High-resolution solar radiation data correction method and system based on deep learning

Abstract

The invention discloses a high-resolution solar radiation data correction method and system based on deep learning, and the method comprises the steps: obtaining meteorological data and digital elevation model data, and carrying out the preprocessing; dynamically optimizing parameter configuration of the WRF model, and simulating and outputting high-temporal-spatial-resolution meteorological data as training data; fusing solar radiation flux values in meteorological reanalysis data and near-real-time product meteorological data as label data of a correction model; inputting the high-temporal-spatial-resolution meteorological data and the label data into a pre-constructed multi-scale feature fused deep residual coding and decoding network model for training, and obtaining an optimal parameter model based on multiple loop iterations; and correcting solar radiation flux data output by the WRF mode by using the optimal parameter model. According to the method, the problem that the radiation variable value output by the WRF in the traditional numerical weather mode is consistent with the real data trend but deviates is effectively solved, and the reliability of meteorological data simulation is enhanced.

Description

Technical Field

[0001] The present invention relates to the field of intelligent meteorological technology, and particularly to a method and system for correcting high-resolution solar radiation data based on deep learning. Background Art

[0002] In recent years, solar radiation, as an important factor affecting multiple fields such as climate, energy production, and agricultural irrigation, has received extensive attention. Accurately predicting the solar radiation flux is of great significance for multiple fields such as climate simulation, energy production, and agricultural irrigation. Traditional numerical weather prediction (NWP) models, such as WRF (Weather Research and Forecasting Model), have been widely used in meteorological prediction. However, in the simulation of solar radiation flux, although models such as WRF can better capture the changing trend of radiation, due to the limitations of their physical parameterization schemes, there are often deviations between the simulation results and the actual observed data, especially when the spatio-temporal resolution of solar radiation is relatively high. Therefore, it is crucial to post-correct the radiation data output by numerical weather models.

[0003] To address this issue, the correction of model output has become an important means to improve simulation accuracy. Traditional correction methods mostly rely on statistical models and empirical formulas, and adjust and correct by comparing simulation results with observed data. These methods are usually limited by the sparsity of data and computing power, and it is difficult to handle solar radiation data with high spatio-temporal resolution. The correction based on statistical methods often cannot fully capture the complex non-linear relationship between model output and actual observation.

[0004] The application of deep learning technology in the correction of meteorological model output has gradually emerged. Deep learning models, especially methods based on neural networks, have powerful data processing and pattern recognition capabilities, and can automatically extract features from a large amount of meteorological and radiation data to capture complex non-linear relationships. However, existing deep learning methods still face many challenges in the process of model output correction, including high requirements for computing resources, long model training time, and difficulties in effectively dealing with the complexity of different geographical and meteorological characteristics. Summary of the Invention

[0005] In view of the above existing problems, the present invention is proposed.

[0006] Therefore, the present invention provides a method and system for correcting high-resolution solar radiation data based on deep learning to solve the problems of high requirements for computing resources, long model training time, and difficulties in effectively dealing with the complexity of different geographical and meteorological characteristics in the prior art.

[0007] To solve the above technical problems, the present invention provides the following technical solutions:

[0008] In a first aspect, the present invention provides a method for correcting high-resolution solar radiation data based on deep learning, including: obtaining meteorological data and digital elevation model data, and preprocessing the obtained data; setting the regional nested grid and resolution in the WRF preprocessing configuration file according to the characteristics of the simulation area, dynamically screening out the optimal parameterization scheme combination, and writing the scheme into the WRF running configuration file; using the preprocessed meteorological data and digital elevation model data as the initial field to drive the WRF model to generate high spatio-temporal resolution meteorological data; setting different fusion weights according to geographical and meteorological characteristics, and fusing the solar radiation flux values in the meteorological reanalysis data and near-real-time product meteorological data as the labeled data of the correction model; inputting the high spatio-temporal resolution meteorological data and the fused labeled data into a pre-constructed deep residual encoder-decoder network model with multi-scale feature fusion for training, and obtaining the optimal parameter model based on multiple loop iterations; using the optimal parameter model to correct the solar radiation flux data output by the WRF model.

[0009] As a preferred embodiment of the method for correcting high-resolution solar radiation data based on deep learning according to the present invention, wherein: the setting of the regional nested grid and resolution in the WRF preprocessing configuration file, dynamically screening out the optimal parameterization scheme combination, and writing the scheme into the WRF running configuration file includes:

[0010] Determine the grid range according to the simulation area, set the number of nested grid layers and resolution, and configure parameters in the WRF preprocessing file, the parameters including grid position, size, resolution, time step and number of model layers;

[0011] Use an automated script to dynamically select different shortwave radiation and boundary layer parameterization scheme combinations in the preset physical parameterization scheme combination, randomly select multiple different simulation time ranges to test each parameterization scheme combination, and obtain the simulation results;

[0012] Collect the corresponding observation data, calculate the mean absolute error and Pearson correlation coefficient between the simulation results and the observation data, and select the physical scheme combination with the smallest comprehensive index as the optimal parameterization scheme combination according to the comprehensive index formula introducing the correlation threshold, and write the scheme into the WRF running configuration file.

[0013] As a preferred embodiment of the method for correcting high-resolution solar radiation data based on deep learning according to the present invention, wherein: the fusion of the solar radiation flux values in the meteorological reanalysis data and near-real-time product meteorological data as the labeled data of the correction model includes:

[0014] Resample the digital elevation model data to the meteorological reanalysis grid to obtain an elevation data set matching the meteorological reanalysis grid;

[0015] Calculate the geographical feature weighting coefficient w according to the altitude and latitude of the grid points geo , the formula is:

[0016]

[0017] where ALT is the altitude value of the grid point, and ALT average is the average altitude of the simulation area, and LAT is the latitude value of the grid point;

[0018] Calculate the meteorological feature weighting coefficient w using the cloud cover data in the meteorological reanalysis data met , the formula is:

[0019] w met = 1 - CLDC

[0020] where CLDC is the cloud cover value of the grid point;

[0021] According to the geographical feature weighting coefficient and the meteorological feature weighting coefficient, fuse the radiation data of the meteorological reanalysis data and the near-real-time product meteorological data, and the formula is expressed as:

[0022] D fused (i,j) = w geo (i,j)·w met (i,j)·D ra (i,j)+(1 - w geo (i,j)·w met (i,j))·D rt (i,j)

[0023] where (i,j) is the spatial coordinate of each grid point, D ra is the radiation value in the meteorological reanalysis data, D rt is the radiation value of the near-real-time product meteorological data, D fused is the fused radiation value.

[0024] As a preferred solution of the high-resolution solar radiation data correction method based on deep learning described in the present invention, wherein: the pre-constructed deep residual encoder-decoder network model with multi-scale feature fusion includes:

[0025] Modify the convolution kernel parameters of the first layer of the residual backbone network, and expand the input channels of the residual backbone network to the number of meteorological element channels;

[0026] Embed the DFWM dynamic feature weighting module in each stage of the encoder, generate channel weights by introducing the channel attention mechanism using global average pooling and fully connected layers, weight the input feature map, and adaptively allocate the contribution weights of the enhanced radiation-related features, as shown in the following formula:

[0027]

[0028] Among them, F out is the output feature map, and F in is the input feature map. is the per-channel product, GAP() is the global average pooling, FC i () is the fully connected layer, ReLU() is the rectified linear activation function, and σ() is the sigmoid activation function;

[0029] In the decoding stage, the low-level high-resolution features of the encoder and the high-level semantic features are fused through cross-layer skip connections. Bilinear interpolation is used to restore the spatial resolution. The MPFEU multi-path feature enhancement unit is introduced, and lightweight residual blocks are embedded in the skip connections. Local details and global context features are extracted through parallel convolutional branches, as shown in the following formula:

[0030] F mpfe = Cat(Conv 3×3 (F in ), Conv 5×5 (F in ))

[0031] Among them, F mpfe is the output feature map of the multi-path feature enhancement unit, F in is the input feature map, Conv i×i () is the i×i convolution operation, and Cat() is the channel dimension concatenation;

[0032] A progressive output layer is adopted to gradually compress the number of channels through multi-level convolutions while keeping the spatial resolution unchanged, and a single-channel radiation correction value is output.

[0033] As a preferred solution of the high-resolution solar radiation data correction method based on deep learning described in the present invention, wherein: the obtaining the optimal parameter model based on multiple loop iterations includes:

[0034] Adopt a multi-objective joint optimization strategy, introduce a dynamic weight adjustment mechanism, and automatically balance the weights of each loss term according to the training stage. The mixed loss function consists of the mean square error, structural similarity, and radiation conservation physical constraint terms;

[0035] Adopt a validation loss-guided mixed scheduler, combine the ReduceLROnPlateau learning rate adjustment optimizer with the cosine annealing algorithm, dynamically balance the learning rate decrease speed, adaptively adjust the learning rate, and then obtain the optimal parameter model.

[0036] As a preferred solution of the high-resolution solar radiation data correction method based on deep learning described in the present invention, wherein: the preprocessing of the obtained data includes:

[0037] Convert the radiation cumulative values in meteorological reanalysis data and near-real-time product meteorological data into instantaneous values;

[0038] Divide the meteorological reanalysis and near-real-time product grid data into multiple sub-grids, detect non-zero outliers within each sub-grid according to the three-sigma rule, and use the interpolation method of adjacent spatio-temporal points to repair non-zero outliers;

[0039] Use the bicubic interpolation method to adjust the spatial resolution of meteorological reanalysis data and near-real-time product meteorological data to be consistent with the WRF model grid resolution.

[0040] As a preferred solution of the high-resolution solar radiation data correction method based on deep learning described in the present invention, wherein: the preprocessing of the acquired data further includes:

[0041] Integrate the digital elevation model data in tile format and convert it into a binary format that can be directly read by the WRF model.

[0042] In a second aspect, the present invention provides a high-resolution solar radiation data correction system based on deep learning, including:

[0043] A data preprocessing module, configured to acquire meteorological data and digital elevation model data, and preprocess the acquired data;

[0044] A configuration optimization module, configured to set the regional nested grid and resolution in the WRF preprocessing configuration file according to the characteristics of the simulation area, dynamically screen out the optimal parameterization scheme combination, and write the scheme into the WRF running configuration file;

[0045] A data processing and fusion module, configured to use the preprocessed meteorological data and digital elevation model data as the initial field to drive the WRF model to generate high spatio-temporal resolution meteorological data; set different fusion weights according to geographical and meteorological characteristics, and fuse the solar radiation flux values in meteorological reanalysis data and near-real-time product meteorological data as the label data of the correction model;

[0046] A training optimization module, configured to input the high spatio-temporal resolution meteorological data and the fused label data into a pre-constructed deep residual encoding and decoding network model with multi-scale feature fusion for training, and obtain the optimal parameter model based on multiple loop iterations;

[0047] A correction output module, configured to correct the solar radiation flux data output by the WRF model using the optimal parameter model.

[0048] In a third aspect, the present invention provides an electronic device, including:

[0049] A memory and a processor;

[0050] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the high-resolution solar radiation data correction method based on deep learning are implemented.

[0051] In a fourth aspect, the present invention provides a computer-readable storage medium storing computer-executable instructions, and when the computer-executable instructions are executed by a processor, the steps of the high-resolution solar radiation data correction method based on deep learning are implemented.

[0052] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0053] ① The present invention deeply integrates multi-source meteorological and terrain collaborative features, constructs training samples by combining multi-dimensional data such as WRF model output and terrain elevation, solves the problem of limited representation ability of single meteorological data, and significantly improves the correction accuracy of solar radiation flux in complex terrain areas.

[0054] ② The lightweight multi-scale feature fusion mechanism adopted by the present invention uses hierarchical residual connection and dynamic channel compression technology. While ensuring the spatial details of the high-resolution radiation field, it reduces the model calculation amount to half of the traditional U-Net, realizing real-time correction and business deployment.

[0055] ③ The present invention uses a physics-driven training strategy. By embedding a radiation energy conservation constraint equation, it breaks through the bottleneck of poor physical consistency of traditional data-driven models, effectively reduces the monthly average error volatility of the correction results, and greatly improves the long-term reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0057] Figure 1 It is a schematic diagram of the overall process logic of the method according to an embodiment of the present invention;

[0058] Figure 2 It is a structural diagram of the RadCorrNet post-correction model of the method according to an embodiment of the present invention;

[0059] Figure 3 It is an internal structural diagram of the encoder module of the method according to an embodiment of the present invention;

[0060] Figure 4Internal structure diagram of the decoder module of the method according to an embodiment of the present invention;

[0061] Figure 5 Convergence performance comparison chart of different correction models of the method according to an embodiment of the present invention;

[0062] Figure 6 MSE experimental comparison chart of different correction models of the method according to an embodiment of the present invention;

[0063] Figure 7 SSIM experimental comparison chart of different correction models of the method according to an embodiment of the present invention. Detailed implementation manners

[0064] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following detailed description of the specific implementation manners of the present invention will be given in conjunction with the accompanying drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention.

[0065] Embodiment 1

[0066] Refer to Figures 1-4 An embodiment of the present invention provides a high-resolution solar radiation data correction method based on deep learning, as Figure 1 shown, which specifically includes the following steps:

[0067] S100: Obtain meteorological data and digital elevation model data, and preprocess the obtained data;

[0068] S200: According to the characteristics of the simulation area, set the regional nested grid and resolution in the WRF preprocessing configuration file, dynamically screen out the optimal parameterization scheme combination, and write the scheme into the WRF running configuration file;

[0069] S300: Use the preprocessed meteorological data and digital elevation model data as the initial field to drive the WRF model to generate high spatio-temporal resolution meteorological data;

[0070] S400: Set different fusion weights according to geographical and meteorological characteristics, and fuse the solar radiation flux values in the meteorological reanalysis data and near-real-time product meteorological data as the label data of the correction model;

[0071] S500: Construct a deep residual encoder-decoder network model with multi-scale feature fusion, that is, the RadCorrNet radiation variable post-correction model;

[0072] S600: Input the high spatiotemporal resolution meteorological data and the fused label data into the RadCorrNet radiation variable post-correction model for training, and obtain the optimal parameter model based on multiple loop iterations;

[0073] S700: Use the optimal parameter model to correct the solar radiation flux data output by the WRF model.

[0074] It should be noted that: ① The present invention deeply fuses multi-source meteorological and terrain collaborative features, constructs training samples by jointly using multi-dimensional data such as WRF model output and terrain elevation, solves the problem of limited representation ability of single meteorological data, and significantly improves the correction accuracy of solar radiation flux in complex terrain areas; ② The lightweight multi-scale feature fusion mechanism adopted by the present invention uses hierarchical residual connection and dynamic channel compression technology. While ensuring the spatial details of the high-resolution radiation field, it reduces the model calculation amount to half of the traditional U-Net, realizing real-time correction and operational deployment; ③ The present invention uses a physics-driven training strategy. By embedding the radiation energy conservation constraint equation, it breaks through the bottleneck of poor physical consistency of traditional data-driven models, effectively reducing the monthly average error volatility of the correction results and greatly improving the long-term reliability.

[0075] In the embodiment of the present application, the above step S100 includes the following sub-steps A1 to A2:

[0076] In A1: Obtain meteorological data such as meteorological reanalysis data, near-real-time product meteorological data, and historical irradiance data of photovoltaic power stations, and digital elevation model data; the meteorological reanalysis data includes NCEP data and ERA5-Land data, the near-real-time product meteorological data is CLDAS-V2.0 data, and the numerical elevation model data is SRTM's DEM data;

[0077] In A2: Preprocess the obtained multi-source data such as meteorology and terrain, use the formula to convert the cumulative value of the radiation variable in the ERA5-Land data into an instantaneous value, check and remove the non-zero outliers in the converted ERA5-Land data and CLDAS-V2.0 data, adjust the spatial resolution of the ERA5-Land data and CLDAS-V2.0 data to be consistent with the WRF model grid, and integrate the tiled DEM data and convert it into a binary format directly readable by the WRF model;

[0078] Specifically, convert the radiation cumulative value in the meteorological reanalysis data and the near-real-time product meteorological data into an instantaneous value, and the formula is expressed as:

[0079]

[0080] where SR h is the cumulative solar radiation amount in a certain hour, and SR h-1is the cumulative solar radiation in the previous 1 hour, and SR is the instantaneous radiation flux;

[0081] Specifically, the meteorological reanalysis and near-real-time product grid data are divided into 100 sub-grids (10x10). Non-zero outliers are detected within each sub-grid according to the three-times standard deviation rule, and non-zero outliers are repaired using the interpolation method of neighboring spatio-temporal points, as shown in the following formula:

[0082]

[0083] where N i,j represents the set of neighboring points of the grid point (i,j), and x i,j represents the valid value of the neighboring grid point. |N i,j | is the number of valid values among the neighboring points, and x repaired is the repaired value;

[0084] Specifically, the bicubic interpolation method is used to adjust the spatial resolution of the meteorological reanalysis data and near-real-time product meteorological data to be consistent with the WRF model grid resolution, as shown in the following formula:

[0085]

[0086] where f(x i ,y j ) is the data value of the neighboring 16 grid points, f(x,y) is the data value of the target grid after interpolation, and the weight function w(i,j) is determined by the bicubic kernel function h(t), as shown in the following formula:

[0087] w(i,j) = h(|x - x i |)·h(|y - y j |)

[0088]

[0089] where |x - x i | is the horizontal distance between the target grid point and the neighboring grid point, |y - y j | is the vertical distance between the target grid point and the neighboring grid point, and t represents the relative distance between the target point and the grid point;

[0090] Specifically, the digital elevation model data in tile format is integrated and converted into a binary format that can be directly read by the WRF model.

[0091] In the embodiment of the present application, in step S200, according to the characteristics of the simulation area, the nested grid and resolution in the WRF preprocessing configuration file are set, and the best combination of different shortwave radiation and boundary layer parameterization schemes among 12 preset physical parameterization scheme combinations is dynamically selected by using an automated script, and this scheme is written into the WRF running configuration file; among them, the preset parameterization scheme combinations are 12 different parameterization scheme combinations including the WSM6 cloud microphysics parameterization scheme, the Kain-Fritsch cumulus convection parameterization scheme, the RRTMG longwave radiation parameterization scheme, 3 shortwave radiation parameterization schemes, 4 planetary boundary layer parameterization schemes, and the Noah LSM land surface process parameterization scheme; among them, the 3 shortwave radiation parameterization schemes are Dudhia, RRTMG, and Goddard, and the 4 planetary boundary layer parameterization schemes are YSU, MYJ, ACM2, and BL; step S200 specifically includes the following sub-steps B1 to B3:

[0092] In B1: Determine the grid range according to the simulation area, set the number of nested grid layers and resolution, and configure the parameters in the WRF preprocessing file, where the parameters include grid position, size, resolution, time step, and number of model layers;

[0093] In B2: Dynamically select different shortwave radiation and boundary layer parameterization scheme combinations among the preset physical parameterization scheme combinations by using an automated script, randomly select multiple different simulation time ranges to test each parameterization scheme combination, and obtain the simulation results;

[0094] In B3: Collect the corresponding observation data, calculate the mean absolute error and Pearson correlation coefficient between the simulation results and the observation data, and select the physical scheme combination with the smallest comprehensive index as the optimal parameterization scheme combination according to the comprehensive index formula introducing the correlation threshold, and write the scheme into the WRF running configuration file, where the calculation formula of the comprehensive index is:

[0095]

[0096] Wherein, MAE norm is the normalized mean absolute error value, r is the Pearson correlation coefficient between the simulation data and the observation data, and MR is the comprehensive index value.

[0097] It should be noted that the above step S200 ensures the high precision and reliability of the model under specific geographical and meteorological conditions, provides an accurate configuration for generating high spatio-temporal resolution meteorological data, and improves the accuracy of solar radiation flux correction in complex terrain areas.

[0098] In the embodiment of the present application, the above step S400 includes the following sub-steps D1 to D4:

[0099] In D1: Resample the digital elevation model data to the meteorological reanalysis grid to obtain an altitude dataset that matches the meteorological reanalysis grid;

[0100] In D2: Calculate the geographical feature weighting coefficient w according to the altitude and latitude of the grid points geo , and the formula is:

[0101]

[0102] where ALT is the altitude value of the grid point, ALT average is the average altitude of the simulation area, and LAT is the latitude value of the grid point;

[0103] In D3: Use the cloud cover data in the meteorological reanalysis data to calculate the meteorological feature weighting coefficient w met , and the formula is:

[0104] w met = 1 - CLDC

[0105] where CLDC is the cloud cover value of the grid point;

[0106] In D4: According to the geographical feature weighting coefficient and the meteorological feature weighting coefficient, fuse the radiation data of the meteorological reanalysis data and the near-real-time product meteorological data, and the formula is expressed as:

[0107] D fused (i,j) = w geo (i,j)·w met [[ID=?]] ra (i,j)·D geo (i,j)+(1 - w met (i,j)·w rt (i,j))·D ra (i,j)

[0108] where (i,j) is the spatial coordinate of each grid point, D ra is the radiation value in the meteorological reanalysis data, D rt is the radiation value of the near-real-time product meteorological data, D fused is the fused radiation value.

[0109] It should be noted that the above step S400 fuses the solar radiation flux values in the meteorological reanalysis data and the near-real-time product meteorological data by setting different fusion weights, making full use of the advantages of multi-source data, improving the accuracy and representativeness of the labeled data, and thus enhancing the precision and reliability of the final solar radiation flux correction result.

[0110] In the embodiment of the present application, the above step S500 constructs a deep residual encoder-decoder network model for multi-scale feature fusion, as Figure 2 There seems to be an unclear tag in the original text at line 38 where the tag ra is shown. This might be an error in the original. The translation is done as accurately as possible based on the available text.As shown, it is the RadCorrNet radiation variable post-correction model, which specifically includes the following sub-steps E1 to E4:

[0111] In E1: Modify the convolutional kernel parameters of the first layer of the residual backbone network, and expand the input channels of the residual backbone network to the number of meteorological element channels;

[0112] In E2: As Figure 3 shown, embed the DFWM dynamic feature weighting module in each stage of the encoder. By introducing the channel attention mechanism, use global average pooling and fully connected layers to generate channel weights, weight the input feature map, and adaptively allocate and strengthen the contribution weights of radiation-related features, as shown in the following formula:

[0113]

[0114] Among them, F out is the output feature map, F in is the input feature map, is the per-channel product, GAP() is global average pooling, FC i () is the fully connected layer, ReLU() is the rectified linear activation function, and σ() is the sigmoid activation function;

[0115] In E3: As Figure 4 shown, in the decoding stage, fuse the low-level high-resolution features and high-level semantic features of the encoder through cross-layer skip connections, use bilinear interpolation to restore the spatial resolution, introduce the MPFEU multi-path feature enhancement unit, embed lightweight residual blocks in the skip connections, and extract local details and global context features through parallel convolutional branches, as shown in the following formula:

[0116] F mpfe = Cat(Conv 3×3 (F in ), Conv 5×5 (F in ))

[0117] Among them, F mpfe is the output feature map of the multi-path feature enhancement unit, F in is the input feature map, Conv i×i () is the i×i convolution operation, and Cat() is channel dimension concatenation;

[0118] In E4: Adopt a progressive output layer, gradually compress the number of channels through multiple levels of convolution while keeping the spatial resolution unchanged, and output a single-channel radiation correction value.

[0119] In the embodiment of the present application, the above step S600 includes the following sub-steps F1 to F2:

[0120] In F1: A multi-objective joint optimization strategy is adopted, introducing a dynamic weight adjustment mechanism to automatically balance the weights of each loss term during the training phase. The hybrid loss function consists of the mean squared error, structural similarity, and a physical constraint term for radiative conservation, and the formula is:

[0121]

[0122] α + β + γ = 1

[0123] where, L total is the hybrid loss value, L MSE is the mean squared error value, L SSIM is the structural similarity value, is the radiative value output by the correction model, y true is the radiative label value, and α, β, and γ are the weights of the MSE loss, 1 - SSIM loss, and the physical constraint term for radiative conservation, respectively;

[0124] In F2: A hybrid scheduler guided by the validation loss is adopted, combining the ReduceLROnPlateau learning rate adjustment optimizer and the cosine annealing algorithm to dynamically balance the learning rate decay speed and adaptively adjust the learning rate, thereby obtaining an optimal parameter model.

[0125] It should be noted that the above steps not only improve the learning ability and prediction accuracy of the model, but also effectively capture the characteristics of solar radiation flux changes under complex terrains, achieving high-precision correction results and real-time operational deployment.

[0126] In the embodiment of the present application, the above step S700 corrects the solar radiation flux data output by the WRF model using the optimal parameter model obtained based on multiple loop iterations.

[0127] It should be noted that the above step S700 effectively improves the accuracy and reliability of the original simulation data. Especially in complex terrain areas, through the physically driven training strategy and multi-source data fusion, the long-term error fluctuations are significantly reduced, enhancing the stability and precision of solar radiation flux prediction.

[0128] Embodiment 2

[0129] Referring to Figures 5-7 , based on the previous embodiment, this embodiment provides an application example of the high-resolution solar radiation data correction method and system based on deep learning to verify and illustrate the technical effects adopted in this method.

[0130] As Figure 5The figure shows the comparison of the convergence performance of different correction models. During the convergence process of the model RadCorrNet provided by the present invention, its loss value rapidly decreases as the number of training epochs increases and tends to be stable within a short period of time, showing good convergence performance. RadCorrNet can quickly adapt and stabilize in fewer training epochs. Although the loss values of the Unet, DenseNet, and Res-Unet models also decrease, their decrease rates are relatively gentle, and there are still certain fluctuations in the later stage of training. This convergence trend indicates that the model RadCorrNet provided by the present invention has better convergence speed and stability than other common models in the radiation data correction task, reflecting its efficiency and advantages in practical applications.

[0131] As Figure 6 The figure shows the comparison of the MSE (Mean Squared Error) of different correction models. The MSE value of the model RadCorrNet provided by the present invention always remains at the lowest level during most time steps and changes relatively smoothly, demonstrating its efficiency and stability in the radiation data correction process. The MSE values of the Res-Unet, DenseNet, and Unet models show large fluctuations at some time steps and the overall error is relatively high. Generally speaking, the model RadCorrNet provided by the present invention has obvious advantages in processing radiation data and can provide more accurate correction results throughout the time period.

[0132] As Figure 7 The figure shows the comparison of the SSIM (Structural Similarity Index Measure) of different correction models. The SSIM value of the model RadCorrNet provided by the present invention always remains high and stable throughout the time period, significantly superior to other models. The SSIM value of RadCorrNet remains above 0.9 for most of the time, indicating that the model can effectively maintain the consistency and accuracy of the spatial distribution of solar radiation data. The SSIM values of the Res-Unet, DenseNet, and Unet models are relatively low and fluctuate greatly, and their performance in the spatial distribution structure is not as good as the model RadCorrNet provided by the present invention. The performance of RadCorrNet in maintaining the spatial distribution structure and correcting meteorological data ensures the reliability of the correction results in space.

[0133] Therefore, from the above experimental results, it can be seen that the present invention deeply integrates multi-source meteorological and terrain collaborative features, constructs training samples by combining multi-dimensional data such as WRF model output and terrain elevation, solves the problem of limited representation ability of single meteorological data, and significantly improves the correction accuracy of solar radiation flux in complex terrain areas; the lightweight multi-scale feature fusion mechanism adopted by the present invention uses hierarchical residual connection and dynamic channel compression technology, while ensuring the spatial details of the high-resolution radiation field, reduces the model calculation amount to half of the traditional U-Net, and realizes real-time correction and business deployment; the present invention uses a physics-driven training strategy, breaks through the bottleneck of poor physical consistency of traditional data-driven models by embedding a radiation energy conservation constraint equation, effectively reduces the monthly average error volatility of the correction results, and greatly improves the long-term reliability.

[0134] Embodiment 3

[0135] This embodiment provides a high-resolution solar radiation data correction system based on deep learning, including:

[0136] A data preprocessing module, configured to obtain meteorological data and digital elevation model data, and preprocess the obtained data;

[0137] A configuration optimization module, configured to set the regional nested grid and resolution in the WRF preprocessing configuration file according to the characteristics of the simulation area, dynamically screen out the optimal parameterization scheme combination, and write the scheme into the WRF operation configuration file;

[0138] A data processing and fusion module, configured to use the preprocessed meteorological data and digital elevation model data as the initial field to drive the WRF model to generate high spatio-temporal resolution meteorological data; set different fusion weights according to geographical and meteorological characteristics, and fuse the solar radiation flux values in meteorological reanalysis data and near-real-time product meteorological data as the label data of the correction model;

[0139] A training optimization module, configured to input the high spatio-temporal resolution meteorological data and the fused label data into a pre-constructed deep residual encoder-decoder network model with multi-scale feature fusion for training, and obtain the optimal parameter model based on multiple loop iterations;

[0140] A correction output module, configured to correct the solar radiation flux data output by the WRF model using the optimal parameter model.

[0141] It should be noted that the technical solution of the system for correcting high-resolution solar radiation data based on deep learning belongs to the same concept as the technical solution of the above-mentioned method for correcting high-resolution solar radiation data based on deep learning. For the details not described in detail in the technical solution of the system for correcting high-resolution solar radiation data based on deep learning in this embodiment, reference can be made to the description of the technical solution of the above-mentioned method for correcting high-resolution solar radiation data based on deep learning.

[0142] The above-mentioned unit modules can be embedded in the processor in the computer device in hardware form or be independent of it, or can be stored in the memory in the computer device in software form, so as to facilitate the processor to call and execute the operations corresponding to the above-mentioned modules.

[0143] This embodiment also provides an electronic device, which includes a processor, a memory, a communication interface, a display screen, and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be realized through WIFI, a carrier network, NFC (Near Field Communication), or other technologies. The computer program, when executed by the processor, implements the method for correcting high-resolution solar radiation data based on deep learning. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covering the display screen, or can also be a button, a trackball, or a touchpad provided on the shell of the computer device, or can also be an external keyboard, a touchpad, or a mouse, etc.

[0144] This embodiment also provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by the processor, it implements the method proposed in the above-mentioned embodiment.

[0145] The storage medium proposed in this embodiment and the method proposed in the above-mentioned embodiment belong to the same inventive concept. For the technical details not described in detail in this embodiment, reference can be made to the above-mentioned embodiment, and this embodiment has the same beneficial effects as the above-mentioned embodiment.

[0146] From the above description of the embodiments, those skilled in the art can clearly understand that the present invention can be implemented by means of software and the necessary general-purpose hardware. Of course, it can also be implemented by hardware, but in many cases the former is a better implementation. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disc of a computer, etc., including several instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute the method of the embodiments of the present invention.

[0147] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.

[0148] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present application can be implemented in various computer languages.

[0149] The present application is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram, as well as the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the functions specified in one Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0150] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to operate in a particular manner, such that the instructions stored in the computer-readable memory produce a manufacture including an instruction device that implements the functions specified in one or more of the processes Figure 1 or blocks. Figure 1 The functions specified in one or more of the processes

[0151] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, so that the instructions executed on the computer or other programmable apparatus provide steps for implementing the functions specified in one or more of the processes Figure 1 or blocks. Figure 1 The functions specified in one or more of the blocks.

[0152] Although the preferred embodiments of the present application have been described, additional changes and modifications can be made by those skilled in the art once they learn of the basic inventive concept. Therefore, the appended claims are intended to be construed to include the preferred embodiments as well as all changes and modifications that fall within the scope of the present application.

[0153] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalent technologies, the present application is also intended to include these modifications and variations.

Claims

1. A method for correcting high-resolution solar radiation data based on deep learning, characterized in that, Including: Obtain meteorological data and digital elevation model data, and preprocess the obtained data; According to the characteristics of the simulation area, set the regional nested grid and resolution in the WRF preprocessing configuration file, dynamically screen out the optimal parameterization scheme combination, and write the scheme into the WRF running configuration file; Use the preprocessed meteorological data and digital elevation model data as the initial field to drive the WRF model to generate high spatiotemporal resolution meteorological data; Set different fusion weights according to geographical and meteorological characteristics, and fuse the solar radiation flux values in the meteorological reanalysis data and near-real-time product meteorological data as the labeled data for the correction model; Input the high spatiotemporal resolution meteorological data and the fused labeled data into a pre-constructed deep residual encoder-decoder network model with multi-scale feature fusion for training, and obtain the optimal parameter model based on multiple loop iterations; Use the optimal parameter model to correct the solar radiation flux data output by the WRF model.

2. The high-resolution solar radiation data correction method based on deep learning according to claim 1, wherein The setting of the regional nested grid and resolution in the WRF preprocessing configuration file, dynamically screening out the optimal parameterization scheme combination, and writing the scheme into the WRF running configuration file includes: Determine the grid range according to the simulation area, set the number of nested grid layers and resolution, and configure parameters in the WRF preprocessing file, where the parameters include grid position, size, resolution, time step, and model layer; Use an automated script to dynamically select different shortwave radiation and boundary layer parameterization scheme combinations in the preset physical parameterization scheme combination, randomly select multiple different simulation time ranges to test each parameterization scheme combination, and obtain simulation results; Collect corresponding observation data, calculate the mean absolute error and Pearson correlation coefficient between the simulation results and the observation data, and select the physical scheme combination with the smallest comprehensive index as the optimal parameterization scheme combination according to the comprehensive index formula introducing the correlation threshold, and write the scheme into the WRF running configuration file.

3. The high-resolution solar radiation data correction method based on deep learning according to claim 2, characterized in that, The fusing of the solar radiation flux values in the meteorological reanalysis data and near-real-time product meteorological data as the labeled data for the correction model includes: Resample the digital elevation model data to the meteorological reanalysis grid to obtain an elevation data set matching the meteorological reanalysis grid; Calculate the geographical feature weighting coefficient w based on the altitude and latitude of the grid points geo , and the formula is: Among them, ALT is the altitude value of the grid point, and ALT average is the average altitude of the simulation area, and LAT is the latitude value of the grid point; Calculating the meteorological characteristic weighting coefficient w using cloud cover data in meteorological reanalysis data met , and the formula is: w met = 1 - CLDC Wherein, CLDC is the cloud cover value of the grid point; According to the geographical feature weighting coefficient and meteorological feature weighting coefficient, fuse the radiation data of the meteorological reanalysis data and near-real-time product meteorological data, and the formula is expressed as: D fused (i,j) = w geo (i,j) · w met (i,j) · D ra (i,j) + (1 - w geo (i,j) · w met (i,j)) · D rt (i,j) where (i,j) are the spatial coordinates of each grid point, and D ra is the radiation value in the meteorological reanalysis data, and D rt is the radiation value in the near-real-time product meteorological data, and D fused is the fused radiation value.

4. The high-resolution solar radiation data correction method based on deep learning according to claim 3, characterized in that The pre-constructed deep residual encoder-decoder network model with multi-scale feature fusion includes: Modify the first layer convolution kernel parameters of the residual backbone network, and expand the input channels of the residual backbone network to the number of meteorological element channels; Embed the DFWM dynamic feature weighting module in each stage of the encoder, generate channel weights by introducing the channel attention mechanism using global average pooling and fully connected layers, weight the input feature map, and adaptively allocate and strengthen the contribution degree weights of radiation-related features, as shown in the following formula: Among them, F out is the output feature map, F in is the input feature map, is the per-channel product, GAP() is the global average pooling, FC i () is the fully connected layer, ReLU() is the rectified linear activation function, σ() is the sigmoid activation function; In the decoding stage, the low-level high-resolution features of the encoder and the high-level semantic features are fused through cross-layer skip connections. Bilinear interpolation is used to restore the spatial resolution. The MPFEU (Multi-Path Feature Enhancement Unit) is introduced, and lightweight residual blocks are embedded in the skip connections. Local details and global context features are extracted through parallel convolutional branches, as shown in the following formula: F mpfe = Cat(Conv 3×3 (F in ), Conv 5×5 (F in )) Among them, F mpfe is the output feature map of the multi-path feature enhancement unit, and F in is the input feature map. Conv i×i () represents the i×i convolution operation, and Cat() represents the channel dimension concatenation; A progressive output layer is adopted to gradually compress the number of channels through multiple-level convolutions while keeping the spatial resolution unchanged, and a single-channel radiative correction value is output.

5. The high-resolution solar radiation data correction method based on deep learning according to claim 4, characterized in that, The obtaining of the optimal parameter model based on multiple loop iterations includes: Adopt a multi-objective joint optimization strategy, introduce a dynamic weight adjustment mechanism, and automatically balance the weights of each loss term according to the training stage. The mixed loss function consists of mean square error, structural similarity, and a physical constraint term of radiative conservation; Adopt a validation loss-guided hybrid scheduler, combine the ReduceLROnPlateau learning rate adjustment optimizer with the cosine annealing algorithm, dynamically balance the learning rate decline speed, adaptively adjust the learning rate, and thus obtain the optimal parameter model.

6. The high-resolution solar radiation data correction method based on deep learning according to claim 1, wherein The preprocessing of the obtained data includes: Convert the radiation cumulative values in the meteorological reanalysis data and near-real-time product meteorological data into instantaneous values; Divide the meteorological reanalysis and near-real-time product grid data into multiple sub-grids, detect non-zero outliers in each sub-grid according to the three-sigma rule, and use the interpolation method of neighboring spatio-temporal points to repair the non-zero outlier points; Use the bicubic interpolation method to adjust the spatial resolutions of the meteorological reanalysis data, near-real-time product meteorological data to be consistent with the WRF model grid resolution.

7. The method for correcting high-resolution solar radiation data based on deep learning according to claim 6, wherein The preprocessing of the obtained data further includes: Integrate the digital elevation model data in tile format and convert it into a binary format that can be directly read by the WRF model.

8. A high-resolution solar radiation data correction system based on deep learning, which applies the method according to any one of claims 1 to 7, characterized in that, Include: A data preprocessing module, which is used to obtain meteorological data and digital elevation model data and preprocess the obtained data; A configuration optimization module, which is used to set the regional nested grid and resolution in the WRF preprocessing configuration file according to the characteristics of the simulation area, dynamically screen out the optimal parameterization scheme combination, and write the scheme into the WRF running configuration file; A data processing and fusion module, which is used to use the preprocessed meteorological data and digital elevation model data as the initial field to drive the WRF model to generate high spatio-temporal resolution meteorological data; Set different fusion weights according to geographical features and meteorological features, and fuse the solar radiation flux values in the meteorological reanalysis data and near-real-time product meteorological data as the labeled data of the correction model; A training optimization module, which is used to input the high spatio-temporal resolution meteorological data and the fused labeled data into a pre-constructed deep residual encoder-decoder network model with multi-scale feature fusion for training, and obtain the optimal parameter model based on multiple loop iterations; A correction output module, which is used to correct the solar radiation flux data output by the WRF model using the optimal parameter model.

9. An electronic device, comprising a memory and a processor, characterized in that: The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having computer-executable instructions stored thereon, characterized in that: When the computer-executable instructions are executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images