Atmospheric mode forecast data comprehensive post-processing method and system based on artificial intelligence

Abstract

The embodiment of the invention provides an atmospheric mode forecast data comprehensive post-processing method and system based on artificial intelligence. The method is applied to the technical field of data processing, and comprises the following steps: establishing a U-Net neural network model for multi-mode integrated forecasting, adjusting hyper-parameters of the U-Net neural network model, and training by using a data set to obtain a trained U-Net neural network model; constructing a structure for a deep super-resolution downscaling MSRLapN neural network model according to elements, adjusting hyper-parameters of the MSRLapN neural network model, and performing training by using the data set to obtain a trained MSRLapN neural network model; and for the characteristic elements of the rainfall and the cloud cover, calculating the characteristic elements under different grades by adopting a graded loss function. In this way, the problems that in the prior art, a scheme is difficult to implement, and implementation cost is high; the calculation amount is large, and simulation and configuration are difficult.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and in particular to an artificial intelligence-based comprehensive post-processing method and system for atmospheric model forecast data. Background Art

[0002] Global atmospheric numerical models (GAMs) are mathematical and physical models used to simulate various physical and chemical processes in the atmosphere. They are currently one of the primary methods for forecasting atmospheric environmental factors (including temperature, humidity, wind speed, and air pressure). While GAMs accurately predict these factors, their output has a relatively low spatial resolution, making it difficult to reflect convective-scale weather variations and regionally differentiated climate characteristics. However, forecasts from different GAMs often differ, reducing accuracy and uncertainty.

[0003] Therefore, improving the horizontal resolution of model data is crucial. Generally, three methods are used to improve model forecast data: 1) developing a high-resolution global atmospheric forecast model; 2) using a regional atmospheric forecast model; and 3) developing novel statistical downscaling algorithms. While method 1 is relatively difficult to implement, method 2, which can be embedded in a global model and run independently using the global model's output as boundary conditions, offers advantages such as physical interpretability and immunity to observational data. However, this method is computationally intensive and difficult to simulate and configure. Method 3, statistical downscaling, involves fusing observational data to learn a mapping relationship, mapping the Earth system model to a high resolution. In recent years, with the rapid development and application of deep neural network methods, image downscaling based on convolutional neural network super-resolution models has shown great potential for downscaling meteorological elements. Therefore, to improve the accuracy of global atmospheric numerical model forecast data, reduce forecast uncertainty, and expand forecast efficiency and application scope, it is crucial to integrate multiple model forecast data sets into a more accurate set. Furthermore, the resulting multi-model ensemble forecast data exhibits enhanced accuracy and stability, making it more valuable for practical applications.

[0004] Therefore, existing technologies have significant shortcomings. First, the technical solution is difficult to implement and the implementation cost is high. Second, the global model is not accurately defined, resulting in large computational workloads and difficulty in simulation and configuration. Third, statistical downscaling, which learns mapping relationships by fusing observational data, is inefficient and requires further improvement in its intelligence. Summary of the Invention

[0005] The present invention provides an artificial intelligence-based comprehensive post-processing method and system for atmospheric model forecast data. Aiming at the defects of various existing technologies, the present invention solves the technical problems that the technical solutions in the existing technologies are difficult to implement and have high implementation costs; the global model definition is inaccurate, resulting in a large amount of calculation and difficulty in simulation and configuration; and the statistical downscaling is inefficient in learning mapping relationships by fusing observation data, and the level of intelligence needs to be further improved.

[0006] According to a first aspect of the present invention, there is provided a method for comprehensive post-processing of atmospheric model forecast data based on artificial intelligence, comprising:

[0007] By constructing a dataset by elements, a U-Net neural network model for multi-mode integrated forecasting is established, the hyperparameters of the U-Net neural network model are adjusted, and the dataset is used for training to obtain a trained U-Net neural network model;

[0008] Construct a dataset for training the MSRLapN neural network model, construct the structure of the MSRLapN neural network model for deep super-resolution downscaling by elements, adjust the hyperparameters of the MSRLapN neural network model, and use the dataset for training to obtain the trained MSRLapN neural network model;

[0009] For the characteristic elements of precipitation and cloud cover, a graded loss function is used to calculate the characteristic elements at different levels.

[0010] According to a second aspect of the present invention, there is provided an artificial intelligence-based comprehensive post-processing system for atmospheric model forecast data, comprising:

[0011] The first model construction module is used to construct a data set by dividing the elements, establish a U-Net neural network model for multi-mode integrated forecasting, adjust the hyperparameters of the U-Net neural network model, and train it using the data set to obtain a trained U-Net neural network model;

[0012] The second model construction module is used to construct a data set for training the MSRLapN neural network model, construct the structure of the deep super-resolution downscaling MSRLapN neural network model by elements, adjust the hyperparameters of the MSRLapN neural network model, and train it using the data set to obtain a trained MSRLapN neural network model;

[0013] The loss function design module is used to calculate the characteristic elements of precipitation and cloud cover at different levels using a graded loss function.

[0014] Compared with the prior art, the advantages and positive effects achieved by the present invention are:

[0015] The results of different global atmospheric numerical model forecasts in the present invention often have certain differences. In order to improve the accuracy of model forecast data, reduce forecast uncertainty, and enhance forecast efficiency and application scope, integrating multiple sets of model forecast data into more accurate model forecast data is of great significance. Compared with the two major types of multi-model integrated forecasting technologies, namely deterministic forecasting and probabilistic forecasting, the multi-model integrated forecasting technology based on machine learning algorithms in the present invention can obtain higher weather forecasting skills; compared with convolutional neural networks, BP neural networks and LSTM neural networks do not consider the spatial correlation of model forecast data; the multi-model integrated forecasting technology based on U-Net proposed in the present invention is more suitable for spatially continuous two-dimensional and three-dimensional model forecast data, and can better learn the spatial correlation of model forecast data, and obtain multi-model integrated forecast data with higher forecast skills while ensuring spatial correlation. The applied deep learning-based downscaling network model has the characteristics of lightness and efficiency, but most of the objects considered are two-dimensional image data. In the field of meteorological elements, three-dimensional data downscaling also has important uses. Conventional color images usually include three color channels: R, G, and B, while the images corresponding to meteorological three-dimensional data are multi-channel. For three-dimensional meteorological element forecast data, before and after applying the super-resolution (SR) method, the changes in meteorological elements such as temperature and humidity between adjacent slices are continuous in three-dimensional space; the downscaling of wind speed is not only related to wind speed, but also to temperature, humidity, etc.; there are internal correlations between meteorological element forecast images of multiple resolutions, etc.; and simply extending the two-dimensional image downscaling model to multi-channel downscaling applications fails to make good use of this type of meteorological domain knowledge. The present invention applies the MSRLapN neural network model to improve the spatial resolution of three-dimensional meteorological forecast images, which can better extract feature information and achieve better three-dimensional downscaling effects by integrating meteorological domain knowledge.

[0016] The present invention comprehensively applies conventional atmospheric elements of the ground and pressure layers, proving that both models have certain application value in different elements. The applied elements include: ground elements (including 2m temperature, 2m humidity, 10m meridional wind, 10m zonal wind, total precipitation, low cloud cover, etc.) and pressure layer elements (including air temperature, humidity, wind, etc.).

[0017] It should be understood that the contents described in the summary of the invention are not intended to limit the key or important features of the embodiments of the present invention, nor are they intended to limit the scope of the present invention. Other features of the present invention will become readily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The above and other features, advantages and aspects of the embodiments of the present invention will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. The accompanying drawings are provided for a better understanding of the present invention and do not constitute a limitation of the present invention. In the accompanying drawings, the same or similar reference numerals represent the same or similar elements, among which:

[0019] Figure 1 A flowchart of an artificial intelligence-based integrated post-processing method for atmospheric model forecast data according to an embodiment of the present invention is shown;

[0020] Figure 2 A schematic diagram of the PSPPNN network structure according to an embodiment of the present invention is shown;

[0021] Figure 3 A schematic diagram of a multi-mode integrated network structure according to an embodiment of the present invention is shown;

[0022] Figure 4 A schematic diagram of the network structure of the MSRLapN neural network model according to an embodiment of the present invention is shown;

[0023] Figure 5 A block diagram of an artificial intelligence-based atmospheric model forecast data comprehensive post-processing system according to an embodiment of the present invention is shown;

[0024] Figure 6 A flowchart of business operation according to an embodiment of the present invention is shown;

[0025] Figure 7 A schematic diagram of the 0-24h test results according to an embodiment of the present invention is shown;

[0026] Figure 8 A schematic diagram of 24-48h test results according to an embodiment of the present invention is shown;

[0027] Figure 9 A schematic diagram of the 48-72h test results according to an embodiment of the present invention is shown;

[0028] Figure 10 A schematic diagram of the 0-24h test results according to an embodiment of the present invention is shown;

[0029] Figure 11 A schematic diagram of 24-48h test results according to an embodiment of the present invention is shown;

[0030] Figure 12 A schematic diagram of the 48-72h test results according to an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0031] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0032] In this document, the term "and / or" simply describes a relationship between related objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A exists alone, A and B exist simultaneously, or B exists alone. Furthermore, the character " / " in this document generally indicates that the related objects are in an "or" relationship.

[0033] Figure 1 FIG. 1 is a flow chart showing an artificial intelligence-based integrated post-processing method 100 for atmospheric model forecast data in an embodiment of the present invention. Figure 1 As shown, the method 100 includes:

[0034] S110: constructing a data set by dividing the elements, establishing a U-Net neural network model for multi-mode integrated forecasting, adjusting the hyperparameters of the U-Net neural network model, and training it using the data set to obtain a trained U-Net neural network model.

[0035] Optionally, in some embodiments, the process of constructing a data set by elements includes, where the elements include: ground meteorological elements such as 2m temperature, 2m humidity, 10m meridional wind, 10m zonal wind, ground air pressure, total precipitation and low cloud cover, and pressure layer elements such as temperature, humidity, meridional wind, zonal wind of 100, 150, 200, 250, 300, 400, 500, 600 and 700hPa pressure layers.

[0036] The present invention constructs data sets according to different elements, unifies the horizontal resolution of atmospheric model forecast data and true value data by bilinear interpolation, unifies the vertical resolution of atmospheric model forecast data and true value data by cubic spline interpolation, and performs normalization on the input (atmospheric model forecast data) and true value (ERA5 reanalysis data) of training and test data respectively. The normalization method is Z-Score normalization, and the formula is as follows, where z i is the sample data after standardization, x i is the original sample data, is the sample mean, and S is the sample standard deviation.

[0037]

[0038] The present invention stores the constructed dataset in a binary file format dedicated to NumPy with the file suffix npy to facilitate reading for subsequent modeling.

[0039] The process of building a U-Net neural network specifically includes:

[0040] The U-Net neural network constructed in the present invention includes an input layer, 4 downsampling layers, 4 upsampling layers and an output layer. Downsampling is achieved by the maximum pooling operation, upsampling is achieved by the deconvolution operation, and downsampling and upsampling are directly connected through skip connections. The specific structure is as follows Figure 3 shown.

[0041] The present invention uses the summary function of the torchkeras module to view the structural information of the U-Net neural network of the present invention as shown in Table 1.

[0042] Table 1 Structural information of the U-Net neural network

[0043]

[0044]

[0045]

[0046]

[0047]

[0048] The process of iterative parameter adjustment training includes: for training different elements, it is necessary to adjust the hyperparameters to converge the error between the model output data and the true value, and obtain the optimal model solution for the element. The hyperparameters involved in the present invention are shown in Table 2.

[0049] Table 2 U-Net model training hyperparameters

[0050]

[0051]

[0052] The process of executing the U-Net neural network to realize the multi-model integration function includes: after the model training is completed, different elements obtain a set of model parameters respectively. For a certain element, the execution script is run, the input is 5 sets of atmospheric model forecast data, and the output is the integration result of the model forecast data of the element.

[0053] It should be noted that, in the embodiment, for the multi-mode integrated forecasting method, compared with the two major types of multi-mode integrated forecasting technologies, namely deterministic forecasting and probabilistic forecasting, the multi-mode integrated forecasting technology based on machine learning algorithm can obtain higher weather forecasting skills; compared with convolutional neural networks, BP neural networks and LSTM neural networks do not consider the spatial correlation of model forecast data, and BP neural networks and LSTM neural networks process data into a set of vectors including several features or a set of time series vectors including several features. The multi-mode integrated forecasting technology based on U-Net proposed in the present invention is more suitable for spatially continuous two-dimensional and three-dimensional model forecast data, can better learn the spatial correlation of model forecast data, and obtain multi-mode integrated forecast data with higher forecasting skills under the premise of ensuring spatial correlation.

[0054] The present invention is about the deep super-resolution downscaling method. Most of the downscaling network models based on deep learning consider the object as two-dimensional image data. In the field of meteorological elements, three-dimensional data downscaling also has important uses. Conventional color images usually include three color channels: R, G, and B, while the images corresponding to meteorological three-dimensional data are multi-channel. For three-dimensional meteorological element forecast data, before and after applying the super-resolution (SR) method, the changes in meteorological elements such as temperature and humidity between adjacent slices are continuous in three-dimensional space; the downscaling of wind speed is not only related to wind speed, but also to temperature, humidity, etc.; there are internal correlations between meteorological element forecast images of multiple resolutions, etc.; and simply extending the two-dimensional image downscaling model to multi-channel downscaling applications fails to better apply this type of meteorological field knowledge. The present invention applies the MSRLapN neural network model to improve the spatial resolution of three-dimensional meteorological forecast images. The model can better extract feature information, and at the same time, the three-dimensional downscaling effect of integrating meteorological field knowledge is better.

[0055] The present invention combines the U-Net network structure, the ConvLSTM network structure, and the MSRLapN network to obtain a comprehensive processing method with multi-modal integration, intelligent correction, and super-resolution downscaling functions. U-Net: U-Net refers to the U-Net convolutional neural network. 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 repairs the detailed information and spatial dimension of the data. The convolutional long short-term memory (ConvLSTM) module is a deep learning architecture that combines a convolutional neural network (CNN) and a long short-term memory network (LSTM). The core of ConvLSTM lies in its cell structure, which has been modified based on LSTM by replacing the fully connected layer in LSTM with a convolutional layer. The full name of the MSRLapN neural network model is the Multi-Scale Residual Laplacian Pyramid Network (MSRLapN), a three-dimensional meteorological element deep learning downscaling model. It constructs a multi-scale residual module (MSRB) to automatically extract forecast features from multiple 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 iterative method, learns to correct the downscaling forecast error based on large sample historical data. The model focuses on solving the following computational problems: (1) New model forecast data synthesis post-processing network: A new network structure is proposed, called the model forecast data synthesis post-processing neural network (PSPPNN, such as Figure 2 (as shown), U-Net, ConvLSTM and MSRLapN are organically combined, with n sets of model forecast data as input and two processing results as output, namely the corrected multi-model integrated forecast data and the downscaled refined model forecast data. (2) Improve the rationality and accuracy of neural network processing model forecast data: In order to improve the rationality of neural network model correction, pressure layer data, ground data and terrain data are added to the model input to obtain more reasonable and accurate refined forecast data. (3) Design a reasonable loss function for characteristic elements: For factors such as precipitation and cloud cover, the loss function is calculated in stages. Taking precipitation as an example, the loss value is calculated according to the weights of heavy rain, moderate rain and light rain to ensure the accuracy of characteristic elements at different levels.

[0056] The present invention uses the U-Net neural network model to achieve multi-mode integrated forecasting, which can obtain more accurate model data, and uses the MSRLapN neural network model to achieve super-resolution downscaling of basic meteorological elements to improve the spatial resolution of three-dimensional meteorological forecast images. The present invention realizes multi-mode integration and super-resolution downscaling of basic meteorological elements in the ground and pressure layers based on a deep learning model.

[0057] S120: Construct a data set for training the MSRLapN neural network model, construct the structure of the MSRLapN neural network model for deep super-resolution downscaling by elements, adjust the hyperparameters of the MSRLapN neural network model, and use the data set for training to obtain a trained MSRLapN neural network model.

[0058] Optionally, in some embodiments, the elements of the dataset constructed by elements include: surface meteorological elements such as 2m temperature, 2m humidity, 10m meridional wind, 10m zonal wind, surface air pressure, total precipitation, and low cloud cover, as well as pressure layer elements such as temperature, humidity, meridional wind, and zonal wind at 100, 150, 200, 250, 300, 400, 500, 600, and 700hPa. The process includes: constructing datasets according to different elements, using a bilinear interpolation method to interpolate high-resolution (horizontal resolution of 0.045°, approximately 5km) three-dimensional atmospheric real-time fusion grid data into low-resolution grid data with spatial resolutions of 0.09°, 0.18°, and 0.36°, respectively. The data is normalized using the following formula: normalized is the sample data after standardization, X is the original sample data, and X min is the minimum value in the data sample, X max is the maximum value in the data sample.

[0059]

[0060] The present invention stores the constructed dataset in a binary file format dedicated to NumPy with the file suffix npy to facilitate reading for subsequent modeling.

[0061] The process of constructing the MSRLapN neural network by elements of the present invention includes: the model divides the downscaling process into two branches: an image reconstruction branch and a feature extraction branch. Figure 4 The LapSRN network structure is given. MSRLapN includes several sets of repeated network structure layers, and a specific scaling factor is set for different layers to amplify them. Figure 3Only the first level of the model structure is shown for the low-resolution to high-resolution reconstruction process. In practice, the number of levels can be increased or decreased based on different needs. The model takes high-spatial-resolution terrain data and low-spatial-resolution meteorological data as input. For each input (low-resolution climate data and high-resolution terrain data), a single convolution layer is used to first extract the hidden features. Because the terrain data and low-resolution climate data have different sizes, these transformed data need to be concatenated into a multi-layer 3D image to achieve uniform size. All transposed convolutional layers in MSRLapN are replaced by subpixel convolutional layers.

[0062] At each level of the feature extraction branch, there is a MSRB and a subpixel convolution layer. The 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, the MSRB first includes a convolution layer with a convolution kernel size of 1×1, which can greatly reduce the number of feature layer parameters and ensure that the number of input and output feature layers of the MSRB remains unchanged. Figure 3 As shown, the output of the subpixel convolution layer will serve as the input of two completely different network layers, one is to reconstruct the residual information at the same level, and the other is used for feature extraction of the next layer. In order to effectively reduce the computational complexity, parameters of different layers are shared. For the data reconstruction branch, it includes a subpixel convolution layer and two ordinary convolution layers. The input data is upsampled by the subpixel convolution layer. It should be noted that, through experiments, it is found that the convolution layer must be passed before the subpixel layer, otherwise the training process will be difficult to converge; the other convolution layer uses the extracted features as input and the high-resolution data reconstructed by the subpixel layer as output. The generated high-resolution data can be used as an intermediate result for reconstruction of the next level. Using the summary function of the torchkeras module, the structural information of the MSRLapN neural network of the present invention is viewed as shown in Table 3.

[0063] The parameters of the model in Table 3 are: 682062

[0064]

[0065]

[0066] For the image reconstruction branch in the network model, the present invention uses deconvolution operations to improve the resolution of ground meteorological elements (ConvTranspose2d-18, ConvTranspose2d-36 and ConvTranspose2d-54), and uses bilinear upsampling to improve the resolution of pressure layer elements (Upsample-18, Upsample-36 and Upsample-54). The meteorological elements with improved resolution obtain a higher peak signal-to-noise ratio.

[0067] The iterative training process involves adjusting hyperparameters to converge the error between the model output data and the true value for each factor, thereby obtaining the optimal model solution for that factor. The hyperparameters involved in this invention are shown in Table 4.

[0068] Table 4 U-Net model training hyperparameters

[0069]

[0070]

[0071] The present invention implements the super-resolution downscaling function using the MSRLapN neural network. After model training is completed, different elements are assigned a set of model parameters. For a particular element, a script is run, with low-resolution (0.5° spatial resolution) multi-model integrated forecast data as input and a refined forecast data product (0.045° spatial resolution) for that element as output. The model data with a spatial resolution of 0.5° is first interpolated onto a 0.36° grid so that the data grid can reach a spatial resolution of 0.045° after 8x super-resolution downscaling using the MSRLapN neural network model. Finally, a set of refined model forecast data is obtained.

[0072] S130: For characteristic elements of precipitation and cloud cover, a graded loss function is used to calculate the characteristic elements at different grades.

[0073] Optionally, in some embodiments, the loss function (Loss Function), also known as the error function or cost function, is a core concept in machine learning. It is mainly used to measure how well the algorithm fits the data, that is, to evaluate the degree of inconsistency between the predicted value of the model and the true value. The loss function is a non-negative real-valued function, usually expressed as L(Y, f(x)), where Y represents the true value and f(x) represents the predicted value of the model. For characteristic elements such as precipitation, low cloud cover and total cloud cover, the present invention introduces the classification standards for testing precipitation and cloud cover into the setting of the loss function, so that the training of the neural network model is more suitable for multi-mode integration and correction of meteorological elements.

[0074] (1) The loss function of precipitation elements is set up, and the 3-hour cumulative precipitation is divided into four levels: light rain, moderate rain, heavy rain and rainstorm according to the precipitation classification table 5.

[0075] Table 5 Precipitation classification table

[0076] grade light rain moderate rain heavy rain rainstorm Threshold (mm) 0.1-10 10-25 25-50 >50

[0077] The loss function of precipitation elements is set as:

[0078] y Loss_TRP =w1·MSE1+w2·MSE2+w3·MSE3+w4·MSE4

[0079] MSE1, MSE2, MSE3 and MSE4 are the mean squared errors (MSE) between the forecast values and the actual values of light rain, moderate rain, heavy rain and torrential rain, respectively. W1, W2, W3 and w4 are the corresponding weights of the errors, w1+w2+W3+w4=1.

[0080] The loss function of cloud cover factor is set. For total cloud cover, low cloud cover and other factors, they are classified into sunny with few clouds (cloud cover 0-40%), cloudy (50-70%) and overcast (80-100%), and cloud cover classification table 6.

[0081] Table 6 Cloud cover classification table

[0082] grade Sunny with few clouds partly cloudy cloudy day Threshold (%) 0-40 50-70 80-100

[0083] The loss function of the cloud cover element is set as:

[0084] y Loss_TFP =w1·MSE1+w2·MSE2+w3·MSE3

[0085] MSE1, MSE2, and MSE3 are the mean squared errors (MSE) between the forecast values and the actual values for sunny, cloudy, and overcast days, respectively. w1, W2, and w3 are the corresponding weights of the errors, w1+w2+w3=1.

[0086] To design a reasonable loss function for characteristic factors, the value of the weight w needs to be adjusted according to the actual situation. Taking precipitation as an example, if there is no heavy rain in winter, w4 will be set to zero. In summer, heavy rain disasters occur frequently, so w4 should be appropriately increased to ensure accurate forecast of heavy rain disasters more accurately.

[0087] The above is an introduction to the method embodiment. The following further illustrates the solution of the present invention through a system embodiment.

[0088] Figure 5FIG. 2 shows a block diagram of an atmospheric model forecast data integrated post-processing system 200 based on artificial intelligence according to an embodiment of the present invention. Figure 2 As shown, the system 200 includes:

[0089] A first model building module 210 is configured to construct a dataset by factoring, establish a U-Net neural network model for multi-mode integrated forecasting, adjust hyperparameters of the U-Net neural network model, and train the model using the dataset to obtain a trained U-Net neural network model;

[0090] A second model construction module 220 is used to construct a data set for training the MSRLapN neural network model, construct the structure of the deep super-resolution downscaling MSRLapN neural network model by elements, adjust the hyperparameters of the MSRLapN neural network model, and train it using the data set to obtain a trained MSRLapN neural network model;

[0091] The loss function design module 230 is used to calculate the characteristic elements of precipitation and cloud cover at different levels using a graded loss function.

[0092] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the described module can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0093] The above is an introduction to the method embodiment. The following is a further explanation of the solution of the present invention through business operation and accuracy verification embodiments.

[0094] Operational operation of multi-model integrated forecasting and deep super-resolution downscaling processes, such as Figure 6 As shown;

[0095] Forecast data reported from 00:00 on July 1, 2023, to 00:00 on July 24, 2023, were selected as the training set, with 600 training samples. Forecast data reported from 00:00 on July 25, 2023, to 00:00 on July 28, 2023, were selected as the testing set, with 125 testing samples. The training to testing sample ratio was approximately 8:2. The trained model parameters were used to produce refined forecast data products for July 2023, and these refined forecast products were tested using station observations.

[0096] The observation data from ground and sounding stations are used as the verification source, and the test results of the refined forecast data products of the test set are as follows.

[0097] 2m relative humidity test results, Figure 7 0-24h test results are shown; Figure 824-48h test results are shown; Figure 9 Shown are the 48-72h test results; Figure 10 0-24h test results are shown; Figure 11 24-48h test results are shown; Figure 12 Results of 48-72h assays are shown.

[0098] From the above test results, it can be seen that the refined forecast products output by the comprehensive post-processing neural network of model forecast data have higher spatial resolution and higher accuracy than the five sets of input model forecast data products.

[0099] It should be understood that the various forms of the processes shown above can be used to reorder, add, or delete steps. For example, the steps described in the present invention can be performed in parallel, sequentially, or in a different order, as long as the desired results of the technical solutions disclosed in the present invention can be achieved. This is not limited herein.

[0100] The above specific embodiments do not limit the scope of protection of the present invention. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention are intended to be included within the scope of protection of the present invention.

Claims

1. A comprehensive post-processing method for atmospheric model forecast data based on artificial intelligence, characterized in that: include: By constructing a dataset by elements, a U-Net neural network model for multi-mode integrated forecasting is established, the hyperparameters of the U-Net neural network model are adjusted, and the dataset is used for training to obtain a trained U-Net neural network model; Construct a dataset for training the MSRLapN neural network model, construct the structure of the MSRLapN neural network model for deep super-resolution downscaling by elements, adjust the hyperparameters of the MSRLapN neural network model, and use the dataset for training to obtain the trained MSRLapN neural network model; For the characteristic elements of precipitation and cloud cover, a graded loss function is used to calculate the characteristic elements at different levels.

2. The method according to claim 1, characterized in that The elements of the sub-element construction data set include: 2m temperature, 2m humidity, 10m meridional wind, 10m zonal wind, surface air pressure, total precipitation and low cloud cover surface meteorological elements and 100, 150, 200, 250, 300, 400, 500, 600 and 700hPa pressure layer elements of temperature, humidity, meridional wind and zonal wind; The method of constructing a dataset by element-by-element includes: unifying the horizontal resolution of atmospheric model forecast data and true value data by bilinear interpolation; unifying the vertical resolution of atmospheric model forecast data and true value data by cubic spline interpolation; performing Z-Score normalization on the input and true value of training and test data respectively; and storing the constructed dataset in NumPy binary file format with the file suffix "npy".

3. The method according to claim 1, characterized in that The U-Net neural network model established for multi-mode integrated forecasting includes: an input layer, 4 downsampling layers, 4 upsampling layers and an output layer, wherein downsampling is achieved through a maximum pooling operation, upsampling is achieved through a deconvolution operation, and downsampling and upsampling are directly connected through a shortcut connection.

4. The method according to claim 1, wherein The adjustment of the hyperparameters of the U-Net neural network model includes training for different elements. By adjusting the hyperparameters, the error between the model output data and the true value converges to obtain the optimal model solution for the element. After the model training is completed, different elements respectively obtain a set of model parameters. For a certain element, an execution script is run, with the input being 5 sets of atmospheric model forecast data and the output being the integrated result of the element model forecast data.

5. The method according to claim 1, wherein The elements of the data set for training the MSRLapN neural network model include: 2m temperature, 2m humidity, 10m meridional wind, 10m zonal wind, surface air pressure, total precipitation and low cloud cover surface meteorological elements, and temperature, humidity, meridional wind and zonal wind pressure layer elements of 100, 150, 200, 250, 300, 400, 500, 600 and 700hPa pressure layers; The dataset for training the MSRLapN neural network model is constructed, including: Using the bilinear interpolation method, the 5km three-dimensional atmospheric real-time fusion grid data with a horizontal resolution of 0.045° was interpolated into low-resolution grid data with spatial resolutions of 0.09°, 0.18°, and 0.36°, respectively. The data were normalized and the constructed dataset was stored in the NumPy binary file format with the file suffix npy.

6. The method according to claim 1, characterized in that The structure of the deep super-resolution downscaling MSRLapN neural network model includes: an image reconstruction branch and a feature extraction branch; At each level of the feature extraction branch, there is a MSRB and a subpixel convolution layer. The MSRB performs multi-scale feature extraction, and the subpixel convolution layer has an upsampling scale of 2. The MSRB includes a convolution layer with a convolution kernel size of 1×1. The output of the subpixel convolution layer is used as input for two different network layers: one to reconstruct residual information at the same level, and the other for feature extraction at the next level. The data reconstruction branch consists of a subpixel convolution layer and two ordinary convolution layers. The input data is upsampled by the subpixel convolution layer; it passes through the convolution layer before the subpixel layer; the other convolution layer takes the extracted features as input and outputs the high-resolution data reconstructed by the subpixel layer; the generated high-resolution data is used as an intermediate result for reconstruction at the next level; In the image reconstruction branch of the deep super-resolution downscaling MSRLapN neural network model, the ground meteorological elements are improved in resolution using deconvolution operations, and the pressure layer elements are improved in resolution using bilinear upsampling, so that the meteorological elements with improved resolution obtain a higher peak signal-to-noise ratio.

7. The method according to claim 1, characterized in that The adjustment of the hyperparameters of the MSRLapN neural network model includes: for training different elements, it is necessary to adjust the hyperparameters so that the error between the model output data and the true value converges to obtain the optimal model solution for the element; The MSRLapN neural network is executed to realize the super-resolution downscaling function. After the model training is completed, different elements obtain a set of model parameters. For a certain element, the execution script is run, the multi-model integrated forecast data with a spatial resolution of 0.5° is input, and the spatial resolution of the element is output as 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 fineness of 0.045° spatial resolution after 8-fold super-resolution downscaling of the MSRLapN neural network model, and finally obtain the model forecast data.

8. The method according to claim 1, characterized in that The loss function for calculating the characteristic elements at different levels includes: The loss function is used to measure how well the algorithm fits the data, that is, to evaluate the degree of inconsistency between the predicted value of the model and the true value; the grading standards for testing precipitation elements and cloud cover elements are introduced into the setting of the loss function, making the training of the neural network model suitable for multi-mode integration and correction of meteorological elements.

9. The method according to claim 8, characterized in that The loss function of the precipitation element is set to divide the 3-hour cumulative precipitation into 4 levels: light rain, moderate rain, heavy rain and rainstorm according to the precipitation classification table; The loss function of precipitation elements is set as: <h2 style=";text-align:left;direction:ltr">y<h2 style=";text-align:left;direction:ltr"> Loss_TRP <h2 style=";text-align:left;direction:ltr"> = w1 MSE1 + w2 MSE2 + w3 MSE3 + w4 MSE4 MSE1, MSE2, MSE3 and MSE4 are the mean square errors (MSEs) between the forecast values and the actual values for light rain, moderate rain, heavy rain and torrential rain, respectively. w1, w2, w3 and w4 are the corresponding weights of the errors, w1+w2+w3+w4=1. The loss function of the loss cloud cover element is set as follows: for total cloud cover and low cloud cover elements, the cloud cover for sunny days is 0-40%, for cloudy days is 50-70%, and for cloudy days is 80-10%. The loss function of the cloud cover element is set as follows: <h2 style=";text-align:left;direction:ltr">y<h2 style=";text-align:left;direction:ltr"> Loss_TFP <h2 style=";text-align:left;direction:ltr"> = w1 MSE1 + w2 MSE2 + w3 MSE3 MSE1, MSE2, and MSE3 are the mean square errors (MSEs) between the forecast values and the actual values for sunny, partly cloudy, and overcast days, respectively. w1, w2, and w3 are the corresponding weights of the errors, w1+w2+w3=1.

10. An artificial intelligence-based comprehensive post-processing system for atmospheric model forecast data, characterized in that: include: The first model construction module is used to construct a data set by dividing the elements, establish a U-Net neural network model for multi-mode integrated forecasting, adjust the hyperparameters of the U-Net neural network model, and train it using the data set to obtain a trained U-Net neural network model; The second model construction module is used to construct a data set for training the MSRLapN neural network model, construct the structure of the deep super-resolution downscaling MSRLapN neural network model by elements, adjust the hyperparameters of the MSRLapN neural network model, and train it using the data set to obtain a trained MSRLapN neural network model; The loss function design module is used to calculate the characteristic elements of precipitation and cloud cover at different levels using a graded loss function.

Images