Artificial intelligence downscaling method and system for meteorological elements

Abstract

The embodiment of the invention provides an artificial intelligence downscaling method and system for meteorological elements. The method is applied to the technical field of meteorological data processing, and comprises the following steps: acquiring low-resolution data, and preprocessing and enhancing the low-resolution data; an MSRLapN model is used to realize depth super-resolution downscaling; after deep super-resolution downscaling is realized, terrain data and climate state data are introduced to be input into an intelligent correction model, and the construction of the intelligent correction model comprises the steps of constructing data, constructing a U-Net neural network according to elements, performing parameter adjustment iterative training and executing the U-Net neural network to realize a multi-mode integration function. In this way, the technical problems that meteorological data after downscaling cannot accurately reflect actual meteorological conditions, and underestimation of the maximum value often occurs for special weather phenomena can be solved.

Description

Technical Field

[0001] The present disclosure relates to the field of meteorological data processing, and in particular to an artificial intelligence downscaling method and system for meteorological elements. Background Art

[0002] With the frequent occurrence of global climate change and extreme weather events, accurate weather forecasts are of great significance to disaster prevention and mitigation, agricultural management, environmental protection and other fields. At present, global atmospheric numerical models can already accurately predict atmospheric environmental elements. However, the spatial resolution of these model outputs is relatively low, and it is often difficult to capture convective-scale weather changes and regionally differentiated climate characteristics. Therefore, improving the horizontal resolution of model data has become a key issue that needs to be urgently addressed in the field of meteorological forecasting.

[0003] At present, there are three main methods to improve the resolution of model forecast data:

[0004] Developing a high-resolution global atmospheric forecast model: This approach can theoretically provide the highest resolution, but it is difficult to implement, requires extremely high computing resources, and has a complex model construction and verification process.

[0005] Use regional atmospheric forecast models: This type of model can be embedded in the global model or run independently using the output of the global model as a boundary condition. It has the advantages of physical interpretability and is not affected by observational data. However, regional atmospheric forecast models are computationally intensive and difficult to simulate and configure, especially when dealing with large-scale complex terrain and climate characteristics.

[0006] Develop new statistical downscaling calculations: Statistical downscaling methods learn the mapping relationship from Earth system models to high resolution by fusing observational data. In recent years, with the rapid development of deep neural network technology, super-resolution models based on convolutional neural networks (CNNs) have achieved remarkable results in image downscaling and have gradually been applied to downscaling of meteorological elements. However, directly applying image downscaling algorithms to meteorological data faces many challenges, such as insufficient terrain constraints and weak data temporal characteristics. In addition, for special weather phenomena such as precipitation and cloud cover, existing downscaling models based on convolutional neural networks often underestimate the maximum values. This is because these weather phenomena are sparsely distributed in space, which makes it easy for model predictions to be biased.

[0007] In the process of meteorological data downscaling, the images corresponding to three-dimensional meteorological data are multi-channel. Compared with the RGB three channels of conventional color images, three-dimensional meteorological data contains more complex meteorological elements (such as temperature, humidity, wind speed, etc.). The changes of these elements in three-dimensional space are continuous and there are complex correlations between them. Therefore, simply extending a two-dimensional image downscaling model to multi-channel downscaling applications is difficult to fully utilize the knowledge in the meteorological field, resulting in the downscaled meteorological data being unable to accurately reflect the actual meteorological conditions. Summary of the Invention

[0008] The present disclosure provides an artificial intelligence downscaling method and system for meteorological elements, which solves the technical problems existing in the prior art in meteorological data downscaling, such as insufficient spatial resolution, insufficient topographic constraints, and underestimation of the maxima of special weather phenomena.

[0009] According to the first aspect of the present disclosure, there is provided an artificial intelligence downscaling method for meteorological elements, including:

[0010] Obtain low-resolution data, perform preprocessing and enhancement processing on the low-resolution data. The preprocessing includes denoising, normalization, and missing value processing; the enhancement processing includes using bilinear interpolation to improve the resolution of the low-resolution data.

[0011] Use the MSRLapN model to achieve deep super-resolution downscaling, including constructing data, constructing the MSRLapN neural network by element, parameter adjustment and iterative training, and executing the MSRLapN neural network to achieve the downscaling function.

[0012] After achieving deep super-resolution downscaling, introduce topographic data and climatological data into the intelligent correction model, and the intelligent correction model is implemented by the U-Net neural network.

[0013] According to the second aspect of the present disclosure, there is provided an artificial intelligence downscaling system for meteorological elements, including:

[0014] A data acquisition module for obtaining low-resolution data, performing preprocessing and enhancement processing on the low-resolution data. The preprocessing includes denoising, normalization, and missing value processing; the enhancement processing includes using bilinear interpolation to improve the resolution of the low-resolution data.

[0015] A scale adjustment module for using the MSRLapN model to achieve deep super-resolution downscaling, including constructing data, constructing the MSRLapN neural network by element, parameter adjustment and iterative training, and executing the MSRLapN neural network to achieve the downscaling function.

[0016] The model correction module is used to implement the intelligent correction model by introducing topographic data and climatological data after deep super-resolution downscaling. The intelligent correction model is implemented by the U-Net neural network. The construction of the intelligent correction model includes constructing data, constructing the U-Net neural network for each element, adjusting parameters for iterative training, and executing the U-Net neural network to achieve the multi-mode integration function.

[0017] Compared with the prior art, the advantages and beneficial effects obtained by the present disclosure are as follows:

[0018] By introducing the MSRLapN model, the present disclosure realizes the deep super-resolution downscaling of meteorological data. This model can automatically extract deep features from low-resolution data and significantly improve the resolution and accuracy of the downscaled data through multi-layer convolution and residual learning. By introducing the adjustment term δ in the denormalization process, the problem of underestimation commonly existing in convolutional neural network models for these small-probability events is effectively solved, further improving the downscaling accuracy. During the downscaling process, the present invention innovatively introduces topographic data, making the downscaled data more conform to the actual topographic features and improving the accuracy and reliability of the data in complex terrain areas. By fusing climatological data, the present invention enables the downscaled data to not only have high resolution but also better reflect the climate characteristics on a long time scale, providing strong support for climate prediction and disaster prevention and mitigation. Compared with traditional regional atmospheric prediction models, the SRRN model proposed by the present invention significantly reduces the computational complexity and has higher computational efficiency. This enables the model to process a large amount of meteorological data faster and meet the requirements of real-time forecasting and decision support. In summary, the present disclosure not only improves the downscaling accuracy and computational efficiency but also enhances the fusion ability of topographic constraints and climatological data, providing more accurate and reliable technical support for meteorological forecasting and disaster prevention and mitigation.

[0019] It should be understood that the content described in the summary of the invention section is not intended to limit the key or important features of the embodiments of the present disclosure, nor to limit the scope of the present disclosure. Other features of the present disclosure will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In combination with the accompanying drawings and with reference to the following detailed description, the above and other features, advantages, and aspects of the embodiments of the present disclosure will become more apparent. The drawings are used to better understand the solution and do not constitute a limitation to the present disclosure. In the drawings, the same or similar reference numerals represent the same or similar elements, where:

[0021] Figure 1 FIG. shows a flowchart of an artificial intelligence downscaling method for meteorological elements according to an embodiment of the present disclosure;

[0022] Figure 2Shows a schematic diagram of the SRRN model network structure according to an embodiment of the present disclosure;

[0023] Figure 3 Shows a schematic diagram of the network structure of the MSRLapN model with the spatial resolution increased by 2 times according to an embodiment of the present disclosure;

[0024] Figure 4 Shows a schematic diagram of the intelligent correction network structure based on U-Net according to an embodiment of the present disclosure;

[0025] Figure 5 Shows a block diagram of an artificial intelligence downscaling system for meteorological elements according to an embodiment of the present disclosure. Detailed implementation manners

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present disclosure clearer, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present disclosure. Apparently, the described embodiments are some but not all of the embodiments of the present disclosure. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present disclosure without creative efforts shall fall within the scope of protection of the present disclosure.

[0027] In addition, the term "and / or" in this article is merely a description of the association relationship between associated objects, indicating that three relationships may exist. For example, A and / or B may represent: A exists alone, A and B exist simultaneously, and B exists alone. In addition, the character " / " in this article generally represents an "or" relationship between the associated objects before and after.

[0028] The present disclosure provides an artificial intelligence downscaling method and system for meteorological elements. Based on a new neural network model, a super-resolution refinement network (SRRN) is used to achieve super-resolution downscaling for meteorological data. This model can refine high-resolution data while performing downscaling, enabling the output data to have better spatial correlation and temporal characteristics: a new meteorological data super-resolution downscaling model: This neural network is called the super-resolution refinement network (SRRN), which can further refine the output data while achieving super-resolution downscaling; a correction method based on feature data: Terrain data and climatological data are used as auxiliary elements to correct high-resolution meteorological data, making the downscaled super-resolution meteorological data conform to terrain constraints and temporal variation characteristics; a systematic deviation adjustment method: For the downscaling of special weather phenomena such as precipitation and cloud cover, extreme values in space usually have systematic deviations. This patent proposes a systematic deviation adjustment method. It solves the technical problems in the prior art that the downscaled meteorological data cannot accurately reflect the actual meteorological conditions and often underestimates extreme values for special weather phenomena.

[0029] The super-resolution refinement network (SRRN) in this embodiment mainly includes a super-resolution downscaling module and a correction part considering terrain constraints. Among them, the deep super-resolution part of meteorological data is implemented by the MSRLapN model, and the correction part considering terrain constraints is implemented by the U-Net neural network. The SRRN model structure is as Figure 2 shown.

[0030] Figure 1 FIG. shows a schematic flowchart of an artificial intelligence downscaling method 100 for meteorological elements in an embodiment of the present disclosure, as Figure 1 shown, and the method 100 includes:

[0031] S110: Obtain low-resolution data, perform preprocessing and enhancement processing on the low-resolution data. The preprocessing includes denoising, standardization, and missing value processing; the enhancement processing includes using bilinear interpolation to improve the resolution of the low-resolution data.

[0032] Optionally, in some embodiments, first, obtain low-resolution data from meteorological observations or numerical models; denoising: Adopt an adaptive filtering technique, combine the spatio-temporal characteristics of meteorological elements, and dynamically adjust the filtering parameters;

[0033] Missing value processing: Introduce an interpolation algorithm based on the correlation of meteorological elements, and use the data relationship of adjacent spatio-temporal points to dynamically fill in the missing values;

[0034] Standardization: The piecewise standardization technique is adopted to map the data of different meteorological elements to a unified scale; by analyzing the distribution characteristics of meteorological elements, the standardization parameters are dynamically adjusted.

[0035] Secondly, for data augmentation processing, an improved bilinear interpolation algorithm is used, and combined with the spatial distribution characteristics of meteorological elements, the interpolation weights are dynamically adjusted;

[0036] Finally, the data after preprocessing and augmentation processing is output.

[0037] It should be noted that in the embodiment, denoising refers to adaptive denoising based on a graph neural network, modeling meteorological data as a graph structure, and using the features of nodes and edges for dynamic denoising;

[0038]

[0039] In the formula, H (l) represents the node feature matrix of the l-th layer; represents the adjacency matrix with self-loops added; A represents the adjacency matrix; I represents the identity matrix; represents the degree matrix; W (l) represents the weight matrix of the l-th layer; σ represents the activation function (such as ReLU); represents the matrix the element at the i'-th row and i'-th column in; represents the element at the intersection of the i'-th row and the j-th column in the matrix.

[0040] For missing value processing, a variational autoencoder is used to learn the latent distribution of meteorological data through a generative model and dynamically generate missing values:

[0041]

[0042] In the formula, ELBO represents the evidence lower bound (optimization objective); q(z|x) represents the latent variable distribution output by the encoder; p(x|z) represents the data distribution generated by the decoder; KL represents the KL divergence; μ(x), σ 2 (x) represents the mean and variance output by the encoder; p(z) represents the prior distribution (standard normal distribution).

[0043] For the standardization process, manifold learning is adopted to map meteorological data to a low-dimensional manifold space to capture its non-linear structure:

[0044]

[0045]

[0046] In the formula, p j|i′ represents the point x in the high-dimensional spacei′ and x j similarity; q j|i′ represents the point y in the low-dimensional space i′ and y j similarity; represents the bandwidth parameter of the Gaussian kernel; Cost represents the optimization objective (KL divergence).

[0047] The data augmentation process uses a quantum computing-inspired interpolation algorithm, leveraging the superposition and entanglement of quantum states to optimize the interpolation weights:

[0048]

[0049] In the formula, |ψ> represents the quantum state; α i′ represents the complex weight; |x i′ > represents the ground state; θ i′ represents the phase angle; f(x, y) represents the interpolation estimate value.

[0050] Through the above hierarchical preprocessing and enhancement processing, the quality of low-resolution data can be significantly improved, providing high-quality data input for subsequent deep super-resolution downscaling. This process is technically different from existing methods and has significant innovation and practicality; the data output after preprocessing and enhancement processing is provided for subsequent use by the downscaling model; the output data includes not only the original meteorological elements but also their potential features and manifold structure information; the introduction of graph neural networks, variational autoencoders, manifold learning, and quantum computing-inspired methods significantly enhances the frontier nature of the algorithm; by capturing the potential structure and non-linear relationships of the data, the accuracy and robustness of the downscaling model are improved.

[0051] S120: Use the MSRLapN model to achieve deep super-resolution downscaling, including constructing data, constructing the MSRLapN neural network for each element, parameter adjustment and iterative training, and executing the MSRLapN neural network to achieve the downscaling function.

[0052] The full name of the MSRLapN model is the Multi-Scale Residual Laplacian Pyramid Network (MSRLapN) three-dimensional meteorological element deep learning downscaling model. This model constructs a multi-scale residual module (MSRB) to automatically extract prediction features from various meteorological elements in three-dimensional space; introduces multi-scale pyramid technology from the field of machine learning to describe the multi-scale interaction of meteorological elements; then, through the super-resolution reconstruction cyclic iteration method, the error of the downscaling prediction is corrected based on large-sample historical data.

[0053] Optionally, in some embodiments, a data set is constructed by elements first. The elements include: elements such as low cloud cover, total cloud cover, 3-hour accumulated precipitation, etc. Data sets are constructed separately according to different elements. Using the bilinear interpolation method, the three-dimensional atmospheric real-time fusion grid data with high resolution (horizontal resolution of 0.045°' approximately 5 km) is interpolated into low-resolution grid data with spatial resolutions of 0.09°, 0.18°, and 0.36° respectively. Then, the data is normalized. The formula is as follows, where X normalized is the standardized sample data, X is the original sample data, X min is the minimum value in the data sample, X max is the maximum value in the data sample;

[0054]

[0055] The constructed data set is stored in the binary file format dedicated to NumPy, with the file suffix of npy, which is convenient for subsequent model reading.

[0056] Secondly, an MSRLapN neural network is constructed by elements. The downscaling process is divided into two branches: an image reconstruction branch and a feature extraction branch. Figure 3 The LapSRN network structure is given. MSRLapN contains several groups of repeated network structure layers, and a specific scale factor is set to enlarge them for different levels. The model takes high-spatial-resolution terrain data and low-spatial-resolution meteorological data as inputs, and for each input (low-resolution climate data and high-resolution terrain data), its hidden layer features are first extracted through a single-layer convolution. Since the sizes of the terrain data and the low-resolution climate data are different, these transformed data need to be concatenated into a multi-layer three-dimensional image to achieve size unification. In MSRLapN, all transposed convolution layers are replaced by subpixel convolution layers.

[0057] On each level of the feature extraction branch, there is an MSRB and a subpixel convolution layer respectively. Among them, MSRB performs multi-scale feature extraction and can effectively avoid model degradation and accelerate convergence. The upsampling scale of the subpixel convolution layer is set to 2. Specifically, MSRB first includes a convolution layer with a convolution kernel size of 1×1, which can greatly reduce the number of parameters in the feature layer and ensure that the number of input and output feature layers of MSRB remains unchanged. As Figure 3As shown, the output of the subpixel convolutional layer will serve as the input for two distinct network layers. One is to reconstruct the residual information at the same level, and the other is for feature extraction in the next layer. To effectively reduce the computational complexity, the parameters of different layers are shared. For the data reconstruction branch, it includes a subpixel convolutional layer and two ordinary convolutional layers. The input data is upsampled by the subpixel convolutional layer. It should be noted that through experiments, this paper found that a convolutional layer must be passed before the subpixel layer; otherwise, the training process is difficult to converge. The other convolutional layer takes the extracted features as input and outputs the high-resolution data reconstructed by the subpixel layer. The generated high-resolution data can be used as an intermediate result for reconstruction at the next level.

[0058] Using the summary function of the torchkeras module, the structural information of the MSRLapN neural network of the present invention is viewed as follows. The following parameters are used to increase the resolution of the input data by 8 times. Among them, Conv2d-1 to ConvTranspose2d-18 are a group to double the resolution, Conv2d-19 to ConvTranspose2d-36 are a group to double the resolution, and Conv2d-37 to ConvTranspose2d-54 are a group to double the resolution.

[0059] Taking Conv2d-1 to ConvTranspose2d-18 as an example, this module includes an MSRN module and a reconstruction module. The MSRN module is Conv2d-1 to Conv2d-15, which includes three parts: head, body, and tail. The head is the single-layer convolution of Conv2d-1, the body is Conv2d-2 to Conv2d-11 layers, which includes an MSRB module to extract the spatial feature information in the data, and the tail is Conv2d-12 to Conv2d-15, where PixelShuffle realizes the function of doubling the resolution; the reconstruction module is Conv2d-16 to ConvTranspose2d-18 layers, where ConvTranspose2d realizes the function of doubling the resolution.

[0060] The parameters of Table 1 are: 682062

[0061]

[0062]

[0063] For the image reconstruction branch in the network model, bilinear upsampling (Upsample-18, Upsample-36, and Upsample-54) is used to improve the resolution of each element such as cloud cover and precipitation. The meteorological elements after resolution improvement have a higher peak signal-to-noise ratio.

[0064] Then, parameter adjustment and iterative training are performed. For the training of different elements, it is necessary to adjust the hyperparameters to make the error between the model output data and the true value converge, and obtain the optimal solution of the model for this element. The hyperparameters involved are shown in Table 2.

[0065] Table 2 Hyperparameters for MSRLapN Model Training

[0066]

[0067] Finally, the MSRLapN neural network is executed to achieve the super-resolution downscaling function. After the model training is completed, a set of model parameters are obtained for different elements. For a certain element, run the execution script. The input is the low-resolution (spatial resolution of 0.5°) multi-model ensemble forecast data, and the output is the refined forecast data product of this element (spatial resolution of 0.045°). Among them, it is necessary to first interpolate the model data with a spatial resolution of 0.5° to a 0.36° grid so that the data grid can reach the fine level of a spatial resolution of 0.045° after 8-fold super-resolution downscaling by the MSRLapN model, and finally obtain a set of refined model forecast data.

[0068] S130: After implementing the deep super-resolution downscaling, introduce topographic data and climatological data into the intelligent correction model. The intelligent correction model is implemented by the U-Net neural network. The construction of the intelligent correction model includes constructing data, constructing the U-Net neural network for each element, parameter adjustment and iterative training, and executing the U-Net neural network to achieve the multi-model integration function.

[0069] U-Net refers to the U-Net convolutional neural network. The U-Net consists of two parts: an encoder and a decoder, which are connected by skip connections to form a U-shaped structure. The encoder gradually reduces the spatial dimension of the input data, and the decoder gradually restores the detailed information and spatial dimension of the data.

[0070] Optionally, in some embodiments, first, a dataset is constructed by elements, where the elements include total cloud cover, low cloud cover, cumulative precipitation, and other elements. To better learn the impact of complex terrain information on meteorological data, after achieving deep super-resolution downscaling, the present invention further introduces terrain data and climatological data into the intelligent correction model. Climatological data, that is, the 30-year climate average state, is the average data of 30 years. It reflects the average state of the climate on a relatively long time scale and is an important basis for evaluating climate change, conducting climate prediction, and formulating climate policies.

[0071] The climatological data is obtained by calculating the average value. Taking low cloud cover as an example, the formula for climatological low cloud cover is as follows, X t is the low cloud cover at time t, and there are n low cloud covers at time t over the years. The climatological low cloud cover is the average result of the data at n times;

[0072]

[0073] Datasets are constructed separately by different elements. The horizontal resolutions of the atmospheric model forecast data and the true value data are unified through bilinear interpolation; then the vertical resolutions of the atmospheric model forecast data and the true value data are unified through cubic spline interpolation; finally, normalization processing is performed on the input (atmospheric model forecast data) and the true value (ERA5 reanalysis data) of the training and test data respectively. The normalization method is Max-Min Scaling, and the formula is as follows, where X normalized is the standardized sample data, X is the original sample data, X min is the minimum value in the data sample, X max is the maximum value in the data sample; for elements such as cloud cover and relative humidity that are percentages, X max takes the value of 100, and X min takes the value of 0:

[0074]

[0075] The constructed dataset is stored in the binary file format dedicated to NumPy, with the file suffix of npy, which is convenient for reading in subsequent modeling.

[0076] Secondly, a U-Net neural network is constructed by elements. The constructed U-Net neural network includes an input layer, 4 layers of downsampling, 4 layers of upsampling, and an output layer. The downsampling is achieved through the Max Pooling operation, and the upsampling is achieved through the Deconvolution operation. The downsampling and upsampling are directly connected through Skip Connections. The specific structure is as Figure 4As shown below. Using the summary function of the torchkeras module, the structural information of the U-Net neural network of the present invention is viewed as follows. Among them, Conv2d-1 to LeakyReLU-7 are the input layers of the model, named BaseConv, and MaxPoo12d-8 to LeakyReLU-15 are the combination of Max Pooling and BaseConv as the first downsampling layer. MaxPoo12d-8 to LeakyReLU-39 include 4 groups of downsampling layers. Conv2d-40 to Conv2d-116 are the ConvLSTM neural network implemented by fully convolutional layers. ConvTranspose 2d-117 to LeakyReLU-126 are the upsampling layer_4 composed of the combination of ConvTranspose and BaseConv. ConvTranspose2d-127 to LeakyReLU-156 contain 3 groups of upsampling layers, and Conv2d-157 is the output layer.

[0077] Table 3 torchkeras module parameters

[0078]

[0079]

[0080]

[0081]

[0082] Then, parameter adjustment and iterative training are carried out. For the training of different elements, it is necessary to adjust the hyperparameters to make the error between the model output data and the true value converge, and obtain the optimal solution of the model for this element. The hyperparameters involved in the present invention are shown in Table 4.

[0083] Table 4 U-Net model training hyperparameters

[0084]

[0085] Finally, the U-Net neural network is executed to implement the multi-mode integration function. After the model training is completed, a set of model parameters are obtained for different elements. For a certain element, run the execution script, the input is the low-resolution mode forecast data, and the output is the corrected high-resolution mode forecast data for this element.

[0086] For the downscaling and correction algorithms of special weather phenomena (such as precipitation, cloud cover, etc.) based on convolutional neural networks, there are often systematic biases. Generally, when the cloud cover reaches more than 80%, it is a small-probability situation spatially, and cloudy weather is a small-probability event temporally. Precipitation also has similar characteristics. Moderate rain and heavy rain are both small-probability events in arid regions. Therefore, the results output by the convolutional neural network model generally underestimate the above-mentioned small-probability events. Therefore, the present invention proposes a simple method to solve this problem by adding an adjustment term δ during the data normalization process.

[0087] Before model training, the data is normalized, and y i is the result of mapping the original data to 0 to 1. In order to adjust the systematic bias of special weather phenomena, the systematic bias is added to the inverse normalization formula when calling the network model to generate data. The specific formula is as follows:

[0088] Y i = y i *(X max - X min + δ)+ X min

[0089] In the formula, X max and X min are the maximum and minimum values of the input data among all samples, y i is the output result of the neural network model, and Y i is the inverse-normalized data value, and δ is the empirical value for adjusting the systematic bias.

[0090] The above is the introduction of the method embodiment. The following further illustrates the solution of the present disclosure through a system embodiment.

[0091] Figure 5 FIG. shows a block diagram of an artificial intelligence downscaling system 200 for meteorological elements according to an embodiment of the present disclosure. As Figure 5 shown, the system 200 includes:

[0092] A data acquisition module 210, configured to acquire low-resolution data, perform preprocessing and enhancement processing on the low-resolution data. The preprocessing includes denoising, standardization, and missing value processing; the enhancement processing includes using bilinear interpolation to improve the resolution of the low-resolution data.

[0093] A scale adjustment module 220, configured to implement deep super-resolution downscaling by using the MSRLapN model, including constructing data, constructing the MSRLapN neural network for each element, parameter adjustment and iterative training, and executing the MSRLapN neural network to implement the downscaling function.

[0094] The model correction module 230 is used to implement the intelligent correction model by introducing terrain data and climate state data after deep super-resolution downscaling. The intelligent correction model is implemented by a U-Net neural network. The construction of the intelligent correction model includes constructing data, constructing the U-Net neural network for each element, parameter adjustment and iterative training, and executing the U-Net neural network to implement the multi-mode integration function.

[0095] Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working process of the described module can refer to the corresponding process in the foregoing method embodiment and will not be elaborated herein.

[0096] The above is the introduction of the system embodiment. The following further illustrates the solution of the present disclosure through the model result verification embodiment.

[0097] The present invention applies the SRRN model to the downscaling and correction of special weather phenomena, produces refined forecast data by applying the systematic deviation adjustment method proposed by the present invention, and finally conducts accuracy verification on the refined forecast products of low cloud amount, total cloud amount, and 3-hour accumulated precipitation. The present invention selects the last five days of two typical months (January and July) in winter and summer as the test set, and obtains the following verification results for the forecast periods of 0-24 hours, 24-48 hours, and 48-72 hours.

[0098] (1) Verification results of low cloud amount for the 0-24-hour forecast period

[0099] Table 5 shows the accuracy rates of the 24-hour forecasts of low cloud amount in January and July 2023. In the 24-hour forecast period of January, the forecast accuracy rate for clear sky with few clouds (0-4 tenths) is 0.34, and the forecast accuracy rates for mostly cloudy (5-7 tenths) and overcast (8-10 tenths) are relatively low, indicating that there are certain errors in the forecast system. For the 24-hour forecast period of July, the forecast accuracy rate for clear sky with few clouds is 0.43, which reflects that the prediction of low cloud amount for clear and mostly cloudy weather in summer is poor. In contrast, the forecast accuracy rate for overcast (greater than 7 tenths) is 0.58, indicating that the forecast system has a relatively high accuracy in predicting overcast days.

[0100] Overall, the refined forecast results perform best in predicting clear sky with few clouds in January and are excellent in predicting overcast days in July. However, the simulation results for low cloud amounts of 5-7 tenths in both January and July are not ideal.

[0101] Table 5 Statistical table of accuracy rates of 24-hour forecasts of low cloud amount in January and July 2023

[0102]

[0103] (2) Verification results of low cloud cover for the 24 - 48h forecast period

[0104] Table 6 shows the accuracy rates of the 48 - hour forecast of low cloud cover in January and July 2023. For the 48 - hour forecast in January, the accuracy rate of the low cloud cover forecast with 0 - 4 tenths is 0.34, while the accuracy rates of the low cloud cover forecasts with 5 - 7 tenths and 8 - 10 tenths are relatively low, which may be related to the weather characteristics in January, as sunny weather is more common. For the 48 - hour forecast in July, the accuracy rate of the low cloud cover with 0 - 4 tenths is 0.43, and the accuracy rate of the low cloud cover forecast with 5 - 7 tenths is not good. In contrast, the accuracy rate of the low cloud cover forecast with 8 - 10 tenths is as high as 0.58, indicating a relatively high accuracy in predicting overcast days.

[0105] From the above analysis, it can be seen that the refined forecasting system performs best in forecasting clear - sky and few - cloud conditions in January, and is excellent in forecasting overcast days in July. However, in forecasting cloudy weather, the forecasting accuracy is not ideal in both January and July. At the same time, this seasonal difference also shows that the changes in weather systems have a significant impact on forecasting accuracy.

[0106] Table 6 Statistical table of the accuracy rates of the 48 - hour forecast of low cloud cover in January and July 2023

[0107]

[0108] (3) Verification results of low cloud cover for the 48 - 72h forecast period

[0109] Table 7 shows the 72 - hour forecast accuracy rates of low cloud cover in January and July 2023. Compared with the 24 - hour forecast, in the 72 - hour forecast in January, the accuracy rate of the low cloud cover with 0 - 4 tenths drops from 0.34 to 0.31, the accuracy rate of the low cloud cover forecast with 5 - 7 tenths increases slightly, and the accuracy rate of the low cloud cover with 8 - 10 tenths remains unchanged. In the 72 - hour forecast in July, the accuracy rates of the low cloud cover with 0 - 4 tenths and 5 - 7 tenths do not change, and the forecasting ability needs to be improved urgently. The accuracy rate of the low cloud cover with 8 - 10 tenths is relatively high, indicating that the forecasting system has better accuracy in forecasting overcast days.

[0110] Comprehensively analyzing the 24 - hour, 48 - hour, and 72 - hour forecast periods in January and July, the accuracy rates fluctuate within different forecast periods, but the overall changes are within an acceptable range. In January, the accuracy rates of the low cloud cover forecasts with 0 - 4 tenths and 5 - 7 tenths decrease, and the accuracy rate for overcast days is stable. In July, the accuracy rates of the low cloud cover forecasts with 0 - 4 tenths and 5 - 7 tenths are still relatively low, but the forecasting of overcast days is very accurate. The results show that the performance of the forecasting model varies under different months and cloud cover conditions.

[0111] Table 7 Statistical table of the accuracy rates of the 72 - hour forecast of low cloud cover in January and July 2023

[0112]

[0113] (4) Total cloud cover inspection results for the 0 - 24h forecast period

[0114] Table 8 shows the accuracy rates of the 24 - hour forecast of total cloud cover in January and July 2023. For the 24 - hour forecast in January, the forecast accuracy rate for clear sky with few clouds (0 - 4 oktas) is 0.40, indicating relatively high accuracy in forecasting clear weather. The forecast accuracy rate for partly cloudy weather (5 - 7 oktas) is 0.41, slightly lower than that for clear sky with few clouds, and the forecast accuracy rate for overcast sky (8 - 10 oktas) is 0.38. For the 24 - hour forecast in July 2023, the forecast accuracy rate for clear sky with few clouds (0 - 4 oktas) increases to 0.59, the forecast accuracy rate for partly cloudy weather (5 - 7 oktas) is 0.56, and the forecast accuracy rate for overcast sky (8 - 10 oktas) significantly increases to 0.64. The above results show that the forecasting system's ability to predict clear weather in summer has been enhanced.

[0115] Based on the accuracy rate data of the 24 - hour forecast in January and July, it can be seen that there are differences in the forecasting accuracy of the forecasting system under different months and weather conditions. In January, the performance of the forecasting system is relatively close when predicting clear sky with few clouds and partly cloudy weather, but the accuracy rate decreases when predicting overcast sky. In July, the accuracy rate of the forecasting system increases under all weather conditions, especially when predicting overcast sky.

[0116] Table 8 Statistical table of the accuracy rates of the 24 - hour forecast of total cloud cover in January and July 2023

[0117]

[0118] (5) Total cloud cover inspection results for the 24 - 48h forecast period

[0119] Table 9 shows the accuracy rates of the 48 - hour forecast of total cloud cover in January and July 2023. For the 48 - hour forecast in January 2023, the accuracy rate for clear sky with few clouds (0 - 4 oktas) drops from 0.43 to 0.41, showing a slight decrease; the accuracy rate for partly cloudy (5 - 7 oktas) remains at 0.40, showing no obvious change compared with the 24 - hour forecast; the accuracy rate for overcast sky (8 - 10 oktas) drops from 0.38 to 0.38, and the forecasting performance remains stable. In contrast, the overall performance of the 48 - hour forecast in July 2023 improves. Although the accuracy rate for clear sky with few clouds (0 - 4 oktas) drops from 0.59 to 0.52, it is still at a relatively high level; the accuracy rate for partly cloudy (5 - 7 oktas) drops from 0.56 to 0.52; the accuracy rate for overcast sky (8 - 10 oktas) drops from 0.64 to 0.61, but it is still within an acceptable range.

[0120] By synthesizing the accuracy data of the 48-hour forecast lead times for January and July, it can be seen that there are differences in the forecast accuracy of the forecasting system under different months and different cloud cover conditions. In January, the accuracy of the forecasting system slightly decreased when predicting clear sky with few clouds and cloudy weather, but its performance remained stable when predicting overcast sky. In July, the accuracy of the forecasting system increased under all weather conditions, especially showing excellent performance when predicting overcast sky.

[0121] Table 9 Statistical Table of the Accuracy of the 48-hour Forecast Lead Time for Total Cloud Cover in January and July 2023

[0122]

[0123] (6) Test Results of the Total Cloud Cover for the 48 - 72h Forecast Lead Time

[0124] In the 72-hour forecast lead time in January 2023, as the forecast lead time extended, the forecast accuracy of the total cloud cover decreased. The change in the forecast accuracy of clear sky with few clouds (0 - 4 oktas) was not significant; the accuracy of cloudy sky (5 - 7 oktas) decreased from 0.40 to 0.38, indicating that the prediction accuracy of the system for cloudy weather decreased after the extension of the forecast lead time; the accuracy of overcast sky (8 - 10 oktas) increased from 0.38 to 0.44, showing that the performance of the forecasting system was relatively stable when predicting overcast sky. Similarly, in the 72-hour forecast in July, the forecast accuracies of clear sky with few clouds (0 - 4 oktas) and cloudy sky (5 - 7 oktas) also decreased slightly, but were still within an acceptable range overall.

[0125] By comprehensively analyzing the accuracy data of the 72-hour forecast lead times for January and July, it can be seen that there are differences in the performance of the forecasting system under different months and different cloud cover conditions. In January, the accuracy of the forecasting system decreased when predicting clear sky with few clouds and cloudy weather, but its performance improved when predicting overcast sky; in July, the forecasting system showed excellent performance when predicting overcast sky, and the forecast accuracies under other cloud cover conditions were also relatively stable. Generally speaking, these results indicate that it is of great significance to optimize and adjust the forecasting model according to seasonal weather characteristics, which helps to improve the prediction accuracy under different meteorological conditions.

[0126] Table 10 Statistical Table of the Accuracy of the 72-hour Forecast Lead Time for Total Cloud Cover in January and July 2023

[0127]

[0128] (7) Test Results of the 3-hour Cumulative Precipitation for the 0 - 24h Forecast Lead Time

[0129] The TS scores of heavy, moderate, light rain and rainstorm events occurring at each station at each time were statistically calculated after classifying the 24-hour precipitation of the refined forecast products (forecast lead time of 0 - 24h) from January 24th to 28th, 2023 and from July 24th to 28th, 2023 according to the precipitation grade, and compared with the ground observation data. As shown in the following table, light rain events mainly occurred in the last five days of July and January, and the TS score distributions were 0.32 and 0.13 respectively; relatively few moderate rain events occurred in the last five days of July and January, and the TS scores were also relatively low, 0.02 and " / " respectively; no heavy rain and rainstorm events occurred in the last five days of July and January, and the TS scores were 0 and " / ", where " / " represents an invalid value.

[0130] Table 11 Spatial and Temporal Statistics of TS Scores in January and July 2023 (Forecast Lead Time 0 - 24h)

[0131]

[0132] (8) Test Results of 3-hour Cumulative Precipitation for Forecast Lead Time of 24 - 48h

[0133] The TS scores of heavy, moderate, light rain and rainstorm events occurring at each station at each time were statistically calculated after classifying the 24-hour precipitation of the refined forecast products (forecast lead time of 24 - 48h) from January 24th to 28th, 2023 and from July 24th to 28th, 2023 according to the precipitation grade, and compared with the ground observation data. As shown in the following table, light rain events mainly occurred in the last five days of July and January, and the TS score distributions were 0.38 and 0.09 respectively; relatively few moderate rain events occurred in the last five days of July and January, and the TS scores were also relatively low, 0.1 and " / " respectively; no heavy rain and rainstorm events occurred in the last five days of July and January, and the TS scores were 0 and " / ", where " / " represents an invalid value. Generally speaking, the TS scores of each classification for the 24 - 48h lead time are worse than those for the 0 - 24h lead time.

[0134] Table 12 Spatial and Temporal Statistics of TS Scores in January and July 2023 (Forecast Lead Time 24 - 48h)

[0135]

[0136] (9) Test Results of 3-hour Cumulative Precipitation for Forecast Lead Time of 48 - 72h

[0137] The TS scores of heavy, moderate, light rain and rainstorm events occurring at each station within each time period were statistically calculated after classifying the refined forecast products (forecast lead time of 48 - 72h) from January 24th to 28th, 2023 and July 24th to 28th, 2023 and the 24h precipitation of ground observation data according to precipitation levels, as shown in the following table. Light rain events mainly occurred in the last five days of July and January, and the TS score distributions were 0.36 and 0.03 respectively; relatively few moderate rain events occurred in the last five days of July, and the TS scores were also relatively low, both being 0.02; no moderate rain events occurred in the last five days of January. No heavy rain or rainstorm events occurred in the last five days of July and January, and the TS scores were 0 and " / ", where " / " represents an invalid value. From the overall trend, the TS scores of each classification level for the 0 - 24h, 24 - 48h, and 48 - 72h lead times decreased gradually, but the TS score for 24 - 48h in January was slightly lower than that for 48 - 72h.

[0138] Table 13 Spatiotemporal Statistics of TS Scores in January and July 2023 (Forecast Lead Time 48 - 72h)

[0139]

[0140]

[0141] The present invention proposes a super - resolution downscaling for meteorological data based on a new neural network model called the Super - Resolution Refinement Network (SRRN). Compared with the regional atmospheric forecast model, this model has low computational complexity and high computational efficiency. In order to correct the problems commonly existing in the convolutional neural network model, such as the lack of terrain constraints and the underestimation of small - probability situations, the present invention proposes a correction network considering terrain constraints and an anti - normalization method for adjusting systematic biases.

[0142] It should be understood that various forms of the processes shown above can be used, with steps reordered, added, or deleted. For example, the steps recorded in this disclosure can be executed in parallel, sequentially, or in a different order, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved. This is not limited herein.

[0143] The above - mentioned specific embodiments do not constitute a limitation to the protection scope of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub - combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure shall be included within the protection scope of this disclosure.

Claims

1. An artificial intelligence downscaling method for meteorological elements, characterized in that: include: Obtain low-resolution data and perform preprocessing and enhancement on the low-resolution data. The preprocessing includes denoising, standardization, and missing value processing. The enhancement process involves increasing the resolution of low-resolution data using bilinear interpolation; Using the MSRLapN model to achieve deep super-resolution downscaling, including constructing data, constructing the MSRLapN neural network by elements, iterative training of parameter adjustment, and executing the MSRLapN neural network to achieve downscaling function; After achieving deep super-resolution downscaling, terrain data and climate data are introduced into the intelligent correction model, and the intelligent correction model is implemented by the U-Net neural network.

2. The method according to claim 1, characterized in that The preprocessing includes denoising, standardization and missing value processing; Missing value processing: introduce an interpolation algorithm based on the correlation of meteorological elements, and use the data relationship between adjacent time and space points to dynamically fill in missing values; Standardization: Use segmented standardization technology to map the data of different meteorological elements to a unified scale; dynamically adjust the standardization parameters by analyzing the distribution characteristics of meteorological elements.

3. The method according to claim 2, characterized in that The denoising means adaptive denoising based on graph neural network, which models the meteorological data into a graph structure and performs dynamic denoising using the features of nodes and edges; In the formula, H (l) Represents the node feature matrix of the lth layer; represents the adjacency matrix with self-loop added; A represents the adjacency matrix; I represents the identity matrix; represents the degree matrix; W ( l ) represents the weight matrix of the lth layer; σ represents the activation function (such as ReLU); Representation Matrix The element in the i′th row and i′th column of ; represents the element at the intersection of the i′th row and the jth column in the matrix; Missing value processing uses a variational autoencoder to learn the potential distribution of meteorological data through a generative model and dynamically generate missing values: Where ELBI represents the lower bound of evidence; q(z|x) represents the latent variable distribution output by the encoder; p(x|z) represents the data distribution generated by the decoder; KL represents the KL divergence; μ(x), σ 2 (x) represents the mean and variance of the encoder output; p(z) represents the prior distribution; The normalization process uses manifold learning to map the meteorological data into a low-dimensional manifold space to capture its nonlinear structure: In the formula, p j|i′ Represents the point x in high-dimensional space i′ and x j Similarity of q j|i′ Represents the midpoint y in the low-dimensional space i′ and j similarity; represents the bandwidth parameter of the Gaussian kernel; Cost represents the optimization objective.

4. The method according to claim 1, characterized in that: The data enhancement process adopts an improved bilinear interpolation algorithm, combines the spatial distribution characteristics of meteorological elements, and dynamically adjusts the interpolation weights; In the formula, |ψ> represents the quantum state; α i′ represents a complex weight; |x i′ > indicates the ground state; θ i′ represents the phase angle; f(x, y) represents the interpolation estimate.

5. The method according to claim 1, characterized in that The sub-element construction data set includes: low cloud cover, total cloud cover and 3-hour cumulative precipitation elements; constructing data sets according to different elements, using the bilinear interpolation method, the high-resolution three-dimensional atmospheric real-time fusion grid data are interpolated into low-resolution grid data with spatial resolutions of 0.09°, 0.18° and 0.36° respectively; normalizing the data; The constructed dataset is stored in a binary file format dedicated to NumPy with the file suffix npy, which is convenient for reading in subsequent modeling.

6. The method according to claim 1, characterized in that The sub-element construction of the MSRLapN neural network divides the downscaling process into two branches: an image reconstruction branch and a feature extraction branch; The image reconstruction branch sets a specific scale factor to enlarge it; it takes high spatial resolution terrain data and low spatial resolution meteorological data as input, and for each input, first extracts its hidden layer features through a single layer convolution, and then concatenates the transformed data into a multi-layer three-dimensional image to achieve size unification; At each level of the feature extraction branch, there is an MSRB and a subpixel convolution layer, where the MSRB performs multi-scale feature extraction and the upsampling scale of the subpixel convolution layer is set to 2; the MSRB first contains a convolution layer with a convolution kernel size of 1×1; the output of the subpixel convolution layer will be used as the input of two different network layers, one is to reconstruct the residual information at the same level, and the other is used for feature extraction in the next layer; for the data reconstruction branch, it contains a subpixel convolution layer and two ordinary convolution layers, and the input data is upsampled by the subpixel convolution layer; the other convolution layer takes the extracted features as input and outputs the high-resolution data reconstructed by the subpixel layer.

7. The method according to claim 1, characterized in that The parameter adjustment iterative training is described, and the error between the model output data and the true value is converged by adjusting the hyperparameters, so as to obtain the optimal model solution for the element, and execute the MSRLapN neural network to realize the super-resolution downscaling function. After the model training is completed, different elements respectively obtain a set of model parameters, and for a certain element, the execution script is run, the input is low-resolution multi-mode integrated forecast data, and the output is a refined forecast data product of the element; the model data with a spatial resolution of 0.5° is interpolated to a 0.36° grid so that the data grid can reach a spatial resolution of 0.045° after 8-fold super-resolution downscaling of the MSRLapN model, and finally the model forecast data is obtained.

8. The method according to claim 1, characterized in that The construction of the intelligent correction model includes constructing data, constructing a U-Net neural network by elements, iterative training of parameter adjustment, and executing the U-Net neural network to realize multi-mode integration function; Construct datasets by elements, including total cloud cover, low cloud cover, and accumulated precipitation; construct datasets by different elements, unify the horizontal resolution of atmospheric model forecast data and true value data through bilinear interpolation; unify the vertical resolution of atmospheric model forecast data and true value data through cubic spline interpolation; normalize the input and true value of training and test data respectively; store the constructed datasets in a binary file format dedicated to NumPy; Secondly, a U-Net neural network is constructed element by element. The constructed U-Net neural network includes an input layer, 4 downsampling layers, 4 upsampling layers and an output layer. Downsampling is achieved through maximum pooling operations, upsampling is achieved through deconvolution operations, and downsampling and upsampling are directly connected through shortcut connections.

9. The method according to claim 8, characterized in that The parameter adjustment iterative training is performed to converge the error between the model output data and the true value by adjusting the hyperparameters, thereby obtaining the optimal solution of the model for the factor; Execute the U-Net neural network to realize the multi-mode integration function. After the model training is completed, different elements obtain a set of model parameters. For a certain element, run the execution script, the input is the low-resolution model forecast data, and the output is the high-resolution model forecast data after the element is corrected.

10. An artificial intelligence downscaling system for meteorological elements, characterized in that: include: The data acquisition module is used to acquire low-resolution data and perform preprocessing and enhancement on the low-resolution data. The preprocessing includes denoising, standardization and missing value processing. The enhancement process involves increasing the resolution of low-resolution data using bilinear interpolation; The scale adjustment module is used to implement deep super-resolution downscaling using the MSRLapN model, including constructing data, constructing the MSRLapN neural network by elements, iterative training of parameter adjustment, and executing the MSRLapN neural network to achieve downscaling function; The model correction module is used to introduce terrain data and climate data into the intelligent correction model after deep super-resolution downscaling. The intelligent correction model is implemented by the U-Net neural network. The construction of the intelligent correction model includes constructing data, constructing the U-Net neural network by elements, adjusting parameters for iterative training, and executing the U-Net neural network to realize multi-mode integration functions.

Images