ECMWF mode element deviation correction method based on AttUnet

Abstract

The invention provides an ECMWF mode element deviation correction method based on AttUnet, and belongs to the field of weather forecast. According to the method, a deep convolutional neural network is used for outputting four earth surface elements including 2m temperature, earth surface air pressure, 2m specific humidity and 10m wind in an ECMWF mode, and deviation correction is carried out on an ECMWF mode output result with the resolution of 0.125 degrees. According to the method, firstly, a feature library is constructed according to ECMWF mode output elements, on this basis, an XgBoost tool is used for carrying out feature analysis on a sample library, the importance of the sample library is sorted, meanwhile, factors are further screened in combination with artificial experience, and terrain information is considered. And then prediction and deviation correction are carried out by using a deep convolutional neural network to obtain a more accurate 0.125-degree grid product. Meanwhile, in the model debugging stage, the model has good forecasting performance on extreme disastrous weather by optimizing the learning rate and the loss function.

Description

Technical Field

[0001] The present invention application relates to the field of meteorological data processing, and particularly relates to a method for correcting ECMWF model element deviations based on AttUnet. Background Art

[0002] Both the inherent errors of meteorological models and the initial value errors are important factors causing deviations in weather forecasts, and these errors will have different manifestations in different regions, seasons, and weather processes. Therefore, at least one error correction is required before actual forecasting.

[0003] Traditional error correction methods have many deficiencies, such as complex processes, low computational efficiency, and unsatisfactory correction effects. In addition, the accuracy of the subjective correction method by forecasters is also limited by personal knowledge and experience. In order to improve the accuracy of model interpretation and reduce the workload of subjective correction, developing an objective and better model interpretation algorithm has become a relatively urgent matter in current operations. Existing model interpretation methods can be roughly divided into the perfect prognostic (PP) method or the model output statistics (MOS) method for predicting meteorological element values. However, these methods are based on a multiple linear regression model and cannot reasonably describe the non-linear relationship between meteorological elements, geographical elements, and the prediction object. Therefore, the accuracy of meteorological element prediction is relatively poor. The former completes the correction by establishing a linear or simple non-linear statistical model between the observed value and the model forecast value, while the latter realizes it by establishing a linear or simple non-linear statistical model between the observed value and a set of related numerical forecast estimates of atmospheric variables. Most of these interpretation methods are multiple linear regression models, which cannot reasonably describe the non-linear relationship between meteorological elements, geographical elements, and the forecast object, have poor adaptability, and the accuracy of element prediction needs to be improved. With the continuous development of numerical weather forecasting and its interpretation application technology, using deep learning technology to mine multi-dimensional and multi-source information in meteorological big data and developing an interpretation method of artificial intelligence technology to further improve the refined forecasting level of meteorological elements is the development trend of the application of numerical forecast products in recent years. Summary of the Invention

[0004] The present invention application provides a method for correcting ECMWF model element deviations based on AttUnet to solve the technical problem of how to improve the accuracy of meteorological element deviation correction.

[0005] To solve the above technical problem, the present invention application provides a method for correcting ECMWF model element deviations based on AttUnet, including:

[0006] Obtain historical model forecast data, label data, and geographic static data; among them, the historical model forecast data is ECMWF model 0-72-hour forecast data; the label data is HRCLDAS data, and the label data includes four meteorological elements: two-meter temperature element, two-meter specific humidity element, surface air pressure, and ten-meter wind element; the geographic static data is DEM data;

[0007] Based on the ECMWF model data, preliminarily determine relevant characteristic factors and construct a preselected characteristic factor library; among them, the characteristic factors include geopotential height at each vertical level, zonal wind at each vertical level, meridional wind at each vertical level, vertical velocity at each vertical level, divergence at each vertical level, vorticity at each vertical level, specific humidity at each vertical level, temperature at each vertical level, 2-meter temperature, 2-meter specific humidity, 2-meter dew point temperature, ten-meter u-wind component, ten-meter v-wind component, surface air pressure, ground temperature, surface albedo, atmospheric column water content, total cloud cover, and precipitation;

[0008] Use the XgBoost tool to rank the characteristic factors of the four meteorological elements according to the significance analysis scoring requirements for the characteristic factor library, and screen to obtain characteristics that meet the scoring requirements;

[0009] Combined with physical laws and the screening operations input by forecasters, further screen the characteristics that meet the scoring requirements to obtain the element correction characteristics of the four meteorological elements;

[0010] According to the above element correction characteristics, process the ECMWF model data and HRCLDAS data into the required data set, and screen the data samples at each time of the data set to eliminate the sample data with data quality not meeting the requirements, and obtain the elimination result data; and perform normalization processing on the elimination result data to obtain the standardized result;

[0011] Use the standardized result to train a deep learning model based on the AttUnet architecture to obtain a target bias correction model;

[0012] Use the target bias correction model to correct the bias of the four meteorological elements output by the model.

[0013] As an optimal solution, after performing normalization processing on the data set, through correlation analysis, eliminate the samples with a deviation between ECMWF data and HRCLDAS data greater than the preset deviation threshold.

[0014] Compared with the prior art, the present invention application has the following beneficial effects:

[0015] The present invention application provides a method for correcting ECMWF model element deviations based on AttUnet. Starting from ECMWF forecast data and HRCLDAS data, the present invention application establishes a method and device for correcting surface elements based on ECMWF forecast data and deep learning methods. This method fully explores the non-linear relationship of model deviations in historical big data samples, and through training a deep learning model, learns the non-linear relationship between meteorological elements, geographical elements and forecast objects, so that the model deviations are better corrected.

[0016] In addition, when training the model, since extreme weather belongs to small samples and the ECMWF forecast effect is poor, without reinforcement, the trained deep learning model cannot predict extreme weather well, and even shows a worse effect than the original EC forecast. Therefore, this patent application designs special loss functions (using an exponential form loss function for 2m specific humidity to increase the extreme value weight, and using a hyperbolic function for 10m meridional wind / zonal wind to increase the extreme value weight) to strengthen extreme weather cases, so as to achieve better performance in forecasting extreme weather. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 : A schematic flow chart of an embodiment of the method for correcting ECMWF model element deviations based on AttUnet provided by the present invention application.

[0018] Figure 2 : A schematic principle diagram of an application example of the XGBoost model provided by the present invention application.

[0019] Figure 3 : A bar chart of the correlation coefficient of an application example of 2m temperature provided by the present invention application.

[0020] Figure 4 : A schematic principle diagram of an embodiment of AttUnet provided by the present invention application.

[0021] Figure 5 : Provided by the present invention application Figure 4 Symbol schematic diagram.

[0022] Figure 6 : A schematic flow chart of another embodiment of the method for correcting ECMWF model element deviations based on AttUnet provided by the present invention application.

[0023] Figure 7 : A schematic flow chart of an embodiment of the method for predicting four elements of the ECMWF model based on AttUnet provided by the present invention application.

[0024] Figure 8: It is a schematic structural diagram of an embodiment of the ECMWF model four-factor prediction device based on AttUnet provided by this invention application. Detailed implementation manners

[0025] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without making creative efforts belong to the scope of protection of the present invention.

[0026] Embodiment 1:

[0027] As Figure 1 shown, Embodiment 1 of this invention application provides an application example of a four-factor deviation correction method based on AttUnet for the ECMWF model. Each step is described in detail as follows:

[0028] Step 1: Collect HRCLDAS reanalysis (1km * 1km) data and DEM data (90m * 90m) and process them into area-averaged (12.5km * 12.5km) grid data as the label data set. The specific data processing method is as follows:

[0029] (1) Process the HRCLDAS 1km data into 500m resolution, and then use the area-averaging method to count to 12.5km * 12.5km. The calculation formula is as follows:

[0030]

[0031] Where: t (i,j) is the original HRCLDAS reanalysis data (0.01 degree), T (i,j) is the processed 12.5km resolution label data set, where i and j respectively represent the subscripts of the grid row and column numbers;

[0032] (2) First, interpolate the downloaded DEM data (data source: Geospatial Data Cloud) to 0.01 degree resolution using bilinear interpolation, and then use the area-averaging method to count to 12.5km resolution. The calculation formula is as follows:

[0033]

[0034] Where: t (i,j) is the original DEM data (0.001 degree), T (i,j) is the processed 12.5km resolution DEM data set, and i and j respectively represent the subscripts of the grid row and column numbers.

[0035] Step 2: According to meteorological professional knowledge, initially determine the input features related to each correction factor and construct a factor library. Then process the ECMWF data for one year to construct a feature screening data set. The initial factor library is as follows: Select the factors that may be used to construct a preliminary factor library. The factors are as follows: Geopotential height (200hPa, 500hPa, 700hPa, 850hPa, 1000hPa), zonal wind (200hPa, 500hPa, 700hPa, 850hPa, 1000hPa), meridional wind (200hPa, 500hPa, 700hPa, 850hPa, 1000hPa), vertical velocity (200hPa, 500hPa, 700hPa, 850hPa, 1000hPa), divergence (200hPa, 500hPa, 700hPa, 850hPa, 1000hPa), vorticity (200hPa, 500hPa, 700hPa, 850hPa, 1000hPa), specific humidity (200hPa, 500hPa, 700hPa, 850hPa, 1000hPa), temperature (200hPa, 500hPa, 700hPa, 850hPa, 1000hPa), 2m temperature, 2m specific humidity, 2m dew point temperature, 10m_U, 10m_V, surface pressure, surface temperature, surface albedo, atmospheric column water content, total cloud cover, precipitation.

[0036] Step 3: Based on the above preliminary factor library for one year, randomly select 500,000 grid point sample data to construct a training data set for the tree model. Then construct a simple XGBoost model, as Figure 2 shown.

[0037] Use the above constructed XGBoost model for training. During the training process, the model will calculate the information gain of each feature and use it as the label when splitting nodes. Since the node splitting order is directly related to the model training effect, it can be used as the feature importance ranking. Adjust the learning rate, number of training rounds, maximum depth, regularization, etc. After sufficient training, obtain the feature importance ranking.

[0038] Step 4: According to the feature ranking selected by XGBoost, with the help of physical laws and the subjective experience of forecasters, further screen the features to obtain the final factor correction features, such as the correction features of the two-meter temperature factor, two-meter specific humidity factor, surface pressure factor, ten-meter u-wind component factor, and ten-meter v-wind component factor.

[0039] For example, as shown in the following table:

[0040]

[0041]

[0042] Table 1 Meteorological elements and corresponding characteristic factors

[0043] Step 5: Process the collected ECMWF data, pre-processed HRCLDAS and DEM data for several years into sample data sets, and then remove samples with poor quality (including large model forecast deviations, label data errors, and poor label data quality) based on data quality. The main removal criteria are as follows:

[0044] (1) Based on the historical extreme values of the observed data, the HRCLDAS reanalysis data that exceeded the historical extreme value range were eliminated; the correlation coefficient of the HRCLDAS data of two adjacent time periods was calculated, and the data with a correlation coefficient less than 0.9 were eliminated (the spatial texture structure was very different and was generally suspected to be erroneous data).

[0045] (2) Perform correlation analysis on the forecasted revised meteorological elements and the label data set, and remove samples with low correlation coefficients (poor forecast quality). Take 2m temperature as an example. Figure 3 shown.

[0046] Finally, the labeled dataset is divided into training set, validation set and test set according to the ratio of 7:1.5:1.5.

[0047] For the removed label data set, the Min-Max normalization method can be used for further processing:

[0048] x a =(x i -min(x)) / (max(x)-min(x));

[0049] In the formula, min(x) represents the minimum value in the data set before standardization, max(x) represents the maximum value in the data set before standardization, and x i represents the i-th element in the data set before standardization, x a Represents the value after normalization.

[0050] Step 6: Select the AttUnet model as the basic model, such as Figures 4 to 5 As shown in the figure, the Att-Unet model network is a deep learning model with a U-shaped structure. The difference from Unet is that Unet only performs splicing operations at the same level of the decoder and encoder, while AttUnet adds an AG structure based on the Unet structure. This structure has the characteristic of weighting part of the encoding of the encoder at the same level. The AG structure of the AttUnet model also has the characteristic of attention, and the activation value can be adjusted by automatically learning parameters.

[0051] When using the AttUnet network model for training, the learning rate is adjusted exponentially to enable the model to converge to the optimal value earlier. The early stopping method is used to control the number of model iterations and avoid overfitting of the model.

[0052] Step 7: Use RMSE to evaluate the prediction results, and use POD, FAR, and CSI to evaluate the accuracy of extreme weather forecasts. According to the four evaluation metrics, adjust the model parameters and loss function to optimize the model performance. Finally, the loss functions for each variable are determined as follows:

[0053] Loss functions for 2m temperature and surface pressure:

[0054]

[0055] In the formula, MSE is the loss function of the 2m temperature element or surface pressure element, and y i represents the true value of the i-th characteristic factor of the 2m temperature element or surface pressure element, represents the predicted value of the i-th characteristic factor of the 2m temperature element or surface pressure element.

[0056] The loss function for 2m specific humidity is:

[0057]

[0058] In the formula, loss represents the loss function of the 2m specific humidity element, and y i represents the true value of the i-th characteristic factor of the 2m specific humidity element, represents the predicted value of the i-th characteristic factor of the 2m specific humidity element.

[0059] The loss functions for 10m_u and 10m_v are:

[0060]

[0061] In the formula, loss represents the loss function of the 10m u-wind component element or 10m v-wind component element, and y i represents the true value of the i-th characteristic factor of the 10m u-wind component element or 10m v-wind component element, represents the predicted value of the i-th characteristic factor of the 10m u-wind component element or 10m v-wind component element.

[0062] Step 8: Based on the real-time ECMWF model output data, using the trained model parameters above, the correction of 2m temperature, surface pressure, 2m humidity, and 10m wind can be achieved.

[0063] Example 2:

[0064] Please refer to Figure 6 , Figure 6A method for correcting ECMWF model element deviations based on AttUnet provided for this invention application, including steps S101 to S107; wherein,

[0065] Step S101, obtain historical model forecast data, label data, and geographical static data; wherein, the historical model forecast data is: ECMWF model 0-72 hour forecast data; the label data is HRCLDAS data, and the label data includes four meteorological elements: two-meter temperature element T 2m , two-meter specific humidity element Q 2m , surface air pressure SP, and ten-meter wind element, and the ten-meter wind element specifically includes ten-meter u-wind component element U 10 and / or ten-meter v-wind component element V 10 ; the geographical static data is DEM data.

[0066] In this step, ECMWF model data refers to the European Centre for Medium-Range Weather Forecasts, abbreviated as ECMWF, and is translated into Chinese as the European Centre for Medium-Range Weather Forecasts model data.

[0067] HRCLDAS refers to the High Resolution China Meteorological Administration Land Data Assimilation System, abbreviated as HRCLDAS, and is translated into Chinese as the High Resolution Land Data Assimilation System.

[0068] DEM refers to the Digital Elevation Model, abbreviated as DEM, and is translated into Chinese as the Digital Elevation Model.

[0069] In a preferred implementation manner, before obtaining the historical model forecast data, label data, and geographical static data, it further includes: through resolution processing, averaging the HRCLDAS data and DEM data onto a 12.5 km regional grid.

[0070] Specifically, the calculation formula for performing the resolution processing on the HRCLDAS data is specifically:

[0071]

[0072] wherein, t (i,j) is the initial data of the HRCLDAS data, T (i,j)The HRCLDAS data after the resolution processing, where i represents the subscript of the row number of the grid and j represents the subscript of the column number of the grid. Alternatively, bilinear interpolation is used to interpolate the HRCLDAS data to a resolution of 0.01 degrees, and then the method of area averaging is used to statistically calculate it to a resolution of 12.5 km.

[0073] The specific calculation formula for the resolution processing of the DEM data is as follows:

[0074]

[0075] where t (i,j) is the initial DEM data with a resolution of 0.001 degrees, and T (i,j) is the DEM data after the resolution processing, where i represents the subscript of the row number of the grid and j represents the subscript of the column number of the grid.

[0076] Step S102: Initially determine the characteristic factors according to the ECMWF model data and construct a preselected characteristic factor library.

[0077] The characteristic factors include geopotential height at each vertical level (200 hPa, 500 hPa, 700 hPa, 850 hPa, 1000 hPa), zonal wind at each vertical level (200 hPa, 500 hPa, 700 hPa, 850 hPa, 1000 hPa), meridional wind at each vertical level (200 hPa, 500 hPa, 700 hPa, 850 hPa, 1000 hPa), vertical velocity at each vertical level (200 hPa, 500 hPa, 700 hPa, 850 hPa, 1000 hPa), divergence at each vertical level (200 hPa, 500 hPa, 700 hPa, 850 hPa, 1000 hPa), vorticity at each vertical level (200 hPa, 500 hPa, 700 hPa, 850 hPa, 1000 hPa), specific humidity at each vertical level (200 hPa, 500 hPa, 700 hPa, 850 hPa, 1000 hPa), temperature at each vertical level (200 hPa, 500 hPa, 700 hPa, 850 hPa, 1000 hPa), 2m temperature, 2m specific humidity, 2m dew point temperature, 10m_U (ten-meter u-wind component), 10m_V (ten-meter v-wind component), surface air pressure, ground temperature, surface albedo, atmospheric column water content, total cloud cover, and precipitation, etc.

[0078] Step S103: Use the XgBoost tool to rank the characteristic factors of the four meteorological elements according to the significance analysis scoring requirements, and screen out the characteristics that meet the scoring requirements.

[0079] For example, through the XGBoost model, the information gain or factor importance of each factor can be calculated and sorted, and several factors with the largest information gain or the highest factor importance are used as the characteristic factors of the meteorological elements.

[0080] When training the XGBoost model, the information gain of each factor can be used as the annotation during node splitting (because the node splitting order is directly and closely related to the model training effect). At the same time, the learning rate, number of training rounds, maximum depth, regularization, etc. of the XGBoost model can be adjusted to ensure that the performance of the XGBoost model meets the requirements.

[0081] Step S104: Further screen the features that meet the requirements by combining physical laws and the screening operations input by the forecaster to obtain the element correction features of the four meteorological elements.

[0082] For example, for each meteorological element, the corresponding characteristic factors are shown in the following table:

[0083]

[0084]

[0085] Table 1 Meteorological elements and corresponding characteristic factors

[0086] The obtained element correction features of the four meteorological elements are specifically:

[0087] The characteristic factors corresponding to the two-meter temperature element are ground temperature, surface albedo, total water content in the atmospheric column, two-meter temperature characteristic factor, and DEM;

[0088] The characteristic factors corresponding to the two-meter specific humidity element are 850 hPa specific humidity, two-meter specific humidity characteristic factor, and DEM;

[0089] The characteristic factors corresponding to the surface air pressure element are 500 hPa geopotential height, 850 hPa geopotential height, SP, surface air pressure characteristic factor, and DEM;

[0090] The characteristic factors corresponding to the ten-meter wind element are 850 hPa meridional wind, 850 hPa zonal wind, surface air pressure characteristic factor, ten-meter u-wind component characteristic factor, ten-meter v-wind component characteristic factor, and DEM.

[0091] Step S105: According to the above element correction features, process the ECMWF model data and HRCLDAS data into the required data set, screen the data samples of each time step in the data set, remove the sample data with data quality not meeting the requirements to obtain the removed result data; and perform normalization processing on the removed result data to obtain the standardized result.

[0092] In this step, the HRCLDAS data with data quality not meeting the requirements is removed, specifically as follows:

[0093] By calculating the correlation coefficient between the HRCLDAS data at the previous and current times, the HRCLDAS data with a correlation coefficient less than 0.9 is removed;

[0094] The calculation formula of the correlation coefficient includes:

[0095]

[0096] where r is the calculation result of the correlation coefficient, x and y respectively represent the data at the previous and current times, represents the average value of a certain data set, represents the average value of another data set, i represents the number of the characteristic factor, and n is the total number of the characteristic factors.

[0097] In another preferred embodiment, after the normalization processing of the data set, through correlation analysis, the samples with a deviation between the ECMWF data and the HRCLDAS data greater than the preset deviation threshold are removed.

[0098] Before or after the normalization processing of the data set, the data outside the preset historical extreme value range can be removed.

[0099] In a preferred implementation scheme, the normalization processing of the data set is specifically as follows:

[0100] The data set is standardized according to the following formula:

[0101] x a =(x i -min(x)) / (max(x)-min(x));

[0102] In the formula, min(x) represents the minimum value in the data set before the standardization processing, max(x) represents the maximum value in the data set before the standardization processing, x i represents the i-th element in the data set before the standardization processing, and x a represents the value after the standardization processing.

[0103] Step S106: Using the standardization result, train a deep learning model based on the AttUnet architecture to obtain a target deviation correction model.

[0104] In this step, a deep learning model based on the AttUnet architecture is adopted, such as Figure 4As shown, the Att-Unet model network is a deep learning model with a U-shaped structure. The difference from Unet is that Unet only performs concatenation operations at the same level of the decoder and encoder, while AttUnet adds an AG structure to the Unet structure. This structure has the characteristic of weighting some encodings of the encoder at the same level. The AG structure of the AttUnet model also has an attention characteristic, which can adjust the activation value by automatically learning parameters. In the figure, conv represents convolution, Relu and sigmoid are function names. Among them, Relu refers to the Rectified Linear Unit function (abbreviated as Relu), and sigmoid is often used as an activation function for neural networks due to its properties such as monotonic increase and monotonic increase of the inverse function. Multiply represents multiplication operation, skip connection represents skip connection, also known as Residual Connection, which is an important component in the deep neural network architecture. Its basic idea is to directly connect the input to the output in some layers of the network to allow information to jump between different layers. This connection is usually achieved through an addition operation, adding the input and the output; Up Sampling represents upsampling.

[0105] Furthermore, for Figure 4 the meanings of the arrows and symbols in Figure 5 as shown, where conv Layer represents the convolutional layer, Pool Layer represents the pooling layer, copy and Crop refers to concatenation (generally used for feature fusion), subpixellayer refers to the sub-pixel layer or sub-pixel layer, and Attention Gate refers to the attention gate.

[0106] Preferably, before training the deep learning model based on the AttUnet architecture, the normalization results can be divided into a training set, a validation set, and a test set according to a ratio of 7:1.5:1.5.

[0107] Training the deep learning model based on the AttUnet architecture to obtain a target bias correction model specifically includes:

[0108] When training the deep learning model, the early stopping method and simulated annealing strategy are used to optimize the model training parameters. RMSE is used to evaluate the output of the deep learning model. At the same time, different loss functions are designed according to different element features, and POD, CSI, and FAR are used to evaluate the extreme weather of the output of the deep learning model. According to the results of multiple evaluation metrics, the loss function and learning rate of the deep learning model are adjusted and optimized to obtain the target bias correction model.

[0109] In a preferred embodiment, the convergence condition of the deep learning model can be set according to the loss functions of the two-meter temperature element, the two-meter specific humidity element, the surface air pressure element, the ten-meter u-wind component element, and the ten-meter v-wind component element. At the same time, evaluation indicators such as POD, FAR, and CSI can also be used to evaluate the predicted values to determine whether the performance of the prediction model meets the requirements.

[0110] For example, the two-meter temperature element and the surface air pressure element can adopt the following loss functions:

[0111]

[0112] In the formula, MSE is the loss function of the two-meter temperature element or the surface air pressure element, and y i represents the true value of the i-th characteristic factor of the two-meter temperature element or the surface air pressure element, represents the predicted value of the i-th characteristic factor of the two-meter temperature element or the surface air pressure element.

[0113] The two-meter specific humidity element can adopt the following loss function:

[0114]

[0115] In the formula, loss represents the loss function of the two-meter specific humidity element, and y i represents the true value of the i-th characteristic factor of the two-meter specific humidity element, represents the predicted value of the i-th characteristic factor of the two-meter specific humidity element.

[0116] The ten-meter u-wind component element and the ten-meter v-wind component element can adopt the following loss functions:

[0117]

[0118] In the formula, loss represents the loss function of the ten-meter u-wind component element or the ten-meter v-wind component element, and y i represents the true value of the i-th characteristic factor of the ten-meter u-wind component element or the ten-meter v-wind component element, represents the predicted value of the i-th characteristic factor of the ten-meter u-wind component element or the ten-meter v-wind component element.

[0119] As described above, the present application designs loss functions for the two-meter temperature element, the two-meter specific humidity element, the surface air pressure element, the ten-meter u-wind component element, and the ten-meter v-wind component element respectively. When the present application faces some extreme weather with small samples, it can strengthen these individual cases of extreme weather and effectively optimize the prediction effect.

[0120] Step S107: Use the target deviation correction model to correct the deviation of the four meteorological elements output by the model.

[0121] In this step, the meteorological four elements output by the target correction data pattern of the target product can be corrected for deviation through the output of the target deviation correction model.

[0122] Example 3:

[0123] Please refer to Figure 7 , Figure 7 A four-element prediction method of the ECMWF model based on AttUnet provided for this invention application, including step S201 to step S202; wherein,

[0124] Step S201, obtain the meteorological model output data of the prediction object in real time, and call the pre-trained prediction model.

[0125] In this embodiment, the prediction object can be a specific weather forecasting software or product. The meteorological model output data can be the meteorological model output data of ECMWF (European Centre for Medium-Range Weather Forecasts, abbreviated as ECMWF, Chinese translated as European Centre for Medium-Range Weather Forecasts).

[0126] In a preferred implementation manner, the ECMWF model element deviation correction method based on AttUnet in this embodiment can be applied to a computer device, which includes but is not limited to smart phones, laptop computers, tablet computers, and desktop computers.

[0127] The pre-trained prediction model can be called from a server or a data center. In this embodiment, the prediction model is called from a server or a data center and combined with the local meteorological model output data for the prediction in step S202.

[0128] Step S202, input the meteorological model output data into the prediction model, and based on the output of the prediction model, obtain the predicted values of at least one meteorological element.

[0129] Wherein, the prediction model is based on AttUnet, and the types of the meteorological elements include two-meter temperature element, two-meter specific humidity element, surface air pressure element, ten-meter u-wind component element, and ten-meter v-wind component element.

[0130] The prediction model is trained based on a number of first sample meteorological model data, second sample meteorological model data, and surface elevation data, and the data sources of the first sample meteorological model data and the second sample meteorological model data are different.

[0131] In one embodiment, the first sample meteorological model data may be derived from the above-mentioned ECMWF meteorological model output data. The second sample meteorological model data may be derived from HRCLDAS (High Resolution China Meteorological Administration Land Data Assimilation System, abbreviated as HRCLDAS, Chinese translation is High Resolution Land Data Assimilation System).

[0132] The surface elevation data may adopt a digital elevation model (DEM), which is a digital model that uses a regular grid to represent the surface elevation. It describes the altitude of the terrain surface through a set of ordered numerical arrays and is widely used in terrain analysis, hydrological simulation, urban planning and other fields.

[0133] Since the first sample meteorological model data and the second sample meteorological model data come from different sources, the two can be combined for verification to ensure the overall accuracy of the sample data set, improve the quality of the sample data, and thereby improve the performance of the prediction model.

[0134] In a preferred embodiment, the training method of the prediction model includes:

[0135] The first sample meteorological model data and the second meteorological model historical data of the same characteristic factors of each meteorological element are obtained, and the deviation value between the characteristic factors of the first sample meteorological model data and the second meteorological model historical data is calculated; according to the deviation value, the data in the second meteorological model historical data whose deviation value is greater than a preset threshold is eliminated, and the second sample meteorological model data is obtained according to the elimination result; according to the first sample meteorological model data, the second sample meteorological model data and the surface elevation data, a sample data set is constructed; based on the sample data set, a basic model is trained, and when the output of the basic model meets the preset convergence condition, the prediction model is obtained.

[0136] Exemplarily, the training of the basic model based on the sample data set is specifically: dividing the sample data set into a training set, a validation set and a test set according to a preset ratio (for example, 7:1.5:1.5), and using the test set to train the basic model, using the validation set to adjust the parameters of the model, and using the test set to evaluate the final performance of the model to verify the generalization ability of the model.

[0137] In a preferred embodiment, before obtaining the first sample meteorological model data and the second meteorological model historical data of the same characteristic factor, the ECMWF model element deviation correction method based on AttUnet includes:

[0138] Obtain the historical data of the first meteorological model;

[0139] Determine the characteristic factors of the meteorological elements according to the historical data of the first meteorological model;

[0140] Screen the historical data of the first meteorological model according to the characteristic factors to obtain the first sample meteorological model data.

[0141] Exemplarily, ECMWF data for one year can be obtained, and the factors that may be used are screened out, such as, but not limited to, geopotential height (200hPa, 500hPa, 700hPa, 850hPa, 1000hPa), zonal wind (200hPa, 500hPa, 700hPa, 850hPa, 1000hPa), meridional wind (200hPa, 500hPa, 700hPa, 850hPa, 1000hPa), vertical velocity (200hPa, 500hPa, 700hPa, 850hPa, 1000hPa), divergence (200hPa, 500hPa, 700hPa, 850hPa, 1000hPa), vorticity (200hPa, 500hPa, 700hPa, 850hPa, 1000hPa), specific humidity (200hPa, 500hPa, 700hPa, 850hPa, 1000hPa), temperature (200hPa, 500hPa, 700hPa, 850hPa, 1000hPa), 2m temperature, 2m specific humidity, 2m dew point temperature, 10m_U (ten-meter u-wind component), 10m_V (ten-meter v-wind component), surface pressure, surface temperature, surface albedo, atmospheric column water content, total cloud cover and precipitation, etc.

[0142] Further, the determining the characteristic factors of the meteorological elements according to the historical data of the first meteorological model includes: calculating the information gain of each factor of the historical data of the first meteorological model (about 500,000 grid point sample data) through a preset XGBoost model; determining the characteristic factors of the meteorological elements according to each information gain.

[0143] For example, through the XGBoost model, the information gain or factor importance of each factor can be calculated and sorted, and several factors with the largest information gain or the highest factor importance are used as the characteristic factors of the meteorological elements.

[0144] When the XGBoost model is trained, the information gain of each factor can be used as the criterion for node splitting (since the node splitting order is directly and closely related to the model training effect), and at the same time, the learning rate, number of training rounds, maximum depth, regularization, etc. of the XGBoost model can be adjusted to ensure that the performance of the XGBoost model meets the requirements.

[0145] According to the factors with higher information gain or importance selected, the final characteristic factors of meteorological elements can be obtained by combining or not combining physical laws and the subjective experience of forecasters.

[0146] For example, the meteorological element can be the two-meter temperature element T 2m and the two-meter specific humidity element Q 2m the surface air pressure element SP, the ten-meter u-wind component element U 10 and / or the ten-meter v-wind component element V 10 .

[0147] For each meteorological element, the corresponding characteristic factors are shown in the following table:

[0148]

[0149] Table 1 Meteorological elements and corresponding characteristic factors

[0150] Preferably, before constructing the sample data set according to the first sample meteorological model data, the second sample meteorological model data, and the surface elevation data, it further includes:

[0151] Processing the resolution of the initial surface data according to the following formula to obtain the surface elevation data:

[0152]

[0153] where t (i,j) is the initial surface data, T (i,j) is the surface elevation data with the processed resolution, i represents the row subscript of the grid, and j represents the column subscript of the grid. Or bilinear interpolation is used to interpolate the initial surface data to a resolution of 0.01 degrees, and then the regional average method is used to statistically calculate it to a resolution of 12.5 km.

[0154] Before obtaining the first sample meteorological model data and the second meteorological model historical data with the same characteristic factors for each meteorological element, it further includes:

[0155] Processing the resolution of the initial data of the second meteorological model historical data according to the following formula to obtain the second meteorological model historical data:

[0156]

[0157] Among them, t (i,j) is the initial data of the second meteorological model historical data, and T (i,j) is the processed second meteorological model historical data with the resolution. i represents the subscript of the row number of the grid, and j represents the subscript of the column number of the grid.

[0158] In this way, the initial data (1km * 1km) of the collected second meteorological model historical data and the initial surface data (90m * 90m) can be processed into regional average 12.5km * 12.5km grid data for constructing the sample dataset. It can be understood that the surface elevation data can be used as one of the characteristic factors.

[0159] Furthermore, for the second meteorological model historical data, further processing can be performed according to the observed historical extreme values. For example, data exceeding the historical extreme value range can be removed, or data with a correlation coefficient less than 0.9 (with a large difference in spatial texture structure, generally suspected error data) can be removed.

[0160] Exemplarily, the calculation formula of the correlation coefficient includes:

[0161]

[0162] Among them, r is the calculation result of the correlation coefficient, and x and y represent the data of the previous and subsequent time steps respectively, represents the average value of a certain dataset, represents the average value of another dataset, i represents the number of the characteristic factor, and n is the total number of characteristic factors.

[0163] In addition, before training the prediction model, the second meteorological model historical data can also be standardized, for example, using the Min - Max normalization method:

[0164] x a =(x i - min(x)) / (max(x) - min(x));

[0165] In the formula, min(x) represents the minimum value in the dataset before standardization, max(x) represents the maximum value in the dataset before standardization, x i represents the i - th value in the dataset before standardization, and x a represents the value after standardization (corresponding to the i - th value in the dataset before the above - mentioned standardization).

[0166] In a preferred embodiment, the convergence condition can be set according to the loss functions of the two-meter temperature element, the two-meter specific humidity element, the surface air pressure element, the ten-meter u-wind component element, and the ten-meter v-wind component element. At the same time, evaluation indicators such as POD, FAR, and CSI can also be used to evaluate the predicted values to determine whether the performance of the prediction model meets the requirements.

[0167] Training the basic model based on the sample data set, and obtaining the prediction model when the output of the basic model meets the preset convergence condition, including: training the basic model according to the sample data set; obtaining the prediction model when the loss functions of the two-meter temperature element, the two-meter specific humidity element, the surface air pressure element, the ten-meter u-wind component element, and the ten-meter v-wind component element meet the convergence condition.

[0168] For example, the two-meter temperature element and the surface air pressure element can adopt the following loss functions:

[0169]

[0170] In the formula, MSE is the loss function of the two-meter temperature element or the surface air pressure element, and y i represents the true value of the i-th characteristic factor of the two-meter temperature element or the surface air pressure element, represents the predicted value of the i-th characteristic factor of the two-meter temperature element or the surface air pressure element.

[0171] The two-meter specific humidity element can adopt the following loss function:

[0172]

[0173] In the formula, loss represents the loss function of the two-meter specific humidity element, and y i represents the true value of the i-th characteristic factor of the two-meter specific humidity element, represents the predicted value of the i-th characteristic factor of the two-meter specific humidity element.

[0174] The ten-meter u-wind component element and the ten-meter v-wind component element can adopt the following loss functions:

[0175]

[0176] In the formula, loss represents the loss function of the ten-meter u-wind component element or the ten-meter v-wind component element, and y i represents the true value of the i-th characteristic factor of the ten-meter u-wind component element or the ten-meter v-wind component element, represents the predicted value of the i-th characteristic factor of the ten-meter u-wind component element or the ten-meter v-wind component element.

[0177] As described above, the present application designs loss functions for the two-meter temperature element, two-meter specific humidity element, surface air pressure element, ten-meter u-wind component element, and ten-meter v-wind component element respectively. When the present application faces some extreme weather with small samples, it can strengthen these individual cases of extreme weather and effectively optimize the prediction effect.

[0178] Exemplarily, for specific convergence conditions, when the loss functions of multiple meteorological elements calculated above are all less than a preset loss threshold, at the same time, evaluation indicators such as POD, FAR, and CSI can be further combined to determine that the prediction model achieves convergence to ensure that the performance of the prediction model reaches the optimal.

[0179] Preferably, as Figure 4 shown, the basic model can adopt the AttUnet model. The Att-Unet model network is a deep learning model with a U-shaped structure. The difference from Unet is that Unet only performs splicing operations at the same level of the decoder and encoder, while AttUnet adds an AG structure to the Unet structure. This structure has the characteristic of weighting some encodings of the encoder at the same level. The AG structure of the AttUnet model also has an attention characteristic and can adjust the activation value by automatically learning parameters. In the figure, conv represents convolution, Relu and sigmoid are function names, where Relu refers to the Rectified Linear Unit (abbreviated as Relu), and sigmoid is often used as an activation function of a neural network due to its properties such as monotonic increase and monotonic increase of the inverse function. Multiply represents multiplication operation, skip connection represents skip connection, also known as residual connection, which is an important component in the deep neural network architecture. Its basic idea is to directly connect the input to the output in some layers of the network to allow information to jump between different layers. This connection is usually achieved through an addition operation, adding the input and the output; Up Sampling represents upsampling.

[0180] Furthermore, for Figure 4 the meanings of the arrows and symbols in Figure 5 shown, where conv Layer represents the convolutional layer, Pool Layer represents the pooling layer, copy and Crop refers to splicing (generally used for feature fusion), subpixellayer refers to the sub-pixel layer or sub-pixel layer, and Attention Gate refers to the attention gate.

[0181] Correspondingly, as Figure 8As shown in the figure, the present invention application also provides an ECMWF mode four-element prediction device 400 based on AttUnet, including a calling module 401 and a prediction module 402; wherein,

[0182] The calling module 401 is used to obtain the meteorological mode output data of the prediction object in real time and call a pre-trained prediction model;

[0183] The prediction module 402 is used to input the meteorological mode output data into the prediction model and obtain the predicted values of at least one meteorological element based on the output of the prediction model;

[0184] Among them, the prediction model is trained based on a number of first sample meteorological mode data, second sample meteorological mode data and surface elevation data. The data sources of the first sample meteorological mode data and the second sample meteorological mode data are different, and the first sample meteorological mode data is ECMWF data.

[0185] As a preferred solution, the ECMWF mode four-element prediction device 400 based on AttUnet further includes a training module, and the training module is used for:

[0186] Obtain the first sample meteorological mode data and the second meteorological mode historical data of the same characteristic factors of each meteorological element, and according to the deviation value of the characteristic factors of the first sample meteorological mode data and the second meteorological mode historical data, eliminate the data in the second meteorological mode historical data whose deviation value is greater than a preset threshold, and obtain the second sample meteorological mode data according to the elimination result;

[0187] Construct a sample data set according to the first sample meteorological mode data, the second sample meteorological mode data and the surface elevation data;

[0188] Train a basic model based on the sample data set, and when the output of the basic model meets the preset convergence condition, obtain the prediction model.

[0189] As a preferred solution, the ECMWF mode four-element prediction device 400 based on AttUnet further includes a screening module. Before the training module obtains the first sample meteorological mode data and the second meteorological mode historical data of the same characteristic factors, the screening module is used for:

[0190] Obtain the first meteorological mode historical data;

[0191] Determine the characteristic factors of the meteorological elements according to the first meteorological mode historical data;

[0192] Screen the first meteorological mode historical data according to the characteristic factors to obtain the first sample meteorological mode data.

[0193] As a preferred solution, the screening module determines characteristic factors of the meteorological elements according to the first historical meteorological pattern data, including:

[0194] The screening module calculates the information gain of each factor of the first historical meteorological pattern data through a preset XGBoost model;

[0195] According to each of the information gains, the characteristic factors of the meteorological elements are determined.

[0196] As a preferred solution, the meteorological element prediction further includes a first resolution processing module, and the first resolution processing module is used before constructing a sample data set according to the first sample meteorological pattern data, the second sample meteorological pattern data, and the surface elevation data:

[0197] Process the resolution of the initial surface data according to the following formula to obtain the surface elevation data:

[0198]

[0199] where t (i,j) is the initial surface data, T (i,j) is the surface elevation data with the resolution processed, i represents the row subscript of the grid, and j represents the column subscript of the grid.

[0200] As a preferred solution, the meteorological element prediction further includes a second resolution processing module, and the second resolution processing module is used before obtaining the first sample meteorological pattern data and the second historical meteorological pattern data with the same characteristic factors of each meteorological element:

[0201] Process the resolution of the initial data of the second historical meteorological pattern data according to the following formula to obtain the second historical meteorological pattern data:

[0202]

[0203] where t (i,j) is the initial data of the second historical meteorological pattern data, T (i,j) is the second historical meteorological pattern data with the resolution processed, i represents the row subscript of the grid, and j represents the column subscript of the grid.

[0204] As a preferred solution, the types of the meteorological elements include two-meter temperature element, two-meter specific humidity element, surface air pressure element, ten-meter u-wind component element, and ten-meter v-wind component element; the convergence condition is set according to the loss functions of the two-meter temperature element, two-meter specific humidity element, surface air pressure element, ten-meter u-wind component element, and ten-meter v-wind component element;

[0205] The training module trains a basic model based on the sample data set, and obtains the prediction model when the output of the basic model meets a preset convergence condition, including:

[0206] The training module trains the basic model according to the sample data set;

[0207] When the loss functions of the two-meter temperature element, two-meter specific humidity element, surface air pressure element, ten-meter u-wind component element, and ten-meter v-wind component element meet the convergence condition, the prediction model is obtained.

[0208] Correspondingly, the present invention application also provides a terminal device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the ECMWF model element deviation correction method based on AttUnet or the ECMWF model four-element prediction method based on AttUnet.

[0209] The so-called processor may be a Central Processing Unit (CPU), or may also be other general-purpose processors, Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field-Programmable Gate Arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The processor is the control center of the terminal, and connects various parts of the entire terminal through various interfaces and lines.

[0210] The memory can be used to store the computer program. By running or executing the computer program stored in the memory and invoking the data stored in the memory, the processor realizes various functions of the terminal. The memory mainly includes a program storage area and a data storage area. Among them, the program storage area can store an operating system, application programs required for at least one function (such as a sound playback function, an image playback function, etc.); the data storage area can store data created according to the use of the mobile phone (such as audio data, phone book, etc.). In addition, the memory can include high-speed random access memory and can also include non-volatile memory, such as a hard disk, memory, plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, Flash Card, at least one magnetic disk storage device, flash memory device, or other volatile solid-state storage devices.

[0211] Correspondingly, the present invention application also provides a computer-readable storage medium, which includes a stored computer program. When the computer program runs, it controls the device where the computer-readable storage medium is located to execute the AttUnet-based ECMWF model element deviation correction method or the AttUnet-based ECMWF model four-element prediction method.

[0212] Among them, if the module integrated in the AttUnet-based ECMWF model four-element prediction device / terminal is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, to implement all or part of the processes in the above-mentioned embodiment methods of the present invention, it can also be completed by a computer program instructing relevant hardware. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, the steps of the above-mentioned various method embodiments can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in the form of source code, object code, executable file, or some intermediate form, etc. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disc, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc.

[0213] Compared with the prior art, the present invention application has the following beneficial effects:

[0214] The present invention application provides a method for correcting ECMWF model element deviations based on AttUnet. Starting from ECMWF forecast data and HRCLDAS data, the present invention application establishes a method and device for correcting surface elements based on ECMWF forecast data and deep learning methods. This method fully explores the non-linear relationship of model deviations in historical big data samples, and by training a deep learning model, learns the non-linear relationship between meteorological elements, geographical elements and forecast objects, so that the model deviations are better corrected.

[0215] In addition, when training the model, since extreme weather belongs to small samples and the ECMWF forecast effect is poor, without reinforcement, the trained deep learning model cannot predict extreme weather well, and even performs worse than the original EC forecast effect. Therefore, this patent application designs special loss functions (the exponential form loss function is used for the 2m specific humidity to increase the extreme value weight, and the hyperbolic function is used for the 10m meridional wind / zonal wind to increase the extreme value weight) to strengthen extreme weather cases, so as to achieve better performance in forecasting extreme weather.

[0216] The specific embodiments described above have further detailed the purpose, technical solutions and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not used to limit the protection scope of the present invention. In particular, for those skilled in the art, any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for correcting the element deviation of the ECMWF model based on AttUnet, characterized in that, include: Obtain historical model forecast data, label data and geographic static data; wherein the historical model forecast data is: ECMWF model 0-72 hour forecast data; the label data is HRCLDAS data, and the label data contains four meteorological elements: two-meter temperature element, two-meter specific humidity element, surface air pressure and ten-meter wind element; the geographic static data is DEM data; Preliminarily determine relevant characteristic factors based on the ECMWF model data, and build a pre-selected characteristic factor library; wherein, the characteristic factors include potential height of each vertical altitude layer, zonal wind of each vertical altitude layer, meridional wind of each vertical altitude layer, vertical velocity of each vertical altitude layer, divergence of each vertical altitude layer, vorticity of each vertical altitude layer, specific humidity of each vertical altitude layer, temperature of each vertical altitude layer, 2-meter temperature, 2-meter specific humidity, 2-meter dew point temperature, ten-meter u wind component, ten-meter v wind component, surface air pressure, ground temperature, surface albedo, atmospheric column water content, total cloud cover and precipitation; Use the XgBoost tool to sort the importance of the characteristic factors of the four meteorological elements according to the significance analysis scoring requirements of the characteristic factor library, and screen out the characteristics whose scores meet the requirements; Combining the physical laws with the screening operation input by the forecaster, the features whose scores meet the requirements are further screened to obtain the element correction features of the four meteorological elements; According to the above-mentioned correction characteristics of the elements, the ECMWF model data and the HRCLDAS data are processed into the required data sets, and the data samples of each time of the data sets are screened, and the sample data whose data quality does not meet the requirements are eliminated to obtain the elimination result data; and the elimination result data are normalized to obtain the standardized results; Using the standardized results, a deep learning model based on the AttUnet architecture is trained to obtain a target deviation correction model; The target deviation correction model is used to correct the deviations of the four meteorological elements output by the model.

2. The method for correcting the ECMWF model element deviation based on AttUnet according to claim 1, wherein, The element correction characteristics of the four meteorological elements are specifically: The characteristic factors corresponding to the two-meter temperature elements are ground temperature, surface albedo, total water content of the atmospheric column, two-meter temperature characteristic factors and DEM; The characteristic factors corresponding to the two-meter specific humidity element are 850hPa specific humidity, two-meter specific humidity characteristic factor and DEM; The characteristic factors corresponding to the surface pressure elements are 500hPa potential height, 850hPa potential height, SP, surface pressure characteristic factor and DEM; The characteristic factors corresponding to the ten-meter wind element are 850hPa meridional wind, 850hPa zonal wind, surface air pressure characteristic factor, ten-meter u wind component characteristic factor, ten-meter v wind component characteristic factor and DEM.

3. The method for correcting the ECMWF model element deviation based on AttUnet according to claim 1, wherein, The normalization process of the elimination result data is specifically as follows: The dataset was normalized according to the following formula: x a = (x i - min(x)) / (max(x) - min(x)); Where, min(x) represents the minimum value in the dataset before normalization, max(x) represents the maximum value in the dataset before normalization, x i represents the i-th element in the dataset before normalization, x a represents the value after normalization.

4. The ECMWF model element deviation correction method based on AttUnet according to claim 1, characterized in that The deep learning model based on the AttUnet architecture is trained to obtain a target deviation correction model, specifically: When training a deep learning model, the early stopping method and simulated annealing strategy are used to optimize the model training parameters. RMSE is used to evaluate the output of the deep learning model. At the same time, different loss functions are designed according to different element features, and POD, CSI, and FAR are used to evaluate the extreme weather of the output of the deep learning model. According to the results of multiple evaluation metrics, the loss function and learning rate of the deep learning model are adjusted and optimized to obtain the target bias correction model.

5. The method for correcting the ECMWF model element deviation based on AttUnet according to claim 4, wherein Before optimizing the deep learning model using the loss function, early stopping method, and simulated annealing strategy, it also includes: Dividing the standardized results into a training set, a validation set, and a test set according to the ratio of 7:1.5:1.

5.

6. The method for correcting the ECMWF model element deviation based on AttUnet according to claim 1, wherein Before obtaining the historical model forecast data, label data, and geographical static data, it also includes: Through resolution processing, the HRCLDAS data and DEM data are averaged onto a 12.5 km regional grid.

7. The method for correcting the ECMWF model element deviation based on AttUnet according to claim 6, wherein The specific calculation formula for performing the resolution processing on the HRCLDAS data is: where t (i,j) is the initial data of the HRCLDAS data, T (i,j) is the HRCLDAS data after the resolution processing, i represents the row subscript of the grid, and j represents the column subscript of the grid.

8. The method for correcting ECMWF model element deviation based on AttUnet according to claim 6, wherein, The specific calculation formula for performing the resolution processing on the DEM data is: where t (i,j) is the initial DEM data with a resolution of 0.001 degrees, and T (i,j) is the DEM data processed at the said resolution, where i represents the row subscript of the grid and j represents the column subscript of the grid.

9. The method for correcting the ECMWF model element deviation based on AttUnet according to claim 6, wherein, The screening of the data samples at each time step of the dataset is specifically: By calculating the correlation coefficient between the HRCLDAS data at the previous and subsequent time steps, the HRCLDAS data with a correlation coefficient less than 0.9 is removed; The calculation formula of the correlation coefficient includes: Among them, r is the calculation result of the correlation coefficient, and x and y represent the data at different times before and after respectively. represents the average value of a certain data set. represents the average value of another data set, i represents the number of the characteristic factor, and n is the total number of the characteristic factors.

10. A method for correcting ECMWF model element deviations based on AttUnet as described in claim 1, characterized in that Before or after normalizing the dataset, the data outside the preset historical extreme value range is removed.

Images