A Spatial Downscaling Method for Weather Forecasting Models Based on Terrain Classification Super-Resolution Model
By using the terrain classification-based super-resolution model SRBTC, the mapping relationship between different scales is learned and multi-terrain prediction results are fused, which solves the problems of insufficient computational efficiency and accuracy in existing technologies and realizes spatial downscaling of efficient and high-precision weather forecast models.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ZHEJIANG UNIV
- Filing Date
- 2023-03-17
- Publication Date
- 2026-05-26
AI Technical Summary
Existing downscaling methods struggle to balance computational efficiency and accuracy, and traditional super-resolution models neglect the differences in resolution data from different weather forecast models during training, making the models unsuitable for use at different scales.
The super-resolution model SRBTC based on terrain classification is adopted. By training the super-resolution model and the terrain classification model under different terrains, the mapping relationship between different scales is learned. The prediction results of multiple terrains are fused using terrain similarity probability, and the data features are processed by combining the specified value scaling method and the split merging method.
It achieves detailed feature reconstruction of high-resolution data, improves computational efficiency and accuracy, and enhances the spatial generalization ability of the model, enabling it to adapt to different terrain and meteorological conditions.
Smart Images
Figure CN116579228B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of numerical simulation, specifically relating to a spatial downscaling method for weather forecast models based on a terrain classification super-resolution model. Background Technology
[0002] As modern technology advances towards higher precision and sophistication, the demands for accuracy and speed in weather forecasting are increasing. High-resolution model simulations provide more detailed information and more closely approximate real-world scenarios, making them crucial for weather forecasting. In recent years, various downscaling methods have been proposed to improve the accuracy and speed of weather forecasts, with deep learning-based downscaling showing significant potential. For example, application number 201910322146.0 uses meteorological station observation data to correct the interpolation results of the original data, thereby obtaining high-resolution data. Application number 201911361694.0 uses spatiotemporally downscaled data input into an ensemble forecasting module to obtain high-resolution ensemble forecast results; its spatiotemporal downscaling method is a traditional interpolation-based downscaling method. Application No. 202210284987.9 integrates ECMWF fine-grid data and CLDAS temperature data, extracts features through a deep learning network, and uses bilinear interpolation to obtain relevant data when the first and second preset thresholds are met. Passing the third preset threshold verifies the data, resulting in high-resolution data. Application No. 202210815192.6 obtains high-resolution data based on dynamic downscaling, then combines it with observational data to obtain a model for predicting precipitation. However, the downscaling methods used in the above patents are mostly pattern-based dynamic downscaling methods or interpolation-based downscaling methods. Dynamic downscaling methods can obtain high-resolution data with more details, but their computational efficiency is low; interpolation-based downscaling methods, while computationally efficient, cannot obtain high-frequency information in the region, i.e., their accuracy is low. Currently, deep learning-based downscaling methods can balance downscaling computational efficiency and accuracy, and have great development potential. Application No. 201910878609.1 interpolates the target image, obtains feature maps through a feature extraction network, and fuses them to obtain downscaled data. Application No.: 201910879807.X. Based on raw meteorological data, a deep learning super-resolution model is used for feature extraction, and then feature fusion is used to obtain high-resolution results. However, the low-resolution data in the training process of the above super-resolution model is obtained by downsampling high-resolution data. Therefore, the super-resolution model trained in this way is difficult to reflect the mapping relationship between multi-scale meteorological simulations. At the same time, it ignores the feature that the data samples of different resolutions obtained by dynamic downscaling of weather forecast models have large differences, and overly optimistically applies the super-resolution model at a certain magnification to different scales.
[0003] Traditional downsampling methods often result in low-resolution data that differ significantly from simulation results, making it difficult to train a model that reflects the mapping relationship between different scales of meteorological simulations using pairs of high-resolution and low-resolution data obtained through downsampling. Summary of the Invention
[0004] The purpose of this invention is to solve the problems existing in the current technology and to provide a spatial downscaling method for weather forecasting models based on a terrain classification super-resolution model.
[0005] The purpose of this invention is to implement a spatial downscaling method for weather forecasting models based on a terrain classification super-resolution model through the following technical solution. The steps of this method are as follows:
[0006] (1) Under m different terrains, the meteorological element samples under m terrains are obtained by performing n (n>1) nested dynamic downscaling calculations through weather forecast models. The area represented by the data in the 1st, 2nd, ..., nth layers gradually decreases while the resolution gradually increases.
[0007] (2) Extract meteorological element data samples of different resolutions in the nth layer area under the same terrain in step (1), select two sets of data of different resolutions, and divide them into training set, validation set and test set according to the time dimension;
[0008] (3) Use the different meteorological element data in the training set and validation set in step (2) to train the super-resolution model respectively. The high-resolution data sample is used as the label of the super-resolution model, and the low-resolution data is interpolated to be the data with the same resolution as the label as the input of the super-resolution model. The model is then tested on the test set to finally obtain the super-resolution models under m different terrains.
[0009] (4) Obtain terrain data under m types of terrain, pass through m classification model to obtain terrain classification model, and when using this terrain classification model, select the similarity probability between the predicted terrain and m types of terrain through the softmax layer as the weight of the multi-terrain prediction result.
[0010] (5) The meteorological data to be predicted is predicted separately using the super-resolution models under the m different terrains obtained in step (3). The terrain data of the area to be predicted is predicted using the terrain classification model obtained in step (4). The similarity probability between the area to be predicted and the m different terrains is output. The similarity probability between the m terrains is used to weight and sum the m prediction results of the super-resolution models to obtain the fusion prediction result of the m super-resolution models.
[0011] Furthermore, in step (1), in the n-layer nested dynamic downscaling calculation, the simulation results of the first layer provide boundary conditions for the second layer calculation, the simulation results of the second layer provide boundary conditions for the third layer simulation results, and so on, until the simulation results of the nth layer are finally obtained, and the calculation ends here.
[0012] Furthermore, in step (1), the n-layer nested dynamic downscaling calculation results have the same time interval and time step.
[0013] Furthermore, in step (1), the terrain regions in the m different terrains refer to the target regions of the climate model dynamic downscaling, i.e., the nth nested region.
[0014] Furthermore, in step (3), the data features of the low-resolution data samples are processed by the specified value scaling method. Specifically, the low-resolution data samples are normalized in a conventional manner, and the high-resolution data samples are normalized in the same way as the low-resolution data samples using the feature data of the low-resolution data samples, so as to ensure the relative size relationship between the low-resolution data samples and the high-resolution data samples. Because only the data features of the low-resolution data samples can be provided during the model application stage, the prediction results of the model trained by the specified value scaling method have smaller errors on the validation set after normalization using the data features of the low-resolution data samples.
[0015] Furthermore, in step (4), when using the terrain classification model, the final classification category is not output. Instead, the similarity probability between the terrain to be predicted after passing through the softmax layer and the terrain corresponding to the m super-resolution models is output, and the sum of the m probabilities is 1.
[0016] Furthermore, in step (5), considering the different physical drivers of multi-scale simulation, the model is applied according to the correspondence between resolutions. If both resolutions before and after downscaling are within ±10% of the existing model's corresponding resolution, it is considered that the existing super-resolution model and terrain classification model can still be used. If the two resolutions before and after downscaling are not within ±10% of the existing model's corresponding resolution, then the super-resolution model and terrain classification model at that resolution need to be trained separately to obtain data that is closer to the real situation.
[0017] Furthermore, in step (5), if the size of the data array to be predicted is different from the input requirements of the super-resolution model and the terrain classification model and cannot be predicted directly, the area to be predicted is split into multiple arrays of the same size as the model input using the split-merge method, predicted separately, and then the arrays are spliced together to obtain the downscaling result of the entire area.
[0018] The beneficial effects of this invention are as follows: The downscaling method described in this invention learns the mapping relationship between different scales of dynamic downscaling of climate models through a deep learning super-resolution model, and applies it to low-resolution climate model samples, enabling the reconstruction of detailed features that are only present in high-resolution data. On the one hand, this invention has the advantage of high accuracy in dynamic downscaling with fine grid simulation results; on the other hand, it greatly improves computational efficiency. Furthermore, because the SRBTC model proposed in this invention considers the adaptability of the super-resolution model to different terrains, it has good spatial generalization ability. Attached Figure Description
[0019] Figure 1 This is a schematic diagram illustrating the training of a multi-terrain super-resolution model.
[0020] Figure 2 A schematic diagram of a terrain classification model;
[0021] Figure 3 This is a schematic diagram of the SRBTC model structure;
[0022] Figure 4 Flowchart for scaling to a specified value;
[0023] Figure 5 This is a schematic diagram illustrating the relationship between the nested layers of WRF dynamic downscaling in the embodiment;
[0024] Figure 6 This is a flowchart illustrating the training process of the super-resolution model under a single terrain in the embodiment.
[0025] Figure 7 This is a schematic diagram of the PRUSR super-resolution model structure in the embodiment;
[0026] Figure 8 This is a comparison chart of the performance of the PRUSR super-resolution model in the embodiments. Detailed Implementation
[0027] To facilitate understanding and implementation of the present invention by those skilled in the art, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.
[0028] This invention proposes a super-resolution model based on deep learning to learn the mapping relationship between different scales, which is a downscaling model at the corresponding scale. It also proposes a terrain classification-based super-resolution model—the SRBTC model. This model considers the significant differences in meteorological conditions across different regions, providing terrain similarity probabilities through a terrain classification model, and then fusing the prediction results of super-resolution models under multiple terrains, thereby improving the model's spatial generalization ability. The proposed method of extracting data from the same geographic space at different resolutions for model training in dynamic downscaling has greater application potential. Furthermore, it should be noted that weather forecast models have significant differences in calculation results at different scales. The purpose of using a super-resolution model is to learn the mapping relationship between different layers. Even with the same downscaling ratio, due to the significant differences in features at different scales, it is not suitable to simply transfer and apply the model based on the ratio. If there is a need for downscaling between specific scales, it is necessary to first determine if there are models with similar scales among the already trained models. If so, they can be transferred and applied; otherwise, dynamic downscaling calculations based on climate models are required to obtain data at the corresponding scale, and then the corresponding super-resolution model is trained.
[0029] This invention first obtains the original data through dynamic downscaling nested calculations of weather forecast models, then extracts data of different resolutions from the same geographic space, selects two sets of data with different spatial resolutions, and divides them into training set, validation set, and test set. The training set and validation set are used for model training. When the model passes the validation on the test set, this model is used as the super-resolution model for that terrain. The above operation is repeated on different terrains to obtain super-resolution models suitable for each terrain. A terrain classification model is then trained. The terrain data of the area to be predicted is input into the terrain classification model to obtain the similarity probability with the terrain corresponding to each super-resolution model. The prediction results of the super-resolution models under each terrain are fused by the similarity probability to obtain the final downscaling result.
[0030] This invention mainly constructs a spatial downscaling method for weather forecasting models based on a terrain classification super-resolution model, achieving spatial downscaling that balances computational efficiency and accuracy. The method steps are as follows:
[0031] (1) Under m different terrains, the weather forecast model is used to perform nested dynamic downscaling calculations of n (n>1) layers (the target area of the climate model dynamic downscaling) to obtain meteorological element samples under m terrains. The area represented by the data of the 1st, 2nd, ..., nth layers gradually decreases while the resolution gradually increases. In the nested dynamic downscaling calculations of the nth layer, there are the same time intervals and time steps. The simulation results of the 1st layer provide boundary conditions for the calculation of the 2nd layer, the simulation results of the 2nd layer provide boundary conditions for the simulation results of the 3rd layer, ..., and so on, until the simulation results of the nth layer are finally obtained, and the calculation ends.
[0032] (2) Extract meteorological element data samples of different resolutions from the nth layer region under the same terrain in step (1), such as Figure 1 As shown, two sets of data with different resolutions are selected and divided into training set, validation set and test set according to the time dimension. This step of extracting meteorological element data samples with different resolutions in the same geographic space is one of the features of this invention. Unlike the low-resolution data obtained by downsampling in the traditional way, which is a pseudo-low-resolution data, this invention extracts data with different resolutions of meteorological models as low-resolution data and high-resolution data. The super-resolution model established in this way can better reflect the physical laws between different scales.
[0033] (3) Use the different meteorological element data in the training set and validation set in step (2) to train the super-resolution model respectively. The high-resolution data sample is used as the label of the super-resolution model, and the low-resolution data is interpolated to be the data with the same resolution as the label as the input of the super-resolution model. The model is then tested on the test set to finally obtain the super-resolution models under m different terrains.
[0034] Because the features of data samples obtained from different scales of meteorological models at different resolutions vary significantly—for example, the maximum and minimum values of samples commonly used for data normalization—are often obtained by training models on low- and high-resolution data samples using conventional normalization methods, followed by denormalization to obtain prediction results. However, the prediction process only provides the data features of low-resolution data samples for denormalization, leading to a significant discrepancy between the prediction results and the actual high-resolution data. Therefore, this invention proposes a specified value scaling method to approximate this problem. The application of this method is detailed in the appendix. Figure 4 Specifically, the method involves: performing regular normalization on low-resolution data samples; then applying the same normalization operation to high-resolution data samples using the feature data from the low-resolution samples to maintain the relative size relationship between them. Since only the data features of the low-resolution samples are available during model application, the model trained using the specified value scaling method shows smaller errors on the validation set when normalized using these low-resolution features, as evidenced by metrics such as MSE, R, MAPE, and MAE. Alternatively, interpolating the low-resolution data before inputting it into the super-resolution model for training leverages the idea of transfer learning. This approach combines the basic features of interpolation with the ability to capture high-frequency information from high-resolution simulations through the super-resolution model, ensuring that this method outperforms interpolation results.
[0035] (4) Due to significant regional differences in meteorological conditions, it is difficult to guarantee the applicability of a super-resolution model trained on a single terrain type to all terrain types. Therefore, it is necessary to train super-resolution models for different terrain types to ensure the accuracy of prediction results. For example... Figure 2As shown, this invention obtains terrain data under m types of terrain, passes through an m-classification model, and obtains a terrain classification model. When using this terrain classification model, the final classification category is not output. Instead, the similarity probability between the terrain to be predicted and the m types of terrain is output through the softmax layer as the weight of the multi-terrain prediction results. The sum of the m probabilities is 1. The higher the similarity between the area to be predicted and a certain type of terrain, the higher the weight is given to the super-resolution model prediction result under the corresponding terrain.
[0036] (5) Figure 3 As shown, by combining the super-resolution model and the terrain classification model—SRBTC model, the meteorological data to be predicted is predicted separately using the super-resolution models under the m different terrains obtained in step (3). The terrain data of the area to be predicted is used to output the similarity probability between the area to be predicted and the m different terrains using the terrain classification model obtained in step (4). The similarity probability between the m terrains is used to weight and sum the m prediction results of the super-resolution model to obtain the fusion prediction result of the m super-resolution models.
[0037] Considering the different physical drivers of multi-scale simulations, the model is applied based on the correspondence between resolutions. If both the resolutions before and after downscaling are within ±10% of the corresponding resolution of the existing model, then the existing super-resolution model and terrain classification model can still be used. If the two resolutions before and after downscaling are not within ±10% of the corresponding resolution of the existing model, then the super-resolution model and terrain classification model at that resolution need to be trained separately to obtain data that is closer to the real situation.
[0038] Even if the resolution of the data to be predicted meets the requirements of the SRBTC model, it cannot be predicted directly because the size of the data array to be predicted is different from the input required by the model. The usual method is to change the size of the data array to be predicted. However, this method will change the resolution of the original array. Therefore, this invention adopts the split and merge method to split the region to be predicted into multiple arrays of the same size as the model input, predict them separately, and then concatenate the arrays to obtain the downscaling result of the entire region.
[0039] One embodiment of this invention provides a spatial downscaling method for the WRF (Weather Research and Forecasting) mesoscale weather forecasting model based on a terrain classification super-resolution model. The specific steps are as follows:
[0040] 1) The mesoscale weather forecasting model WRF performs dynamic downscaling calculations through a three-layer nesting approach, obtaining meteorological data at grid points corresponding to different regions in three layers: D01, D02, and D03. D03 represents the target topographic region. The geospatial relationships represented by the three-layer nesting are shown in the appendix. Figure 5 Meteorological data contains information in dimensions such as time, longitude, latitude, and altitude;
[0041] 2) Extract data from three nested grids in the same geographic space. In this embodiment, the data of D01, D02 and D03 corresponding to D03 are extracted. They represent data of different resolutions in the same geographic space. In this embodiment, the time resolution is the same, but the spatial resolution is different. The data at a specified time and height are similar to image data with multiple channels. Therefore, the image super-resolution technology in deep learning can be transferred to learn the mapping relationship between data of different resolutions.
[0042] 3) In this embodiment, model data at near-ground altitudes, including 2m temperature (T2), 2m humidity (rh2), and 10m wind speed (wspd10), were selected as meteorological variables. High-resolution data in resolution D03 and the corresponding data in D02 in D03 were used as low-resolution data (see attached). Figure 5 Experiments were conducted by dividing high-resolution and low-resolution data for each variable into training, validation, and test sets in a ratio of 8:1:1. The training and validation sets were fed into the PRUSR super-resolution model (the PRUSR model is a self-developed super-resolution model, and its network structure is shown in the appendix). Figure 7 The model is trained on the test set, and the test set is used to validate the super-resolution model. Successful validation indicates successful training; failure requires modification and retraining until it passes validation on the test set. See the appendix for the basic process. Figure 6 The trained PRUSR super-resolution model possesses excellent reconstruction capabilities, significantly outperforming interpolation-based spatial downscaling methods. The spatial downscaling method of this invention can obtain more high-frequency information missing from coarse-grid simulation results. See the attached figure for a comparison of the PRUSR super-resolution model's performance. Figure 8 .
[0043] 4) Regarding the PRUSR super-resolution model shown in 3), it is a deep learning model based on CNN, which incorporates an improved U-net (see...). Figure 7 The network shown in b) and the pyramid network (see Figure 7 As shown in (a), more low-level information and different features can be obtained through multi-channel feature extraction. At the same time, pre-activated residual blocks with the BN layer removed are added to the network (see...). Figure 7As shown in c), the addition of residual blocks allows deeper layers of the model to receive better parameter updates. The activation function used is LeakyReLU (the negative half-axis slope is set to 0.2 in this embodiment) because many physical quantities are different from image pixel values and have a large number of negative values. Regarding the model training in 3), this example uses the deep learning framework Tensorflow. First, the low-resolution data samples in the training data need to be normalized. For the high-resolution data samples, a specified value scaling method is used, and the data format is converted into the form (batch, h, w, c), which is the required format for model input. Here, batch corresponds to time in the pattern data, h and w correspond to latitude and longitude respectively, and c represents different variables. Then, the optimizer of the model is set to Adam, and the loss function is mean squared error (MSE). Before feeding the low-resolution data into the PRUSR model, the low-resolution data needs to be converted into high-resolution data through bilinear interpolation. The training epochs are 200, and the batch size is 16.
[0044] 5) The generality of deep learning models is a major challenge. In order to improve the generalization ability of super-resolution models, this invention proposes the SRBTC model. In this embodiment, it is necessary to obtain the meteorological simulation results of each typical terrain through WRF dynamic downscaling. Based on these data, the above process of training super-resolution models under single terrain is repeated to obtain super-resolution models under each terrain.
[0045] 6) Obtain the typical terrain data in 5), train the terrain classification model using the Googlenet classification model, save the parameters of the trained Googlenet classification model, and when using the terrain classification model, directly output the similarity probability between the area to be predicted and the typical terrain through the softmax layer. The sum of the terrain similarity probabilities is 1.
[0046] 7) Input the meteorological data to be predicted (coarse-grid simulation results) into the super-resolution models trained in 5) for each terrain type to obtain the corresponding prediction results. Then, use the terrain similarity probabilities obtained in 6) to perform a weighted summation of these prediction results to obtain the final prediction result, which is the downscaled result of the coarse-grid simulation results. The above-mentioned fusion model combining the super-resolution model and the terrain classification model for each typical terrain is the SRBTC model described in this invention. See Appendix. Figure 3 ;
[0047] 8) For the SRBTC models (super-resolution model and terrain classification model) already trained in 6) and 7), in practice, the size of the coarse grid data array to be scaled down and its represented resolution often do not match the array size and resolution required by the model. If the resolution before and after the target scaling down is not significantly different from the existing model, scaling down can be achieved by adding interpolation. If the resolution before and after the target scaling down differs significantly from the resolution required by the model, a model for the corresponding situation should be trained instead of directly using an unsuitable existing model. For example, if the target is to scale down the coarse grid simulation results with a resolution of 10km to fine grid data of 2.5km, and the model in this embodiment is a 9km to 3km model, the 10km data can be interpolated to 9km data first, and then the 9km data can be scaled down. The data was initially set to 3km, and then interpolated to a 2.5km finer grid, thus achieving a downscaling from 10km to 2.5km. If the goal is to downscale from 27km to 9km, although the downscaling ratio is the same as the existing 9km to 3km model, the corresponding physical drivers between the multiple scales are quite different, so the existing model cannot be used for downscaling calculations from 27km to 9km. Even if the resolution of the coarse grid to be downscaled is consistent with the input required by the SRBTC model, if the size of the data array to be downscaled is different from the size required by the SRBTC model, downscaling cannot be performed directly. To address this problem, this invention divides the array into multiple groups of the same size as the input array required by the SRBTC model, performs predictions on each group separately, and initially splices them together to form the downscaling result for the entire region.
[0048] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained through equivalent substitution or transformation fall within the protection scope of the present invention.
Claims
1. A spatial downscaling method for weather forecasting models based on a terrain classification super-resolution model, characterized in that, The steps of this method are as follows: (1) In m Weather forecasting models are used to forecast different terrains. n Layer-nested dynamic downscaling calculation yields m Meteorological element samples under various terrain types, where the 1st, 2nd, ..., n The area represented by the layer data gradually decreases, while the resolution gradually increases; (2) Extract the first number of the same terrain from step (1) respectively. n Data samples of different meteorological elements at different resolutions in the layer region were selected. Two sets of data at different resolutions were chosen and divided into training set, validation set and test set according to the time dimension. (3) Using different meteorological element data from the training and validation sets in step (2), the super-resolution model is trained. High-resolution data samples are used as labels for the super-resolution model, and low-resolution data are interpolated to have the same resolution as the labels as inputs to the super-resolution model. The model is then tested on the test set to obtain the final result. m Super-resolution models for different terrains; (4) Obtain m Topographic data under various terrain types, after m The classification model yields a terrain classification model. When applying this terrain classification model, the output is selected based on the terrain to be predicted passing through the softmax layer. m The similarity probability between different terrain types is used as the weight for the fusion of multi-terrain prediction results; (5) Use the meteorological data to be predicted obtained in step (3) m Different super-resolution models are used to predict different terrain types. The terrain data of the area to be predicted is then used with the terrain classification model obtained in step (4) to output the area to be predicted and the terrain classification model obtained in step (4). m The similarity probability between different terrains is used m The similarity probability between different terrain types is important for the super-resolution model. m The prediction results are weighted and summed to obtain m The fusion prediction results of several super-resolution models.
2. The spatial downscaling method for weather forecast models based on terrain classification super-resolution models according to claim 1, characterized in that, In step (1), n In the nested dynamic downscaling calculation, the simulation results of the first layer provide boundary conditions for the second layer, the simulation results of the second layer provide boundary conditions for the third layer, and so on, until the final result is obtained. n The simulation results are now complete, and the calculation is finished.
3. The spatial downscaling method for weather forecast models based on a terrain classification super-resolution model according to claim 1, characterized in that, In step (1), n The results of nested dynamic downscaling calculations have the same time interval and time step.
4. The spatial downscaling method for weather forecast models based on terrain classification super-resolution models according to claim 1, characterized in that, In step (1), m The topographic regions in different terrain types refer to the target regions of climate model dynamic downscaling, i.e., the nth nested region.
5. The spatial downscaling method for weather forecast models based on a terrain classification super-resolution model according to claim 1, characterized in that, In step (3), the data features of low-resolution data samples are processed by the specified value scaling method. Specifically, the low-resolution data samples are normalized in a regular manner, and the high-resolution data samples are normalized in the same way as the low-resolution data samples using the feature data of the low-resolution data samples to ensure the relative size relationship between the low-resolution data samples and the high-resolution data samples. Because only the data features of low-resolution data samples can be provided during the model application stage, the prediction results of the model trained by the specified value scaling method have smaller errors on the validation set after normalization using the data features of low-resolution data samples.
6. The spatial downscaling method for weather forecast models based on terrain classification super-resolution models according to claim 1, characterized in that, In step (4), when using the terrain classification model, the final classification category is not output; instead, the predicted terrain after passing through the softmax layer is output. m The similarity probability of the terrain corresponding to each super-resolution model. m The sum of the probabilities is 1.
7. The spatial downscaling method for weather forecast models based on a terrain classification super-resolution model according to claim 1, characterized in that, In step (5), considering the different physical drivers of multi-scale simulation, the model is applied according to the correspondence between resolutions. If both resolutions before and after downscaling are within ±10% of the existing model's corresponding resolution, it is considered that the existing super-resolution model and terrain classification model can still be used. If the two resolutions before and after downscaling are not within ±10% of the existing model's corresponding resolution, then the super-resolution model and terrain classification model at that resolution need to be trained separately to obtain data that is closer to the real situation.
8. The spatial downscaling method for weather forecast models based on terrain classification super-resolution models according to claim 1, characterized in that, In step (5), if the size of the data array to be predicted is different from the input requirements of the super-resolution model and the terrain classification model and cannot be predicted directly, the splitting and merging method is used to split the area to be predicted into multiple arrays of the same size as the model input, predict them separately, and then splice the arrays to obtain the downscaling result of the entire area.