling.

5. The regional cloud field collaborative prediction and correction method based on multi-site sky-imaging network according to claim 1, characterized in that, In step S3, the graph structure modeling includes converting the multi-source observation data of the station into node features in the graph network, and modeling by the graph neural network (GNN); step S3 includes the following steps: S31: node construction and feature extraction, The system regards each station as a node in the graph network, and the features of each node include the cloud image information of the local station and other meteorological related features such as wind speed, temperature and humidity. These features provide local observation data of the station and input data for subsequent graph neural network modeling, The cloud map data is a panoramic image obtained by a sky imager, which is converted into features available for the GNN model after image processing and preprocessing. The color, brightness and corresponding cloud classification information of each pixel in the image, including cumulus, stratus and cirrus, are used as input features of the nodes. These image features are extracted using a deep convolutional neural network (CNN). Assuming that the image features of a certain site are F i The node feature is represented as: F i = CNN(I i ) (10) wherein I i represents the sky image obtained by the station i, the CNN is a convolutional neural network used to extract the cloud layer information in the image, and through the convolutional network, the features of each region in the image are encoded into a vector representation as the features of the nodes; S32: construction of graph structure, Once the node features are extracted, the graph structure needs to be constructed. The core idea of graph neural network is to transmit information through the connection relationship between nodes, i.e. edge. Each node of the graph represents an observation station, and the weight of the edge is dynamically calculated according to the spatial relationship and meteorological conditions between the stations, The edges between sites represent the physical or meteorological correlation between sites, and the weight of the edge reflects the similarity between sites and its influence on the evolution of the cloud field, assuming the distance between site i and site j is d ij , and the difference in wind speed between the two is v ij , then the weight of the edge w ij is represented as: where a and b are the weights of adjustment factors, l scale and v scale are the scale factors of distance and wind speed, respectively, to ensure that the calculated weights have appropriate units and ranges. This formula combines spatial distance and meteorological factors, making the edge weights truly reflect the relationship between sites. S33: spatio-temporal feature modeling and information transmission, After the graph structure is constructed, the graph neural network will start to model the cloud field. The features of each node are interacted with other nodes through the message passing mechanism of the graph neural network. Through graph convolution operation, each node will update its features according to the information of its neighbor nodes, and learn more comprehensive cloud field features; The graph convolution network (GCN) updates the node features by the following formula: wherein, represents the feature of node i at the k+1 layer, is the set of neighbor nodes of node i, w ij is the weight of the edge, is the feature of node j at the k layer, b is the bias term, and σ is the activation function.

6. The regional cloud field collaborative prediction and correction method based on multi-site sky-imaging network according to claim 1, characterized in that, Step S4 further includes: After building the fusion spatio-temporal convolution (ST-CNN) and graph neural network (GNN) model for prediction, the cross-station information is shared through the message passing mechanism to realize real-time warning; The real-time warning predicts the movement of strong convective cloud clusters or large-scale cloud clusters to give early warning to downstream stations.

7. The regional cloud field cooperative prediction and correction method based on multi-site sky-imaging network according to claim 6, characterized in that, Step S4 includes the following steps: S41: construction of spatio-temporal graph convolution network, The spatio-temporal graph convolutional network (ST-GCN) is a model combining graph neural networks and spatio-temporal convolution, which can capture information in the spatial and temporal dimensions. In the graph neural network, the graph convolution operation transmits information through the adjacency relationship of the nodes. In the graph convolutional network, the node feature update of each layer is represented by the following formula: where, is the feature of node i at the k+1 layer, w ij is the weight of the edge, is the feature of neighbor node j at the k layer, is the neighbor set of node i, σ is the activation function, and b is the bias term. This formula represents the update of the current node's feature by aggregating the information of the neighbor nodes. In the spatio-temporal graph neural network, a time dimension convolution operation is also introduced. The spatio-temporal convolution network captures the evolution of the cloud layer in the time dimension by introducing a time convolution based on the graph convolution. The spatio-temporal convolution update formula is as follows: where T is the time dimension, λ t is the weight of the time step, representing the contribution of each time step to the node features; S42: Joint learning of graph convolution and time dependence, A spatio-temporal weighted graph convolution mechanism is designed, so that each time step has different influence on the update of the current node. For a node i, the update process of the node considers the information of the neighbor nodes, and also dynamically adjusts the weight according to the historical time characteristics. This process is represented by the following formula: where λ t The degree of influence on the node features is dynamically adjusted according to each time step. S43: Spatio-temporal prediction and cloud field evolution, The output prediction result of the spatio-temporal graph neural network is represented by the following formula: wherein, is the cloud field prediction result at the future time, X is the input feature of the node, A is the adjacency matrix of the graph, T is the time step, f ST-GCN is the prediction function of the spatio-temporal graph neural network model; S44: Establishing an early warning mechanism and real-time adjustment, When the system predicts the movement of a strong convective cloud cluster or a large-scale cloud cluster, it can give an early warning to the downstream station.

8. The regional cloud field collaborative prediction and correction method based on multi-site sky-imaging network according to claim 1, characterized in that, In step S5, the multi-source observation data fusion and dynamic correction includes using weighted average, Kalman filtering or Bayesian inference method to correct the inconsistency between the data. Step S5 includes the following steps: S51: Using a multi-source data fusion method, The observation data from different sensors are combined to make up for their respective shortcomings by utilizing the complementarity of different data sources. The input data of the cloud field prediction system includes cloud image data captured by a sky imager, wind speed, temperature, humidity meteorological parameters provided by a weather station, and radar data. The fusion of these data is achieved by using a weighted average method, Kalman filtering method or Bayesian inference method; First, the data from each data source is preprocessed and standardized so that different data types have the same dimension and data scale. Then, Kalman filtering is used for data fusion. Kalman filtering is a recursive estimation method that updates the state estimation based on the previous prediction result and the current observation result, and calculates an optimal estimation value. S52: Establishing a dynamic correction mechanism, The correction mechanism includes two aspects: one is to fill in the missing data, and the other is to correct the abnormal data. In the processing of missing data, interpolation method or completion method based on adjacent station data is used. For the correction of outliers, the identification and processing are carried out by a statistical model-based method, assuming that the cloud image data z of a certain station i In the case of abnormal deviation at a certain moment, the system uses the historical data model or the observation data of the adjacent station for correction. By calculating the cloud image similarity between the station and the adjacent station, when the deviation exceeds a certain threshold, the system will correct the data according to the characteristics of the adjacent station: where z′ i is the modified data, z i is the original data, w ij is the edge weight between the station i and the neighboring station j, is the neighbor node set of the node i, and a is the correction coefficient. S53: Establishing an adaptive correction and feedback mechanism; The core idea of adaptive correction is to dynamically adjust the correction strategy and weight of each station according to the real-time cloud field prediction error. The implementation of the feedback mechanism depends on the real-time monitoring of the prediction error by the system. When there is a large deviation between the cloud field prediction result output by the system and the actual observation result, the model will automatically feedback the error information and adjust the data fusion and correction strategy. Assuming that the prediction result The error between the actual result y and the prediction result is δ. The system adjusts the weight of the data source through the following formula: wherein, is the weight for the next time, is the weight for the current time, γ is the learning rate, and δ is the current prediction error; S54: Multi-dimensional data collaborative optimization, The system performs collaborative optimization on multi-dimensional data, and introduces cloud image, meteorological data, radar data, wind speed, radiation, and terrain auxiliary information. In the collaborative optimization process of multi-dimensional data, the system combines different dimensional data through the following weighted model: wherein Z 优化 is the optimized data, Z i is the i-th data source, w i is the weight of each data source.

9. The regional cloud field collaborative prediction and correction method based on multi-site sky-imaging network according to claim 1, characterized in that, Step S6 further includes the following steps: After generating the missing cloud image using the generative adversarial network, the cloud image spatio-temporal consistency optimization is generated; Step S6 further includes the following steps: S61: Identifying and positioning the missing cloud image data, Firstly, the system identifies and locates the missing part in the input data. In the sky imager and weather sensor, the system determines whether the cloud image at a certain time is missing or abnormal. If the cloud image data at a certain time is obviously inconsistent with the data before and after that time, the system marks that time as "missing" state, The identification of missing data is based on the difference of image features, assuming that the cloud image data at a certain moment is I t Difference comparison with the cloud image data of adjacent moments: ΔI t = ||I t -I t-1 ||2 (25); wherein I t and I t-1 are cloud image data at the current time and the previous time, respectively, ΔI t represents the difference between the images, and by calculating the image difference, if ΔI t exceeds a set threshold, it is considered that the data at that time is missing or abnormal; S62: supplement the missing data based on the generative adversarial network (GAN), After identifying the missing cloud image data, the system intelligently supplements the missing part through the generative adversarial network (GAN). In the GAN, the generator G aims to generate new cloud images based on existing data, while the discriminator D determines whether the generated cloud images are real. The loss function of the generative adversarial network is represented by the following formula: where p data is the distribution of real cloud image data, p model is the distribution of data generated by the generator, I fake is the generated fake data, G(Ifake) is the cloud image generated by the generator, and D is the discriminator, which outputs the probability that the image is real. Through repeated training, the generator G learns to generate high-quality cloud images based on the input existing data, while the discriminator D ensures that the generated cloud images are similar to the real cloud images. S63: time and space consistency optimization, When generating missing cloud image data, the consistency of the generated image with time and space must be considered. When generating missing data, ensure that the generated cloud image is consistent with the evolution of the cloud field at the previous and next time points. The system designs a time and space consistency optimization module. The above optimization process is controlled by the following time and space consistency loss function: where I i and I j denote the i-th and j-th pixel in the cloud image, is the neighborhood region, L spatial is the spatial consistency loss, is the temporal consistency loss.

10. The method of claim 1, wherein, In step S7, the regional cloud field prediction result output includes outputting cloud coverage, cloud cluster position and shape information at different sites and different time points. Step S7 includes the following steps: S71: generate regional cloud field prediction results, Post-processing of cloud field data, this processing process includes but is not limited to: Spatial interpolation: interpolation methods include bilinear interpolation and Kriging interpolation; Time smoothing: the system uses a time smoothing algorithm to smooth the prediction results; Assuming that the cloud field prediction data output by the space-time graph neural network is Y After interpolation and smoothing at a certain time t, the final cloud field prediction result is Y t , which is represented as: wherein, is the original prediction result of the ST-GCN model, Interp(·) is an interpolation operation, and Smooth(·) is a time smoothing operation. After interpolation and smoothing, Y t is the processed, spatially and temporally continuous cloud field prediction result; S72: establish cloud cluster movement path prediction and early warning triggering mechanism, In the prediction of the moving path of the cloud cluster, the system predicts the next moving trend of the cloud cluster based on the cloud field data at the previous moment, in combination with the wind speed and airflow direction factors. It is assumed that the position of a certain cloud cluster at time t is P t According to the wind speed and airflow information, the system can predict the position P t+1 of the cloud cluster at the next moment t+1 by the following formula: P t+1 = P t + v t · Δt (31); where P t is the cloud position at the current time, v t is the cloud movement speed at the current time, and Δt is the time interval, According to the movement path and predicted speed of the cloud cluster, the system generates early warnings for different regions. Assuming that the system detects that a cloud cluster will affect the light conditions of a wind and light power generation base in a short time, the system will issue an early warning a certain time before the cloud cluster arrives. The early warning mechanism is triggered by the following conditions: Wherein, Time to Impact (P t ) target ) is the estimated moving time of the cloud cluster from the current position P t to the target area P target , T warning is the early warning time, when the early warning trigger condition is met, the system will output an early warning signal, prompting that the wind and light base may be affected by the cloud layer, and suggesting the corresponding scheduling measures; S73: establish wind and light base scheduling and response mechanism, When the system issues a warning signal, the wind and light power generation base adjusts and responds according to the prediction results of the cloud field. The scheduling system will adjust the operation state of the photovoltaic and wind power generation equipment in advance according to the predicted movement path and prediction time of the cloud cluster. The scheduling system makes prediction and scheduling decisions through the following formula: P adjusted = P forecasted (1 - f cloud coverage (P target , t)) (33). where P adjusted is the adjusted power production, P forecasted is the pre-forecasted power production, f cloud coverage (P target , t) is the change in cloud cover function over time t at the target location P target . S74: result visualization and report generation.

Images