An ECMWF model element bias correction method based on AttUnet

Abstract

The application provides an ECMWF mode element bias correction method based on AttUnet, and belongs to the meteorological prediction field. The method uses a deep convolutional neural network in 4 elements of ECMWF mode output surface, 2m temperature, surface pressure, 2m specific humidity and 10m wind, and performs bias correction on the 0.125 degree resolution ECMWF mode output result. The method first constructs a feature library according to the ECMWF mode output elements, and then uses the XgBoost tool to analyze the sample library and sort the importance, and further screens the factors combined with artificial experience and considers the terrain information. Then, the deep convolutional neural network is used for prediction and bias correction to obtain more accurate 0.125 degree grid products. At the same time, in the model debugging stage, by optimizing the learning rate and the loss function, the model has good prediction performance for extreme disastrous weather.

Description

Technical Field

[0001] This invention application relates to the field of meteorological data processing, and in particular to a method for correcting element biases in the ECMWF model based on AttUnet. Background Technology

[0002] Errors inherent in meteorological models and errors in initial values ​​are significant factors contributing to deviations in weather forecasts. These errors manifest differently in different regions, seasons, and weather events. Therefore, at least one error correction is necessary before making an actual forecast.

[0003] Traditional error correction methods have many shortcomings, such as complex processes, low computational efficiency, and unsatisfactory correction results. Furthermore, forecasters' subjective correction methods are limited by their individual knowledge and experience. Therefore, developing more objective and effective model interpretation algorithms has become an urgent operational priority to improve model interpretation accuracy and reduce the workload of subjective correction. Existing model interpretation methods can be broadly categorized into perfect prognostic (PP) methods and model output statistics (MOS) methods for predicting meteorological element values. However, these methods are based on multiple linear regression models, which cannot adequately describe the nonlinear relationships between meteorological elements, geographical elements, and the predicted object, resulting in poor accuracy in meteorological element prediction. The former corrects errors by establishing linear or simple nonlinear statistical models between observed values ​​and model forecast values, while the latter establishes linear or simple nonlinear statistical models between observed values ​​and a set of related atmospheric variable numerical forecast estimates. Most of these interpretation methods are based on multiple linear regression models, which cannot adequately describe the nonlinear relationships between meteorological elements, geographical elements, and the predicted object, exhibiting poor adaptability and requiring improvement in element prediction accuracy. With the continuous development of numerical weather prediction and its interpretation and application technologies, the development trend of numerical weather prediction product applications in recent years is to utilize deep learning technology to mine multidimensional and multi-dimensional information in meteorological big data, develop artificial intelligence technology interpretation methods, and further improve the level of refined forecasting of meteorological elements. Summary of the Invention

[0004] This invention application provides an ECMWF model feature bias correction method based on AttUnet to solve the technical problem of how to improve the accuracy of meteorological feature bias correction.

[0005] To address the aforementioned technical problems, this invention provides a method for correcting ECMWF model feature bias based on AttUnet, comprising:

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

[0007] Based on the ECMWF model data, relevant characteristic factors were initially determined, and a pre-selected characteristic factor library was constructed. These characteristic factors include: geopotential height at each vertical altitude level, zonal wind at each vertical altitude level, meridional wind at each vertical altitude level, vertical velocity at each vertical altitude level, divergence at each vertical altitude level, vorticity at each vertical altitude level, specific humidity at each vertical altitude level, temperature at each vertical altitude level, 2-meter temperature, 2-meter specific humidity, 2-meter dew point temperature, 10-meter u-wind component, 10-meter v-wind component, surface air pressure, surface temperature, surface albedo, atmospheric column water content, total cloud cover, and precipitation.

[0008] The XgBoost tool was used to sort the feature factors of the four meteorological elements according to the significance analysis scoring requirements, and the features that meet the scoring requirements were selected.

[0009] By combining physical laws with the filtering operations input by forecasters, the features that meet the requirements of the scoring are further filtered to obtain the element correction features of the four meteorological elements.

[0010] Based on the above-mentioned element correction features, the ECMWF model data and HRCLDAS data are processed into the required dataset, and the data samples of each time period in the dataset are screened to remove the sample data that does not meet the data quality requirements, so as to obtain the removal result data; and the removal result data is normalized to obtain the standardized result.

[0011] Using the standardized results, a deep learning model based on the AttUnet architecture is trained to obtain a target bias correction model;

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

[0013] As a preferred embodiment, after normalizing the dataset, samples whose deviation between ECMWF data and HRCLDAS data is greater than a preset deviation threshold are removed through correlation analysis.

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

[0015] This invention application provides a method for correcting ECMWF model feature biases based on AttUnet. Starting from ECMWF forecast data and HRCLDAS data, this application establishes a method and apparatus for correcting surface features based on ECMWF forecast data and deep learning methods. This method fully explores the nonlinear relationships of model biases in historical big data samples. By training a deep learning model, it learns the nonlinear relationships between meteorological elements, geographical elements, and forecast objects, resulting in better correction of model biases.

[0016] Furthermore, during model training, extreme weather events, due to their small sample size, result in poor ECMWF forecast performance. Without reinforcement, the trained deep learning model cannot predict extreme weather effectively, and may even perform worse than the original EC forecast. Therefore, this patent application designs special loss functions (an exponential loss function is used to increase extreme value weights for 2m specific humidity, and a hyperbolic function is used to increase extreme value weights for 10m meridional / zonal winds) to reinforce extreme weather cases, thereby achieving better performance in forecasting extreme weather events. Attached Figure Description

[0017] Figure 1 : A flowchart illustrating an embodiment of the ECMWF mode element deviation correction method based on AttUnet provided in this application.

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

[0019] Figure 3 : A bar chart illustrating the correlation coefficient of a 2-meter temperature application example provided in this invention application.

[0020] Figure 4 : A schematic diagram of an embodiment of AttUnet provided in this application.

[0021] Figure 5 Provided for this invention application Figure 4 A schematic diagram of the symbols.

[0022] Figure 6 : A schematic flowchart of another embodiment of the ECMWF mode element deviation correction method based on AttUnet provided in this application.

[0023] Figure 7 : A flowchart illustrating an embodiment of the ECMWF pattern four-element prediction method based on AttUnet provided in this invention application.

[0024] Figure 8: A schematic diagram of an embodiment of the ECMWF model four-element prediction device based on AttUnet provided in this application. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Example 1:

[0027] like Figure 1 As shown, Embodiment 1 of this invention application provides an application example of the ECMWF model four-element deviation correction method based on AttUnet. The steps are detailed below:

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

[0029] (1) The HRCLDAS 1km data was processed to a 500m resolution, and then the average of the regions was calculated to a total of 12.5km*12.5km. The calculation formula is as follows:

[0030] ;

[0031] in: The original HRCLDAS reanalysis data (0.01 degrees) This is the processed 12.5km resolution labeled dataset, where i and j represent the grid row and column indices, respectively;

[0032] (2) First, the downloaded DEM data (data source: Geospatial Data Cloud) is interpolated to a resolution of 0.01 degrees using bilinear interpolation, and then statistically analyzed to a resolution of 12.5 km using regional averaging. The calculation formula is as follows:

[0033] ;

[0034] in: This is the original DEM data (0.001 degrees). The dataset is the processed 12.5km resolution DEM dataset, where i and j represent the grid row and column indices, respectively.

[0035] Step 2: Based on meteorological expertise, preliminarily determine the input characteristics related to each correction element and construct a factor library. Then, process one year's ECMWF data to construct a feature selection dataset. The preliminary factor library is as follows: Select the factors that may be used to construct a pre-selected factor library, including: 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), and divergence (200hPa). a. 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 air pressure, surface temperature, surface albedo, atmospheric column water content, total cloud cover, precipitation.

[0036] Step 3: Based on the factor pre-selection library from the previous year, randomly select 500,000 gridded sample data points to construct the training dataset for the tree model. Then, construct a simple XGBoost model, such as... Figure 2 As shown.

[0037] The XGBoost model constructed above is used for training. During training, the model calculates the information gain of each feature and uses it as a label when splitting nodes. Since the node splitting order is directly related to the model's training performance, it can be used to rank feature importance. After adjusting the learning rate, number of training epochs, maximum depth, regularization, etc., and training sufficiently, the factor importance ranking is obtained.

[0038] Step 4: Based on the feature sorting selected by XGBoost, further feature screening is carried out with the help of physical laws and forecasters' subjective experience to obtain the final element correction features, such as the correction features of 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.

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

[0040]

[0041] Table 1 Meteorological elements and corresponding characteristic factors

[0042] Step 5: Process the collected ECMWF data from several years, along with preprocessed HRCLDAS and DEM data, into a sample dataset. Then, based on data quality, remove samples with poor quality (including those with large model prediction bias, incorrect label data, and poor label data quality). The main removal criteria are as follows:

[0043] (1) Based on the historical extreme values ​​of the observation data, remove the HRCLDAS reanalysis data that exceeds the historical extreme value range; calculate the correlation coefficient between two adjacent HRCLDAS data and remove the data with a correlation coefficient less than 0.9 (the spatial texture structure is significantly different and is generally suspected data).

[0044] (2) Perform correlation analysis between the forecasted corrected meteorological elements and the labeled dataset, and remove samples with low correlation coefficients (poor forecast quality). For example, consider 2m temperature. Figure 3 As shown.

[0045] Finally, the labeled dataset was divided into training set, validation set and test set according to the ratio of 7:1.5:1.5.

[0046] For the label dataset after removal, the Min-Max normalization method can be used for further processing:

[0047] x a = (x i -min(x)) / (max(x)-min(x));

[0048] In the formula, min(x) represents the minimum value in the dataset before standardization, and max(x) represents the maximum value in the dataset before standardization. i x represents the i-th feature in the dataset before standardization. a This represents the value after standardization.

[0049] Step 6: Select the AttUnet model as the base model, such as... Figures 4 to 5 As shown, the Att-Unet model is a U-shaped deep learning model. The difference between Att-Unet and Unet is that Unet only performs concatenation operations at the same level as the decoder and encoder, while Att-Unet adds an AG structure to the Unet structure. This structure has the characteristic of weighting the partial encodings of the encoder at the same level. The AG structure of the Att-Unet model also has attention features, and the activation values ​​can be adjusted through automatically learned parameters.

[0050] When training the AttUnet network model, the learning rate is adjusted exponentially to allow the model to converge to the optimal value earlier. Early stopping is used to control the number of model iterations and avoid overfitting.

[0051] Step 7: Evaluate the forecast results using RMSE, and simultaneously evaluate the accuracy of extreme weather forecasts using POD, FAR, and CSI. Based on these four evaluation metrics, adjust the model parameters and loss function to optimize the model's performance. The final loss functions for each variable are as follows:

[0052] 2m temperature and surface air pressure loss function:

[0053] ;

[0054] In the formula, MSE is the loss function of the two-meter temperature element or the surface air pressure element, and y i This represents the true value of the i-th sample of a two-meter temperature element or a surface air pressure element. This represents the predicted value of the i-th sample of the two-meter temperature element or surface air pressure element.

[0055] The specific humidity loss function for 2m is:

[0056] ;

[0057] In the formula, loss represents the loss function of the two-meter specific humidity factor, y i This represents the true value of the i-th sample of the two-meter specific humidity factor. This represents the predicted value of the i-th sample of the two-meter specific humidity factor.

[0058] The loss functions for 10m_u and 10m_v are:

[0059] ;

[0060] In the formula, loss represents the loss function of the 10-meter u-wind component or the 10-meter v-wind component, and y i This represents the true value of the i-th sample of the 10-meter u-wind component or the 10-meter v-wind component. This represents the predicted value of the i-th sample of the 10-meter u-wind component or the 10-meter v-wind component.

[0061] Step 8: Based on the real-time ECMWF model output data, using the trained model parameters mentioned above, corrections can be made for 2m temperature, surface air pressure, 2m humidity, and 10m wind.

[0062] Example 2:

[0063] Please refer to Figure 6 , Figure 6 The present invention application provides a method for correcting ECMWF model feature bias based on AttUnet, comprising steps S101 to S107; wherein,

[0064] Step S101: Obtain historical model forecast data, label data, and geostatic 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 the four meteorological elements: two-meter temperature element T. 2m Two-meter specific humidity element Q 2m Surface air pressure (SP) and 10-meter wind element, the 10-meter wind element specifically includes the 10-meter u-wind component (U). 10 And / or 10-meter wind component element V 10 The geographic static data is DEM data.

[0065] In this step, ECMWF model data refers to European Centre for Medium-Range Weather Forecasts (ECMWF).

[0066] HRCLDAS refers to the High Resolution China Meteorological Administration Land Data Assimilation System, abbreviated as HRCLDAS.

[0067] DEM stands for Digital Elevation Model.

[0068] In a preferred embodiment, before acquiring the historical model forecast data, label data, and geographic static data, the method further includes: averaging the HRCLDAS data and DEM data onto a 12.5km regional grid through resolution processing.

[0069] Specifically, the calculation formula for the resolution processing of the HRCLDAS data is as follows:

[0070] ;

[0071] Among them, t (i,j) For the initial data of the HRCLDAS data, T (i,j) The HRCLDAS data is processed to the specified resolution, where i represents the row index of the grid and j represents the column index of the grid. Alternatively, bilinear interpolation can be used to interpolate the HRCLDAS data to a resolution of 0.01 degrees, and then the data can be statistically analyzed to a resolution of 12.5 km using a regional averaging method.

[0072] The specific formula for calculating the resolution processing of the DEM data is as follows:

[0073] ;

[0074] Among them, t (i,j) This is the initial DEM data, with a resolution of 0.001 degrees, T (i,j) The data is DEM data processed at the aforementioned resolution, where i represents the row index of the grid and j represents the column index of the grid.

[0075] Step S102: Based on the ECMWF pattern data, the feature factors are initially determined, and a pre-selected feature factor library is constructed.

[0076] The characteristic factors include the potential height (200hPa, 500hPa, 700hPa, 850hPa, 1000hPa) at each vertical altitude, the zonal wind (200hPa, 500hPa, 700hPa, 850hPa, 1000hPa) at each vertical altitude, the meridional wind (200hPa, 500hPa, 700hPa, 850hPa, 1000hPa) at each vertical altitude, the vertical velocity (200hPa, 500hPa, 700hPa, 850hPa, 1000hPa) at each vertical altitude, and the divergence (200hPa, 500hPa, 700hPa, 850hPa, 1000hPa) at each vertical altitude. The data included: 0 hPa, 1000 hPa), vorticity at each vertical altitude (200 hPa, 500 hPa, 700 hPa, 850 hPa, 1000 hPa), specific humidity at each vertical altitude (200 hPa, 500 hPa, 700 hPa, 850 hPa, 1000 hPa), temperature at each vertical altitude (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, surface temperature, surface albedo, atmospheric column water content, total cloud cover, and precipitation.

[0077] Step S103: Using the XgBoost tool, the feature factor library is sorted by importance according to the significance analysis scoring requirements, and features that meet the scoring requirements are selected.

[0078] For example, the XGBoost model can be used to calculate the information gain or importance of each factor and rank them, and the factors with the largest information gain or the highest factor importance can be used as the feature factors of the meteorological element.

[0079] During training, the XGBoost model can use the information gain of each factor as a label when splitting nodes (because the order of node splitting is directly related to the training effect of the model). 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.

[0080] Step S104: Combining physical laws with the filtering operation input by the forecaster, the features that meet the requirements of the scoring are further filtered to obtain the element correction features of the four meteorological elements.

[0081] For example, the corresponding characteristic factors for each meteorological element are shown in the table below:

[0082]

[0083] Table 1 Meteorological elements and corresponding characteristic factors

[0084] The obtained element correction features for the four meteorological elements are specifically as follows:

[0085] The characteristic factors corresponding to the two-meter temperature element are ground temperature, surface albedo, total atmospheric water content, two-meter temperature characteristic factor, and DEM;

[0086] The characteristic factors corresponding to the two-meter specific humidity element are 850hPa specific humidity, two-meter specific humidity characteristic factor and DEM;

[0087] The characteristic factors corresponding to the surface pressure elements are 500hPa geopotential height, 850hPa geopotential height, SP, surface pressure characteristic factors and DEM.

[0088] 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.

[0089] Step S105: Based on the above-mentioned element correction features, the ECMWF model data and HRCLDAS data are processed into the required dataset, and the data samples of each time period in the dataset are screened to remove sample data that does not meet the data quality requirements, so as to obtain the removal result data; and the removal result data is normalized to obtain the standardized result.

[0090] In this step, the removal of HRCLDAS data that does not meet the data quality requirements specifically involves:

[0091] By calculating the correlation coefficient between HRCLDAS data at different time points, HRCLDAS data with a correlation coefficient less than 0.9 are removed.

[0092] The formula for calculating the correlation coefficient includes:

[0093] ;

[0094] Where r is the correlation coefficient, and x and y represent the data at different time points, respectively. Represents the average value of a dataset. represents the average value of another dataset, i represents the feature factor number, and n is the total number of feature factors.

[0095] In another preferred embodiment, after the dataset is normalized, samples whose deviation from ECMWF data and HRCLDAS data is greater than a preset deviation threshold are removed through correlation analysis.

[0096] Before or after the normalization process is performed on the dataset, data outside the preset historical extreme value range can be removed.

[0097] In a preferred embodiment, the normalization process for the dataset specifically includes:

[0098] The dataset is standardized according to the following formula:

[0099] x a = (x i -min(x)) / (max(x)-min(x));

[0100] In the formula, min(x) represents the minimum value in the dataset before standardization, and max(x) represents the maximum value in the dataset before standardization. i x represents the i-th feature in the dataset before standardization. a This represents the value after standardization.

[0101] Step S106: Using the standardized results, train the deep learning model based on the AttUnet architecture to obtain the target bias correction model.

[0102] This step employs a deep learning model based on the AttUnet architecture, such as... Figure 4As shown, the Att-Unet model is a U-shaped deep learning model. The difference between Att-Unet and Unet is that Unet only performs concatenation operations at the same level of the decoder and encoder, while Att-Unet adds an AG structure to the Unet structure. This structure has the characteristic of weighting the partial encodings of the encoder at the same level. The AG structure of the Att-Unet model also has attention features, and activation values ​​can be adjusted through automatically learned parameters. In the figure, conv represents convolution, ReLU and sigmoid are function names, where ReLU refers to the Linear Rectification Function (RRC), and sigmoid is often used as an activation function in neural networks due to its monotonically increasing properties and the monotonically increasing properties of its inverse function. multiply represents multiplication, and skip connection represents a skip connection, also known as a residual connection, which is an important component in deep neural network architectures. The basic idea is to connect the input directly to the output in certain layers of the network to allow information to jump between different layers. This connection is usually achieved through an addition operation, which adds the input and the output together; Up Sampling stands for upsampling.

[0103] Furthermore, for Figure 4 The meaning of the arrows and symbols in the text, such as Figure 5 As shown, conv Layer represents a convolutional layer, Pool Layer represents a pooling layer, copy and Crop refers to concatenation (generally used for feature fusion), subpixellayer refers to a subpixel layer, and Attention Gate refers to attention gating.

[0104] Preferably, before training the deep learning model based on the AttUnet architecture, the standardized results can be divided into a training set, a validation set, and a test set in a ratio of 7:1.5:1.5.

[0105] The training of the deep learning model based on the AttUnet architecture to obtain the target bias correction model is specifically as follows:

[0106] When training the deep learning model, early stopping and simulated annealing strategies are used to optimize the model training parameters. RMSE is used to evaluate the output of the deep learning model. Different loss functions are designed according to different feature characteristics. POD, CSI and FAR are used to evaluate the extreme weather output of the deep learning model. Based on the results of multiple evaluation indicators, the loss function and learning rate of the deep learning model are adjusted and optimized to obtain the target bias correction model.

[0107] In a preferred embodiment, the convergence condition of the deep learning model can be set according to the loss function 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. At the same time, evaluation indicators such as POD, FAR and CSI can be used to evaluate the predicted values ​​to determine whether the performance of the prediction model meets the requirements.

[0108] For example, the following loss function can be used for two-meter temperature data and surface air pressure data:

[0109] ;

[0110] In the formula, MSE is the loss function of the two-meter temperature element or the surface air pressure element, and y i This represents the true value of the i-th sample of a two-meter temperature element or a surface air pressure element. This represents the predicted value of the i-th sample of the two-meter temperature element or surface air pressure element.

[0111] The following loss function can be used for the two-meter specific humidity factor:

[0112] ;

[0113] In the formula, loss represents the loss function of the two-meter specific humidity factor, y i This represents the true value of the i-th sample of the two-meter specific humidity factor. This represents the predicted value of the i-th sample of the two-meter specific humidity factor.

[0114] The following loss functions can be used for the 10-meter u-wind component and the 10-meter v-wind component:

[0115] ;

[0116] In the formula, loss represents the loss function of the 10-meter u-wind component or the 10-meter v-wind component, and y i This represents the true value of the i-th sample of the 10-meter u-wind component or the 10-meter v-wind component. This represents the predicted value of the i-th sample of the 10-meter u-wind component or the 10-meter v-wind component.

[0117] As mentioned above, this 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. When facing some small-sample extreme weather, this application can strengthen the forecast for these extreme weather cases and effectively optimize the forecast effect.

[0118] Step S107: Use the target deviation correction model to correct the deviation of the four meteorological elements output by the model.

[0119] This step can correct the deviation of the four meteorological elements output by the target product's target correction data model using the output of the target deviation correction model.

[0120] Example 3:

[0121] Please refer to Figure 7 , Figure 7 This invention application provides a method for predicting four elements of ECMWF patterns based on AttUnet, comprising steps S201 to S202; wherein,

[0122] Step S201: Obtain the meteorological model output data of the predicted object in real time and call the pre-trained prediction model.

[0123] In this embodiment, the predicted object can be a specific weather forecasting software or product. The meteorological model output data can be ECMWF (European Centre for Medium-Range Weather Forecasts) meteorological model output data.

[0124] In a preferred embodiment, the ECMWF mode feature deviation correction method based on AttUnet described in this embodiment can be applied to a computer device, including but not limited to smartphones, laptops, tablets, and desktop computers.

[0125] The pre-trained prediction model can be retrieved from a server or data center. In this embodiment, the prediction model is retrieved from the server or data center and combined with local meteorological model output data for prediction in step S202.

[0126] Step S202: Input the meteorological model output data into the prediction model, and obtain the predicted value of at least one meteorological element based on the output of the prediction model.

[0127] The prediction model is based on AttUnet, and 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.

[0128] The prediction model is trained based on several first sample meteorological model data, second sample meteorological model data and surface elevation data. The first sample meteorological model data and the second sample meteorological model data have different data sources.

[0129] In one implementation, the first sample meteorological model data may be derived from the ECMWF meteorological model output data mentioned above. The second sample meteorological model data may be derived from HRCLDAS (High Resolution China Meteorological Administration Land Data Assimilation System).

[0130] The surface elevation data can be obtained using a Digital Elevation Model (DEM), which is a digital model that represents surface elevation using a regular grid. It describes the elevation of the terrain surface through an ordered array of numerical values ​​and is widely used in fields such as terrain analysis, hydrological simulation, and urban planning.

[0131] Since the first and second sample meteorological model data come from different sources, they can be combined for verification to ensure the overall accuracy of the sample dataset, improve the quality of the sample data, and thus improve the performance of the prediction model.

[0132] In a preferred embodiment, the training method of the prediction model includes:

[0133] First sample meteorological model data and second meteorological model historical data with the same characteristic factors of each meteorological element are obtained. The deviation value between the characteristic factors of the first sample meteorological model data and the second meteorological model historical data is calculated. Based on the deviation value, data in the second meteorological model historical data with a deviation value greater than a preset threshold are removed, and the second sample meteorological model data is obtained based on the removal result. A sample dataset is constructed based on the first sample meteorological model data, the second sample meteorological model data and the surface elevation data. The basic model is trained based on the sample dataset. When the output of the basic model meets the preset convergence condition, the prediction model is obtained.

[0134] For example, training the base model based on the sample dataset specifically involves dividing the sample dataset into a training set, a validation set, and a test set according to a preset ratio (e.g., 7:1.5:1.5), training the base model using the test set, adjusting the model's parameters using the validation set, evaluating the model's final performance using the test set, and verifying the model's generalization ability.

[0135] In a preferred embodiment, before acquiring the first sample meteorological model data and the second meteorological model historical data with the same characteristic factors, the ECMWF model feature bias correction method based on AttUnet includes:

[0136] Acquire historical data from the first weather model;

[0137] Based on historical data from the first meteorological model, the characteristic factors of the meteorological elements are determined;

[0138] The historical data of the first meteorological model are filtered according to the aforementioned feature factors to obtain the first sample meteorological model data.

[0139] For example, one year's worth of ECMWF data can be obtained, and potentially useful factors can be selected, 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), and divergence (200hPa, 500hPa, 700hPa, 850hPa, 1000hPa). The following parameters are used to measure atmospheric pressure, surface temperature, surface albedo, atmospheric column water content, total cloud cover, and precipitation.

[0140] Further, determining the characteristic factors of the meteorological elements based on the historical data of the first meteorological model includes: calculating the information gain of each factor in the historical data of the first meteorological model (approximately 500,000 grid sample data points) using a preset XGBoost model; and determining the characteristic factors of the meteorological elements based on each of the information gains.

[0141] For example, the XGBoost model can be used to calculate the information gain or factor importance of each factor and rank them, and the factors with the largest information gain or the highest factor importance can be used as the feature factors of the meteorological element.

[0142] During training, the XGBoost model can use the information gain of each factor as a label when splitting nodes (because the order of node splitting is directly related to the training effect of the model). 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.

[0143] Based on the selected factors with high information gain or importance, the final characteristic factors of meteorological elements can be obtained, with or without combining physical laws and forecasters' subjective experience.

[0144] For example, meteorological elements can be the two-meter temperature element T. 2m Two-meter specific humidity element Q 2m Surface air pressure element SP, 10-meter wind component element U 10 And / or 10-meter wind component element V 10 .

[0145] The corresponding characteristic factors for each meteorological element are shown in the table below:

[0146]

[0147] Table 1 Meteorological elements and corresponding characteristic factors

[0148] Preferably, before constructing the sample dataset based on the first sample meteorological model data, the second sample meteorological model data, and the surface elevation data, the method further includes:

[0149] The surface elevation data is obtained by processing the resolution of the initial surface data according to the following formula:

[0150] ;

[0151] Among them, t (i,j) For the initial surface data, T (i,j) The surface elevation data is processed to a higher resolution, where i represents the row index of the grid and j represents the column index of the grid. Alternatively, bilinear interpolation can be used to interpolate the initial surface data to a resolution of 0.01 degrees, and then regional averaging can be used to statistically analyze it to a resolution of 12.5 km.

[0152] Before acquiring the first sample meteorological model data and the second meteorological model historical data of the same characteristic factors of each meteorological element, the following steps are also included:

[0153] The historical data of the second meteorological model are obtained by processing the initial data resolution according to the following formula:

[0154] ;

[0155] Among them, t (i,j) This is the initial data for the historical data of the second meteorological model, T (i,j) The resolution is processed historical data of the second weather model, where i represents the row index of the grid and j represents the column index of the grid.

[0156] In this way, the initial data (1km*1km) and initial surface data (90m*90m) of the collected historical data from the second meteorological model can be processed into regionally averaged 12.5km*12.5km grid data for the construction of the sample dataset. It can be understood that surface elevation data can be used as a feature factor.

[0157] Furthermore, for historical data from the second meteorological model, further processing can be performed based on 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 can be removed (data with large differences in spatial texture structure is generally suspected data).

[0158] For example, the formula for calculating the correlation coefficient includes:

[0159] ;

[0160] Where r is the correlation coefficient, and x and y represent the data at different time points, respectively. Represents the average value of a dataset. represents the average value of another dataset, i represents the feature factor number, and n is the total number of feature factors.

[0161] In addition, historical data from the second weather model can be standardized before training the prediction model, for example, by using the Min-Max normalization method:

[0162] x a = (x i -min(x)) / (max(x)-min(x));

[0163] In the formula, min(x) represents the minimum value in the dataset before standardization, and max(x) represents the maximum value in the dataset before standardization. i x represents the i-th value in the dataset before standardization. a This represents the value after standardization (corresponding to the i-th value in the dataset before standardization).

[0164] In a preferred embodiment, the convergence condition can be set according to the loss function 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. At the same time, evaluation indicators such as POD, FAR and CSI can be used to evaluate the predicted value to determine whether the performance of the prediction model meets the requirements.

[0165] The step of training the basic model based on the sample dataset and obtaining the prediction model when the output of the basic model meets the preset convergence condition includes: training the basic model based on the sample dataset; and obtaining the prediction model 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.

[0166] For example, the following loss function can be used for two-meter temperature data and surface air pressure data:

[0167] ;

[0168] In the formula, MSE is the loss function of the two-meter temperature element or the surface air pressure element, and y i This represents the true value of the i-th sample of a two-meter temperature element or a surface air pressure element. This represents the predicted value of the i-th sample of the two-meter temperature element or surface air pressure element.

[0169] The following loss function can be used for the two-meter specific humidity factor:

[0170] ;

[0171] In the formula, loss represents the loss function of the two-meter specific humidity factor, y i This represents the true value of the i-th sample of the two-meter specific humidity factor. This represents the predicted value of the i-th sample of the two-meter specific humidity factor.

[0172] The following loss functions can be used for the 10-meter u-wind component and the 10-meter v-wind component:

[0173] ;

[0174] In the formula, loss represents the loss function of the 10-meter u-wind component or the 10-meter v-wind component, and y i This represents the true value of the i-th sample of the 10-meter u-wind component or the 10-meter v-wind component. This represents the predicted value of the i-th sample of the 10-meter u-wind component or the 10-meter v-wind component.

[0175] As mentioned above, this 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. When facing some small-sample extreme weather, this application can strengthen the forecast for these extreme weather cases and effectively optimize the forecast effect.

[0176] For example, for specific convergence conditions, when the loss functions of the multiple meteorological elements calculated above are all less than the preset loss thresholds, the prediction model can be further judged to have converged by combining evaluation indicators such as POD, FAR, and CSI, so as to ensure that the performance of the prediction model has reached the optimal level.

[0177] Preferably, such as Figure 4 As shown, the basic model can be the AttUnet model, which is a U-shaped deep learning model. The difference between AttUnet and 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 the partial encodings of the encoder at the same level. The AG structure of the AttUnet model also has attention characteristics, and the activation values ​​can be adjusted by automatically learning parameters. In the figure, conv represents convolution, ReLU and sigmoid are function names, where ReLU refers to the Linear Rectification Function (RRC), and sigmoid is often used as an activation function in neural networks due to its monotonically increasing properties and the monotonically increasing properties of its inverse function. multiply represents multiplication, and skip connection represents a skip connection, also known as a residual connection, which is an important component in deep neural network architectures. The basic idea is to connect the input directly to the output in certain layers of the network to allow information to jump between different layers. This connection is usually achieved through an addition operation, which adds the input and the output together; Up Sampling stands for upsampling.

[0178] Furthermore, for Figure 4 The meaning of the arrows and symbols in the text, such as Figure 5 As shown, conv Layer represents a convolutional layer, Pool Layer represents a pooling layer, copy and Crop refers to concatenation (generally used for feature fusion), subpixellayer refers to a subpixel layer, and Attention Gate refers to attention gating.

[0179] Correspondingly, such as Figure 8As shown, this invention application also provides an ECMWF pattern four-element prediction device 400 based on AttUnet, including a calling module 401 and a prediction module 402; wherein,

[0180] The calling module 401 is used to acquire the meteorological model output data of the predicted object in real time and call the pre-trained prediction model.

[0181] The prediction module 402 is used to input meteorological model output data into the prediction model, and obtain the predicted value of at least one meteorological element based on the output of the prediction model.

[0182] The prediction model is trained based on several first sample meteorological model data, second sample meteorological model data and surface elevation data. The first sample meteorological model data and the second sample meteorological model data have different data sources. The first sample meteorological model data is ECMWF data.

[0183] As a preferred embodiment, the ECMWF pattern four-element prediction device 400 based on AttUnet further includes a training module, which is used for:

[0184] First sample meteorological model data and second meteorological model historical data with the same characteristic factors of each meteorological element are obtained. Based on the deviation value of the characteristic factors of the first sample meteorological model data and the second meteorological model historical data, data in the second meteorological model historical data with a deviation value greater than a preset threshold are removed. The second sample meteorological model data is obtained based on the removal result.

[0185] A sample dataset is constructed based on the first sample meteorological model data, the second sample meteorological model data, and the surface elevation data.

[0186] The basic model is trained based on the sample dataset, and the prediction model is obtained when the output of the basic model meets the preset convergence condition.

[0187] As a preferred embodiment, the ECMWF model four-element prediction device 400 based on AttUnet further includes a screening module. Before the training module acquires the first sample meteorological model data and the second meteorological model historical data with the same feature factors, the screening module is used to:

[0188] Acquire historical data from the first weather model;

[0189] Based on historical data from the first meteorological model, the characteristic factors of the meteorological elements are determined;

[0190] The historical data of the first meteorological model are filtered according to the aforementioned feature factors to obtain the first sample meteorological model data.

[0191] As a preferred embodiment, the screening module determines the characteristic factors of the meteorological elements based on historical data from the first meteorological model, including:

[0192] The filtering module calculates the information gain of each factor in the historical data of the first meteorological model using a preset XGBoost model.

[0193] Based on the information gain described above, the characteristic factors of the meteorological elements are determined.

[0194] As a preferred embodiment, the meteorological element prediction further includes a first resolution processing module, which is used before constructing the sample dataset based on the first sample meteorological model data, the second sample meteorological model data, and the surface elevation data:

[0195] The surface elevation data is obtained by processing the resolution of the initial surface data according to the following formula:

[0196] ;

[0197] Among them, t (i,j) For the initial surface data, T (i,j) The surface elevation data is processed to a higher resolution, where i represents the row index of the grid and j represents the column index of the grid.

[0198] As a preferred embodiment, the meteorological element prediction further includes a second resolution processing module, which is used before acquiring the first sample meteorological model data and the second meteorological model historical data of the same characteristic factors of each meteorological element:

[0199] The historical data of the second meteorological model are obtained by processing the initial data resolution according to the following formula:

[0200] ;

[0201] Among them, t (i,j) This is the initial data for the historical data of the second meteorological model, T (i,j) The resolution is processed historical data of the second weather model, where i represents the row index of the grid and j represents the column index of the grid.

[0202] As a preferred embodiment, the meteorological elements include two-meter temperature elements, two-meter specific humidity elements, surface air pressure elements, ten-meter u-wind component elements, and ten-meter v-wind component elements; the convergence conditions are set according to the loss functions of the two-meter temperature elements, two-meter specific humidity elements, surface air pressure elements, ten-meter u-wind component elements, and ten-meter v-wind component elements.

[0203] The training module trains the base model based on the sample dataset, and obtains the prediction model when the output of the base model meets the preset convergence condition, including:

[0204] The training module trains the base model based on the sample dataset;

[0205] The prediction model is obtained 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 conditions.

[0206] Accordingly, this 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 AttUnet-based ECMWF model element deviation correction method or the AttUnet-based ECMWF model four-element prediction method.

[0207] The processor can be a Central Processing Unit (CPU), or 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. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the terminal, connecting various parts of the terminal via various interfaces and lines.

[0208] The memory can be used to store the computer program. The processor implements various functions of the terminal by running or executing the computer program stored in the memory and calling the data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0209] Accordingly, this application also provides a computer-readable storage medium, which includes a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to execute the ECMWF model element deviation correction method based on AttUnet or the ECMWF model four-element prediction method based on AttUnet.

[0210] The module integrated into the ECMWF pattern four-element prediction device / terminal based on AttUnet, if implemented as a software functional unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0211] Compared with the prior art, this invention application has the following beneficial effects:

[0212] This invention application provides a method for correcting ECMWF model feature biases based on AttUnet. Starting from ECMWF forecast data and HRCLDAS data, this application establishes a method and apparatus for correcting surface features based on ECMWF forecast data and deep learning methods. This method fully explores the nonlinear relationships of model biases in historical big data samples. By training a deep learning model, it learns the nonlinear relationships between meteorological elements, geographical elements, and forecast objects, resulting in better correction of model biases.

[0213] Furthermore, during model training, extreme weather events, due to their small sample size, result in poor ECMWF forecast performance. Without reinforcement, the trained deep learning model cannot predict extreme weather effectively, and may even perform worse than the original EC forecast. Therefore, this patent application designs special loss functions (an exponential loss function is used to increase extreme value weights for 2m specific humidity, and a hyperbolic function is used to increase extreme value weights for 10m meridional / zonal winds) to reinforce extreme weather cases, thereby achieving better performance in forecasting extreme weather events.

[0214] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for correcting feature bias in an ECMWF model based on AttUnet, characterized in that, include: Acquire historical model forecast data, label data, and geostatic data; wherein, the historical model forecast data is: ECMWF model 0-72 hour forecast data; the label data is HRCLDAS data, which includes four meteorological elements: two-meter temperature element, two-meter specific humidity element, surface air pressure, and ten-meter wind element; the geostatic data is DEM data; Based on the ECMWF model data, relevant characteristic factors were initially determined, and a pre-selected characteristic factor library was constructed. These characteristic factors include: geopotential height at each vertical altitude level, zonal wind at each vertical altitude level, meridional wind at each vertical altitude level, vertical velocity at each vertical altitude level, divergence at each vertical altitude level, vorticity at each vertical altitude level, specific humidity at each vertical altitude level, temperature at each vertical altitude level, two-meter temperature, two-meter specific humidity, two-meter dew point temperature, ten-meter u-wind component, ten-meter v-wind component, surface air pressure, surface temperature, surface albedo, atmospheric column water content, total cloud cover, and precipitation. The XgBoost tool was used to sort the feature factors of the four meteorological elements according to the significance analysis scoring requirements, and the features that meet the scoring requirements were selected. By combining physical laws with the filtering operations input by forecasters, the features that meet the requirements of the scoring are further filtered to obtain the element correction features of the four meteorological elements. Based on the above-mentioned element correction features, the ECMWF model data and HRCLDAS data are processed into the required dataset, and the data samples of each time period in the dataset are screened to remove the sample data that does not meet the data quality requirements, so as to obtain the removal result data; and the removal result data is normalized to obtain the standardized result. Using the standardized results, a deep learning model based on the AttUnet architecture is trained to obtain a target bias correction model; The target bias correction model is used to correct the biases of the four meteorological elements output by the model. The two-meter specific humidity element adopts the following loss function: ； In the formula, loss1 represents the loss function of the two-meter specific humidity factor, y i This represents the true value of the i-th sample of the two-meter specific humidity factor. This represents the predicted value of the i-th sample of the two-meter specific humidity factor; The ten-meter wind element adopts the following loss function: ； In the formula, loss2 represents the loss function for the 10-meter wind element. This represents the true value of the i-th sample of the 10-meter wind element. This represents the predicted value of the i-th sample of the 10-meter wind element.

2. The ECMWF model feature deviation correction method based on AttUnet as described in claim 1, characterized in that, The element correction features of the four meteorological elements are as follows: The characteristic factors corresponding to the two-meter temperature element are ground temperature, surface albedo, total atmospheric water content, two-meter temperature characteristic factor, 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 geopotential height, 850hPa geopotential height, two-meter temperature, 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 ECMWF model feature deviation correction method based on AttUnet as described in claim 1, characterized in that, The normalization process for the removed data is specifically as follows: The dataset is standardized according to the following formula: x a =(x i -min(x)) / (max(x)-min(x)); In the formula, min(x) represents the minimum value in the dataset before standardization, and max(x) represents the maximum value in the dataset before standardization. i x represents the i-th feature in the dataset before standardization. a This represents the value after standardization.

4. The ECMWF model feature deviation correction method based on AttUnet as described in claim 1, characterized in that, The training of the deep learning model based on the AttUnet architecture to obtain the target bias correction model is specifically as follows: When training the deep learning model, early stopping and simulated annealing strategies are used to optimize the model training parameters. RMSE is used to evaluate the output of the deep learning model. Different loss functions are designed according to different feature characteristics. POD, CSI and FAR are used to evaluate the extreme weather output of the deep learning model. Based on the results of multiple evaluation indicators, the loss function and learning rate of the deep learning model are adjusted and optimized to obtain the target bias correction model.

5. The ECMWF model feature deviation correction method based on AttUnet as described in claim 4, characterized in that, Before optimizing the deep learning model using a loss function, early stopping, and simulated annealing, the following steps are also included: The standardized results are divided into training, validation, and test sets in a ratio of 7:1.5:1.

5.

6. The ECMWF model feature deviation correction method based on AttUnet as described in claim 1, characterized in that, Before acquiring the historical model forecast data, label data, and geographic static data, the process also includes: Through resolution processing, the HRCLDAS data and DEM data are averaged onto a 12.5km area grid.

7. The ECMWF model feature deviation correction method based on AttUnet as described in claim 6, characterized in that, The specific formula for calculating the resolution processing of the HRCLDAS data is as follows: ； Among them, t (i,j) For the initial data of the HRCLDAS data, T (i,j) The HRCLDAS data is processed at the aforementioned resolution, where i represents the row index of the grid and j represents the column index of the grid.

8. The ECMWF model feature deviation correction method based on AttUnet as described in claim 6, characterized in that, The specific formula for calculating the resolution processing of the DEM data is as follows: ； Among them, t (i,j) This is the initial DEM data, with a resolution of 0.001 degrees, T (i,j) The data is DEM data processed at the aforementioned resolution, where i represents the row index of the grid and j represents the column index of the grid.

9. The ECMWF model feature deviation correction method based on AttUnet as described in claim 6, characterized in that, The filtering of data samples from each time period in the dataset specifically involves: By calculating the correlation coefficient between HRCLDAS data at different time points, HRCLDAS data with a correlation coefficient less than 0.9 are removed. The formula for calculating the correlation coefficient include: ； Where R is the correlation coefficient, and X and Y represent the data at different time points, respectively. Represents the average value of a dataset. represents the mean of another dataset, i represents the sample number, and N is the total number of samples used to calculate the correlation coefficient.

10. The ECMWF model feature deviation correction method based on AttUnet as described in claim 1, characterized in that, Before or after normalizing the dataset, data outside the preset historical extreme value range are removed.

Images