A solar radiation prediction method and device based on multi-source satellite data fusion, electronic equipment, and storage medium

Abstract

The present application relates to the technical field of meteorological observation, and more particularly to a solar radiation prediction method and device based on multi-source satellite data fusion, electronic equipment and storage medium, the solar radiation prediction method based on multi-source satellite data fusion provided by the present application firstly introduces the infrared brightness temperature data of the stationary satellite, and constructs an end-to-end deep learning prediction model, effectively overcoming the dark start prediction blind area caused by the lack of visible light before and after sunrise, realizing continuous prediction throughout the day, especially during the dawn period, and expanding the prediction coverage capability. Secondly, by calibrating the regional adaptability parameters of the physical inversion model, the local accuracy of the historical reference radiation field is improved, providing a more reliable training basis for the prediction model. Finally, the output high spatio-temporal resolution and physically consistent solar radiation prediction product can directly support photovoltaic power station output prediction and power grid climbing management and other key businesses, enhancing the practicality and reliability of the prediction results.

Description

Technical Field

[0001] This invention relates to the field of meteorological observation technology, and in particular to a method, device, electronic equipment, and storage medium for predicting solar radiation based on multi-source satellite data fusion. Background Technology

[0002] With the rapid development and large-scale application of photovoltaic power generation technology, high-precision and timely ultra-short-term forecasts of total solar radiation (GHI) are of great significance for the safe and stable operation of power systems and the efficient consumption of new energy. Currently, mainstream GHI forecasting methods mainly include numerical weather prediction, physical inversion methods based on satellite visible light channels, and statistical extrapolation methods based on historical observation data. However, each of these methods has limitations: numerical weather prediction has a long update cycle and limited spatiotemporal resolution, making it difficult to effectively capture the rapid evolution of cloud formations; satellite inversion methods based on visible light channels perform well during the daytime, but they rely on effective solar illumination conditions and cannot provide reliable input during the transition from night to dawn; statistical methods based on historical data are limited by their time-series extrapolation capabilities and are insufficient in responding to abrupt changes under complex meteorological conditions.

[0003] The core technical bottleneck in current GHI prediction lies in the "dark start" blind zone, which occurs approximately 0-4 hours before and after sunrise. During this period, the visible light channel cannot effectively retrieve the cloud index, causing traditional physical models to fail and prediction sequences to be interrupted. This period is precisely the critical stage for photovoltaic power plants to ramp up their output from zero; the lack of prediction directly impacts grid dispatch and ramp-up management. Although satellite infrared channels can acquire cloud top brightness temperature information around the clock, and there is a certain physical correlation between this information and surface irradiance, current technologies have not yet established an effective mechanism for fusing infrared and visible light data, nor do they possess a prediction model capable of uniformly processing all-day, multi-scale spatiotemporal characteristics. This has resulted in the long-standing inability to systematically solve the "dark start" problem.

[0004] Therefore, there is an urgent need to develop a new GHI prediction method that can integrate multi-source satellite observations, has regional adaptive capabilities, and combines advanced spatiotemporal deep learning extrapolation technology. This method can overcome the data gaps during the "dark start" period and achieve continuous, stable, and high-precision ultra-short-term solar radiation forecasts from night to day, providing reliable technical support for photovoltaic power prediction and smart grid dispatch. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide a method, device, electronic device, and storage medium for predicting solar radiation based on multi-source satellite data fusion.

[0006] In a first aspect, embodiments of the present invention provide a solar radiation prediction method based on multi-source satellite data fusion, the method comprising:

[0007] Acquire geostationary satellite multispectral remote sensing data, infrared remote sensing data, and ground observation data;

[0008] Based on visible light data from geostationary satellite multispectral remote sensing data and historical radiation sequences obtained through a physical inversion model, key parameters of the physical inversion model are optimized by combining ground observation data to generate a spatiotemporally continuous reference radiation field adapted to the region.

[0009] Spatiotemporal continuous reference radiation field and infrared brightness temperature data from geostationary satellites are spatiotemporally standardized and feature-encoded, and fused to form a standardized model input tensor containing historical radiation sequences, infrared features and time period information;

[0010] The standardized model input tensor is fed into a trained deep learning prediction model. The deep learning prediction model adaptively fuses the spatiotemporal dependencies of multi-source features through an encoder-decoder architecture and directly outputs a future time period and a time-continuous prediction field of the total solar radiation on the Earth's surface.

[0011] In conjunction with the first aspect, the steps for obtaining historical radiation sequences based on visible light data and through a physical inversion model include:

[0012] Based on visible light data, ground observation data are geocalibrated and atmospherically corrected to obtain apparent reflectance data;

[0013] Based on apparent reflectance data and a pre-set surface reference reflectance, a cloud index reflecting cloud cover is calculated.

[0014] The cloud index is converted into a cloud transmittance function that characterizes the degree of attenuation of solar radiation by clouds;

[0015] Based on the sun's position, atmospheric parameters, and surface information, the theoretical total solar radiation under clear skies is calculated using a clear-sky radiation model. The theoretical total solar radiation under clear skies is then corrected using a cloud transmittance function to obtain a preliminary historical radiation sequence.

[0016] In conjunction with the first aspect, the steps for optimizing key parameters of the physical inversion model using ground observation data and generating a regionally adapted spatiotemporally continuous reference radiation field include:

[0017] Based on ground observation data and combined with the clear sky radiation model, the actual clear sky index is calculated.

[0018] The actual clear sky index and cloud index are spatiotemporally matched and compared to obtain the comparison results.

[0019] Based on the comparison results, the key functions related to cloud index and radiation transmittance in the physical inversion model were regionalized and calibrated using a fitting method.

[0020] The inversion is re-executed using the calibrated model parameters to generate a spatiotemporally continuous reference radiation field that is adapted to the region.

[0021] In conjunction with the first aspect, the infrared brightness temperature data includes the infrared brightness temperature data of channel 7 and channel 13 of the geostationary satellite;

[0022] The steps involved in spatiotemporally standardizing and feature-encoding the spatiotemporally continuous reference radiation field and the infrared brightness temperature data from geostationary satellites, and fusing them to form a standardized model input tensor containing historical radiation sequences, infrared features, and time period information, include:

[0023] Spatial grid unification and temporal resolution alignment are performed on spatiotemporally continuous reference radiation field and infrared brightness temperature data to obtain spatiotemporally consistent radiation field and brightness temperature data.

[0024] A normalization method is used to map the values ​​of the reference radiation field and infrared brightness temperature data to a preset range; the scaling parameters used in the normalization are statistically obtained and saved based on the training dataset and reused during prediction.

[0025] The timestamp information of the data is periodically encoded to generate time features that characterize the daily cycle of solar radiation.

[0026] The normalized reference radiation field, infrared brightness temperature data, and time features are spliced ​​along the feature channel dimension to form a unified five-dimensional spatiotemporal tensor, which serves as the input tensor for the standardized model.

[0027] In conjunction with the first aspect, the steps of periodically encoding the timestamp information of the data to generate time features characterizing the daily cycle of solar radiation include:

[0028] Extract the timestamp information corresponding to the reference radiation field or infrared brightness temperature data to obtain the value representing a specific time of day;

[0029] Convert the numerical value into a continuous time variable that changes cyclically within a preset period;

[0030] By inputting continuous time variables into a set of periodic functions with fixed phase relationships, a time feature vector containing multiple components is generated to encode the diurnal cycle of solar radiation.

[0031] In conjunction with the first aspect, the encoder-decoder architecture of deep learning prediction models includes:

[0032] The multi-branch coding unit is used to extract three types of features in parallel from the input tensor of the standardized model: the first coding branch extracts the spatiotemporal evolution features of the historical radiation sequence to obtain the main feature tensor; the second coding branch extracts the auxiliary meteorological features of the infrared brightness temperature data to obtain the auxiliary feature tensor; and the third coding branch extracts and expands the encoded time period information to obtain the time embedding feature tensor. Each coding branch is based on a three-dimensional convolutional structure to achieve layer-by-layer downsampling.

[0033] The spatiotemporal feature fusion unit is used to adaptively fuse the main feature tensor, auxiliary feature tensor, and temporal embedding feature tensor. The spatiotemporal feature fusion unit uses a dual-branch attention mechanism and is guided by the auxiliary feature tensor to perform weighted modulation of multi-source features to enhance the sensitivity to key meteorological areas.

[0034] The decoding and reconstruction unit is used to recover the spatial resolution step by step through upsampling operations based on the fused features. During the decoding process, the decoding and reconstruction unit introduces spatial detail features from the shallow layer of the multi-branch coding unit and embeds temporal codes representing future moments to generate a normalized radiation field sequence for future time periods.

[0035] The denormalized output unit is used to restore the normalized radiation field sequence to its true physical dimensions through a linear inverse transformation based on the minimum and maximum values ​​of the training data, and output the final predicted field of total solar radiation on the Earth's surface.

[0036] In conjunction with the first aspect, the training steps for a deep learning prediction model include:

[0037] The deep learning prediction model is trained using a training sample set that includes historical reference radiation fields, corresponding infrared brightness temperature data, and future reference radiation field labels.

[0038] During training, a composite objective function combining pixel-level error and sequence consistency loss is used;

[0039] The training process is optimized by applying dynamic learning rate adjustment and gradient clipping strategies until the model converges.

[0040] Secondly, embodiments of the present invention also provide a solar radiation prediction device based on multi-source satellite data fusion, the device comprising:

[0041] The acquisition module is used to acquire geostationary satellite multispectral remote sensing data, infrared remote sensing data, and ground observation data;

[0042] The generation module is used to obtain historical radiation sequences based on visible light data from geostationary satellite multispectral remote sensing data and through a physical inversion model. It also optimizes the key parameters of the physical inversion model by combining ground observation data to generate a spatiotemporally continuous reference radiation field that is adapted to the region.

[0043] The fusion module is used to perform spatiotemporal standardization and feature encoding on the spatiotemporal continuous reference radiation field and the infrared brightness temperature data of geostationary satellites, and fuse them to form a standardized model input tensor containing historical radiation sequences, infrared features and time period information.

[0044] The prediction module is used to input the standardized model input tensor into the trained deep learning prediction model. The deep learning prediction model adaptively fuses the spatiotemporal dependencies of multi-source features through an encoder-decoder architecture and directly outputs the predicted field of total solar radiation on the Earth's surface for future time periods that are continuous in time.

[0045] Thirdly, the present invention provides an electronic device, the electronic device including a memory and a processor, the memory being used to store a computer program, and the processor running the computer program to cause the electronic device to perform the above-described method.

[0046] Fourthly, the present invention provides a readable storage medium storing computer program instructions, which, when read and executed by a processor, perform the above-described method.

[0047] The embodiments of the present invention bring the following beneficial effects: The present invention provides a solar radiation prediction method, device, electronic device, and storage medium based on multi-source satellite data fusion. The method includes: acquiring geostationary satellite multispectral remote sensing data, infrared remote sensing data, and ground observation data; obtaining historical radiation sequences based on visible light data and through a physical inversion model, optimizing key parameters of the physical inversion model by combining ground observation data, and generating a regionally adapted spatiotemporally continuous reference radiation field; performing spatiotemporal standardization and feature encoding on the spatiotemporally continuous reference radiation field and the infrared brightness temperature data of geostationary satellites, fusing them to form a standardized model input tensor containing historical radiation sequences, infrared features, and time period information; inputting the standardized model input tensor into a trained deep learning prediction model, the deep learning prediction model adaptively fusing the spatiotemporal dependencies of multi-source features through an encoder-decoder architecture, and directly outputting a future time period, time-continuous prediction field of total solar radiation on the Earth's surface.

[0048] This invention provides a solar radiation prediction method based on multi-source satellite data fusion. First, by introducing infrared brightness temperature data from geostationary satellites and constructing an end-to-end deep learning prediction model, it effectively overcomes the dark-start prediction blind zone caused by the lack of visible light before and after sunrise, achieving continuous prediction throughout the day, especially during dawn, and expanding forecast coverage. Second, by calibrating the physical inversion model with regional adaptive parameters, the localization accuracy of the historical reference radiation field is improved, providing a more reliable training foundation for the prediction model. Finally, the output high spatiotemporal resolution and physically consistent solar radiation prediction product can directly support key operations such as photovoltaic power plant output prediction and grid ramp management, enhancing the practicality and reliability of the prediction results.

[0049] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained in accordance with the structures particularly pointed out in the description, claims and drawings.

[0050] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0051] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0052] Figure 1 A flowchart illustrating a solar radiation prediction method based on multi-source satellite data fusion, provided as an embodiment of the present invention;

[0053] Figure 2 A flowchart for solar radiation prediction based on multi-source satellite data fusion is provided for embodiments of the present invention.

[0054] Figure 3 A schematic diagram of a solar radiation prediction device based on multi-source satellite data fusion provided in an embodiment of the present invention;

[0055] Figure 4 This is a schematic diagram of the electronic device structure provided in an embodiment of the present invention.

[0056] Figure label:

[0057] 10 - Acquisition module, 20 - Generation module, 30 - Fusion module, 40 - Prediction module;

[0058] 130 - Processor, 131 - Memory, 132 - Bus, 133 - Communication interface. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions 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, 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.

[0060] To facilitate understanding of this embodiment, the technical terms used in this invention will be briefly introduced below.

[0061] Geostationary satellites are meteorological satellites that orbit at an altitude of about 36,000 kilometers above the Earth's equator and have a fixed position relative to the ground, enabling them to continuously observe fixed areas.

[0062] The Heliosat-2 physical model is a widely used physical-empirical model for retrieving total solar radiation from satellite visible light data. It is used to estimate the actual solar irradiance received by the Earth's surface by quantifying the attenuation effect of clouds on solar radiation.

[0063] Infrared remote sensing refers to the technology of using satellites to detect the thermal radiation of objects such as the Earth's surface and cloud tops in the infrared band (usually referring to the electromagnetic band with wavelengths greater than about 1 micrometer). The energy intensity obtained from the detection is calibrated and expressed as brightness temperature, which is the blackbody temperature of the object's radiation equivalent under the satellite's field of view.

[0064] Ground-based observation data refers to the actual value of total solar radiation (GHI) obtained directly from fixed stations on the Earth's surface using specialized radiation measuring instruments (such as a pyranometer). These stations typically form an observation network covering the target area.

[0065] Clear-sky radiation models are a class of fundamental physical models in solar radiation research, used to calculate the theoretical total solar radiation (GHI) that can be received by the Earth's surface under ideal clear-sky (i.e. cloudless) conditions.

[0066] After introducing the technical terms involved in this invention, the application scenarios and design concepts of the embodiments of this invention will be briefly described below.

[0067] Traditional methods for predicting total solar radiation (GHI) on the Earth's surface have a dark start blind zone: satellite inversion based on visible light channels fails around sunrise due to insufficient light, and infrared data is not effectively integrated and utilized, resulting in interruptions in predictions during the dawn transition period, which seriously affects the continuity of photovoltaic power forecasts.

[0068] Based on this, embodiments of the present invention provide a method, apparatus, electronic device, and storage medium for predicting solar radiation based on multi-source satellite data fusion.

[0069] Example 1

[0070] This invention provides a solar radiation prediction method based on multi-source satellite data fusion, combined with... Figure 1 As shown, the method includes:

[0071] S110 acquires geostationary satellite multispectral remote sensing data, infrared remote sensing data, and ground observation data.

[0072] S120 is based on visible light data from geostationary satellite multispectral remote sensing data and obtains historical radiation sequences through a physical inversion model. It combines ground observation data to optimize the key parameters of the physical inversion model and generate a spatiotemporally continuous reference radiation field that is adapted to the region.

[0073] S130 performs spatiotemporal standardization and feature encoding on the spatiotemporal continuous reference radiation field and the infrared brightness temperature data of geostationary satellites, and fuses them to form a standardized model input tensor containing historical radiation sequences, infrared features and time period information.

[0074] S140 inputs the standardized model input tensor into the trained deep learning prediction model. The deep learning prediction model adaptively fuses the spatiotemporal dependencies of multi-source features through an encoder-decoder architecture and directly outputs the predicted field of total solar radiation on the Earth's surface for future time periods and in a time-continuous manner.

[0075] This invention overcomes the prediction blind spot during periods of missing visible light (especially around dawn) by introducing geostationary satellite infrared brightness temperature data as a key input feature. Furthermore, it regionally calibrates key parameters of the physical inversion model using ground observation data, generating a high-precision historical radiation field with strong local adaptability as a prediction benchmark. Based on this, a deep learning prediction model fusing multi-source data is constructed, employing an encoder-decoder architecture to adaptively extract and fuse spatiotemporal features, directly outputting a temporally continuous prediction field of total solar radiation at the Earth's surface for future periods. This avoids error accumulation in traditional serial prediction, ensuring the spatiotemporal consistency and reliability of the prediction results. It provides crucial technical support for photovoltaic power prediction and grid dispatch, ultimately achieving continuous, uninterrupted solar radiation prediction throughout the day.

[0076] In step S110, the geostationary satellite multispectral remote sensing data refers to the Earth-atmosphere system radiation information synchronously acquired by satellite sensors in multiple discrete spectral channels. In this embodiment, observation data from the Himawari-8 or Himawari-9 geostationary meteorological satellites are specifically used. This data also has a temporal resolution of 15 minutes and an original spatial resolution of approximately 2 kilometers, possessing high timeliness and good spatial coverage capabilities, and can effectively support the dynamic capture of cloud evolution and radiation changes.

[0077] Infrared remote sensing data refers to brightness temperature information obtained from infrared band detection, reflecting the thermal radiation characteristics of the Earth's surface and cloud tops. This embodiment utilizes infrared channel data from the Himawari satellite, particularly channel 7 (approximately 3.9 micrometers) and channel 13 (approximately 10.4 micrometers). These channels are sensitive to physical properties such as cloud top height, cloud phase (water cloud or ice cloud), and cloud thickness, and possess all-weather observation capabilities. The data temporal resolution is also 15 minutes, and the original spatial resolution is approximately 2 kilometers. Therefore, during the dark start-up period when visible light is absent, these data become crucial inputs for predictive models to perceive cloud conditions and infer the attenuation of solar radiation.

[0078] Ground-based observation data refers to the measured values ​​of total solar radiation (GHI) directly obtained through ground-based radiation measurement instruments. This embodiment primarily uses GHI data acquired from the ground-based radiation observation network constructed by the China Meteorological Administration and related institutions. Its functions include: firstly, serving as a true reference value to verify the satellite inversion results; and secondly, performing regional adaptive calibration of key parameters in the Heliosat-2 model, eliminating systematic biases in the original model based on European data, thereby generating a benchmark radiation field that better reflects the climate and geographical characteristics of China.

[0079] In step S110 of this embodiment, by coordinating satellite visible light data, infrared data, and ground-based measured data, a multi-source data system covering all time periods and possessing physical consistency and regional adaptability is constructed, laying a reliable data foundation for subsequent high-precision and continuous solar radiation prediction.

[0080] In conjunction with the first aspect, step S120, based on the visible light data and through a physical inversion model, obtains the historical radiation sequence, specifically including:

[0081] S121, Based on the visible light data, the ground observation data is geocalibrated and atmospherically corrected to obtain apparent reflectance data.

[0082] S122, based on the apparent reflectance data and the preset surface reference reflectance, a cloud index reflecting the cloud coverage status is calculated.

[0083] S123, convert the cloud index into a cloud transmittance function that characterizes the degree of cloud attenuation of solar radiation.

[0084] S124. Based on the sun's position, atmospheric parameters, and surface information, the theoretical total solar radiation under clear skies is calculated using a clear-sky radiation model. The theoretical total solar radiation under clear skies is then corrected using the cloud transmittance function to obtain a preliminary historical radiation sequence.

[0085] Steps S121-S124 use satellite visible light data and the Heliosat-2 physical model to invert the total solar radiation (GHI) on the Earth's surface, so as to gradually transform the raw radiation signal observed by the satellite into the actual surface radiation with physical meaning.

[0086] In step S121, the original satellite observation counts (Digital Number, DN) are converted into apparent surface reflectance suitable for physical calculation. First, through geocalibration, using satellite orbit, attitude, and sensor geometry models, the row and column coordinates of each pixel in the image are accurately converted into real-world latitude and longitude geographic coordinates, laying the foundation for subsequent spatial matching of satellite data and ground station observation data. Next, atmospheric correction is performed. Its physical essence is to quantitatively eliminate interference from atmospheric molecules (Rayleigh scattering) and aerosols (scattering and absorption) along the surface-satellite sensor path of solar radiation using a radiative transfer model. After correction, the resulting data is no longer the atmospheric top radiance including atmospheric contributions, but is standardized to apparent reflectance, which characterizes the ratio of solar radiation flux reflected by the surface to the incident solar radiation flux under the assumption of no atmospheric influence. This data is the unique and crucial initial input for all subsequent inversion calculations.

[0087] Step S122 aims to convert the apparent reflectance obtained in the previous step into a quantitative physical parameter reflecting cloud cover, namely the cloud index. The physical basis for this is that the apparent reflectance of cloud-covered areas is typically significantly higher than the reflectance of the Earth's surface under clear skies. To achieve this conversion, a known surface reference reflectance needs to be introduced. This value is usually obtained from long-term historical images of the target area under clear skies, representing the reflectivity of the surface itself. By comparing the current apparent reflectance with the inherent surface reference reflectance of the pixel and calculating it using a specific formula, the cloud index can be obtained. This index is a dimensionless value between 0 and 1, where 0 typically represents completely clear skies and 1 represents complete thick cloud cover. Essentially, this step compresses complex satellite imagery information into a key intermediate variable that can intuitively and quantitatively reflect the amount of cloud cover above each pixel.

[0088] Step S123 converts the cloud index into a cloud transmittance function. The cloud index reflects the degree of cloud cover, but what ultimately affects the amount of solar radiation received by the Earth's surface is the cloud's ability to attenuate radiation, i.e., cloud transmittance. Therefore, a physical or empirical conversion function is needed to map the cloud index to a cloud transmittance function. This function describes the proportion of solar radiation that can penetrate the cloud layer and reach the ground as the cloud index changes from 0 to 1 (the value range is usually 0 to 1, where 1 represents no attenuation). This relationship is usually non-linear because different phases (water clouds, ice clouds), different thicknesses, and different altitudes of clouds have different radiation attenuation intensities per unit cloud index. This conversion function is one of the most critical and regionally variable parameters in the Heliosat-2 model. Its accuracy directly determines the accuracy of the final radiation inversion results and is also the main object for further regional adaptive optimization using ground observation data.

[0089] Step S124 is the final synthesis step of the physical inversion, aiming to integrate astronomical, atmospheric, surface, and cloud information to calculate the actual total solar radiation (GHI) at the Earth's surface. First, based on precise solar position (solar zenith angle, azimuth angle), standard atmospheric profile, and surface elevation, a clear-sky radiation model is used to simulate and calculate the theoretical clear-sky total solar radiation that the Earth's surface should receive under current astronomical and atmospheric conditions, assuming no clouds. This represents the maximum radiation value that the location can reach at that moment. Then, the cloud transmittance function, representing the attenuation effect of clouds, obtained in the previous step, is used as a correction factor and applied to the theoretical clear-sky radiation value. Specifically, the theoretical clear-sky radiation value for each pixel at each moment is multiplied by the cloud transmittance for the corresponding pixel at the corresponding moment. The lower the transmittance (the stronger the cloud attenuation), the smaller the final radiation value. After this correction, the final output is a raster data of the preliminary inverted historical radiation sequence covering the entire target area with continuous time resolution (e.g., 15 minutes). This sequence forms the basis of the historical radiation field, but its accuracy is limited by the universality of the model parameters. Therefore, it will be used as the input for the next step of regional parameter calibration.

[0090] In conjunction with the first aspect, step S120, which involves optimizing the key parameters of the physical inversion model using the ground observation data to generate a region-adapted spatiotemporally continuous reference radiation field, includes:

[0091] S125. Based on the ground observation data and combined with the clear sky radiation model, the actual clear sky index is calculated.

[0092] This step aims to extract true information from ground-based measured data that can be used for model calibration. Specifically, using total solar radiation (GHI) data observed at ground stations, combined with theoretical clear-sky radiation values ​​calculated based on a clear-sky radiation model under the same spatiotemporal conditions, the actual clear-sky index for each station is calculated. This index reflects the ratio of the actual radiation received at the Earth's surface to the ideal clear-sky radiation under real atmospheric conditions (which may include influences such as aerosols). It is a key physical quantity characterizing the atmospheric transparency of a region under cloudless conditions and serves as the baseline true value for subsequent calibration.

[0093] S126, perform spatiotemporal matching and comparison between the actual clear sky index and the cloud index to obtain the comparison result.

[0094] Step S126 is a crucial step in establishing a quantitative relationship between satellite inversion and ground-based measurements. The cloud index obtained from satellite pixel-level inversion is matched with the actual clear sky index at the corresponding geographical location and observation time using high-precision spatiotemporal matching. Since the cloud index primarily reflects the degree of cloud cover, while the actual clear sky index comprehensively reflects the overall atmospheric transmittance, there should be a certain theoretical relationship between the two under clear sky conditions. By systematically comparing the differences in their spatiotemporal distribution, the inversion bias characteristics and statistical regularities of the model in the current region can be obtained, forming a quantitative comparison result and providing data for parameter calibration. The specific implementation involves two closely related aspects: high-precision spatiotemporal matching and systematic comparison analysis.

[0095] First, high-precision spatiotemporal matching is performed to address the inherent differences in spatiotemporal representation of the data sources. Since the actual clear sky index originates from discretely distributed ground stations, while the cloud index is continuous satellite raster data, strict coordinate and temporal alignment is essential. Spatially, cloud index values ​​of surrounding satellite pixels are extracted using the latitude and longitude coordinates of each ground station as the center. Methods such as bilinear interpolation are typically employed to obtain a single cloud index precisely corresponding to the station's location, minimizing errors caused by pixel scale mismatch. Temporally, the ground observation time is finely aligned with the satellite image observation time. Temporal interpolation or selection of the nearest neighbor time ensures that the data participating in the comparison strictly represent the same instant. Furthermore, to ensure physical consistency in the comparison, data from periods when the ground is determined to be clear sky and the solar altitude angle is suitable are typically selected. After these processes, a pairwise matched dataset covering all stations in the target area and spanning a long time series is finally generated.

[0096] Secondly, based on the matched dataset, a systematic comparison and analysis are conducted. The core of this process is to statistically analyze these data pairs (satellite cloud index, actual ground clear sky index), typically visualized as scatter plots. Ideally, for clear sky conditions, the two should exhibit a deterministic functional relationship. By analyzing the overall distribution of data points in this relationship graph, we can intuitively and quantitatively reveal the systematic deviations of the original physical model (whose built-in cloud index-cloud transmittance conversion function determines the theoretical correspondence) when applied in this region. For example, the data points may deviate from the model's theoretical curve in a general and regular manner. The comparison results obtained in this step are essentially a set of quantitative statistical relationship datasets and charts, accurately depicting the actual correlation between satellite observations and ground measurements under local real conditions. This provides an indispensable objective basis for the next step of optimizing model parameters through mathematical fitting.

[0097] S127. Based on the comparison results, the key functions related to cloud index and radiation transmittance in the physical inversion model are regionalized and calibrated using a fitting method.

[0098] This step utilizes the quantitative comparison relationship generated in step S126 to mathematically modify the most sensitive and regionally climate-affected modules of the Heliosat-2 physical model, thereby adapting the model's universal framework to the specific environment of the target region. Specifically, this step targets the conversion function in the model responsible for converting cloud index to cloud transmittance. In the original model, this function was calibrated based on European climate and geographical conditions, and its preset parameter relationships (e.g., linear or nonlinear coefficients) may not be applicable to the scattering characteristics of different types of aerosols, the microphysical properties of common cloud systems, and the reflectivity characteristics of the underlying surface in China. Through fitting methods, specifically using the set of (cloud index, actual clear sky index) data pairs generated in step S126, a new set of function parameters is solved using statistical optimization algorithms (such as least squares). This allows the function to optimally describe the true relationship between the locally observed cloud index and radiation transmittance. This process essentially involves embedding regional ground truth knowledge in reverse and correcting the empirical elements in the physical model, thereby systematically eliminating biases introduced by model transplantation. This is a crucial step in improving the regional accuracy of the inversion results.

[0099] Specifically, the comparison results usually fall into the following categories: ideal matching, systematic deviation, high dispersion, significant nonlinear relationship, and significant regional differences.

[0100] In the example, the characteristics of an ideal matching scenario are: data points are closely distributed along a defined curve (usually close to the theoretical curve of the original model, but may have some translation or rotation), with low dispersion. This indicates that the cloud index retrieved by the satellite and the ground radiation attenuation information have a highly consistent physical correlation locally, and the original model's functional form is basically applicable, requiring only parameter adjustment. The corresponding processing method is to directly use standard fitting methods such as the least squares method to optimize the parameters of the given functional form (such as the transformation function of the original model). This is the most direct and ideal calibration scenario.

[0101] The characteristics of systematic bias are: data points deviate from the original model's theoretical curve in a general and regular manner, such as being generally higher or lower, or having significant differences in slope. However, the dispersion of the data points around the new trend line remains low. This indicates that the parameter system of the original model has a systematic bias in this region, but the physical relationships themselves may still hold. The corresponding handling method is similar to the ideal case, but the necessity of calibration is more prominent. After determining the new parameters through fitting, the systematic error can be significantly corrected. The key is to ensure that the data samples involved in the fitting are sufficient and highly representative.

[0102] The characteristics of high dispersion are: data points are widely distributed and do not form a clear, compact trend line. This can be caused by a variety of factors: inaccurate spatiotemporal matching (e.g., mismatch between station representativeness and pixel scale), local weather phenomena (e.g., local clouds around the station but clear skies after satellite pixel averaging), or significant noise in ground observation data. The corresponding solutions are: first, check and remove obvious outliers (e.g., "pseudo-clear skies" data affected by local clouds); next, evaluate and improve the spatiotemporal matching strategy, such as adjusting the pixel range used for spatial interpolation or adopting a more accurate time alignment method; then, use robust regression methods (RANSAC, Theil-Sen estimator, etc.) for fitting, as these methods are insensitive to outliers and can still estimate a reasonable trend even with high data dispersion; finally, group the data by season or weather type and fit them separately to reduce intra-group dispersion.

[0103] The characteristic of a significant nonlinear relationship is that the data points clearly exhibit a trend, but this trend cannot be well described by the functional form of the original model (such as a simple linear or exponential relationship). This may indicate that the optical properties of clouds in this region are more complex in relation to radiation attenuation. The corresponding approach is to adjust or extend the functional form of the model. For example, adding higher-order terms to the original function, or using a more flexible nonlinear function (such as a piecewise function or a specific form of exponential / logarithmic function) for fitting.

[0104] Significant regional differences are characterized by the following: when data points from different geographical regions (e.g., plateaus and plains, arid and humid areas) are labeled with different colors on the same map, the clusters of data points in different regions exhibit distinct distribution trends or shifts. The corresponding approach is to divide the target region into several climatic or geographical sub-regions, independently execute steps S125-S128 to generate multiple sets of regionalized parameters, and introduce continuous variables related to regional characteristics (e.g., altitude, average aerosol optical thickness) as explanatory variables into the calibration function to establish dynamic parameter relationships.

[0105] Understandably, in the actual operation of step S127, multiple processing strategies mentioned above are usually applied in combination. First, the comparison results are visualized to diagnose the main situation. Then, based on the diagnostic results, data cleaning and matching optimization are performed, and appropriate function forms and fitting algorithms are selected to solve for the parameters. The ultimate goal is to obtain a physically reasonable calibrated function that best fits the local high-quality observation data. This process from data diagnosis to model tuning is crucial to ensuring the accuracy of the regionally adapted reference radiation field.

[0106] S128, Re-execute the inversion using the calibrated model parameters to generate a spatiotemporally continuous reference radiation field adapted to the region.

[0107] This step aims to apply the regionally calibrated and optimized model parameters to the reprocessing of all historical satellite visible light data, thereby systematically generating a high-quality solar radiation benchmark dataset that is fully adapted to the climate and geographical characteristics of the target region. Specifically, firstly, the optimized cloud index-cloud transmittance key transformation function and its parameters, determined through fitting in step S127, are formally updated into the physical inversion model (such as Heliosat-2), replacing the original default parameter set based on calibrations from other regions, thus completing the localization reset of the model. Subsequently, using this updated model, the complete inversion chain, from raw data input to geocalibration, atmospheric correction, cloud index calculation, cloud transmittance conversion, and clear-sky radiation correction, is re-executed for all geostationary satellite visible light images of the target region throughout history. Since the core transformation relationship has been corrected for local cloud characteristics, aerosol conditions, and surface reflectance characteristics, this re-inversion effectively corrects the systematic regional bias of the original model. Finally, this process generates a regionally adapted benchmark radiation field with high spatiotemporal resolution (e.g., 15-minute temporal resolution and approximately 2-kilometer spatial resolution), spatiotemporal continuity, and physical consistency. This product not only provides high-quality training labels and reliable historical state inputs for subsequent deep learning prediction models, but it also constitutes high-value basic data for solar energy resource assessment and climate monitoring, serving as the fundamental data foundation for the entire prediction method to achieve high accuracy.

[0108] In conjunction with the first aspect, the infrared brightness temperature data in step S130 includes the infrared brightness temperature data of channel 7 and channel 13 of the geostationary satellite. Step S130 includes:

[0109] S131 performs spatial grid unification and temporal resolution alignment on the spatiotemporally continuous reference radiation field and infrared brightness temperature data to obtain spatiotemporally consistent radiation field and brightness temperature data.

[0110] S132, a normalization method is used to map the values ​​of the reference radiation field and infrared brightness temperature data to a preset range; wherein, the scaling parameter used for normalization is statistically obtained and saved based on the training dataset and reused during prediction.

[0111] S133, periodically encodes the timestamp information of the data to generate time features that characterize the daily cycle of solar radiation.

[0112] S134 stitches together the normalized reference radiation field, infrared brightness temperature data, and time features along the feature channel dimension to form a unified five-dimensional spatiotemporal tensor, which serves as the input tensor for the standardized model.

[0113] Step S130 is a crucial preprocessing step that systematically processes and merges the regionalized reference radiation field generated by physical inversion with the infrared brightness temperature data from geostationary satellites into a standardized input format that can be directly received by deep learning models. Its core objective is to integrate multi-source observation data from different sources, with varying physical dimensions and spatiotemporal scales, into a structured tensor with unified mathematical specifications and rich physical and temporal information, providing high-quality, consistent input for subsequent spatiotemporal prediction models. In this invention, the geostationary satellite is taken as an example from the Chinese Fengyun-4 series meteorological satellites, whose 7 channels are shortwave infrared channels with a center wavelength of approximately 3.8 micrometers, and whose 13 channels are far-infrared window channels with a center wavelength of approximately 10.7 micrometers. It should be understood that other geostationary meteorological satellites with the same or similar spectral observation capabilities (such as the corresponding channels of the Japanese Himawari satellite and the US GOES series satellites) are also applicable to this step of the invention. Step S130 specifically includes the following four sequentially executed sub-steps:

[0114] Step S131 aims to address the inherent spatial and temporal benchmark differences among multi-source data and is a prerequisite for effective data fusion. The spatial grid unification operation aims to resample or interpolate the spatiotemporally continuous benchmark radiation field and infrared brightness temperature data onto a common geographic projection coordinate system and the same spatial grid.

[0115] Specifically, firstly, all input data (reference radiation field, infrared brightness temperature data) are resampled to a predefined uniform spatial grid. Typically, bilinear interpolation is used to project the original data and interpolate it to the center of each cell in the target grid. The parameters of the target grid (such as geographic projection, spatial extent, and cell size) need to be preset. For example, in this embodiment, the target grid can be set to an equal latitude / longitude grid (GeographicLat / Lon) covering the Chinese region, with a uniform spatial resolution (cell size) of 0.02 degrees (approximately 2 kilometers).

[0116] Secondly, the timeline of all data is unified to the same timestamp sequence. Since the time resolution of the original satellite data may be 10 minutes or 15 minutes, this embodiment interpolates all data to whole time points at 15-minute intervals (such as 00:00, 00:15, 00:30, etc.). For data at non-target times, a time linear interpolation method is used for calculation. Through the above operations, the output consists of two datasets with perfectly matched reference radiance and infrared brightness temperature values ​​at each spatial pixel location within each 15-minute time interval, namely, "spatiotemporally consistent radiation field data and brightness temperature data".

[0117] The purpose of step S132 is to process input features (radiation values ​​in W / m²) that have vastly different physical dimensions and numerical ranges. 2 (Brightness temperature in K), standardized to a uniform numerical range to eliminate the influence of dimensions, accelerate model training convergence, and improve stability. Each feature is linearly mapped to the interval [0, 1], and its calculation formula is as follows:

[0118]

[0119] in, Represents the original feature values. This is the minimum value of the feature in the training samples. This represents the maximum value of the feature in the training samples. This is the result after normalization.

[0120] This process can effectively avoid gradient imbalance caused by different feature scales during model training, thereby improving model convergence speed and prediction robustness.

[0121] Step S133 aims to encode linear, discrete time information into continuous, periodic features that can effectively characterize the daily cycle (24-hour cycle) of solar radiation, enabling deep learning models to easily capture and utilize this key physical law.

[0122] In conjunction with the first aspect, step S133 includes:

[0123] S1331, extract the timestamp information corresponding to the reference radiation field or infrared brightness temperature data to obtain the value representing a specific time of day.

[0124] Understandably, raw satellite data usually comes with precise observation time metadata, but this information exists in the form of structured data (such as UTC time strings or timestamps), which cannot be directly processed by mathematical models. Deep learning models process numerical tensors, so the human-readable time information must first be converted into a machine-calculate numerical form. The core task of this step is to accurately extract key components, such as hours and minutes, that reflect the periodic changes in the sun's position throughout the day from the raw time metadata, while ignoring information such as year, month, and day that are unrelated to the daily cycle.

[0125] In this embodiment, the absolute Coordinated Universal Time (UTC) or local time represented by each data sample (corresponding to a 15-minute time slice) is obtained by parsing the time metadata field of the data file (such as the time variable in a NetCDF or HDF5 file, or time information parsed from the filename). Then, a date and time processing library (such as Python's datetime module) is used to extract the hour and minute integer values ​​from the time object. For example, for the time 2023-07-15 09:30:00, hour=9 and minute=30 are extracted. These discrete integer values ​​form the basis for subsequent mathematical transformations.

[0126] S1332 converts the numerical value into a continuous time variable that changes cyclically within a preset period.

[0127] The principle behind this step is to address the issue of the cyclical nature and continuity of time. Understandably, directly inputting hours and minutes as independent features into the model has two fundamental flaws: first, it fails to reflect the physical proximity of 23:59 and 00:01 (numerically, 23.983 and 0.017 are far apart); second, as discrete ordinal features, the numerical values ​​of hours and minutes are not entirely equivalent to the physical driving force of radiation changes. Therefore, it is necessary to construct a variable that can map the 24 hours of a day into a continuous, cyclical mathematical space.

[0128] In this embodiment, the mapping is accomplished using a linear normalization formula:

[0129]

[0130] in, Converting time into a cumulative number of minutes calculated from midnight achieves the transformation from a two-dimensional discrete representation of "hour:minute" to a one-dimensional continuous quantity of daily minutes. Dividing by the total number of minutes in a day, 1440, normalizes this one-dimensional quantity to the interval [0, 1], resulting in the continuous variable t. This t possesses key properties: cyclicity (t=0 and t≈1 both represent midnight), continuity, and linear growth (approximately linearly correlated with changes in the solar hour angle). It successfully abstracts the physical daily cycle into a mathematically tractable cyclic variable defined on a unit interval, paving the way for further periodic coding.

[0131] S1333 inputs a continuous time variable into a set of periodic functions with a fixed phase relationship to generate a time feature vector containing multiple components to encode the diurnal cycle of solar radiation.

[0132] The principle of this step is to solve how to encode the continuous cyclic time variable t obtained in step S1332 into a feature representation that both preserves its periodicity and is easy for neural networks to learn.

[0133] If we take t itself as input, deep learning prediction models still struggle to automatically recognize its cyclic characteristics because when t jumps from 0.99 back to 0.01, it is discontinuous in the input space. We need an encoding that ensures temporally adjacent points are also adjacent in the feature space, and that the beginning and end of the cycle are completely connected. Sine-cosine coding perfectly achieves this, using trigonometric functions to map the one-dimensional cyclic scalar t to a point on a two-dimensional Cartesian coordinate system (unit circle). Specifically, for each value of t, the sine-coded component is calculated. Cosine coding components :

[0134]

[0135] This pair ( , This constitutes the final time feature vector. As t increases from 0 to 1, the point ( , The system operates at a uniform speed, smoothly, and without interruption for one complete cycle on the unit circle. The endpoint (t=1) and the starting point (t=0) of the time cycle are the same point in the encoding space, completely resolving the discontinuity problem. Two time-close t values ​​(such as t values ​​representing 09:00 and 09:15) also have their corresponding encoding points that are very close in Euclidean distance on the unit circle, which better preserves the time sequence and proximity relationship. Furthermore, on a two-dimensional plane, the diurnal variation of solar radiation intensity (such as the strongest radiation at noon) may be more easily learned by deep learning prediction models as the projection or clustering of encoding points in a specific direction on that circle.

[0136] Finally, for each time step, we obtain a two-dimensional vector[ , This vector is not a simple time label, but a continuous, differentiable mathematical feature rich in periodic phase information. When this feature is provided as part of the input to a deep learning prediction model, the model can learn and utilize the diurnal cycle of solar radiation very naturally and efficiently. This is a crucial feature engineering step in improving prediction accuracy, especially for key transitional periods such as dawn and dusk.

[0137] Step S134 is the final integration of data preparation, aiming to integrate the processed multi-source data (historical radiation field, infrared brightness temperature, time encoding) according to the standard input format of deep learning frameworks (such as PyTorch, TensorFlow) to construct a structured five-dimensional spatiotemporal tensor. The input to this step is the normalized baseline radiation field and normalized infrared brightness temperature data output from S132, and the periodic time feature vector output from S133. Subsequently, three parallel data branches are constructed:

[0138] The first branch is a historical GHI data tensor, which organizes the normalized reference radiation field data from the past L time steps (e.g., L=16, corresponding to 4 hours of history). Since there is only one physical quantity, GHI, its channel number is 1. Therefore, this branch is a tensor with dimensions [B, L, 1, H, W], where B is the batch size, L is the time step size, and H and W are the spatial height and width, respectively.

[0139] The second branch is the satellite infrared auxiliary data tensor, which organizes the normalized brightness temperature data of C infrared channels (e.g., C=2, corresponding to channels 7 and 13) used within the same historical period. This branch is a tensor of dimension [B, L, C, H, W], where C is the number of infrared channels used;

[0140] The third branch is the temporal embedding tensor, which takes the temporal feature vector generated by S133 (such as a 2D [ , The temporal features are expanded in the spatial dimension. That is, the same temporal encoding value is assigned to each time step (L) of each spatial location (H, W). In this way, the temporal features are expanded from a vector of [B, L, 2] to a tensor of [B, L, 2, H, W], which can then be concatenated with spatial data to encode the periodic temporal features of solar radiation.

[0141] Finally, the three data branches are concatenated along the Channel Dimension (the third dimension). The total number of channels after concatenation is... Thus, the final standardized model input tensor is generated, with dimensions [B, L, ... [B, H, W]. Where B is the batch size, which processes multiple independent samples at once to improve efficiency; L is the time step, representing the length of the historical context "seen" by the deep learning prediction model; W is the total number of feature channels, which integrates the historical target variable, auxiliary observation variables, and key temporal context; H is the spatial grid height, and W is the spatial grid width.

[0142] In conjunction with the first aspect, before step S134 and before constructing the five-dimensional spatiotemporal tensor, the data processed in steps S131-S133 needs to undergo standardized sample construction and strict integrity screening to generate a high-quality sample set that meets the training requirements of the spatiotemporal extrapolation model. The specific operation is as follows: Taking any time T0 as the prediction start time, the continuous data from the preceding N hours (T0-Nh to T0) is extracted as the model input period, and the continuous data from the following 4 hours (T0 to T0+Nh) is extracted as the label period for model prediction, with a time resolution of 15 minutes for both. For each sample, its input period must contain complete GHI inversion data (for periods without valid GHI values, such as nighttime, it is uniformly filled with 0) and all infrared channel data; its label period must contain complete GHI inversion data. If any data channel in the input period or label period is missing, the sample corresponding to that time T0 is completely removed from the training set. This screening mechanism ensures that each training sample has strict temporal continuity and physical consistency at both the input and label ends, thus providing a high-quality, unbroken supervised learning foundation for deep learning models. It is a key preprocessing step to ensure the effectiveness of model training and the stability of prediction.

[0143] In conjunction with the first aspect, the encoder-decoder architecture of deep learning prediction models includes:

[0144] The multi-branch coding unit is used to extract three types of features in parallel from the input tensor of the standardized model: the first coding branch extracts the spatiotemporal evolution features of the historical radiation sequence to obtain the main feature tensor; the second coding branch extracts the auxiliary meteorological features of the infrared brightness temperature data to obtain the auxiliary feature tensor; and the third coding branch extracts and expands the encoded time period information to obtain the time embedding feature tensor. Each coding branch is based on a three-dimensional convolutional structure to achieve layer-by-layer downsampling.

[0145] The spatiotemporal feature fusion unit is used to adaptively fuse the main feature tensor, auxiliary feature tensor, and temporal embedding feature tensor. The spatiotemporal feature fusion unit uses a dual-branch attention mechanism and is guided by the auxiliary feature tensor to perform weighted modulation of multi-source features to enhance the sensitivity to key meteorological areas.

[0146] The decoding and reconstruction unit is used to restore the spatial resolution step by step through upsampling operations based on the fused features. During the decoding process, the decoding and reconstruction unit introduces spatial detail features from the shallow layer of the multi-branch coding unit and embeds temporal codes representing future moments to generate a normalized radiation field sequence for future time periods.

[0147] The denormalized output unit is used to restore the normalized radiation field sequence to its true physical dimensions through a linear inverse transformation based on the minimum and maximum values ​​of the training data, and output the final predicted field of total solar radiation on the Earth's surface.

[0148] In this embodiment, the multi-branch encoding unit is used to process input data with different sources and physical meanings separately and efficiently. Specifically, the first encoding branch (main feature) processes the historical radiation sequence. This branch corresponds to the main channel encoder, and its input is the historical GHI data channel in the normalized input tensor constructed in step S130. By performing layer-by-layer downsampling based on a 3D convolution structure, this branch compresses the spatial size while aggregating information along the time dimension, thereby extracting the core high-level semantic features that can characterize the spatiotemporal evolution of solar radiation itself, and outputting the main feature tensor.

[0149] The second encoding branch (auxiliary features) processes infrared brightness temperature data. This branch corresponds to the auxiliary channel encoder, and its input is the infrared channel data in the normalized input tensor. It employs a three-dimensional convolutional downsampling structure similar to the first branch, specifically designed to extract spatiotemporal features of auxiliary meteorological elements such as cloud top temperature and cloud structure. These features are crucial for assessing cloud conditions, especially during the "dark start" period when visible light is absent. Its output is the auxiliary feature tensor.

[0150] The third encoding branch (temporal embedding features) is used to process the encoded temporal periodic information. This branch corresponds to the temporal channel encoder, whose input is the temporal features encoded by sine-cosine in step S133. This branch does not perform spatial downsampling, but rather uses neural network layers (such as fully connected or 1x1 convolutions) to perform nonlinear transformations and feature dimension expansion on this low-dimensional but crucial periodic information, converting it into a temporal embedding feature tensor that matches the spatial feature tensor dimension for subsequent fusion.

[0151] The spatiotemporal feature fusion unit receives output tensors from the three encoding branches mentioned above and adaptively interacts with features to establish correlations between multi-source features, rather than simply concatenating them. Specifically, this module introduces a two-branch spatial attention mechanism, using sigmoid and tanh functions to generate attention weight maps. Notably, this mechanism uses an auxiliary feature tensor (infrared features) rich in cloud information as a guide, dynamically weighting the importance of the main feature tensor (radiation history features) and the temporally embedded feature tensor at different spatial locations. This allows the deep learning prediction model to automatically focus on key areas that significantly affect radiation changes, such as cloud edges and abrupt changes in cloud thickness, achieving adaptive feature fusion and improving the model's ability to identify complex weather conditions.

[0152] Based on the highly abstracted and compressed fusion feature tensor output by the fusion unit, the decoding and reconstruction unit performs progressive high-resolution reconstruction. This process gradually restores the size of the feature map from low resolution to a size consistent with the original input space scale through a series of upsampling operations such as transposed convolution or pixel shuffle, thus achieving spatial dimension reconstruction. To avoid the loss of spatial details due to feature abstraction during upsampling, this unit introduces a skip connection mechanism. Specifically, it directly passes the high-resolution features containing rich local details output by the early layers of the encoding stage (i.e., the shallow network of the multi-branch encoding unit) to the corresponding layers in the decoder. These features are concatenated channel-wise or added element-wise with the upsampled features of the current decoding layer, thereby enhancing texture consistency while restoring the spatial structure and significantly improving the spatial fineness and physical interpretability of the final reconstructed radiation field. At the same time, temporally encoded features representing future target time periods are embedded at each stage or specific layer of decoding. These codes are generated based on the time context of future solar altitude angles, diurnal cycle phases, etc., enabling the decoder to dynamically adapt to the radiation variation patterns at different times, thereby adaptively generating a solar radiation field sequence for the next 0-4 hours that conforms to the diurnal cycle characteristics.

[0153] The denormalized output unit is a crucial final step in generating the final usable product from the deep learning prediction model. Its core function is to inversely convert the radiation field sequence output by the decoding and reconstruction unit within the model, which lies in the normalized numerical range [0,1], back into a sequence with true physical meaning and units (W / m²). 2The predicted total solar radiation (GHI) field of the Earth's surface is an essential bridge for directly comparing, verifying, and applying the model's predictions with actual observation data. The operation of this unit strictly corresponds to and is the reverse of the normalization process performed in the preprocessing stage (such as step S132). Specifically, in the data preparation stage before model training, all input historical baseline radiation field data are linearly mapped to the [0,1] interval using the Min-Max Normalization method, based on the global minimum and global maximum values ​​obtained from the statistics of all samples in the training dataset. Therefore, denormalization is the inverse process of the above transformation. This unit receives the normalized predicted values ​​output by the model. :

[0154]

[0155] It is a normalized prediction field generated by the model, ranging from [0,1].

[0156] It is only the actual minimum and maximum total solar radiation that are pre-statistically collected and stored from the training dataset;

[0157] It is only the actual minimum and maximum total solar radiation that are pre-statistically collected and stored from the training dataset;

[0158] It is the product recovered after inverse transformation, possessing true physical dimensions (W / m). 2 The predicted value of solar radiation.

[0159] In conjunction with the first aspect, the deep learning prediction model described above uses an end-to-end mapping method to directly output the solar radiation prediction field based on the input standardized model input tensor. The generation of the prediction field during the visible light absence period does not depend on the explicit intermediate inversion results of cloud conditions.

[0160] In conjunction with the first aspect, the training steps for a deep learning prediction model include:

[0161] S210 uses a training sample set containing historical reference radiation fields, corresponding infrared brightness temperature data, and future reference radiation field labels to train a deep learning prediction model.

[0162] S220: The deep learning model is trained using a composite objective function that combines pixel-level error and sequence consistency loss, and the training process is optimized by applying dynamic learning rate adjustment and gradient pruning strategies until the model converges.

[0163] First, construct the training sample set.

[0164] Input data (X): The input for each training sample is a normalized model input tensor obtained after complete preprocessing steps S130-S135. This tensor contains three core parts:

[0165] Historical baseline radiation field sequence: for example, the past 4 hours (T0-4h to T0), with a temporal resolution of 15 minutes and spatial coverage of the target area.

[0166] Infrared brightness temperature data sequence for the corresponding time period: infrared channel data (such as channels 07 and 13) that are strictly registered with the historical radiation field in time and space.

[0167] Time period coding: The time features obtained by performing sine-cosine coding on each time step within the above historical period.

[0168] Label data (Y): The true value or label corresponding to each sample is a high-quality reference radiation field for the future target time period (such as T0 to T0+4h) that has undergone the same regional inversion and standardization process, which provides a clear learning target for the model.

[0169] Subsequently, continuous "input-label" pairs are extracted from the timeline using a sliding window approach. GHI data missing during the nighttime period is filled with 0s, and any samples with missing data are strictly removed to ensure the continuity and physical consistency of the training data.

[0170] Next, a composite objective function (loss function) is defined. This embodiment of the invention employs a weighted composite loss function to simultaneously constrain spatial accuracy and temporal rationality:

[0171] Pixel-level error loss (such as mean absolute error, MAE) is calculated using the following formula:

[0172]

[0173] in, These are model predictions. This is the true label value. This loss ensures that the model's output at each spatial cell and each prediction time step is as close as possible to the true value, thus guaranteeing the overall accuracy of the prediction results.

[0174] Temporal feature consistency loss: This loss aims to constrain the smoothness and physical evolution of the predicted sequence over time. For example, the norm of the difference between adjacent time steps in the predicted sequence can be calculated to match the temporal variation of the actual sequence, avoiding drastic, non-physical fluctuations in the prediction results. This is crucial for generating continuous and stable 0-4 hour forecast sequences.

[0175] Composite form: The total loss function is:

[0176] in, These are adjustable weighting coefficients, all greater than 0. By adjusting these two parameters, the model's emphasis on instantaneous accuracy and temporal coherence can be balanced.

[0177] Subsequently, the measurement configuration is optimized to efficiently and stably minimize the aforementioned loss function. Specifically, in this embodiment, the optimizer uses the AdamW optimizer. Compared to the standard Adam, AdamW decouples weight decay (a regularization technique used to prevent overfitting due to excessive model complexity) from the gradient update step, which has been proven to bring better generalization performance and final convergence results.

[0178] Dynamic learning rate adjustment: A cosine annealing scheduling strategy is employed. This strategy causes the learning rate to decrease from its initial value to near zero according to a cosine function curve with the number of training iterations. This arrangement allows for rapid convergence using a larger learning rate in the early stages of training, followed by fine-tuning of parameters using a very small learning rate in the later stages, which helps to find a better solution for the model parameters.

[0179] Training stability assurance: Introducing a gradient clipping mechanism. In deep learning training, gradients may become abnormally large due to network structure or data issues (gradient explosion), leading to training failure. Gradient clipping, by setting a threshold, forces the magnitude of the gradient vector to be limited within a certain range, thereby ensuring the numerical stability of the training process.

[0180] Following this, the prepared training sample set is input into the model in batches, and the optimizer updates the model parameters using backpropagation based on the gradients calculated by the loss function. Throughout the process, performance is continuously monitored on the validation set (a subset of samples from the training data that does not participate in gradient calculation). An early stopping mechanism is employed: when the validation set loss no longer decreases over several consecutive training epochs, it is determined that the model has been sufficiently trained and may begin to overfit the training data; training is then stopped, and the model parameters at the point of optimal validation set performance are saved. This combined strategy effectively balances the model's fitting ability and generalization ability, and is key to obtaining a high-performance, highly reliable prediction model.

[0181] Finally, the trained model was systematically validated and its performance evaluated. A linear interpolation method was used to map the rasterized GHI prediction sequence output by the model to ground observation stations, achieving point-by-point validation. The RMSE, MAE, systematic bias, and correlation coefficients of the interpolated predicted values ​​and the actual station observations were calculated and statistically compared during the dark start-up phase (around sunrise) and the daytime phase to comprehensively evaluate the model's prediction accuracy and stability at different times. The focus was on validating the error improvement during the dark start-up phase, demonstrating that introducing an infrared brightness temperature channel and a unified spatiotemporal extrapolation strategy when visible light is unavailable allows GHI predictions to maintain continuity and acceptable accuracy during the dawn transition period. This confirms that the present invention effectively overcomes the prediction blind spots of traditional methods and achieves all-day, highly continuous ultra-short-term solar radiation prediction capabilities.

[0182] Combination Figure 2 As shown, firstly, multispectral remote sensing data from geostationary satellites is acquired, and visible light and infrared remote sensing channels are extracted from them. The visible light channel data is input into a physical model calibrated with ground observation data parameters to generate a global horizontal irradiance sequence for historical periods (e.g., from T-4h to T). The infrared channel data directly extracts thermal radiation information to characterize cloud and atmospheric states. These two types of data are spatiotemporally registered to form a dual-branch data input, which, together with the parsed time labels, is converted into a periodic feature vector through a time embedding coding module.

[0183] Subsequently, the dual-branch data and time-encoded data are input into a multi-branch encoder module. This module extracts deep spatiotemporal features from the radiation sequence and infrared features through independent convolutional branches, and enhances key regions (such as cloud activity areas) through attention weights. The extracted multi-source features are adaptively fused in the spatiotemporal information fusion module to establish a cross-modal correlation between radiation evolution and infrared signals.

[0184] Finally, the fused high-dimensional features are progressively restored to spatial details through a deconvolutional upsampling decoding module, and skip connections are introduced to pass shallow features, achieving the reconstruction of a high-resolution radiation field from low-resolution features. The system ultimately outputs global horizontal irradiance predictions for the next 0-4 hours with a temporal resolution of up to 15 minutes, realizing end-to-end modeling from multi-source satellite observations to continuous radiation field prediction. This process, by integrating physical inversion and deep learning techniques, balances the physical interpretability of the model with the spatiotemporal continuity of the prediction, especially enhancing the prediction capability during periods of visible light deficiency (such as around dawn).

[0185] Secondly, embodiments of the present invention also provide a solar radiation prediction device based on multi-source satellite data fusion, combined with Figure 3 As shown, the device includes: an acquisition module 10, a generation module 20, a fusion module 30, and a prediction module 40.

[0186] The acquisition module 10 is used to acquire geostationary satellite multispectral remote sensing data, infrared remote sensing data, and ground observation data.

[0187] The generation module 20 is used to obtain historical radiation sequences based on visible light data and through a physical inversion model, optimize the key parameters of the physical inversion model by combining ground observation data, and generate a spatiotemporally continuous reference radiation field adapted to the region.

[0188] The fusion module 30 is used to perform spatiotemporal standardization and feature encoding on the spatiotemporal continuous reference radiation field and the infrared brightness temperature data of geostationary satellites, and fuse them to form a standardized model input tensor containing historical radiation sequences, infrared features and time period information.

[0189] The prediction module 40 is used to input the standardized model input tensor into the trained deep learning prediction model. The deep learning prediction model adaptively fuses the spatiotemporal dependencies of multi-source features through an encoder-decoder architecture and directly outputs the predicted field of total solar radiation on the Earth's surface for future time periods and time continuity.

[0190] Thirdly, embodiments of the present invention provide an electronic device, combined with Figure 4 As shown, the electronic device includes a memory 131 and a processor 130. The memory 131 stores a computer program, and the processor 130 runs the computer program to make the electronic device perform the above-described method.

[0191] Furthermore, combined Figure 4 The electronic device shown also includes a bus 132 and a communication interface 133, with the processor 130, the communication interface 133 and the memory 131 connected via the bus 132.

[0192] The memory 131 may include high-speed random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 133 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc. The bus 132 may be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 4 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0193] Processor 130 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 130 or by instructions in software form. Processor 130 may be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it may also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this invention can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. The software module can reside in a mature storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory 131, and processor 130 reads the information in memory 131 and, in conjunction with its hardware, completes the steps of the method described in the foregoing embodiments.

[0194] Fourthly, embodiments of the present invention provide a readable storage medium storing computer program instructions, which are read and executed by a processor to perform the above-described method.

[0195] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system and apparatus described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0196] Furthermore, in the description of the embodiments of the present invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in the present invention based on the specific circumstances.

[0197] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0198] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0199] Finally, it should be noted that the above embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for predicting solar radiation based on multi-source satellite data fusion, characterized in that, The method includes: Acquire geostationary satellite multispectral remote sensing data, infrared remote sensing data, and ground observation data; Based on the visible light data in the geostationary satellite multispectral remote sensing data and the historical radiation sequence obtained through the physical inversion model, the key parameters of the physical inversion model are optimized by combining the ground observation data to generate a spatiotemporally continuous reference radiation field adapted to the region. The spatiotemporal continuous reference radiation field and the infrared brightness temperature data are unified in spatial grid and aligned in temporal resolution to obtain radiation field data and brightness temperature data with consistent spatiotemporal reference. A normalization method is used to map the values ​​of the reference radiation field and the infrared brightness temperature data to a preset interval. The scaling parameters used for normalization are statistically obtained from the training dataset and saved for reuse during prediction. The timestamp information of the data is periodically encoded to generate a time feature representing the daily cycle of solar radiation. The normalized reference radiation field, infrared brightness temperature data, and time feature are concatenated along the feature channel dimension to form a unified five-dimensional spatiotemporal tensor, which serves as the input tensor for the standardized model. The standardized model input tensor is fed into a trained deep learning prediction model. The deep learning prediction model adaptively fuses the spatiotemporal dependencies of multi-source features through an encoder-decoder architecture and directly outputs a future time period and a time-continuous prediction field of total solar radiation on the Earth's surface. The encoder-decoder architecture of the deep learning prediction model includes: A multi-branch coding unit is used to extract three types of features in parallel from the input tensor of the standardized model: the spatiotemporal evolution features of historical radiation sequences are extracted through the first coding branch to obtain the main feature tensor; the auxiliary meteorological features of infrared brightness temperature data are extracted through the second coding branch to obtain the auxiliary feature tensor; and the time period information after encoding is extracted and expanded through the third coding branch to obtain the time embedding feature tensor. Each coding branch is based on a three-dimensional convolutional structure to achieve layer-by-layer downsampling. The spatiotemporal feature fusion unit is used to adaptively fuse the main feature tensor, the auxiliary feature tensor, and the temporal embedding feature tensor. The feature fusion unit uses a dual-branch attention mechanism and is guided by the auxiliary feature tensor to perform weighted modulation of multi-source features to enhance the sensitivity to key meteorological areas. The decoding and reconstruction unit is used to recover the spatial resolution step by step through upsampling operations based on the fused features. During the decoding process, the decoding and reconstruction unit introduces the shallow spatial detail features from the multi-branch coding unit and embeds the temporal code representing the future time to generate a normalized radiation field sequence for the future time period. The denormalized output unit is used to restore the normalized radiation field sequence to its true physical dimensions through a linear inverse transformation based on the minimum and maximum values ​​of the training data, and output the final predicted field of total solar radiation on the Earth's surface.

2. The method according to claim 1, characterized in that, The steps for obtaining historical radiation sequences based on the visible light data and through a physical inversion model include: Based on the visible light data, the ground observation data is geocalibrated and atmospherically corrected to obtain apparent reflectance data; Based on the apparent reflectance data and the preset surface reference reflectance, a cloud index reflecting cloud coverage is calculated. The cloud index is converted into a cloud transmittance function that characterizes the degree of attenuation of solar radiation by clouds; Based on the sun's position, atmospheric parameters, and surface information, the theoretical total solar radiation under clear skies is calculated using a clear-sky radiation model. The cloud transmittance function is then used to correct the theoretical total solar radiation under clear skies, resulting in a preliminary inverted historical radiation sequence.

3. The method according to claim 2, characterized in that, The steps of optimizing the key parameters of the physical inversion model by combining the ground observation data to generate a region-adapted spatiotemporally continuous reference radiation field include: Based on the aforementioned ground observation data and combined with the clear-sky radiation model, the actual clear-sky index is calculated. The actual clear sky index and the cloud index are spatiotemporally matched and compared to obtain the comparison results. Based on the comparison results, the key functions related to cloud index and radiation transmittance in the physical inversion model are regionalized and calibrated using a fitting method. The inversion is re-executed using the calibrated model parameters to generate a spatiotemporally continuous reference radiation field adapted to the region.

4. The method according to claim 1, characterized in that, The infrared brightness temperature data includes the infrared brightness temperature data of channels 7 and 13 of the geostationary satellite.

5. The method according to claim 1, characterized in that, The training steps of the deep learning prediction model include: The deep learning prediction model is trained using a training sample set that includes historical reference radiation fields, corresponding infrared brightness temperature data, and future reference radiation field labels. During training, a composite objective function combining pixel-level error and sequence consistency loss is used; The training process is optimized by applying dynamic learning rate adjustment and gradient clipping strategies until the model converges.

6. A solar radiation prediction device based on multi-source satellite data fusion, characterized in that, The device includes: The acquisition module is used to acquire geostationary satellite multispectral remote sensing data, infrared remote sensing data, and ground observation data; The generation module is used to obtain historical radiation sequences based on visible light data in the geostationary satellite multispectral remote sensing data and through a physical inversion model, and to optimize the key parameters of the physical inversion model by combining the ground observation data to generate a spatiotemporally continuous reference radiation field that is adapted to the region. The fusion module is used to unify the spatial grid and align the temporal resolution of the spatiotemporally continuous reference radiation field and the infrared brightness temperature data to obtain radiation field data and brightness temperature data with consistent spatiotemporal reference. A normalization method is used to map the values ​​of the reference radiation field and infrared brightness temperature data to a preset interval. The scaling parameters used for normalization are statistically obtained from the training dataset and saved for reuse during prediction. The timestamp information of the data is periodically encoded to generate a time feature representing the daily cycle of solar radiation. The normalized reference radiation field, infrared brightness temperature data, and time feature are concatenated along the feature channel dimension to form a unified five-dimensional spatiotemporal tensor, which serves as the input tensor for the standardized model. The prediction module is used to input the standardized model input tensor into the trained deep learning prediction model. The deep learning prediction model adaptively fuses the spatiotemporal dependencies of multi-source features through an encoder-decoder architecture and directly outputs the predicted field of total solar radiation on the Earth's surface for future time periods and time continuity. The encoder-decoder architecture of the deep learning prediction model includes: A multi-branch coding unit is used to extract three types of features in parallel from the input tensor of the standardized model: the spatiotemporal evolution features of historical radiation sequences are extracted through the first coding branch to obtain the main feature tensor; the auxiliary meteorological features of infrared brightness temperature data are extracted through the second coding branch to obtain the auxiliary feature tensor; and the time period information after encoding is extracted and expanded through the third coding branch to obtain the time embedding feature tensor. Each coding branch is based on a three-dimensional convolutional structure to achieve layer-by-layer downsampling. The spatiotemporal feature fusion unit is used to adaptively fuse the main feature tensor, the auxiliary feature tensor, and the temporal embedding feature tensor. The feature fusion unit uses a dual-branch attention mechanism and is guided by the auxiliary feature tensor to perform weighted modulation of multi-source features to enhance the sensitivity to key meteorological areas. The decoding and reconstruction unit is used to recover the spatial resolution step by step through upsampling operations based on the fused features. During the decoding process, the decoding and reconstruction unit introduces the shallow spatial detail features from the multi-branch coding unit and embeds the temporal code representing the future time to generate a normalized radiation field sequence for the future time period. The denormalized output unit is used to restore the normalized radiation field sequence to its true physical dimensions through a linear inverse transformation based on the minimum and maximum values ​​of the training data, and output the final predicted field of total solar radiation on the Earth's surface.

7. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program and the processor running the computer program to cause the electronic device to perform the method of any one of claims 1 to 5.

8. A storage medium, characterized in that, The storage medium stores computer program instructions, which, when read and executed by a processor, perform the method described in any one of claims 1 to 5.

Images