Short-term irradiance ai forecasting system and method incorporating meteorological satellite data

Abstract

The present application relates to the technical field of irradiance prediction, and discloses a short-term irradiance AI prediction system and method combined with meteorological satellite data. The method collects a historical meteorological satellite remote sensing data sequence and a ground irradiance monitoring data sequence of a target area to form a training sample set. Multi-scale spatio-temporal feature extraction is performed on the historical meteorological satellite remote sensing data sequence to generate a three-dimensional meteorological feature tensor. The ground irradiance monitoring data sequence and the three-dimensional meteorological feature tensor are spatio-temporally aligned and matched to establish a meteorological-irradiance correlation feature matrix. A deep spatio-temporal fusion network model is then constructed, and the meteorological-irradiance correlation feature matrix is input into the model for end-to-end training to obtain a trained irradiance prediction model. Real-time meteorological satellite remote sensing data flow is acquired, real-time three-dimensional meteorological feature tensors are extracted and input into the trained model, and an irradiance prediction value sequence within a preset time window in the future is output.

Description

Technical Field

[0001] This invention relates to the field of irradiance prediction technology, specifically to a short-term irradiance AI prediction system and method that combines meteorological satellite data. Background Technology

[0002] Surface irradiance, as indispensable data for the assessment, planning, and application of solar energy resources, plays a crucial role in several key areas. In solar energy resource assessment, it is a key indicator for measuring the abundance of solar energy resources in a region. Accurate measurement and analysis of surface irradiance provide a clear understanding of the distribution of solar energy resources in different regions, offering solid data support for the development and utilization of solar energy. For example, in western my country, such as Xinjiang and Qinghai, surface irradiance is high, indicating abundant solar energy resources, making them ideal for large-scale solar power plant construction.

[0003] In the site selection process for photovoltaic (PV) power plants, surface irradiance is a core consideration. Accurate irradiance data helps engineers determine the optimal location for the plant, ensuring it receives sufficient sunlight to improve power generation efficiency and reduce costs. A well-located PV power plant can significantly increase power generation compared to a poorly located one, which is crucial for improving energy efficiency and achieving sustainable development.

[0004] For power prediction of photovoltaic (PV) power plants, surface irradiance also plays a decisive role. It is a major factor affecting PV power generation, and its changes directly impact the power generation capacity of PV power plants. Accurate prediction of irradiance allows for advance knowledge of the power generation capacity of PV power plants at different times in the future, which is crucial for grid dispatch and management. Grid operators can use the power prediction results from PV power plants to rationally arrange power production and distribution, ensuring the stable operation of the grid and avoiding shocks to the grid caused by fluctuations in PV power generation.

[0005] Using satellite remote sensing data to retrieve various meteorological parameters around a target site and thus obtain surface irradiance is a common method. However, this method has several problems. First, the retrieval process involves multiple steps and complex physical models, which can easily lead to error accumulation. Each data processing and model calculation may introduce some error, and as the retrieval steps increase, these errors gradually accumulate, eventually resulting in a significant deviation between the retrieved irradiance data and the actual values. Second, the physical modeling parameters are complex, requiring consideration of various factors such as atmospheric composition, cloud characteristics, and surface reflectivity. Accurately obtaining and determining these parameters is difficult, and inaccuracy in any one parameter can affect the accuracy of the retrieval results. Furthermore, the temporal resolution of satellite remote sensing data is usually low, which cannot meet the requirements for real-time monitoring and short-term forecasting of irradiance. For example, some commonly used meteorological satellites may have data update cycles of several hours or even a day. For short-term irradiance forecasting, such temporal resolution is far from sufficient to capture rapid changes in irradiance in a timely manner, resulting in low forecast accuracy and failing to meet the needs of practical applications. Summary of the Invention

[0006] The purpose of this invention is to provide a short-term irradiance AI prediction system and method that combines meteorological satellite data to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides a short-term irradiance AI prediction method combining meteorological satellite data, the method comprising: Historical meteorological satellite remote sensing data sequences and ground irradiance monitoring data sequences of the target area were collected to form a training sample set; Multi-scale spatiotemporal feature extraction is performed on historical meteorological satellite remote sensing data sequences to generate a three-dimensional meteorological feature tensor containing cloud dynamic features, atmospheric transmission features, and surface reflection features. The ground irradiance monitoring data sequence was spatiotemporally aligned and matched with the three-dimensional meteorological feature tensor to establish a meteorological-irradiance correlation feature matrix. A deep spatiotemporal fusion network model is constructed, and the meteorological-irradiance correlation feature matrix is ​​input into the deep spatiotemporal fusion network model for end-to-end training to obtain a trained irradiance prediction model. The system acquires real-time meteorological satellite remote sensing data streams, extracts real-time three-dimensional meteorological feature tensors, inputs them into the trained irradiance prediction model, and outputs a sequence of predicted irradiance values ​​within a preset future time window.

[0008] Preferably, the multi-scale spatiotemporal feature extraction of historical meteorological satellite remote sensing data sequences includes: The historical meteorological satellite remote sensing data sequence is divided into multiple time segments, each of which contains remote sensing image data in the visible light band, infrared band, and water vapor band. Spatial pyramid features were extracted from the remote sensing image data of each band to obtain local meteorological features at different spatial scales. Temporal convolutional networks are used to model the time dependence of local meteorological features in each band, generating multi-band meteorological features with temporal correlation. Multi-band meteorological features are fused through a cross-band attention mechanism to ultimately form a three-dimensional meteorological feature tensor that includes cloud dynamics, atmospheric transmission, and surface reflection.

[0009] Preferably, the spatiotemporal alignment and matching of the ground irradiance monitoring data sequence with the three-dimensional meteorological feature tensor includes: Based on the geographical location information of ground monitoring stations, meteorological feature vectors corresponding to the spatial location are extracted from the three-dimensional meteorological feature tensor. Time interpolation processing is performed on the ground irradiance monitoring data sequence to ensure that its sampling frequency is consistent with the temporal resolution of the meteorological feature vector; The mutual information between meteorological feature vectors and irradiance monitoring data is calculated, and feature dimensions with mutual information exceeding the threshold are selected to form a meteorological-irradiance correlation feature matrix.

[0010] Preferably, the construction of the deep spatiotemporal fusion network model includes: Design a network architecture consisting of a spatiotemporal encoder and a spatiotemporal decoder, wherein the spatiotemporal encoder contains parallel spatial attention branches and temporal convolution branches; Deformable convolutional modules are deployed in the spatial attention branch to capture non-rigid deformation features during cloud movement; A causal dilated convolution structure is used in the temporal convolution branch to ensure that the prediction process conforms to the temporal causality law; The spatiotemporal decoder integrates the output features of the spatial and temporal branches through a gating fusion mechanism to generate irradiance predictions for future time steps.

[0011] Preferably, the training process of the deep spatiotemporal fusion network model includes: The meteorological-irradiance correlation feature matrix is ​​divided into a training set and a validation set, and the mean squared error and correlation coefficient are set as joint loss functions. An adaptive moment estimation algorithm is used to optimize network parameters, and the weight allocation ratio of spatial and temporal branches is dynamically adjusted during training. The prediction performance on the validation set is monitored by an early stopping mechanism, and the training process is terminated when the performance does not improve for several consecutive training cycles.

[0012] Preferably, the real-time acquisition of the current meteorological satellite remote sensing data stream includes: Establish a real-time data interface with the meteorological satellite data center to receive the latest satellite remote sensing observation data; The received raw data is subjected to radiometric calibration and geometric correction to eliminate the effects of sensor noise and projection distortion. The corrected data is spliced ​​with historical data to form a continuous time series, maintaining consistency in the extraction of spatiotemporal features.

[0013] Preferably, the extraction of real-time three-dimensional meteorological feature tensors includes: Real-time data streams are processed using a sliding time window approach, and the data in each window undergoes the same multi-scale feature extraction process as in the training phase. An incremental feature update mechanism is introduced, which recalculates only the feature dimensions corresponding to newly arrived data; The feature caching module retains key features of historical windows, reducing the overhead of redundant calculations.

[0014] Preferably, the output sequence of predicted irradiance values ​​within a future preset time window includes: The prediction time window is divided into multiple time intervals, each interval corresponding to a different prediction confidence level; The prediction results of adjacent time intervals are smoothed to eliminate abrupt changes in the predicted value sequence; Generate structured outputs containing predicted values ​​and confidence intervals for use in downstream system decision-making.

[0015] Preferably, the method further includes: Regularly collect the statistical measures of the deviation between actual observed irradiance data and predicted results; When the deviation exceeds a preset threshold, the model retraining process is triggered to update the parameters of the deep spatiotemporal fusion network model; Establish a predictive performance monitoring dashboard to visualize the trend of prediction accuracy changes at different time scales.

[0016] Preferably, the present invention also includes a short-term irradiance AI prediction system combining meteorological satellite data, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the above-described short-term irradiance AI prediction method combining meteorological satellite data.

[0017] Compared with the prior art, the beneficial effects of the present invention are: This method possesses unique advantages in data processing and feature extraction. When processing historical meteorological satellite remote sensing data sequences, it employs multi-scale spatiotemporal feature extraction technology, enabling in-depth analysis of the data from multiple dimensions and scales. Spatially, it can capture the dynamic changes of clouds at different altitudes and in different regions, accurately acquiring both large-scale cloud movement and subtle changes in localized areas, thus clearly depicting the dynamic characteristics of clouds. In analyzing atmospheric transmission characteristics, this method considers the absorption and scattering of solar radiation by different atmospheric components, as well as the impact of changes in the physical state of the atmosphere at different times and locations on transmittance, comprehensively and accurately grasping atmospheric transmission characteristics. For surface reflection characteristics, it comprehensively considers the differences in reflectivity of different surface types and the impact of changes in surface cover over time on reflectivity, generating extremely comprehensive surface reflection characteristic information. This rich and accurate feature information provides a solid data foundation for subsequent irradiance prediction, enabling the prediction process to be based on more realistic and comprehensive meteorological information, greatly improving the reliability of the prediction.

[0018] Another key advantage of this method is the spatiotemporal alignment and matching of ground irradiance monitoring data sequences with three-dimensional meteorological feature tensors to establish a meteorological-irradiance correlation feature matrix. This matching method allows for in-depth exploration of the potential relationship between ground irradiance and meteorological elements. In the temporal dimension, it reveals the variation patterns of irradiance with meteorological elements at different times, such as how irradiance changes rapidly during periods of rapid cloud movement. In the spatial dimension, it analyzes the correspondence between meteorological conditions and local irradiance at different geographical locations, thereby establishing a more robust and accurate correlation model. This correlation feature matrix effectively integrates the advantages of both types of data, significantly improving data usability and providing more valuable input data for subsequent model training and prediction.

[0019] The deep spatiotemporal fusion network model employs an end-to-end training approach, offering significant advantages. Traditional prediction models often require manual setting of numerous parameters and rules, which is not only time-consuming and labor-intensive but also susceptible to human error. End-to-end training, however, allows the model to automatically learn the complex mapping relationship between weather and irradiance directly from the input data, minimizing manual intervention. During training, the model automatically adjusts its internal parameters and structure to adapt to different weather conditions and data characteristics. When dealing with complex and variable weather data, it can quickly capture key information and patterns, accurately learning how meteorological elements influence irradiance changes. This training method reduces errors caused by manually set rules, improves the model's adaptability and accuracy, enabling it to better cope with various real-world situations and maintain high prediction accuracy under different weather conditions and geographical environments.

[0020] This method acquires real-time meteorological satellite remote sensing data streams and inputs the extracted real-time 3D meteorological feature tensors into a trained irradiance prediction model. This allows for the rapid and accurate output of irradiance prediction sequences within a preset time window, meeting the timeliness requirements of practical applications. In the field of solar power generation, power dispatching needs to adjust power generation plans and power allocation schemes promptly based on irradiance predictions. This method tracks meteorological changes in real time and quickly inputs the latest meteorological information into the model for prediction, enabling power dispatching departments to obtain irradiance predictions for a future period in a short time, allowing them to make advance power generation arrangements and avoid power supply instability caused by untimely predictions. For applications sensitive to irradiance changes, such as using solar energy for greenhouse heating and drying in agricultural production, timely and accurate irradiance predictions can help farmers rationally plan production activities, improve production efficiency, and reduce energy waste. Attached Figure Description

[0021] Figure 1 This is a schematic diagram illustrating the working principle of the short-term irradiance AI prediction method combining meteorological satellite data described in this invention. Figure 2 A flowchart for extracting multi-scale spatiotemporal features from historical meteorological satellite remote sensing data sequences; Figure 3 This is a flowchart for spatiotemporal alignment and matching of ground irradiance monitoring data sequences with three-dimensional meteorological feature tensors. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Please see Figure 1This invention provides a short-term irradiance AI prediction system and method combining meteorological satellite data. The method includes: collecting historical meteorological satellite remote sensing data sequences and ground irradiance monitoring data sequences of a target area to form a training sample set; extracting multi-scale spatiotemporal features from the historical meteorological satellite remote sensing data sequences to generate a three-dimensional meteorological feature tensor containing cloud dynamic features, atmospheric transmission features, and surface reflection features; performing spatiotemporal alignment and matching between the ground irradiance monitoring data sequences and the three-dimensional meteorological feature tensor to establish a meteorological-irradiance correlation feature matrix; constructing a deep spatiotemporal fusion network model, inputting the meteorological-irradiance correlation feature matrix into the deep spatiotemporal fusion network model for end-to-end training to obtain a trained irradiance prediction model; acquiring the current meteorological satellite remote sensing data stream in real time, extracting the real-time three-dimensional meteorological feature tensor and inputting it into the trained irradiance prediction model, and outputting a sequence of predicted irradiance values ​​within a preset future time window. Historical meteorological satellite remote sensing data sequences are derived from meteorological satellite observation platforms, covering multiple bands including visible light, infrared, and water vapor. Ground irradiance monitoring data sequences are sourced from monitoring stations distributed across the ground, ensuring consistency between timestamps and geographic coordinates during data acquisition. The multi-scale spatiotemporal feature extraction process employs a spatial pyramid structure and a temporal convolutional network to capture dynamic changes in cloud movement and atmospheric conditions. Spatiotemporal alignment matching is based on geospatial interpolation and mutual information calculation to eliminate spatiotemporal biases between data. The deep spatiotemporal fusion network model is designed as an encoder-decoder architecture, integrating spatial attention and temporal convolution mechanisms. The training process uses a joint loss function and adaptive optimization algorithms. In the real-time prediction stage, a sliding window is used to process the data stream, achieving incremental feature updates and smooth output.

[0024] Example 1: See Figure 2Multi-scale spatiotemporal feature extraction of historical meteorological satellite remote sensing data sequences involves dividing the data sequences into fixed time segments containing images in visible light, infrared, and water vapor bands. The visible light band reflects cloud reflectance characteristics, the infrared band characterizes temperature distribution, and the water vapor band indicates atmospheric humidity. Time segments are divided at 15-minute or 1-hour intervals, and satellite data timestamps are processed using automated scripts to ensure segment lengths match the satellite revisit cycle. Spatial pyramid feature extraction is performed on the remote sensing images of each band, employing multi-level convolutional operations to capture local meteorological features at different spatial scales. The convolutional operations use pre-trained ResNet or VGG network structures. The spatial pyramid has three scales: a 3×3 convolutional kernel extracts detailed features, a 5×5 convolutional kernel captures intermediate patterns, and a 7×7 convolutional kernel obtains global context. Feature maps at each scale are upsampled and fused to preserve spatial hierarchy information. Local meteorological features are input into a temporal convolutional network for temporal dependency modeling, and dilated convolutional kernels are used to capture long-term dependencies. The network comprises multiple one-dimensional convolutional layers with a kernel size of 3 and dilation factors increasing in increments of 1, 2, 4, and 8. Each convolutional layer is followed by a ReLU activation function. The network analyzes the trend of the feature sequence through time-axis sliding analysis, outputting a multi-band meteorological feature tensor with temporal correlation. The multi-band features are fused using a cross-band attention mechanism. The similarity matrix between bands is calculated and normalized using softmax to obtain the attention weights. The weighted fusion forms a three-dimensional meteorological feature tensor with a dimensional structure of [batch_size, time_steps, height, width, channels]. This tensor includes dynamic cloud features derived from the visible / infrared bands, atmospheric transmission features based on the infrared / water vapor bands, and surface reflection features extracted from the visible band.

[0025] In the segmentation and processing of historical meteorological satellite remote sensing data sequences, the time segment length can be configured to 1 hour, with no overlap between segments to ensure temporal continuity. Data sources include Fengyun or GOES series satellites, and the original data format is HDF or NetCDF. During parsing, missing values ​​are filled by linear interpolation or masking. Visible light band remote sensing image data is valid during the day, while infrared band and water vapor band remote sensing image data are available around the clock. The data processing pipeline uses Python libraries such as xarray or GDAL for reading and preprocessing. After the time segments are segmented, they are stored as multidimensional arrays for easy subsequent feature extraction. The specific implementation of spatial pyramid feature extraction relies on deep learning frameworks such as TensorFlow or PyTorch. Convolutional layer parameters are initialized using a He normal distribution, and weights are fixed during training to avoid overfitting. For local meteorological feature extraction, each band is processed independently to avoid inter-band interference. The number of feature map channels is set to 64, 128, and 256, corresponding to three scales. Max pooling is used to preserve salient features. Multi-scale feature fusion uses bilinear interpolation to upsample low-resolution feature maps to high resolution, then concatenates them with the original features to enhance spatial details. The application of temporal convolutional networks targets the local meteorological feature sequences for each band. The input tensor shape is [time_steps, features]. The temporal convolutional network has four layers, with each layer outputting 128 channels. Dilated convolutions ensure that the receptive field expands without increasing the number of parameters. After modeling time dependencies, the multi-band meteorological feature tensors are dimensionality-reduced through fully connected layers to reduce computational complexity. Causal convolutions are used during temporal convolutional network training to ensure that the temporal order does not leak future information. In the fusion stage of the cross-band attention mechanism, the attention score is calculated using dot product attention. The query, key, and value vectors are obtained from the linear transformation of the multi-band meteorological feature tensor. The number of attention heads is set to 8 to improve the model's expressive power. The fusion output is stabilized through layer normalization training. In the process of generating the three-dimensional meteorological feature tensor, the cloud dynamic features are estimated by optical flow algorithm to estimate cloud displacement, the atmospheric transmission features are approximated by combining the radiative transfer model, and the surface reflection features are normalized to reduce the influence of the solar angle. Finally, the tensor is stored as a floating-point type to support GPU-accelerated computing.

[0026] The multi-scale spatiotemporal feature extraction process is deployed in a distributed computing environment, such as Apache Spark or Dask, to process large-scale historical data. The feature extraction module is encapsulated as a Python class, providing configuration parameter interfaces such as time segment length and convolution kernel size. The feature extraction results are saved in HDF5 format with metadata recording processing parameters. The entire process integrates data verification steps to check feature quality, such as value range and consistency, and triggers reprocessing logic in case of anomalies. The preprocessing of historical meteorological satellite remote sensing data sequences includes radiometric calibration and geometric correction. Radiometric calibration uses satellite-specific coefficients to convert digital values ​​to radiance. Geometric correction uses ground control points to reproject onto an equal latitude and longitude grid. The convolutional operation of spatial pyramid feature extraction is accelerated using CUDA and executed on an NVIDIA GPU. The training of the temporal convolutional network uses historical data to split into training and validation sets. The optimizer is Adam, and the learning rate is set to 0.001. The cross-band attention mechanism is implemented with reference to the Transformer architecture. Location encoding adds time information to enhance sequence modeling capabilities. The three-dimensional meteorological feature tensor is used as input to downstream tasks and, after alignment with ground data, is input into the deep spatiotemporal fusion network model. Time segmentation considers the temporal resolution of satellite data. For example, Fengyun satellites update every 15 minutes, and segmentation matches this frequency to avoid loss of temporal information. Cloud masks are processed during the parsing of visible light band remote sensing image data to distinguish between cloud areas and clear sky areas. Infrared band remote sensing image data is calibrated to brightness and temperature. Atmospheric water content is calculated for water vapor band remote sensing image data. Batch normalization layers are added to the convolutional layers of spatial pyramid feature extraction to accelerate convergence. After local meteorological feature extraction, dimensionality reduction techniques such as PCA are applied to reduce redundancy while retaining the main variance. The dilated convolution configuration of the temporal convolutional network avoids excessive holes that could lead to loss of local information. Feature scaling is performed before multi-band meteorological feature fusion to ensure consistent dimensions across different bands. Gradient pruning is used to prevent explosion during the training of cross-band attention mechanisms. The dimensions of the three-dimensional meteorological feature tensor are adjusted to adapt to the network input requirements, such as adjusting the height and width to a fixed size, achieved through pruning or padding.

[0027] Historical meteorological satellite remote sensing data sequences are stored using spatiotemporal databases such as PostGIS, supporting fast queries. The multi-scale spatiotemporal feature extraction workflow is designed to be scalable, allowing the addition of new bands such as the near-infrared band. The output log of the feature extraction module records the processing status for easy debugging. The time step of the temporal convolutional network is set to 24 hours, covering the daily cycle. The attention weights of the cross-band attention mechanism are visualized to analyze the importance of features. The application of 3D meteorological feature tensors is not limited to irradiance prediction but can also be used for other meteorological forecasting tasks. Computational resource management during the feature extraction process is encapsulated using containerization technologies such as Docker and deployed on a cloud platform for elastic scaling. The historical data archiving strategy regularly cleans up old data, retaining feature tensors to save storage space. Performance monitoring of multi-scale spatiotemporal feature extraction uses metrics such as processing time and memory usage for optimization and iterative improvement. Multi-scale spatiotemporal feature extraction from historical meteorological satellite remote sensing data sequences is a fundamental step in the patented method. The quality of these features directly impacts prediction accuracy. Implementation details are designed based on the characteristics of the actual satellite data. For example, when processing Fengyun-4 satellite data, the convolution kernel size is adjusted to adapt to its spatial resolution. The dilation factor of the temporal convolutional network is dynamically set according to the data sampling rate. The parameters of the cross-band attention mechanism are determined through cross-validation. The generation of the three-dimensional meteorological feature tensor ensures spatiotemporal alignment with ground monitoring data, and subsequent model training depends on the integrity of this tensor. The entire process runs on a Linux server, using multi-threading for parallel band feature extraction. After the feature tensor is output, its shape and value range are verified to avoid outlier propagation. The reproducibility of the implementation method is ensured through code version control such as Git management, and parameter settings and dependent library versions are documented.

[0028] Example 2: See Figure 3The spatiotemporal alignment and matching of ground irradiance monitoring data sequences with a 3D meteorological feature tensor involves extracting meteorological feature vectors corresponding to the spatial location from the 3D meteorological feature tensor based on the geographical location information of the ground monitoring stations. The geographical location information of the ground monitoring stations is accurately recorded by GPS receiving equipment using latitude and longitude coordinates. These coordinates are stored in the system database as key spatial index information and are associated with metadata such as station number and monitoring equipment model. During feature extraction, the system first reads the latitude and longitude coordinates of the target monitoring station, and then calls the spatial index module of the 3D meteorological feature tensor to locate the region to which the coordinates belong in the tensor's spatial grid. The spatial dimension of the 3D meteorological feature tensor is based on equal latitude and longitude grids, with each grid cell corresponding to a fixed range of latitude and longitude. The system determines the target grid and surrounding adjacent grids where the station coordinates are located through coordinate comparison. For each dimension of data in the 3D meteorological feature tensor—cloud dynamics, atmospheric transmission, and surface reflection—feature information of the target grid and surrounding grids is extracted to form a unique original feature set for that station. For example, if the GPS coordinates of a ground monitoring station are (116.3°E, 39.9°N), the system can use spatial indexing to find that these coordinates fall within the (116°-116.5°E, 39.5°-40°N) grid cells of the three-dimensional meteorological feature tensor. Then, it extracts the feature data related to cloud movement speed, atmospheric transmittance coefficient, and surface reflectance from this grid and the adjacent (115.5°-116°E, 39.5°-40°N), (116°-116.5°E, 40°-40.5°N), and (116.5°-117°E, 39.5°-40°N) grids. After integration, it forms the initial meteorological feature vector corresponding to the station.

[0029] Spatial location matching employs a bilinear interpolation algorithm to map grid point features to the precise location of monitoring stations. The interpolation calculation involves a weighted sum of feature values ​​from four adjacent grid points surrounding the station, with the weights determined by the reciprocal of the distance between the station and the grid center. The extraction operation iterates through each time step of the three-dimensional meteorological feature tensor, generating a meteorological feature vector sequence consistent with the number of ground monitoring stations. This meteorological feature vector includes multiple dimensions such as cloud dynamics, atmospheric transmission characteristics, and surface reflection characteristics. Temporal interpolation is performed on the ground irradiance monitoring data sequence. Since the original sampling frequency of the ground irradiance monitoring data sequence may be higher than the satellite data update time, a cubic spline interpolation algorithm is used to resample the irradiance data to a timestamp sequence consistent with the meteorological feature vector. The interpolation process preserves the physical characteristics of the irradiance data, such as non-negativity and smooth variation. The time-aligned ground irradiance monitoring data sequence and the meteorological feature vector have the same number of time points and intervals, forming a data pair with a completely matched time dimension.

[0030] During spatial location matching, the bilinear interpolation algorithm first determines the specific location of the ground monitoring station in the three-dimensional meteorological feature tensor space grid based on its latitude and longitude coordinates. It identifies four adjacent grid cells surrounding the station, denoted as Grid A, Grid B, Grid C, and Grid D, with these four grids diagonally distributed around the station's coordinates. Next, it acquires the target meteorological feature values ​​of the four grid cells at the current time step, while simultaneously calculating the straight-line distance between the station's coordinates and the center of each grid cell. The weight of each grid feature value is determined based on the reciprocal of the distance; that is, the closer the station is to the grid center, the greater the weight of the feature value, and the farther the distance, the smaller the weight, ensuring that the features of nearby grid cells contribute more significantly to the station's feature vector. Subsequently, the feature value of each grid is multiplied by its corresponding weight to obtain the weighted feature value of each grid. Finally, the weighted feature values ​​of the four grids are summed to obtain the interpolation result for the station in the current feature dimension. The extraction operation iterates through each time step of the three-dimensional meteorological feature tensor in chronological order. For each time step, the above interpolation calculation process is repeated to obtain the interpolation results of cloud dynamic features, atmospheric transmission features, and surface reflection features corresponding to each station at each time step. Finally, the interpolation data of all time steps and each feature dimension are integrated to generate a meteorological feature vector sequence with the same number of ground monitoring stations. The temporal resolution of each vector sequence is consistent with that of the three-dimensional meteorological feature tensor.

[0031] The mutual information between meteorological feature vectors and irradiance monitoring data is calculated. The mutual information calculation employs a histogram estimation method to discretize continuous variables into multiple intervals, statistically analyzing the KL divergence between the joint and marginal distributions. The mutual information between each feature dimension and irradiance value is calculated independently, generating a feature importance ranking list. A higher mutual information value indicates a stronger statistical dependence between the feature and irradiance. Feature dimensions with mutual information exceeding a threshold are selected to form a meteorological-irradiance correlation feature matrix. The threshold is set as the median of all feature mutual information values. The selection process retains feature dimensions highly correlated with irradiance changes, such as cloud cover changes, while removing irrelevant features, such as stable surface reflectance. The rows of the meteorological-irradiance correlation feature matrix correspond to time samples, and the columns correspond to the selected feature dimensions. The matrix is ​​stored as a two-dimensional floating-point array for model training. The geographic location information of ground monitoring stations is stored in the spatial database PostGIS, including attributes such as station number, latitude and longitude, and altitude. The grid coordinate system of the three-dimensional meteorological feature tensor adopts an equal latitude and longitude projection, with a grid resolution of 0.1 degrees × 0.1 degrees. Errors caused by projection transformation are handled during spatial location matching. The station's latitude and longitude are converted to the grid coordinate system using a coordinate transformation formula. The bilinear interpolation algorithm is implemented as a numerical calculation function, taking the coordinates of the four corner points and feature values ​​of the grid cell containing the station as input, and outputting the interpolated feature vector. Temporal interpolation filters outliers in the ground irradiance monitoring data sequence, removing abrupt changes caused by sensor malfunctions. Cubic spline interpolation ensures the curve is second-order continuous and differentiable. The interpolation nodes are set to the timestamps of the satellite data. The resampled irradiance data sequence maintains the same physical unit (W / m²) as the original data. 2 Before calculating mutual information, meteorological feature vectors and irradiance monitoring data are standardized to eliminate dimensional differences. The number of histogram intervals is automatically determined based on the sample size using the Sturges formula. The KL divergence is calculated using the logarithmic probability density function. The mutual information value range is normalized to the [0,1] interval for easy comparison. The feature selection threshold can be adjusted according to the application scenario; for example, a higher quantile threshold is used when high precision is required. After constructing the meteorological-irradiance correlation feature matrix, collinearity checks are performed to remove highly correlated features and reduce redundancy. The matrix is ​​stored as an HDF5 format file with metadata describing the meaning of the feature dimensions.

[0032] The spatiotemporal alignment and matching system is deployed on a distributed computing cluster, using Apache Spark to process large-scale site data. Location matching operations are executed in parallel, with each worker node handling feature extraction tasks for a set of monitoring stations. Temporal interpolation utilizes Spark SQL's window functions for efficient time-series resampling. Mutual information calculation is distributed across multiple computing nodes using a MapReduce model, and threshold filtering results are aggregated on the driver node. The system monitoring module records data processing progress and quality control indicators, such as interpolation error statistics and mutual information distribution, triggering an alarm mechanism for abnormal situations. Data from ground monitoring stations is received in real-time via the MQTT IoT protocol, and 3D meteorological feature tensors are loaded from the distributed file system HDFS. The entire alignment and matching process is automated, with daily updates to the meteorological-irradiance correlation feature matrix. In the geographic location information matching stage, terrain shading effects are considered; for mountainous sites, a digital elevation model is additionally introduced to correct interpolation weights. Temporal interpolation sets special markers for nighttime zero-value irradiance data to avoid negative values ​​in the interpolation algorithm. Mutual information calculation is performed in zones based on different weather types, such as calculating separately for sunny and cloudy days to improve the accuracy of feature correlation assessment. The meteorological-irradiance correlation feature matrix is ​​augmented with time-lag features, including meteorological feature vectors from historical time steps. The matrix dimensions are expanded to capture temporal dependencies. After the matrix is ​​constructed, missing values ​​are filled in using the K-nearest neighbor algorithm based on feature values ​​from similar stations.

[0033] When constructing the meteorological-irradiance correlation feature matrix, the selection range of time-lag features is first determined. Combining the 15-minute update frequency of satellite data and the temporal correlation of irradiance changes, historical meteorological feature vectors from the previous 1, 2, and 4 time steps are selected as time-lag features. For example, if the current time step is 10:00 on a certain day (corresponding to the satellite data update time), then meteorological feature vectors from three historical time steps of 9:45, 9:30, and 9:00 need to be included. The feature vector of each historical time step contains cloud dynamic features (such as cloud coverage area ratio and cloud movement speed), atmospheric transmission features (such as atmospheric transmission coefficient), and surface reflection features (such as surface reflectivity). These historical feature vectors are placed side by side with the meteorological feature vector of the current time step and the irradiance monitoring data, expanding the number of columns of the original feature matrix to four times the original number (1 current time step + 3 historical time steps), forming an extended feature matrix that includes time-series dependencies. After the matrix is ​​constructed, if meteorological feature values ​​are missing at some time steps (e.g., cloud movement speed data is missing at a certain station at 9:30 due to satellite data transmission delay), the K-nearest neighbor algorithm is activated to fill in the missing data. First, five reference stations are selected from the system that are geographically adjacent to the missing station (e.g., the difference between latitude and longitude is within 0.5°) and have similar meteorological conditions (e.g., all stations are cloudy at the same time on the same day). The cloud movement speed data at 9:30 of these reference stations are extracted. Then, the weights are determined by calculating the spatial distance between the missing station and each reference station (based on the straight-line distance calculated from latitude and longitude coordinates). The closer the reference station is, the greater its weight. Finally, the cloud movement speed data at 9:30 of the missing station is filled in based on the weighted average result, ensuring that there are no missing values ​​in the entire meteorological-irradiance correlation feature matrix, thus meeting the training data requirements of the subsequent deep spatiotemporal fusion network model.

[0034] The spatiotemporal alignment and matching system is integrated with the data quality management system. Geographical location information of ground monitoring stations is regularly calibrated and maintained. The grid coordinate system of the 3D meteorological feature tensor is consistent with international standards. The mutual information calculation module provides a visualization interface for feature importance, assisting experts in analyzing feature correlations. The meteorological-irradiance correlation feature matrix version management records each update log. System performance optimizations include spatial indexing to accelerate location queries, a time-series database to store historical irradiance data, and a mutual information calculation result caching mechanism to reduce redundant calculations.

[0035] The data sources for the ground irradiance monitoring data sequence include the national meteorological observation network and photovoltaic power station monitoring systems. Data format differences are converted using a unified data model. The grid data of the three-dimensional meteorological feature tensor is stored using a distributed array structure, supporting block-based reading and processing of large-scale data. The spatiotemporal alignment matching algorithm is implemented as a Python package, providing an application programming interface (API) for downstream modules. Matching accuracy is verified through cross-validation with reserved stations, comparing the trends in irradiance data before and after interpolation. Sensitivity analysis of the mutual information threshold determines the optimal feature subset. The dimensionality compression technique of the meteorological-irradiance correlation feature matrix retains over 95% of the original information. The spatiotemporal alignment matching process is configured with different parameters in spring and winter to adapt to the impact of solar altitude angle changes on irradiance. The ground monitoring station geographic location information update mechanism handles station migration. The grid resolution of the three-dimensional meteorological feature tensor supports dynamic adjustment. A sliding window is introduced to analyze the time-varying characteristics of feature correlation in mutual information calculation. The construction of the meteorological-irradiance correlation feature matrix considers spatial correlation, introducing features from neighboring stations to expand the number of columns in the matrix, enhancing the representation of local meteorological conditions.

[0036] Example 3: Constructing a deep spatiotemporal fusion network model involves designing a network architecture composed of a spatiotemporal encoder and a spatiotemporal decoder. The spatiotemporal encoder includes parallel spatial attention branches and temporal convolution branches. The spatial attention branch deploys deformable convolutional modules to capture non-rigid deformation features during cloud movement. The temporal convolution branch employs a causal dilation convolution structure to ensure the prediction process conforms to temporal causality. The spatiotemporal decoder integrates the output features of the spatial and temporal branches through a gating fusion mechanism to generate irradiance predictions for future time steps. The training process of the deep spatiotemporal fusion network model divides the meteorological-irradiance correlation feature matrix into training and validation sets. A joint loss function is set, combining mean squared error and a regularization term. The formula for calculating the joint loss function is: in, This represents the total number of samples in the training set. For sample index, Representing the The true irradiance value of each sample Representative model for the first The predicted irradiance values ​​for each sample. The regularization coefficient is . This represents the total number of parameters in the deep spatiotemporal fusion network model. For parameter index, The representative model The numerical values ​​of each parameter.

[0037] An adaptive moment estimation algorithm is employed to optimize network parameters. This algorithm dynamically adjusts the learning rate of each parameter by calculating the first and second moment estimates of the parameter gradients, thereby improving the stability and efficiency of parameter updates. During training, the weight allocation ratio between the spatial and temporal branches is dynamically adjusted, with the initial spatial branch weight coefficient set to... The time branch weight coefficient is After each training cycle, the prediction error is adjusted based on the validation set. The value of is increased if the spatial branching feature contributes more significantly to the prediction result. Conversely, it decreases. The prediction performance on the validation set is monitored by an early stopping mechanism. A "patience period" is preset. When the loss function value of the validation set does not decrease (i.e. the prediction performance does not improve) within several consecutive training periods, the training process is terminated immediately, and the model parameters of the current training period are saved as the parameters of the final trained irradiance prediction model.

[0038] The deep spatiotemporal fusion network model employs an encoder-decoder structure. The spatiotemporal encoder consists of two parallel branches. The spatial attention branch uses a convolutional attention module to calculate the spatial weights of the feature maps. The deformable convolutional module is implemented as a convolutional layer with learnable offsets, generated by additional convolutional layers. The kernel size is set to 3×3, and the deformable convolutional module captures cloud deformation features. The temporal convolutional branch uses causal dilated convolutions stacked in multiple layers, with the dilation factor set according to a geometric progression. Layer normalization and ReLU activation functions are added after each convolutional layer. The spatiotemporal decoder is designed as a gated recurrent unit structure, taking the fused features from the two branches as input and using fully connected layers to map the output to the predicted value. The training process of the deep spatiotemporal fusion network model uses the meteorological-irradiance correlation feature matrix as training data. The training and validation sets are divided in chronological order to avoid data leakage. The joint loss function is defined as a balance between prediction error and model complexity. in: This represents the numerical value of the joint loss function. Indicates the number of samples. Indicates the sample index. Indicates the first The true irradiance value of each sample Indicates the first The predicted irradiance values ​​for each sample. Represents the regularization coefficient. Indicates the total number of model parameters. Indicates the parameter index. Indicates the first The model parameters are numerically defined. The adaptive moment estimation algorithm is configured with an initial learning rate of 0.001, a momentum parameter of 0.9, and weight distribution ratios adjusted using trainable parameters. The spatial branch output is multiplied by the weight coefficients. The time branch output is multiplied by the weighting factor. , The value is dynamically updated based on the performance of the validation set. The early stopping mechanism sets the patience parameter to 10 training epochs. When the validation set loss does not decrease for 10 consecutive epochs, the model parameters are saved and training is terminated.

[0039] The deep spatiotemporal fusion network model is implemented based on the PyTorch deep learning framework. Network parameters are initialized using a Kaiming normal distribution, with a training batch size of 32 and a training epoch count of 100. The weather-irradiance correlation feature matrix is ​​standardized before input, with a mean of 0 and a standard deviation of 1. Training is performed on an NVIDIA V100 GPU. For each training batch, forward propagation calculates the predicted value, backpropagation calculates the gradient, the optimizer updates the network parameters, and the training log records the loss function value and evaluation metrics. Model checkpoints are saved periodically. The hyperparameters of the deep spatiotemporal fusion network model are determined through grid search. The number of convolutional channels in the spatial attention branch is set to 64, 128, and 256; the number of dilated convolutional layers in the temporal convolution branch is set to 8; the hidden layer dimension of the gated fusion mechanism is set to 512; and the regularization coefficient of the joint loss function is... The initial value of the weight allocation ratio coefficient ω is set to 0.01 and 0.5. During training, the value of ω is dynamically adjusted according to the mean square error of the validation set.

[0040] The inference phase of the deep spatiotemporal fusion network model uses the trained model parameters as input, a real-time 3D meteorological feature tensor, and outputs a sequence of predicted irradiance values ​​for future time windows. During inference, a model evaluation mode is enabled, network weights are fixed, Dropout layers are disabled, and the output results are de-normalized to restore physical dimensions. Deployment of the deep spatiotemporal fusion network model is accelerated using TensorRT, optimizing the computational graph structure to reduce inference latency. The model service is encapsulated as a RESTful API, receiving meteorological data requests and returning prediction results in JSON format. Version management of the deep spatiotemporal fusion network model uses a Git system; training scripts and model weights are stored in a repository. Model performance is evaluated using a test set to calculate root mean square error and mean absolute error.

[0041] The spatial attention branch of the deep spatiotemporal fusion network model implements deformable convolutional modules using a DeformableConvNets structure. Offset convolutional layers use standard 3×3 convolutions with twice the kernel size as the output channels. Offset values ​​are added to the regular convolutional grid coordinates. These deformable convolutional modules capture the non-rigid deformation features of clouds. The causal dilated convolutions in the temporal convolution branch ensure that the output at each time step depends only on the current and historical inputs. Stacked dilated convolutions increase the receptive field, capturing long-term temporal dependencies. The training process of the deep spatiotemporal fusion network model uses gradient clipping to limit the gradient norm and prevent gradient explosion. A cosine annealing strategy is used for learning rate scheduling, gradually reducing the learning rate for smooth convergence. An early stopping mechanism preserves the model parameters with the best performance on the validation set, avoiding overfitting to the training data. The spatiotemporal decoder gating fusion mechanism of the deep spatiotemporal fusion network model is implemented as a sigmoid function-based gating unit, controlling the information flow ratio of spatial and temporal features. The decoder hidden state is initialized with the encoder's output at the last time step. Each decoding time step takes the current hidden state and encoder features as input, outputs the predicted value, and updates the hidden state. The training data augmentation of the deep spatiotemporal fusion network model employs time series flipping and Gaussian noise addition strategies to improve the model's generalization ability. Mixed-precision computation is used to accelerate the training process and reduce memory usage. Interpretive analysis of the deep spatiotemporal fusion network model visualizes spatial attention weights through gradient-weighted class activation mapping, helping to understand the model's decision-making process.

[0042] The deep spatiotemporal fusion network model's architecture supports variable-length inputs, handling sequences of different lengths through a masking mechanism. During training, a teacher-forced strategy is used, employing true values ​​as decoder input to accelerate convergence. In the inference phase, an autoregressive approach is used, using the predicted value from the previous time step as the input for the next time step. The ensemble learning strategy trains multiple models, with weighted averaging of prediction results to improve robustness. The model update mechanism periodically fine-tunes parameters using new data to adapt to changes in data distribution. The anomaly detection module identifies outliers in the input data to avoid erroneous predictions. A monitoring system records model prediction biases and alerts administrators when thresholds are triggered. The regularization term in the deep spatiotemporal fusion network model uses the L2 norm to constrain model parameter values, preventing overfitting to training data. The regularization coefficient γ controls the regularization strength; a larger γ value indicates stronger constraints on model complexity. The training data partitioning employs time-series cross-validation to ensure the temporal continuity of the training and validation sets. Validation set performance serves as the primary criterion for model selection. The hyperparameter tuning of the deep spatiotemporal fusion network model uses the Bayesian optimization algorithm to automatically search for the optimal combination of hyperparameters and improve model performance.

[0043] During the training of the deformable convolutional module of the deep spatiotemporal fusion network model, offset learning employs a separate learning rate setting, with the offset parameter's learning rate set to 0.1 times that of the ordinary convolutional parameter learning rate, stabilizing the training process. The spatial attention branch of the deep spatiotemporal fusion network model calculates the importance weights for each location in the feature map, and the weight values ​​are normalized using the softmax function to highlight key region features. The temporal convolutional branch of the deep spatiotemporal fusion network model uses a gating mechanism to control information flow; the forget gate determines the proportion of historical information retained, and the input gate controls the degree of new information addition. The training process of the deep spatiotemporal fusion network model uses label smoothing technology to convert hard labels into soft labels, improving the model's generalization ability. The inference phase of the deep spatiotemporal fusion network model uses a beam search algorithm to retain multiple candidate prediction sequences and select the optimal prediction result. The deployment environment of the deep spatiotemporal fusion network model is configured with a model service monitoring system to monitor model inference latency and resource usage in real time, ensuring service stability. The version rollback mechanism of the deep spatiotemporal fusion network model ensures rapid recovery to a stable version in the event of a model update failure, maintaining business continuity.

[0044] Example 4: Real-time acquisition of current meteorological satellite remote sensing data streams includes establishing a real-time data interface with the meteorological satellite data center. This interface establishes a long-lived connection with the meteorological satellite data center server based on the HTTP protocol, receiving the latest released satellite remote sensing observation data streams. The data stream transmission format uses the HDF5 standard to encapsulate remote sensing image data and metadata in the visible light, infrared, and water vapor bands. The received raw data undergoes radiometric calibration, converting the digital quantization values ​​of the satellite sensor into radiance values ​​at the top of the atmosphere. The numerical conversion is completed using the calibration coefficient matrix provided by the satellite platform. Geometric correction eliminates image distortion caused by satellite attitude changes and Earth curvature. A polynomial correction model, combined with digital elevation data, is used to reposition each pixel. The corrected data is then stitched together with historical data to form a continuous time series. The stitching operation is based on a unified timestamp and spatial reference system. The new data block and the historical data block are contiguous in the time dimension and maintain complete spatial consistency, forming a seamless spatiotemporal data cube. Refer to Table 1, the real-time data stream processing parameter configuration table, which lists the key parameter settings for system operation.

[0045] Table 1: Real-time Meteorological Satellite Data Stream Processing Parameter Configuration Table Extracting real-time 3D meteorological feature tensors involves processing real-time data streams using a sliding time window approach. The sliding time window size is set to 24 hours, and the window sliding step size is consistent with the satellite data update frequency of 15 minutes. Data within each window undergoes the same multi-scale feature extraction process as in the training phase. The feature extraction network uses the network weight parameters obtained during the training phase to ensure consistency in feature extraction. An incremental feature update mechanism is introduced. This mechanism detects the degree of difference between newly arrived data and processed data, recalculating features only for data regions that have changed, while directly reusing cached feature values ​​for unchanged regions. A feature caching module retains key features from historical windows. This module uses an LRU (Least Recently Used) eviction policy to manage memory space, keeping recently used feature data in the cache and transferring historical features exceeding the cache capacity to a distributed file system.

[0046] The real-time data interface is implemented in a multi-threaded asynchronous processing mode. The main thread is responsible for establishing network connections, while worker threads parse data packets in parallel. The data receiving buffer size is set to 1GB to prevent data overflow. Radiometric calibration processing uses a calibration parameter lookup table provided by the satellite platform to convert raw digital values ​​into radiance values. The calibration parameters are updated quarterly to ensure data accuracy. Geometric correction processing uses a ground control point network. The control point coordinates are derived from a high-precision digital map. The corrected imagery is resampled to a standard latitude and longitude grid, with a unified spatial resolution of 0.1 degrees × 0.1 degrees. The data stitching process undergoes quality checks, verifying the continuity of data timestamps and the integrity of spatial coverage. Data retransmission requests are triggered when missing data is detected.

[0047] When processing real-time data streams using a sliding time window, the window sliding is implemented using a circular buffer, with new data overwriting the oldest historical data while maintaining a constant window size. The multi-scale feature extraction process calls a pre-trained deep convolutional neural network model. The model input is multi-band satellite data within the sliding window, and the output is a 128-dimensional feature vector sequence. An incremental feature update mechanism calculates the difference between two consecutive frames of data, with a difference threshold set to 0.3. If the threshold is exceeded, local feature recalculation is triggered. The feature caching module is designed with a hierarchical storage structure: hot data is stored in the in-memory database Redis, warm data is stored in a solid-state drive array, and cold data is archived in the distributed storage system Ceph. The real-time 3D meteorological feature tensor extraction system is deployed on a cloud computing platform, using containerization technology to encapsulate the processing modules and managing container instances through Kubernetes orchestration tools. The system monitoring module tracks the real-time operation of the data processing pipeline, recording data reception rate, feature calculation latency, and cache hit rate performance metrics. An anomaly handling mechanism detects data stream interruptions or quality anomalies, automatically switching to a backup data source or alerting maintenance personnel. Feature tensor extraction results are distributed to the prediction module via a message queue, achieving decoupled asynchronous processing.

[0048] The real-time data interface supports a reconnection mechanism, automatically attempting to re-establish the connection when the network is interrupted and retransmitting missing data segments after transmission resumes. Radiometric calibration is performed independently for each band: visible light band calibration uses solar irradiance reference values, infrared band calibration uses blackbody radiation reference values, and water vapor band calibration uses an atmospheric absorption model. Geometric correction generates an accuracy report for each image, including root mean square error and maximum displacement error indices. Spatiotemporal consistency checks are implemented during data stitching to ensure smooth numerical transitions between adjacent data blocks in overlapping areas. The sliding time window processing employs a parallel computing architecture, processing multiple time windows simultaneously and utilizing GPUs to accelerate feature extraction calculations. A multi-scale feature extraction network optimizes inference speed, using the TensorRT acceleration library to optimize the model computation graph and reduce inference latency. The incremental feature update mechanism is implemented as a difference detection algorithm, comparing the magnitude of pixel value changes between consecutive frames and performing convolution operations only on areas of change. The feature caching module implements data prefetching, predicting the next required data based on access patterns and loading it into memory in advance to reduce waiting time.

[0049] The real-time 3D meteorological feature tensor extraction system is integrated with a metadata management system. Each feature tensor comes with a complete metadata description, including data source, processing time, and quality identifiers. System performance optimizations include memory pool management to reduce memory allocation overhead and pipelined computation tasks to improve hardware utilization. Fault tolerance mechanisms handle computing node failures, automatically migrating processing tasks to healthy nodes to ensure service continuity. Feature data is compressed and stored to reduce storage space usage, using a lossy compression algorithm to balance storage costs and feature accuracy. The real-time data interface implements flow control to prevent data influx from exceeding processing capacity, employing a token bucket algorithm to smooth the data flow. Radiometric calibration processing incorporates quality control flags to identify anomalous data points discovered during calibration. Geometric correction processing uses parallel computing technology, dividing large-scale images into multiple blocks for parallel correction to improve processing speed. Data stitching algorithms handle data sources with different spatiotemporal resolutions, unifying them to a standard grid through spatial interpolation. Sliding time window processing supports variable window size configuration, dynamically adjusting the window length according to prediction needs. The multi-scale feature extraction network supports dynamically loading different versions of model weights to adapt to different satellite data sources. An incremental feature update mechanism records the feature calculation history, supporting feature backtracking and recalculation. The feature caching module implements a distributed shared cache, allowing multiple computing nodes to access a unified cache cluster to avoid redundant calculations. The real-time 3D meteorological feature tensor extraction system is integrated with a permission management system to control data access permissions and ensure data security. System logs record detailed operation logs, supporting auditing and troubleshooting needs.

[0050] Example 5: Outputting the predicted irradiance sequence within a preset future time window involves dividing the prediction time window into multiple time intervals. These intervals are set to three gradients based on business requirements: 15 minutes, 1 hour, and 4 hours. Each interval corresponds to a different prediction confidence level. The 90% confidence interval for the 15-minute interval is calculated using the probability distribution output by the model; the 80% confidence interval for the 1-hour interval; and the 70% confidence interval for the 4-hour interval. The prediction results for adjacent time intervals are smoothed using the exponentially weighted moving average method in time series filtering algorithms, with a weighting coefficient of 0.3. This smoothing eliminates abrupt changes in the predicted value sequence, generating a structured output containing the predicted values ​​and confidence intervals. The structured output is encapsulated in JSON-LD format and includes a timestamp, predicted irradiance value, upper and lower bounds of the confidence interval, and data quality identifier fields, for use by downstream photovoltaic power plant power prediction systems in decision-making. The system regularly collects statistical deviations between actual observed irradiance data and predicted results. The deviation statistics are calculated using a standardized average deviation index. The deviation calculation process is automatically triggered daily at 00:00 UTC, collecting measured data from 500 ground monitoring stations nationwide and corresponding predicted data for each station, calculating the daily average deviation and root mean square deviation for each station. When the deviation exceeds a preset threshold, a model retraining process is triggered. This preset threshold is dynamically adjusted according to seasonal changes: 15% in summer and 20% in winter. When the deviation exceeds the limit, the system automatically starts the model retraining workflow, extracting training samples from the most recent 90 days from the data warehouse to update the parameters of the deep spatiotemporal fusion network model. A prediction performance monitoring dashboard is established, developed using web visualization technology. It displays the 7-day rolling average prediction accuracy using a line chart, the spatial distribution of regional prediction deviations using a heatmap, and the prediction error statistics under different weather types using a bar chart.

[0051] The prediction time window is divided according to the sunrise and sunset time patterns, dynamically adjusting the window start point and interval length. The summer prediction window runs from 05:00 to 20:00, and the winter prediction window runs from 07:00 to 18:00. The prediction confidence level for each time interval is calculated using a quantile regression model, with the 10%, 50%, and 90% quantiles of the model's output prediction values ​​serving as the confidence interval boundaries. The smoothing algorithm uses different parameters for different time scales: a moving average with a weighting coefficient of 0.5 is used for 15-minute interval predictions, exponential smoothing with a weighting coefficient of 0.3 is used for 1-hour interval predictions, and low-pass filtering with a weighting coefficient of 0.1 is used for 4-hour interval predictions. The smoothed prediction sequence maintains the overall trend while eliminating short-term fluctuations. The structured output format follows the OGC sensor observation standard, with each prediction point including latitude and longitude coordinates, prediction time, irradiance value, uncertainty, and quality control indicators. Downstream systems obtain data via REST API. Actual observed irradiance data is collected through a meteorological data exchange platform. Ground observation data is downloaded daily from the National Meteorological Information Center in CSV format, including station number, observation time, and irradiance value fields. Data quality control is performed before calculating deviation statistics, removing obvious outliers such as noon night values ​​and observations exceeding theoretical maximum values. The standardized average deviation is calculated as the arithmetic mean of the deviations at each station. The model retraining process is a fully automated pipeline, including five stages: data preprocessing, feature engineering, model training, model evaluation, and model deployment. The newly trained model is compared with the old model through A / B testing; if the performance improvement exceeds 5%, the model is automatically replaced in the production environment.

[0052] The predictive performance monitoring dashboard is deployed on a dedicated server using a microservice architecture. The data interface can handle 1000 concurrent requests per second. The dashboard view supports time range selection, region filtering, and metric comparison. Monitoring data is stored in the time-series database InfluxDB, retaining performance metrics from the past three years and supporting historical data queries and trend analysis. An anomaly detection algorithm monitors prediction deviations in real time, sending alarm messages to the system administrator when the deviation exceeds a threshold for three consecutive time points. The prediction time window division algorithm also considers weather conditions, extending the weight of the 4-hour interval under sunny conditions and increasing the sampling frequency by 15 minutes under cloudy or rainy conditions. Time interval switching is automatically determined based on real-time weather cloud maps. Smoothing processing employs adaptive parameters for different seasons: reducing the weight coefficient to preserve detailed features during drastic summer irradiance changes and increasing the weight coefficient to enhance the smoothing effect during gentler winter irradiance changes. Structured output includes metadata descriptions, including the prediction model version, data processing time, and data source identifier, facilitating data quality traceability for downstream systems.

[0053] During the actual irradiance data collection process, missing station data is addressed. When data for a particular station is missing, spatial interpolation is used to estimate the station's value. The interpolation algorithm employs Kriging spatial interpolation, taking into account altitude and topographical factors. Deviation statistics are calculated separately for different weather types: sunny, cloudy, and rainy weather, generating categorized performance evaluation reports. A rollback mechanism is implemented in the model retraining process. When the performance of the new model is inferior to the old model, it automatically reverts to the previous version to ensure system stability. The prediction performance monitoring dashboard provides multi-dimensional analysis capabilities, analyzing prediction performance by geographical region, time scale, and weather type, and supports automatic generation and email delivery of performance reports. The monitoring system sets a performance baseline. When the prediction accuracy consistently falls below the baseline, a root cause analysis process is triggered to check data quality, feature engineering, and model status.

[0054] The prediction time window division supports manual adjustment, allowing photovoltaic power plant operators to customize the prediction interval and window length through the management interface to meet specific business needs. The smoothing algorithm offers multiple filter options, including mean filtering, median filtering, and Kalman filtering, automatically selecting the optimal filter based on prediction error characteristics. Structured output supports multiple data formats, including JSON, XML, and CSV, meeting the integration requirements of different downstream systems. Actual observed irradiance data collection achieves near real-time updates, with ground monitoring data pushed to the central database every 5 minutes, and the frequency of deviation statistics calculation increased to once every 15 minutes, achieving near real-time monitoring of prediction performance. The model retraining process incorporates incremental learning technology, using new data daily to fine-tune the model, reducing the resource consumption of full retraining. The prediction performance monitoring dashboard adds an early warning function, issuing warnings in advance when prediction performance shows a downward trend, reminding maintenance personnel to monitor the system status. Data quality identification in the structured output is calculated based on multiple dimensions, including data integrity, temporal consistency, and spatial consistency scores; data points with scores below a threshold are marked as unreliable data. During the actual irradiance data collection process, data validation rules are implemented to check the reasonableness of the data range and the continuity of time. Invalid data is automatically discarded and logged. In the feature engineering stage of the model retraining process, feature importance is recalculated, feature dimensions with decreasing importance are removed, and newly discovered significant features are added. User access is managed hierarchically for the predictive performance monitoring dashboard; power plant operators can only view data for their designated area, while system administrators can view performance indicators for the entire system.

[0055] The prediction time window is divided to take into account special weather events, such as solar eclipses, sandstorms, and other rare weather phenomena. The time interval is automatically adjusted to a high-frequency 5-minute monitoring interval to ensure the capture of rapid changes in irradiance. The smoothing algorithm automatically reduces the smoothing intensity during sudden weather changes, preserving the detailed features of irradiance changes and avoiding the loss of important information due to over-smoothing. The structured output interface supports a subscription push mode, allowing downstream systems to subscribe to prediction data from specific sites, with new prediction results automatically pushed to designated endpoints. The actual observed irradiance data collection system implements automatic failover, automatically switching to a backup data source when the primary data source is unavailable, ensuring continuous data collection. Deviation statistics calculation considers the impact of day and night alternation, with daytime and nighttime deviations calculated separately to avoid interference from zero values ​​at night on the overall deviation assessment. The model evaluation stage in the model retraining process uses multiple indicators for comprehensive evaluation, including mean absolute error, root mean square error, and correlation coefficient, ensuring improved overall performance of the new model. The prediction performance monitoring dashboard supports multi-terminal access, allowing monitoring data to be viewed on desktop computers, tablets, and mobile phones, meeting monitoring needs in different scenarios.

[0056] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0057] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A short-term irradiance AI prediction method combining meteorological satellite data, characterized in that, Includes the following steps: Historical meteorological satellite remote sensing data sequences and ground irradiance monitoring data sequences of the target area were collected to form a training sample set; Multi-scale spatiotemporal feature extraction is performed on historical meteorological satellite remote sensing data sequences to generate a three-dimensional meteorological feature tensor containing cloud dynamic features, atmospheric transmission features, and surface reflection features. The ground irradiance monitoring data sequence was spatiotemporally aligned and matched with the three-dimensional meteorological feature tensor to establish a meteorological-irradiance correlation feature matrix. A deep spatiotemporal fusion network model is constructed, and the meteorological-irradiance correlation feature matrix is ​​input into the deep spatiotemporal fusion network model for end-to-end training to obtain a trained irradiance prediction model. The system acquires real-time meteorological satellite remote sensing data streams, extracts real-time three-dimensional meteorological feature tensors, inputs them into the trained irradiance prediction model, and outputs a sequence of predicted irradiance values ​​within a preset future time window.

2. The short-term irradiance AI prediction method combining meteorological satellite data according to claim 1, characterized in that, The multi-scale spatiotemporal feature extraction of historical meteorological satellite remote sensing data sequences includes: The historical meteorological satellite remote sensing data sequence is divided into multiple time segments, each of which contains remote sensing image data in the visible light band, infrared band, and water vapor band. Spatial pyramid features were extracted from the remote sensing image data of each band to obtain local meteorological features at different spatial scales. Temporal convolutional networks are used to model the time dependence of local meteorological features in each band, generating multi-band meteorological features with temporal correlation. Multi-band meteorological features are fused through a cross-band attention mechanism to ultimately form a three-dimensional meteorological feature tensor that includes cloud dynamics, atmospheric transmission, and surface reflection.

3. The short-term irradiance AI prediction method combining meteorological satellite data according to claim 2, characterized in that, The spatiotemporal alignment and matching of the ground irradiance monitoring data sequence with the three-dimensional meteorological feature tensor includes: Based on the geographical location information of ground monitoring stations, meteorological feature vectors corresponding to the spatial location are extracted from the three-dimensional meteorological feature tensor. Time interpolation processing is performed on the ground irradiance monitoring data sequence to ensure that its sampling frequency is consistent with the temporal resolution of the meteorological feature vector; The mutual information between meteorological feature vectors and irradiance monitoring data is calculated, and feature dimensions with mutual information exceeding the threshold are selected to form a meteorological-irradiance correlation feature matrix.

4. The short-term irradiance AI prediction method combining meteorological satellite data according to claim 1, characterized in that, The construction of the deep spatiotemporal fusion network model includes: Design a network architecture consisting of a spatiotemporal encoder and a spatiotemporal decoder, wherein the spatiotemporal encoder contains parallel spatial attention branches and temporal convolution branches; Deformable convolutional modules are deployed in the spatial attention branch to capture non-rigid deformation features during cloud movement; A causal dilated convolution structure is used in the temporal convolution branch to ensure that the prediction process conforms to the temporal causality law; The spatiotemporal decoder integrates the output features of the spatial and temporal branches through a gating fusion mechanism to generate irradiance predictions for future time steps.

5. The short-term irradiance AI prediction method combining meteorological satellite data according to claim 4, characterized in that, The training process of the deep spatiotemporal fusion network model includes: The meteorological-irradiance correlation feature matrix is ​​divided into a training set and a validation set, and the mean squared error and correlation coefficient are set as joint loss functions. An adaptive moment estimation algorithm is used to optimize network parameters, and the weight allocation ratio of spatial and temporal branches is dynamically adjusted during training. The prediction performance on the validation set is monitored by an early stopping mechanism, and the training process is terminated when the performance does not improve for several consecutive training cycles.

6. The short-term irradiance AI prediction method combining meteorological satellite data according to claim 1, characterized in that, The real-time acquisition of current meteorological satellite remote sensing data stream includes: Establish a real-time data interface with the meteorological satellite data center to receive the latest satellite remote sensing observation data; The received raw data is subjected to radiometric calibration and geometric correction to eliminate the effects of sensor noise and projection distortion. The corrected data is spliced ​​with historical data to form a continuous time series, maintaining consistency in the extraction of spatiotemporal features.

7. The short-term irradiance AI prediction method combining meteorological satellite data according to claim 6, characterized in that, The extraction of the real-time three-dimensional meteorological feature tensor includes: Real-time data streams are processed using a sliding time window approach, and the data in each window undergoes the same multi-scale feature extraction process as in the training phase. An incremental feature update mechanism is introduced, which recalculates only the feature dimensions corresponding to newly arrived data; The feature caching module retains key features of historical windows, reducing the overhead of redundant calculations.

8. The short-term irradiance AI prediction method combining meteorological satellite data according to claim 1, characterized in that, The output sequence of predicted irradiance values ​​within a future preset time window includes: The prediction time window is divided into multiple time intervals, each interval corresponding to a different prediction confidence level; The prediction results of adjacent time intervals are smoothed to eliminate abrupt changes in the predicted value sequence; Generate structured outputs containing predicted values ​​and confidence intervals for use in downstream system decision-making.

9. The short-term irradiance AI prediction method combining meteorological satellite data according to claim 8, characterized in that, Also includes: Regularly collect the statistical measures of the deviation between actual observed irradiance data and predicted results; When the deviation exceeds a preset threshold, the model retraining process is triggered to update the parameters of the deep spatiotemporal fusion network model; Establish a predictive performance monitoring dashboard to visualize the trend of prediction accuracy changes at different time scales.

10. A short-term irradiance AI prediction system combining meteorological satellite data, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the short-term irradiance AI prediction method combining meteorological satellite data as described in any one of claims 1 to 9.

Images