A method and device for predicting surface solar radiation based on satellite observation data

Abstract

The application discloses a kind of surface solar radiation prediction method and equipment based on satellite observation data, belong to solar radiation remote sensing prediction technical field.The method includes obtaining target observation data and prediction model;Difference is obtained by the target observation data of the adjacent observation time of each spectral channel of the top reflectivity of atmosphere layer, and cloud movement sensing feature is obtained;The input sequence that target observation data and cloud movement sensing feature jointly constitute is input into the prediction model, and the surface solar radiation prediction result is obtained.The application can realize more rapid, more continuous surface solar radiation short-term prediction, and reduce the dependence of intermediate inversion product and multi-source auxiliary data while maintaining good prediction accuracy in actual prediction stage, and has practical application value for solar energy resource monitoring, photovoltaic power generation power estimation and short-time power dispatching.

Description

Technical Field

[0001] This invention belongs to the field of solar radiation remote sensing prediction technology, specifically relating to a method and device for predicting surface solar radiation based on satellite observation data. Background Technology

[0002] Surface solar radiation is a crucial component of the Earth's energy budget and a key parameter in solar energy resource assessment, photovoltaic power generation estimation, and short-term power dispatch. Accurately obtaining short-term changes in surface solar radiation helps determine changes in the output of solar power systems, providing a reference for new energy grid connection and power system operation. Especially under weather conditions with rapidly changing cloud cover, surface solar radiation can fluctuate significantly on a minute-to-hour scale, making short-term solar radiation forecasting of significant practical value. Ground-based observation is an important method for obtaining surface solar radiation. High-temporal-resolution solar radiation data can be obtained through radiation observation stations, but these stations are typically unevenly distributed, making it difficult to meet the needs of large-scale continuous forecasting. Satellite remote sensing, with its wide observation range and strong temporal continuity, has become an important means of obtaining solar radiation-related information. Polar-orbiting satellites can provide high-spatial-resolution observation data, but their long revisit periods make it difficult to continuously capture cloud movement and development. In contrast, geostationary meteorological satellites can conduct high-frequency observations of fixed areas, making them more suitable for monitoring and forecasting short-term solar radiation changes.

[0003] Existing methods for estimating or predicting surface solar radiation mainly include empirical statistical methods, physical methods, numerical weather prediction methods, and data-driven methods. Physical methods typically require multiple intermediate parameters such as cloud optical thickness, effective cloud particle radius, aerosols, water vapor, and ozone, resulting in relatively complex calculation processes. Furthermore, the uncertainty of these intermediate parameters can affect the final solar radiation results. Numerical weather prediction methods can describe atmospheric change processes, but in rapid predictions at minute to hourly scales, they are easily limited by computational timeliness and cloud process simulation errors. While some data-driven methods can establish nonlinear predictive relationships using historical data, they often still rely on multi-source meteorological data, cloud products, or pre-calculated radiation fields, making data preparation complex in practical applications. Summary of the Invention

[0004] In view of this, in order to overcome the above-mentioned technical defects, the present invention provides a method and device for predicting surface solar radiation based on geostationary meteorological satellite observation data, which reduces the input data requirements while maintaining the accuracy of surface solar radiation prediction.

[0005] To achieve the above objectives, the solution adopted by the present invention is:

[0006] A method for predicting surface solar radiation based on satellite observation data includes the following steps:

[0007] Acquire target observation data from geostationary meteorological satellites and trained prediction models;

[0008] The data dimensions of the target observation data include atmospheric top reflectivity, solar zenith angle, observation zenith angle, and relative azimuth angle. The target observation data for each data dimension are sequential values ​​obtained from observations at multiple consecutive times.

[0009] The cloud motion sensing characteristics are obtained by subtracting the target observation data at adjacent observation times for each spectral channel of the atmospheric top reflectance.

[0010] The target observation data and cloud motion sensing features are combined to form an input sequence, which is then input into the prediction model to obtain the surface solar radiation prediction result.

[0011] Furthermore, the atmospheric top reflectance is the first-order multispectral atmospheric top reflectance.

[0012] Furthermore, the prediction model is first pre-trained using the first dataset, and then the pre-trained prediction model is re-trained using the second dataset to obtain the trained prediction model.

[0013] In the first dataset, surface solar radiation output from RSTAR software simulation is used as the label data; in the second dataset, ground-based solar radiation observation data is used as the label data.

[0014] Furthermore, the historical observation data of the target geostationary meteorological satellite and the corresponding ground-based solar radiation observation data are preprocessed, and then the second dataset is constructed using the preprocessed historical observation data and ground-based solar radiation observation data; the preprocessing includes at least one of time unification, data quality screening and standardization processing.

[0015] Furthermore, the calculation process within the prediction model includes:

[0016] The input sequence is convolved using multiple temporal convolutions at different time steps to obtain multiple multi-channel convolutional feature sequences of equal size;

[0017] Multiple convolutional feature sequences are fused to obtain a multi-scale feature sequence;

[0018] Based on the multi-scale feature sequence, multiple branch weights corresponding to the convolutional feature sequence are calculated;

[0019] Based on the multiple branch weights, the convolutional feature sequences corresponding to each branch weight are weighted respectively, and the weighted convolutional feature sequences are summed to obtain multi-scale temporal features;

[0020] The multi-scale temporal features are recalibrated to obtain calibration features;

[0021] The calibration features are input into the control module, which calculates the hidden features at each time step.

[0022] The hidden features at each time step are input into the attention module, and the attention module calculates the attention weights corresponding to each time step.

[0023] Using the attention weights corresponding to each time step, the hidden features at each time step are weighted and summed to obtain the context features;

[0024] The context features are input into the prediction output module, which maps the context features to the predicted values ​​of surface solar radiation at multiple prediction times to obtain the predicted surface solar radiation results.

[0025] Further, based on the multi-scale feature sequence, multiple branch weights corresponding to the convolutional feature sequence are calculated, including:

[0026] The multi-scale feature sequence is subjected to average pooling along the time dimension to obtain the first descriptive feature;

[0027] The first descriptive feature is input into the first fully connected layer and processed by the ReLU function to obtain the intermediate feature;

[0028] The intermediate features are respectively input into multiple second fully connected layers that correspond one-to-one with the temporal convolutions to obtain multiple selection score features;

[0029] The softmax function is used to normalize each selection score feature to obtain multiple branch weights.

[0030] Furthermore, the multi-scale temporal features are recalibrated, including:

[0031] The multi-scale temporal features are subjected to average pooling along the time dimension to obtain the second descriptive features;

[0032] The second descriptive feature is processed sequentially by a one-dimensional convolution along the channel dimension and a sigmoid function to obtain the channel weights;

[0033] The channel weights are multiplied by the corresponding elements of the multi-scale time features to obtain the calibration features.

[0034] Furthermore, the control module is an LSTM.

[0035] Furthermore, the internal calculation process of the attention module includes:

[0036] The hidden features at each time step are input into the third fully connected layer and processed by the tanh function to obtain the attention intermediate features corresponding to each time step.

[0037] The attention intermediate features corresponding to each time step are multiplied by the learnable projection vector to obtain the attention score corresponding to each time step.

[0038] The attention scores are normalized using the softmax function to obtain the attention weights at each time step.

[0039] The present invention also provides a surface solar radiation prediction device based on satellite observation data, including a processor and a memory, wherein the memory stores a computer program, and the processor loads the computer program to execute the method described above.

[0040] The beneficial effects of this invention are:

[0041] The new generation of geostationary meteorological satellites possesses high temporal resolution and multispectral observation capabilities. Their primary observation data includes information on the impact of cloud cover, surface background, and solar geometry on radiation variations. This invention directly utilizes primary observation data from geostationary meteorological satellites to establish a short-term prediction method for surface solar radiation. This reduces reliance on complex auxiliary data and intermediate inversion products, improving the real-time performance and applicability of solar radiation prediction. It offers advantages such as easy data acquisition, fewer intermediate processing steps, continuous prediction timeliness, and strong real-time applicability.

[0042] Compared with existing technologies, this invention can achieve faster and more continuous short-term prediction of surface solar radiation, and maintain good prediction accuracy while reducing reliance on intermediate inversion products and multi-source auxiliary data in the actual prediction stage. It has practical application value for solar energy resource monitoring, photovoltaic power generation estimation and short-term power dispatch. Attached Figure Description

[0043] Figure 1 Verification results of surface solar radiation stations under different prediction lead times for the method provided by this invention;

[0044] Figure 2 The method provided by this invention provides short-term prediction results of regional surface solar radiation under complex cloud system conditions of Typhoon Krosa. Detailed Implementation

[0045] The following embodiments are intended to illustrate the method provided by the present invention in detail with reference to the accompanying drawings. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of methods consistent with some aspects of the schemes described in the appended claims.

[0046] This embodiment provides a method for predicting surface solar radiation based on satellite observation data, including:

[0047] S100, Data Acquisition.

[0048] The collected data consists of two parts: first-level observation data from geostationary meteorological satellites and second-level observation data from ground-based solar radiation.

[0049] The short-term prediction method for surface solar radiation proposed in this invention is theoretically applicable to next-generation geostationary meteorological satellites equipped with multispectral imagers, including Japan's Himawari-8 / 9, China's FY-4A / B, the US GOES-R series, and Europe's Meteosat series. The following explanation uses the Himawari-8 geostationary meteorological satellite as an example; other geostationary meteorological satellites can be substituted based on their corresponding spectral channels and observation geometry parameters. Himawari-8's Level 1 observation data is provided by the Japan Meteorological Agency and distributed through the JAXA P-Tree system. The Advanced Himawari Imager (AHI) onboard Himawari-8 can provide 10-minute continuous multispectral observation data for the Asia-Pacific region. This invention uses the atmospheric top reflectivity of AHI bands 1–6 as the main satellite input variable and simultaneously acquires or calculates geometric observation parameters such as solar zenith angle, observed zenith angle, and relative azimuth angle. Among them, the atmospheric top reflectivity of AHI 1–6 bands is used to characterize the influence of cloud reflection, cloud spatial texture, cloud microphysical properties and surface background conditions on the changes in surface solar radiation; the solar zenith angle, observation zenith angle and relative azimuth angle are used to characterize the influence of the solar incident direction, satellite observation direction and the relative observation geometry of the Sun, Earth and Satellite on the changes in satellite reflectivity and surface solar radiation.

[0050] In this embodiment, during the pre-training phase, the atmospheric radiative transfer model software RSTAR is used to construct simulated surface solar radiation sample data as label data. This simulated sample is used in the physical simulation pre-training process to supplement the deficiencies of real ground-based solar radiation observation samples in terms of spatial distribution and condition coverage. This improves the applicability of the trained model under different regions, seasons, and atmospheric background conditions. In the subsequent model training phase, ground-based solar radiation observation data is used as label data and evaluation reference for model training. Ground-based solar radiation observation data can be surface downward shortwave radiation data obtained from solar radiation observation stations, meteorological observation stations, or ecological observation stations, including but not limited to Baseline Surface Radiation Network (BSRN), China Meteorological Administration (CMA) observation data, and China Ecosystem Research Network (CERN) observation data.

[0051] S200. Optionally, in some embodiments, preprocessing of the data is further included before constructing the second dataset.

[0052] For example, one or more of the following preprocessing steps can be performed on the collected historical geostationary meteorological satellite primary observation data and the corresponding ground-based solar radiation observation data: time unification, spatial matching, quality screening, geometric parameter processing, missing value handling, sample construction, and standardization.

[0053] First, historical geostationary satellite observation data and ground observation data are uniformly converted to UTC time, and corresponding geostationary meteorological satellite observation pixels are matched according to the latitude and longitude of the ground observation stations. Since the time resolution of geostationary meteorological satellite observations is usually on the order of minutes, and the time resolution of Himawari-8 observations in this embodiment is 10 minutes, the ground-based solar radiation observation data are uniformly processed to match the time resolution of satellite observations. For the 1-minute BSRN surface solar radiation observation value used in this embodiment, it is averaged into 10-minute data, so that each target value represents the average surface solar radiation within a 10-minute time period.

[0054] Then, the samples were quality-screened, removing samples with a solar zenith angle greater than 85° to avoid the impact of unstable radiation under low solar altitude conditions on model training and prediction; at the same time, abnormal records with ground-based solar radiation observations less than 0 W / m² or greater than 1360 W / m² were also removed. For geostationary meteorological satellite Level 1 input data, visible light and near-infrared multispectral atmospheric top reflectance were retained, and geometric observation parameters such as solar zenith angle, observed zenith angle, and relative azimuth angle were simultaneously processed; if there were abnormal or missing input features, they were removed before sample construction or filled using linear interpolation between adjacent observation times to maintain the continuity of the time series.

[0055] S300, Model Building.

[0056] After the above processing is completed, a sliding window approach is used to construct a dataset for training the prediction model. Satellite observation data from multiple consecutive historical moments are used as model input, and surface solar radiation observations at corresponding prediction moments are used as label data, thus forming sample pairs from historical satellite observation sequences to corresponding future surface solar radiation sequences. The length of the input and output data time series can be set according to prediction requirements; the prediction lead time can be set to 3h, 6h, 24h, etc. In this embodiment, observation data from 12 consecutive historical moments are used as model input, corresponding to the past 2 hours of satellite observation sequences, and surface solar radiation observations from the subsequent 18 consecutive moments are used as prediction label data, corresponding to the future 3 hours of surface solar radiation sequences. After the sample construction is completed, all station samples are merged and divided into training, validation, and test sets according to time order. The training set is used to train the prediction model to obtain the trained prediction model. Optionally, continuous data can be Z-score normalized. The normalization parameters are calculated only from the training set, and the same parameters are applied to the validation and test sets to avoid data leakage.

[0057] After the prediction model is trained, target observation data from a geostationary meteorological satellite and the trained prediction model are acquired. The input sequence, composed of the target observation data and cloud motion sensing features, is then input into the prediction model, which calculates the predicted surface solar radiation. The specific principles and detailed process are as follows:

[0058] Surface solar radiation on a minute-to-hour scale is primarily influenced by solar incidence conditions, satellite observation geometry, and cloud evolution. The visible and near-infrared multispectral atmospheric top reflectance from geostationary meteorological satellites contains information on cloud reflection, cloud spatial texture, cloud microphysical properties, and surface background conditions. The solar zenith angle, observational zenith angle, and relative azimuth angle characterize the influence of the solar incidence direction, satellite observation direction, and the relative geometric relationship between the Sun, Earth, and satellite on satellite-observed reflectance and surface solar radiation variations, respectively. By using multispectral atmospheric top reflectance along with the aforementioned geometric observation parameters as input, the model can simultaneously utilize cloud state information, observational geometry, and solar incidence conditions to predict future surface solar radiation variations. Unlike methods that require explicit cloud parameter inversion results, numerical weather prediction results, or pre-calculated radiation fields as necessary inputs during the actual prediction phase, this invention directly learns the surface solar radiation variation characteristics from the target geostationary meteorological satellite's first-level observation sequence data during the actual prediction phase, establishing a predictive relationship between satellite observation sequence data and the corresponding future surface solar radiation sequence.

[0059] In continuous geostationary meteorological satellite observations, the movement, development, and dissipation of clouds directly alter the spatiotemporal distribution of solar radiation at the Earth's surface. If predictions rely solely on single-moment observations or simple historical sequences, the model needs to implicitly learn the impact of cloud motion on future radiation changes, which can easily lead to insufficient temporal representation under conditions of rapid cloud movement, cloud boundary changes, and complex cloud system evolution. To improve this problem, this invention introduces cloud motion sensing features based on a first-order multispectral observation sequence. The change in atmospheric top reflectance between consecutive observation moments contains information about cloud movement, development, and dissipation. Therefore, by calculating the change in reflectance of each spectral channel at adjacent observation moments and combining the results from multiple consecutive moments, cloud motion sensing features are constructed. Let the atmospheric top reflectance of the c-th spectral channel at the t-th observation moment be Rc(t), then the change in reflectance of the c-th spectral channel between adjacent observation moments is calculated as follows:

[0060]

[0061] In the formula, Let Rc(t) represent the reflectance change of the c-th spectral channel at observation time t; Rc(t) and Rc(t-1) represent the atmospheric top reflectance of the c-th spectral channel at observation times t and t-1, respectively. After calculating the reflectance changes between consecutive observation times for each spectral channel, the results of multiple consecutive times are combined in chronological order to form cloud motion sensing features, which are then used as model inputs along with the multispectral observation features and geometric observation parameters at the current time. The reflectance change information over consecutive times reflects the changes in cloud boundaries and cloud coverage, thus providing cloud change information for subsequent temporal feature extraction.

[0062] For geostationary meteorological satellite observation sequences with minute-level resolution, cloud-induced variations in surface solar radiation exhibit significant multi-timescale characteristics. Short-term abrupt changes in radiation may occur between adjacent observation times due to rapid cloud boundary movement, while a wider historical window may also contain continuous trends resulting from cloud migration, development, and dissipation. To simultaneously characterize these rapid disturbances and continuous evolution processes, this invention employs a selective multi-scale temporal convolution structure based on cloud motion sensing characteristics to extract cloud-induced radiation variation features within different time windows before LSTM time series modeling. Let the model input sequence be X={X1,X2,…,X…} T}, where the input feature X at each time step t It consists of atmospheric top reflectance from a geostationary meteorological satellite at level one, solar zenith angle, observation zenith angle, relative azimuth angle, and cloud motion sensing features constructed based on reflectance changes over continuous observation periods.

[0063] The input sequence is processed through temporal convolution branches at 3, 5, and 7 time steps to obtain feature sequences U under different time windows. 3U 5 and U 7 The superscript numbers represent the corresponding convolutional window lengths. Each temporal convolutional branch uses a one-dimensional temporal convolution with a stride of 1, and padding is applied to both ends of the sequence to ensure that the length of the output sequence matches the length of the input sequence. This guarantees that features extracted from different time windows can be used for subsequent fusion processing. The outputs of the three branches are first fused by adding corresponding elements, and the temporal description of each channel is calculated along the time dimension. The calculation relationship is as follows:

[0064]

[0065] In the formula, U (3) U (5) and U (7) These represent the convolutional feature sequences extracted by the three temporal convolutional branches, respectively. U represents the fused multi-scale feature sequence, and T represents the length of the input sequence. c,t The first descriptive feature S represents the feature value of the c-th channel at time t in the multi-scale feature sequence. The multi-scale feature sequence is then subjected to average pooling along the time dimension to obtain the first descriptive feature S. c S c This represents the temporal description of the c-th channel. The c channels refer to the feature channels output after convolution calculation, and their number is equal to the number of convolution kernels used in the convolution calculation (e.g., 64). Through this processing, the rapid changes in cloud boundaries within a short time window and the continuous evolution of cloud clusters within a wider time window can be aggregated into a unified feature space. Specifically, in some implementations, during the convolution calculation, all target observation data at the same time are first concatenated into a 15-dimensional vector (6 reflectance values ​​+ 6 calculated cloud motion perception features + 3 geometric angles (solar zenith angle, observed zenith angle, relative azimuth angle)), thus forming an input tensor of shape [T, 15]. In the temporal convolution (Conv1D) operation, the number of input channels of the convolution kernel corresponds to these 15-dimensional features, and the sliding direction is along the time axis T; each convolution kernel simultaneously covers all 15 feature channels within a local time window for weighted calculation, with the 3 geometric angles participating in the calculation as ordinary channels.

[0066] To further determine the contribution of different time windows to the prediction of future surface solar radiation, this invention constructs a channel description vector s from the time descriptions Sc of each channel, and obtains intermediate features z through a first fully connected layer and a ReLU activation function. The intermediate features z are input into three second fully connected layers corresponding to the convolutional branches at time steps 3, 5, and 7, respectively, to generate selection score features q(i) for each branch. The selection score features of each branch are normalized along the branch dimension using a Softmax function to obtain the branch weights corresponding to each channel. Subsequently, the features extracted from different time windows are weighted and fused according to the branch weights to obtain the selected multi-scale temporal features, the calculation relationship of which is as follows:

[0067]

[0068]

[0069]

[0070]

[0071] In the formula, s represents the time description of each channel. c The channel description vector; W s and b s Represents the weight matrix and bias vector of the first fully connected layer that generates the intermediate features, respectively; z represents the intermediate feature obtained after fully connected transformation and ReLU activation; W i and b i q represents the weight matrix and bias vector of the second fully connected layer corresponding to the i-th time window branch, respectively; (i) Let qc represent the selection score feature of the i-th time window branch. (i) This represents the selection score of the vector on the c-th channel; i = 1, 2, 3 correspond to the convolution branches at time steps 3, 5, and 7, respectively; a c (i) Y represents the corresponding branch weight. c,t This represents the multi-scale temporal features after selective fusion. The weights and biases of the fully connected layers mentioned above are learned as model parameters during training. Through this selective weighting process, the model can adaptively enhance time window features that are more conducive to prediction and weaken time window features that contribute less, based on the cloud change state in different samples.

[0072] Since the contributions of first-order multispectral reflectance, geometric observation parameters, and cloud motion sensing features to changes in surface solar radiation vary under different cloud conditions and prediction lead times, this invention sets up a lightweight feature recalibration structure after selecting multi-scale temporal features. This structure first performs channel description (average pooling along the time dimension) on the selected multi-scale temporal features, and then generates channel weights through one-dimensional convolution and the sigmoid function along the channel dimension, adaptively adjusting the feature contributions of different channels. Compared to directly stacking complex network layers, this structure can enhance effective feature representation without significantly increasing the number of model parameters. Its calculation relationship is as follows:

[0073]

[0074]

[0075] In the formula, This represents the time description (second description feature) of the c-th channel. C1D() represents a feature vector composed of time descriptions from all channels, C1D() represents a one-dimensional convolution along the channel dimension, and σ() represents the sigmoid function. This represents the calibration features after recalibration. Through this recalibration process, the model can enhance the multispectral, geometric observation, and cloud motion-related channel features that are strongly correlated with future changes in surface solar radiation, while weakening the channel features that contribute less or are redundant. This improves the short-term prediction feature representation ability while maintaining a relatively compact model structure.

[0076] After completing multi-scale temporal feature extraction and feature recalibration, this invention uses LSTM to model the temporal dependencies of the recalibrated sequence features. Continuous observation sequences from geostationary meteorological satellites contain both information on rapid changes in cloud boundaries at adjacent times and continuous evolutionary information on cloud movement, development, and dissipation within a longer historical window. LSTM selectively retains and updates historical information through input gates, forget gates, output gates, and memory units, used to extract temporal dependencies related to future changes in surface solar radiation from continuous first-order satellite observation sequences. Let... Let represent the recalibrated feature vector input to the LSTM at time t. Then, the calculation relationship of the LSTM is as follows:

[0077]

[0078]

[0079]

[0080]

[0081]

[0082] In the formula, i t f t and o t These represent the input gate, forget gate, and output gate, respectively. t h represents the state of a memory cell. t W represents the hidden feature. i W f W o and W c R represents the input weight matrix. i R f R o and R c Let b represent the cyclic weight matrix. i b f b o and b c The term denoted by is the bias term, σ() represents the Sigmoid function, and ⊙ represents element-wise multiplication. Through this gating structure, the model can retain information related to future radiation changes in historical observation sequences and suppress weakly correlated historical disturbances, thereby mitigating the impact of cloud change information attenuation on short-term predictions of surface solar radiation over longer forecast periods.

[0083] After the LSTM outputs the hidden state sequence (hidden features), this invention further employs a time attention module to identify historical moments that contribute more to multi-step surface solar radiation prediction. For different historical moments, the influence of cloud motion, cloud boundary changes, and solar incidence conditions on future radiation variations varies; by weighting the LSTM hidden states with attention weights, the prediction module can focus more on historical observation information that is more effective for short-term future predictions. The calculation relationship for time attention is as follows:

[0084]

[0085]

[0086]

[0087] In the formula, e t W represents the attention score at the t-th historical moment. a and b a This represents the weight matrix and bias vector of the third fully connected layer, v a T Let a represent the learnable projection vector, tanh() represent the tanh activation function, and a t h represents the attention weight at the t-th historical moment. tLet 'z' represent the hidden features at the corresponding time point, and 'z' represent the contextual features obtained by weighted summation of the hidden features at each historical time point using attention weights. This contextual feature aggregates historical time-series information that is more effective for future predictions. The contextual feature is input into the prediction output module, which maps it to multiple predicted values ​​of surface solar radiation corresponding to the number of time steps to be predicted, thus obtaining a predicted sequence of surface solar radiation. In this embodiment, the number of time steps to be predicted is 18, corresponding to a predicted sequence of surface solar radiation for the next 3 hours with a time resolution of 10 minutes.

[0088] In some implementations, the prediction output module is an existing multi-layer (e.g., 3, 4, 5 or more layers) fully connected neural network. This fully connected neural network includes multiple fully connected layers arranged in series, with non-linear activation functions (e.g., ReLU or sigmoid) applied between adjacent fully connected layers. The number of output nodes in the last fully connected layer of the fully connected neural network is the same as the number of time steps to be predicted. Optionally, in some implementations, random deactivation processing can also be applied between adjacent fully connected layers of the fully connected neural network.

[0089] In summary, the theoretical basis of this invention lies in: utilizing the first-level multispectral atmospheric top reflectivity of geostationary meteorological satellites to characterize cloud reflection, cloud spatial texture, and surface background information; utilizing the solar zenith angle, observed zenith angle, and relative azimuth angle to characterize solar incidence conditions and satellite observation geometry; utilizing cloud motion sensing features to characterize cloud movement and cloud boundary changes between consecutive observation times; and utilizing solar radiation simulation samples constructed using RSTAR to supplement the insufficient coverage of real ground-based observation samples. Based on this, selective multi-scale temporal convolution is used to extract cloud-induced radiation variation features within different time windows; lightweight feature recalibration is used to adjust feature channel contributions; historical sequence dependencies are modeled using LSTM; and key historical moment information is extracted through temporal attention. Thus, in the actual prediction stage, short-term prediction of future multi-timestep surface solar radiation is achieved without requiring real-time operation of atmospheric radiative transfer models, explicit cloud parameter inversion results, or numerical weather prediction results as necessary inputs.

[0090] Unlike existing short-term forecasting methods that rely on cloud parameter inversion, numerical weather prediction results, or pre-calculated radiation fields during actual forecasting, this invention utilizes RSTAR to construct solar radiation simulation samples during the training phase and optimizes the model by combining real ground-based solar radiation observation data. During the actual forecasting phase, it directly uses the geostationary meteorological satellite's first-level multispectral observation sequence and geometric observation parameters as the main inputs, and constructs cloud motion sensing features from reflectance changes between consecutive observation times. This method combines physical simulation sample constraints, cloud motion change information, cloud-induced radiation perturbation features at different time scales, and a lightweight feature recalibration process. While reducing intermediate inversion steps and reliance on auxiliary input data during the actual forecasting phase, it enhances the ability to express rapid cloud movement, cloud boundary changes, and complex cloud system evolution, thereby improving the stability and applicability of surface solar radiation forecasting results over longer short-term forecasting periods.

[0091] Figures 1 and 2 respectively show the site verification results of the present invention under different prediction lead times and the regional prediction results under complex cloud conditions. As can be seen from Figure 1, under the prediction lead times of 1 h, 2 h, and 3 h set in this embodiment, the prediction results of the model of the present invention have a good correspondence with the ground-based observation results, with RMSE values ​​of 81.9 W / m², 104.9 W / m², and 122.2 W / m², respectively. Compared with the prediction skill scores of the baseline short-term prediction method, these are 14.1%, 25.5%, and 33.6%, respectively, indicating that the method can continuously predict surface solar radiation at multiple future times and maintain effective prediction capability as the prediction lead time increases. Figure 2 shows the regional surface solar radiation short-term prediction results of the model of the present invention under the case of Typhoon Krosa, and compares them with ERA5 reanalysis data and CERES satellite radiation products. This figure illustrates that the model of this invention can generate continuous regional surface solar radiation prediction results under complex cloud system conditions, and reflects the changes in low-radiation areas, cloud system boundaries, and spatial gradients caused by typhoon cloud systems. Furthermore, by comparing the spatial distribution with ERA5 and CERES, it can be shown that the model of this invention can effectively characterize the spatial variation characteristics of surface solar radiation under the influence of complex cloud systems at the regional scale. These results demonstrate that this invention, while reducing the reliance on intermediate inversion steps and auxiliary input data in the actual prediction stage, can improve the problem of insufficient representation of rapid cloud movement and cloud system evolution in baseline short-term prediction methods under longer short-term effects, and is suitable for short-term surface solar radiation prediction at both the site-scale and regional scales.

[0092] The embodiments described above are merely illustrative of specific implementations of the present invention, and while the descriptions are detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A method for predicting surface solar radiation based on satellite observation data, characterized in that, Includes the following steps: Acquire target observation data from geostationary meteorological satellites and trained prediction models; The data dimensions of the target observation data include atmospheric top reflectivity, solar zenith angle, observation zenith angle, and relative azimuth angle. The target observation data for each data dimension are sequential values ​​obtained from observations at multiple consecutive times. The cloud motion sensing characteristics are obtained by subtracting the target observation data at adjacent observation times for each spectral channel of the atmospheric top reflectance. The target observation data and cloud motion sensing features are combined to form an input sequence, which is then input into the prediction model to obtain the surface solar radiation prediction result.

2. The method according to claim 1, characterized in that, The atmospheric top reflectance is the first-order multispectral atmospheric top reflectance.

3. The method according to claim 1, characterized in that, First, the prediction model is pre-trained using the first dataset, and then the pre-trained prediction model is re-trained using the second dataset to obtain the trained prediction model. In the first dataset, surface solar radiation output from RSTAR software simulation is used as the label data; in the second dataset, ground-based solar radiation observation data is used as the label data.

4. The method according to claim 3, characterized in that, The historical observation data of the target geostationary meteorological satellite and the corresponding ground-based solar radiation observation data are first preprocessed, and then the second dataset is constructed using the preprocessed historical observation data and ground-based solar radiation observation data; the preprocessing includes at least one of time unification, data quality screening and standardization processing.

5. The method according to claim 1, characterized in that, The calculation process within the prediction model includes: The input sequence is convolved using multiple temporal convolutions at different time steps to obtain multiple multi-channel convolutional feature sequences of equal size; Multiple convolutional feature sequences are fused to obtain a multi-scale feature sequence; Based on the multi-scale feature sequence, multiple branch weights corresponding to the convolutional feature sequence are calculated; Based on the multiple branch weights, the convolutional feature sequences corresponding to each branch weight are weighted respectively, and the weighted convolutional feature sequences are summed to obtain multi-scale temporal features; The multi-scale temporal features are recalibrated to obtain calibration features; The calibration features are input into the control module, which calculates the hidden features at each time step. The hidden features at each time step are input into the attention module, and the attention module calculates the attention weights corresponding to each time step. Using the attention weights corresponding to each time step, the hidden features at each time step are weighted and summed to obtain the context features; The context features are input into the prediction output module, which maps the context features to the predicted values ​​of surface solar radiation at multiple prediction times to obtain the predicted surface solar radiation results.

6. The method according to claim 5, characterized in that, Based on the multi-scale feature sequence, multiple branch weights corresponding to the convolutional feature sequence are calculated, including: The multi-scale feature sequence is subjected to average pooling along the time dimension to obtain the first descriptive feature; The first descriptive feature is input into the first fully connected layer and processed by the ReLU function to obtain the intermediate feature; The intermediate features are respectively input into multiple second fully connected layers that correspond one-to-one with the temporal convolutions to obtain multiple selection score features; The softmax function is used to normalize each selection score feature to obtain multiple branch weights.

7. The method according to claim 5, characterized in that, The recalibration of the multi-scale temporal features includes: The multi-scale temporal features are subjected to average pooling along the time dimension to obtain the second descriptive features; The second descriptive feature is processed sequentially by a one-dimensional convolution along the channel dimension and a sigmoid function to obtain the channel weights; The channel weights are multiplied by the corresponding elements of the multi-scale time features to obtain the calibration features.

8. The method according to claim 5, characterized in that, The control module is an LSTM.

9. The method according to claim 5, characterized in that, The internal calculation process of the attention module includes: The hidden features at each time step are input into the third fully connected layer and processed by the tanh function to obtain the attention intermediate features corresponding to each time step. The attention intermediate features corresponding to each time step are multiplied by the learnable projection vector to obtain the attention score corresponding to each time step. The attention scores are normalized using the softmax function to obtain the attention weights at each time step.

10. A surface solar radiation prediction device based on satellite observation data, characterized in that, It includes a processor and a memory, the memory storing a computer program, and the processor loading the computer program to perform the method as described in any one of claims 1-9.

Images