A weather forecast post-processing method and system for a ground meteorological station ensemble forecast

Abstract

The application discloses a weather forecast post-processing method and system for ground meteorological station ensemble prediction, and the method comprises the following steps: in a disordered zoom-in graph convolution network (DZGCN), a plurality of information extraction blocks are utilized to extract information, then a graph convolution layer pair is utilized to extract supplementary features, and then a prediction block is utilized to predict precipitation; the information extraction block comprises the following steps: a zoom-in module (ZI) is utilized as a gate, a zoom-out module (ZO) is utilized as a filter to extract global-to-local correlation information and local-to-global correlation information, a gating unit is utilized to integrate the extracted information, a graph convolution layer pair is utilized to extract features from the integrated information, then the features are connected with the information input into the information extraction block in a residual manner, and finally batch normalization is performed on the features to output the features. The application can obtain an overall hierarchical prediction skill score in terms of precipitation accuracy and hit rate of different precipitation categories, and has the advantage that the precipitation accuracy of the ground meteorological station is high.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of weather forecasting, and particularly relates to a weather forecasting post-processing method and system for ensemble prediction of a ground meteorological station. BACKGROUND

[0002] The development of weather forecasting relies on the continuous improvement of numerical weather prediction models (NWP models), but the original prediction still has systematic errors that need to be corrected using statistical post-processing methods. Research has found that the average prediction skill of ensemble prediction (with different initial conditions and / or model physics simulations) greatly exceeds that of any member. This indicates that the inter-member or intra-member of ensemble prediction contains potential information that is beneficial for error correction. In recent years, one of the research hotspots is to use deep learning methods to automatically extract multi-variable features to generate more accurate weather forecasts. In the weather forecasting post-processing task, a fully connected network called DRN (Distributed Regression Network) is used to add all member information of the ensemble. Using a normal ANN to calibrate the 2-meter ground temperature gets more accurate prediction. The purpose of this study is to post-process the multi-variable ensemble prediction to generate more accurate precipitation forecasts. The research content is to extract the spatial distribution features, variable correlations, and inter-member potential correlations in the multi-variable ensemble prediction data, and to predict the precipitation values of real ground observation stations with irregular positions. This study has the following challenges: First, the daily observation data after quality monitoring does not lie on the standard grid, and extracting the spatial features of such data and maintaining their spatial consistency requires a unique algorithm. Second, the ensemble prediction has potential correlations in the member dimension, but does not follow the sequential relationship, which cannot be handled by ordinary time series neural networks. Finally, the frequency distribution of precipitation is highly skewed. Therefore, how to improve the accuracy of weather forecasting post-processing of ground meteorological station ensemble prediction is still a key technical problem to be solved. SUMMARY

[0003] The technical problem solved by the present application: In view of the above problems of the prior art, the present application provides a weather forecasting post-processing method and system for ensemble prediction of a ground meteorological station, which can achieve overall hierarchical prediction skill scores in terms of precipitation accuracy and hit rate of different precipitation categories, and has the advantage of high accuracy of precipitation of the ground meteorological station.

[0004] To solve the above technical problems, the technical scheme adopted by the present application is:

[0005] The application discloses a weather forecast post-processing method for ensemble prediction of a ground meteorological station, and relates to the technical field of weather forecast post-processing.

[0006] Optionally, when the multiple information extraction blocks with different network parameters are used to extract information, the information extraction block extracts information by using an upscaling module ZI as a gate to extract irregular position distribution spatial feature information of linear coding and global-to-local correlation information of ensemble prediction member dimensions, using a downscaling module ZO as a filter to extract spatial feature information of linear coding and local-to-global correlation information of member dimensions, using a gating unit to integrate the extracted global-to-local correlation information and the local-to-global correlation information, and then extracting features of the integrated information through a graph convolution layer pair, and finally outputting after residual connection of information input into the information extraction block and batch normalization.

[0007] Optionally, the upscaling module ZI comprises multiple graph convolution layers, and a function expression of any lth graph convolution layer is as follows:

[0008]

[0009] In the above formula, is global-to-local correlation information output by the lth layer of the upscaling module ZI, ReLU is an activation function, and Φ l is a 1*1 convolution kernel of the member dimension, and * is a standard convolution operation, represents global-to-local correlation information output by the (l-1)th layer of the upscaling module ZI, L is the number of layers of the upscaling module ZI, and the input of the first layer of the upscaling module ZI is wherein Linear is linear coding, and X is the ensemble prediction of the ground meteorological station.

[0010] Optionally, the downscaling module ZO comprises multiple graph convolution layers, and a function expression of any lth graph convolution layer is as follows:

[0011]

[0012]

[0013] In the above formula, is the i to (i+K) member slices of the information output by the lth layer of the downscaling module ZO, ReLU is an activation function, is a 1*1 convolution kernel, and ★ is a standard convolution operation, The (l-1)th layer output information of the downsizing module ZO and the lth layer input information are reduced; The local-to-global correlation information output by the lth layer of the downsizing module ZO, and Concat is a concatenation operation, The 1th to (i+κ)th member slice information of the downsizing module ZO extracted by convolution from the output of the upper layer and the 1th to (i+κ)th member slice output by the lth layer, The (ν+1)th to (ν+1+κ)th slice information of the member dimension of the lth layer output of the downsizing module ZO, The (M l-1 -κ)th to M l-1 th slice information of the member dimension of the lth layer output of the downsizing module ZO, κ and ν are the kernel size and step size of the convolution.

[0014] Optionally, the function expression for integrating the global-to-local correlation information and the local-to-global correlation information by the gating unit is:

[0015]

[0016] In the above formula, is the integrated information, g is an activation function of the output, is the global-to-local correlation information of the lth layer output of the upsizing module ZI, is an element product, sigmoid is an activation function, is the local-to-global correlation information output by the downsizing module ZO.

[0017] Optionally, the function expression for predicting the precipitation of the ground meteorological station according to all the extracted features by the prediction block is:

[0018]

[0019] In the above formula, Y is the predicted precipitation of the ground meteorological station, F FC,2 is a fully connected network, mean is an average value, is an intermediate variable, and has:

[0020]

[0021]

[0022]

[0023] In the above formula, F FC,1 is a fully connected network, is an intermediate variable, Stack is a stacking operation, respectively are features obtained by performing supplementary feature extraction on information output by the first to the B-th information extraction blocks through the graph convolution layer, respectively are features obtained by performing supplementary feature extraction on information output by the first to the B-th information extraction blocks through the graph convolution layer, is a graph convolution operation of the graph convolution layer, is a convolution kernel of the graph convolution layer, is information output by the B-th information extraction block, and B is the number of information extraction blocks.

[0024] Optionally, the graph convolution layer uses Chebyshev polynomials as the convolution kernel, and the input feature of the arbitrary l-th layer of the graph convolution layer is:

[0025]

[0026] In the above formula, respectively are slices of the member dimension of the input information of the l-th layer of the graph convolution layer, N is the number of graph nodes, C l-1 is the number of input feature information of the l-th layer of the graph convolution layer, M l-1 is the size of the member dimension of the input information of the l-th layer of the graph convolution layer; and the function expression for generating an arbitrary m-th slice of the l-th layer of the graph convolution layer is:

[0027]

[0028] In the above formula, is the m-th slice of the output feature information of the l-th layer of the graph convolution layer, ReLU is an activation function, is a graph convolution operation of the l-th layer of the graph convolution layer on , is a convolution kernel of the l-th layer of the graph convolution layer, is a convolution kernel of the l-th layer of the graph convolution layer, is the m-th slice of the input feature information of the l-th layer of the graph convolution layer; and the function expression for generating the output feature of the arbitrary l-th layer of the graph convolution layer is:

[0029]

[0030] In the above formula, is the output feature of the l-th layer of the graph convolution layer, Concat is a connection operation, respectively are M l slices of the l-th layer of the graph convolution layer.

[0031] Optionally, the loss function used by the out-of-order zooming graph convolution network DZGCN during training is:

[0032]

[0033] In the above formula, is a loss function, Y is a predicted precipitation of a ground weather station, is a true value of precipitation of a ground weather station, and W is a weight vector, is a Huber loss, and has:

[0034] w i = max(min(f(y i ), 2), 10 -8 ),

[0035]

[0036] In the above formula, w i is the i-th component of the weight vector, max is a maximum value, min is a minimum value, f(y i ) is a function defined by taking the i-th component y i of the true value of precipitation of a ground weather station as an argument, and α1, α2, β1, β2, γ1, γ2 are hyperparameters, α1, α2, β1, β2, γ1, γ2 ∈ [0, 1], y i is the i-th component of the true value of precipitation of a ground weather station.

[0037] In addition, the present application also provides a weather forecast post-processing system for a ground weather station ensemble forecast, comprising a microprocessor and a memory connected to each other, the microprocessor being programmed or configured to execute the weather forecast post-processing method for the ground weather station ensemble forecast.

[0038] In addition, the present application also provides a computer readable storage medium, wherein a computer program for being programmed or configured by a microprocessor to execute the weather forecast post-processing method for the ground weather station ensemble forecast is stored in the computer readable storage medium.

[0039] ​Compared with the prior art, the present application mainly has the following advantages: the present application comprises linearly encoding and inputting the collective forecast of the ground meteorological station into a disordered zoom-in-out graph convolution network DZGCN, using multiple information extraction blocks to extract information respectively in the disordered zoom-in-out graph convolution network DZGCN, then performing supplementary feature extraction through a graph convolution layer, using a prediction block to predict the precipitation of the ground meteorological station according to all the extracted features, and through parallel stacking of multiple information extraction blocks, the disordered zoom-in-out graph convolution network (DZGCN can extract dependent relationships of different aggregation dimension combinations. For example, some modules prefer to extract light precipitation potential layer information, while other modules prefer to extract strong precipitation related information. Finally, the information extraction block uses a single layer GCN to perform supplementary extraction on the potential layer information, and then the prediction block obtains the final output, which can obtain an overall hierarchical prediction skill score in terms of precipitation accuracy and hit rate of different precipitation categories, and has the advantage of high accuracy of the precipitation of the ground meteorological station. BRIEF DESCRIPTION OF DRAWINGS

[0040] Figure 1 It is a basic flow diagram of the method of the embodiment of the present application.

[0041] Figure 2 It is a structure diagram of the disordered zoom-in-out graph convolution network DZGCN in the embodiment of the present application.

[0042] Figure 3 It is a structure diagram of the zoom-in module ZI in the embodiment of the present application.

[0043] Figure 4 It is a structure diagram of the zoom-out module ZO in the embodiment of the present application.

[0044] Figure 5 It is a structure diagram of the i-th component w i of the weight vector in the embodiment of the present application.

[0045] Figure 6 It is a performance diagram of each component in the embodiment of the present application.

[0046] Figure 7 It is a spatial average value diagram of the index in the embodiment of the present application.

[0047] Figure 8 It is a typical case diagram of precipitation in the embodiment of the present application. DETAILED DESCRIPTION

[0048] As Figure 1As shown, the embodiment provides a weather forecast post-processing method for ensemble prediction of a ground meteorological station, which comprises inputting linearly coded ensemble prediction of the ground meteorological station into a disordered zoom-in graph convolution network DZGCN, using multiple information extraction blocks with different network parameters in the disordered zoom-in graph convolution network DZGCN to respectively extract information, then performing supplementary feature extraction through a graph convolution layer (GCU), and finally using a prediction block to predict the precipitation of the ground meteorological station according to all the extracted features. The weather forecast post-processing method for ensemble prediction of the ground meteorological station comprises inputting linearly coded ensemble prediction of the ground meteorological station into a disordered zoom-in graph convolution network DZGCN, using multiple information extraction blocks in the disordered zoom-in graph convolution network DZGCN to respectively extract information, then performing supplementary feature extraction through a graph convolution layer, and finally using a prediction block to predict the precipitation of the ground meteorological station according to all the extracted features. By stacking multiple information extraction blocks in parallel, the disordered zoom-in graph convolution network (DZGCN) can extract dependent relationships of different aggregation dimension combinations. For example, some modules prefer to extract light precipitation potential layer information, while other modules prefer to extract strong precipitation related information. Finally, the information extraction block uses a single-layer GCN to perform supplementary extraction on the potential layer information obtained by the information extraction block, and then the prediction block obtains the final output, which can obtain overall hierarchical prediction skill scores in terms of precipitation accuracy and hit rate of different precipitation categories, and has the advantage of high accuracy of precipitation of the ground meteorological station.

[0049] As shown, Figure 2 In the embodiment, when multiple information extraction blocks with different network parameters are used to respectively extract information, the information extraction block includes: using an upscaling module ZI as a gate to extract irregularly distributed spatial feature information of linear coding and global-to-local correlation information of ensemble prediction member dimensions, using a downsizing module ZO as a filter to extract spatial feature information of linear coding and local-to-global association information of member dimensions, using a gating unit (GU) to integrate the extracted global-to-local correlation information and local-to-global association information, then extracting features of the integrated information through a graph convolution layer (GCU), and finally outputting after residual connection of the information input into the information extraction block and batch normalization (BN). It should be noted that the multiple information extraction blocks have the same structure, but are independently trained during training, so that the network parameters of the multiple information extraction blocks are different, so that similar but different information can be extracted from the linearly coded ensemble prediction of the ground meteorological station, so that dependent relationships of different aggregation dimension combinations can be extracted.

[0050] In this embodiment, the zoom-in module (ZI) and zoom-out module (ZO) are collectively referred to as the zoom module, which aims to extract the potential internal relationships between members. The zoom-in module (ZI) gradually focuses on detailed features by stacking convolutional layers, where cross-member information extraction is achieved through a 1x1 convolution operation spanning the entire member dimension. Figure 3 As shown, the amplification module ZI in this embodiment includes multiple graph convolutional layers, and the function expression of any l-th graph convolutional layer is:

[0051]

[0052] In the above formula, To amplify the global-to-local correlation information output of the l-th layer of the ZI module, ReLU is the activation function, Φ l The kernel is a 1×1 convolution with member dimensions, and * represents the standard convolution operation. This represents the global-to-local correlation information output from the (l-1)th layer of the amplification module ZI, where L is the layer number of the amplification module ZI, and the input of the first layer of the amplification module ZI is... Where Linear represents linear coding, and X represents the ensemble forecast from surface weather stations. Local correlation information. and Dimensions satisfy Where N is the number of graph nodes in the graph convolutional layer, and C l-1 M is the number of features input to the l-th graph convolutional layer. l-1 The size of the member dimension of the input information for the l-th graph convolutional layer.

[0053] The shrinking module ZO adopts a similar idea to the enlarging module ZI, but uses a local receptive field. It gradually expands the receptive field by stacking multiple layers of sliding 1x1 convolutions, allowing the network to gradually extract local member information into global long-range information. For example... Figure 4 As shown, the reduction module ZO in this embodiment includes multiple graph convolutional layers, and the function expression of any l-th graph convolutional layer is:

[0054]

[0055]

[0056] In the above formula, To reduce the number of slices of the output information of the l-th layer of module ZO from the i-th to (i+κ)-th members, ReLU is used as the activation function. The kernel is 1×1, and * represents the standard convolution operation. To reduce the output information of the (l-1)th layer and the input information of the lth layer of module ZO; Concat is the concatenation operation, and ZO is the downsampling module. are the member dimension of the lth layer output of the downsampling module ZO, and are the member dimension of the lth layer output of the downsampling module ZO, and are the member dimension of the lth layer output of the downsampling module ZO, and l-1 are the member dimension of the lth layer output of the downsampling module ZO, and l-1 are the member dimension of the lth layer output of the downsampling module ZO, and The dimension of the local-to-global association information output by the downsampling module ZO satisfies

[0057] The gating mechanism used by the gating unit is very effective in controlling the flow of information between the layers of the temporal convolutional network. In this embodiment, it is used to integrate the extracted global-to-local correlation information and the local-to-global association information. In this embodiment, the function expression for integrating the global-to-local correlation information and the local-to-global association information by the gating unit is:

[0058]

[0059] In the above formula, is the integrated information, g is the activation function of the output, is the global-to-local correlation information of the lth layer output of the upsampling module ZI, is the element product, and sigmoid is the activation function (determines the proportion of information passed to the next layer), is the local-to-global association information output by the downsampling module ZO. In this embodiment, the hyperbolic tangent function tanh is used as the activation function g.

[0060] In this embodiment, the prediction block first stacks multiple possible latent information directly, then encodes it through a fully connected network, and then integrates it through the average operation of the member dimension. The output layer is a linear connection layer. Specifically, in this embodiment, the function expression for predicting the precipitation of the ground meteorological station by the prediction block according to all the extracted features is:

[0061]

[0062] In the above formula, Y is the predicted precipitation of the ground meteorological station, F FC,2 is the fully connected network, mean is the average, is an intermediate variable, and has:

[0063]

[0064]

[0065]

[0066] In the above formula, F FC,1 is a full connection network, is an intermediate variable, Stack is a stacking operation, respectively, are features obtained by performing supplementary feature extraction on the information output by the first to the Bth information extraction blocks through the graph convolution layer, is a feature obtained by performing supplementary feature extraction on the information output by the arbitrary bth information extraction block through the graph convolution layer, and ReLU is an activation function, is a graph convolution operation of the graph convolution layer, is a convolution kernel of the graph convolution layer, is the information (i.e., the latent layer information) output by the bth information extraction block, and B is the number of information extraction blocks. The dimensions of each parameter are as follows:

[0067]

[0068]

[0069] where K is the order of the Chebyshev polynomial, C L is the number of features of and , N is the number of nodes of the graph, and C L ' is the number of features of .

[0070] In spectral analysis, a graph is represented by its corresponding Laplacian matrix. By analyzing the Laplacian matrix and its eigenvalues, the properties of the graph structure can be obtained. The Laplacian matrix L of a graph is defined as: L = D - A, and its normalized form is:

[0071]

[0072] In the above formula, D is a degree matrix, A is an adjacency matrix, I N is an identity matrix, N is the number of nodes of the graph, and the degree matrix D ∈ R N×N is a diagonal matrix, where the i-th diagonal element is:

[0073]

[0074] In the above formula, A ij is the i-th row and j-th column element of the adjacency matrix A.

[0075] The eigenvalue decomposition of the Laplacian matrix is:

[0076] L = UΛUT ,

[0077] In the above formula, is a diagonal matrix, λ0~λ N-1 are eigenvalues of the Laplacian matrix; U is the Fourier basis of the graph: suppose a prediction member X (m) The signal on the entire graph is The graph Fourier transform of a signal x is defined as: According to the properties of the Laplacian matrix, U is an orthogonal matrix, so the corresponding inverse Fourier transform is That is, the graph convolution realizes the operation similar to the classical convolution operator by using the diagonalized linear operator in the Fourier domain. The signal x on the graph G is filtered by the graph convolution kernel g θ The general graph convolution operator can be expressed as:

[0078] g θ * G x=g θ (L)x=g θ (UΛU T )x=Ug θ (Λ)U T x

[0079] In the above formula, G represents the graph convolution operation. Since the convolution operation of the signal in the spatial domain is equal to the product thereof in the spectral domain, the above formula can be understood as transforming g θ and x into the spectral domain by the Fourier transform, then multiplying the transformed results. Then, the Fourier inverse transform is performed to obtain the final result of the convolution operation. When the size of the graph is large, the direct eigenvalue decomposition of the Laplacian matrix is expensive. In order to solve the above technical problem, the graph convolution layer in the embodiment adopts the Chebyshev polynomial as the convolution kernel, and the graph convolution operator thereof can be expressed as:

[0080]

[0081] In the above formula, the parameter is a polynomial coefficient, is an intermediate variable, and has:

[0082]

[0083] In the above formula, λ max is the maximum eigenvalue of the Laplacian matrix, L is the Laplacian matrix, I N is the unit matrix, is the Chebyshev polynomial, and the recursive definition of the Chebyshev polynomial is:

[0084] T k(x) = 2xT k-1 (x) - T k-2 (x),

[0085] In the above formula, T k (x), T k-1 (x) and T k-2 (x) are Chebyshev polynomials at the k, k-1 and k-2th iterations respectively, and T0(x) = 1 and T1(x) = x.

[0086] The graph convolution module uses a rectified linear unit (ReLU) as the final activation function, i.e. ReLU(g θ * G x). In this embodiment, a GCN (Graph Convolutional Network) based on Chebyshev polynomials is used, which is stacked in the member dimension, similar to spatial dimension convolution. The input feature of any lth layer of graph convolution layer is:

[0087]

[0088] In the above formula, T (x) and T (x) are the slices of the input information member dimension of the lth layer of graph convolution layer respectively, N is the number of graph nodes, C l-1 is the input feature number of the lth layer of graph convolution layer, and M l-1 is the member dimension size of the input information of the lth layer of graph convolution layer; and the function expression of the lth layer of graph convolution layer generating any mth slice is:

[0089]

[0090] In the above formula, T (x) is the mth slice of the output feature information of the lth layer of graph convolution layer, ReLU is an activation function, is the graph convolution operation of the lth layer of graph convolution layer on , A is the convolution kernel of the lth layer of graph convolution layer, and X is the mth slice of the input feature information of the lth layer of graph convolution layer; the function expression of the lth layer of graph convolution layer generating the output feature is:

[0091]

[0092] In the above formula, Y is the output feature generated by the lth layer of graph convolution layer, Concat is a connection operation, are the M l slices of the lth layer of graph convolution layer, wherein the dimensions of the respective parameters are:

[0093]

[0094]

[0095] wherein N is the number of nodes of the graph, M l-1 is the member dimension size of the input and output of the l-th layer graph convolutional layer, C l-1 is the feature number of the input of the l-th layer graph convolutional layer, C l is the feature number of the output of the l-th layer graph convolutional layer.

[0096] In this embodiment, the loss function used by the out-of-order zooming graph convolutional network DZGCN in training is as follows:

[0097]

[0098] In the above formula, is the loss function, Y is the predicted precipitation of the ground meteorological station, is the true value of the precipitation of the ground meteorological station, W is the weight vector, is the Huber loss, and has:

[0099] w i = max(min(f(y i ), 2), 10 -8 ),

[0100]

[0101] In the above formula, w i is the i-th component of the weight vector, max is the maximum value, min is the minimum value, f(y i ) is a function defined by the i-th component y i of the true value of the precipitation of the ground meteorological station as the independent variable, and α1, α2, β1, β2, γ1, γ2 are hyperparameters, α1, α2, β1, β2, γ1, γ2 ∈ [0, 1], y i is the i-th component of the true value of the precipitation of the ground meteorological station . Compared with the square error loss, the Huber loss has lower sensitivity to outliers, so the Huber loss is used in the loss function in this embodiment. The Huber loss is an existing loss function, and its function expression is as follows:

[0102]

[0103] In the above formula, δ is a threshold parameter for controlling the range of error loss square, and the value is 0.5 in the embodiment. Due to the high frequency of precipitation and the importance of precipitation in operational forecasting, the weighted Huber loss based on precipitation characteristics is used as the final loss function in the embodiment, and according to the precipitation classification threshold: light rain ∈ [0.1, 10.1) mm·day -1 , moderate rain ∈ [10.1, 25.1) mm·day -1 , heavy rain ∈ [25.1, 50.1) mm·day -1 and heavy rain ∈ [50.1, 999) mm·day -1 , the parameters are set as: α1=1.0α2=0.8, β1=β2=1.0, γ1=1.0γ2=1.5, and the i-th component w i in the weight vector in the embodiment is as shown in Figure 5 . During training of the disordered zoom-in graph convolutional network DZGCN, the training set is given as:

[0104]

[0105] In the above formula, X (i) is the i-th sample of the set forecast X of the ground meteorological station, is the i-th label of the true value of the precipitation of the ground meteorological station , K is the number of samples participating in training (batch size), and the training process of the model is an end-to-end learning process of a group of model parameters based on the true value of the precipitation of the ground meteorological station using the loss function as the objective function L(Θ). The training process of the disordered zoom-in graph convolutional network DZGCN can be represented as:

[0106]

[0107] In the above formula, Θ is the network parameter of the disordered zoom-in graph convolutional network DZGCN.

[0108] The ensemble prediction X of the ground meteorological station in the training of the disordered zooming convolutional network DZGCN in this embodiment comes from the public TIGGE dataset, which is a public dataset provided by ECMWF, collecting and arranging historical data of global scale ensemble prediction products of multiple NWP models run by meteorological centers around the world. This study uses the 50-member ensemble prediction published by ECMWF, which has two initial times: 00:00 UTC and 12:00 UTC. The ensemble prediction of 24 ground variables and 8 high-altitude pressure layer 5 variables is produced every 6 hours. The embodiment of this study collects the ensemble prediction of 6 ground variables of the 0.25°x0.25° grid points in [10.25°N~35°N, 107°E~122.75°E] (see Table 1). The date range is from March 1, 2010 to February 28, 2015. We take all the prediction values of these 5 years, with initial time of UTC00 and UTC12, and lead time of every 6 hours from 24 to 48 hours.

[0109] Table 1: Ensemble prediction of 6 ground variables

[0110] Abbreviation ID Unit Name msl 151 Pa Mean sea level pressure 10u 166 m·s -1 ]]> 10 meter wind U component 10v 167 m·s -1 ]]> 10 meter wind V component 2d 168 K 2 meter dew point temperature 2t 165 K 2 meter temperature tp 228228 [kg·s -2 ]]> Total precipitation

[0111] The observation data of this embodiment comes from the China National Ground Meteorological Station Basic Meteorological Element Daily Value Dataset (V3.0), which contains 2474 basic element daily observation data of pressure, temperature, precipitation, evaporation, relative humidity, wind direction and speed, sunshine duration and 0cm ground temperature from January 1951 to February 2015 distributed in meteorological stations, reference climate stations or comprehensive meteorological stations in China, with a data volume of 21.3GB. We selected the 24-hour cumulative precipitation data of the stations in the region of [10.25°N ~ 35°N, 107°E ~ 122.75°E] from March 1, 2010 to February 28, 2015. We removed all stations without valid values in the dates of this embodiment, but kept the stations without collecting valid data for a short time, and performed quality control to delete values that violate the physical laws on the observation data. Then, we aligned the prediction time of the ensemble prediction data with the ground observation time, and interpolated the ensemble prediction data set to the station by the nearest neighbor (K=4) average method to make the spatial format same. Please note that the original resolution of the ensemble prediction is 0.25°, while the resolution of the station data is 0.01°. Finally, the dataset contains 975 nodes and 3837 samples, and there are 5 flood seasons (May-October) in this region. In this embodiment, the dataset is divided into training set, test set and validation set according to the ratio of 7:2:1. In order to be consistent with the business prediction process, the training set and the validation set are strictly separated in date, that is, the time span is disjointed. In addition, in order to make the model more robust, the dataset is divided by season. That is, this embodiment first divides the dates from March 1, 2010 to February 28, 2015 into 20 quarters (spring March-May, summer June-August, autumn September-November, winter December-February of next year), and the test set selects an average of 10% of the dates in the last 4 seasons, and the validation set selects dates ranked 70%-90% in the last 4 seasons. Before training, this embodiment applies StandardScaler (remove mean and scale each variable to unit variance) and MinMaxScaler (scale each variable to 0 and 1) for preprocessing, and applies their inverse transformation after prediction to obtain the final result, in order to reduce the influence of different variable scale differences. And for the precipitation variable in the prediction data and the observation data, an additional exponential normalization transformation is used before the above two normalizations, and the coefficient of the exponential function is 1. Other normalization factors are calculated by season, that is, there are 6x4=12 groups of normalization parameters in the experiment. The station map of the ground meteorological station is constructed according to the nearest neighbor, K=10, and the distance function is "ll2dist", which is provided by Jeff Barton of Johns Hopkins APL. The prediction variable graph is represented by G=(V,E), where V is the set of nodes, |V|=N, and E is the set of edges. The adjacency matrix of the graph is denoted as A∈R N×N .If v i ,vj ∈ V, (v i , v j ) ∈ E, then A i j = 1, otherwise 0. For each set member m, the point set V has a feature matrix Here D = 6 denotes 6 input face variables of the prediction data. The training of the disordered zoom graph convolutional network DZGCN in this embodiment can be represented by the following conversion:

[0112]

[0113] where X = {X (1) , X (M)} is a prediction "super set" with M members of ground weather stations, each X (m) describes D weather variable values of a specific time N stations. This work does not involve the processing of time series information, so the time symbol is omitted in all representations. F denotes the disordered zoom graph convolutional network DZGCN, and Θ denotes the parameters of the disordered zoom graph convolutional network DZGCN. The precipitation forecast generated by the disordered zoom graph convolutional network DZGCN is denoted by Y, and the precipitation station observation value of the ground weather station is denoted by .

[0114] The disordered zoom graph convolutional network (DZGCN) in this embodiment has 3 layers of information extraction blocks. The kernel size is set to 5, and the stride in the zoom module is 3. The disordered zoom graph convolutional network (DZGCN) in this embodiment is trained using the Adam optimizer on a single node of an HPC (high performance computing) cluster containing 4 Tesla V100, with each GPU maximizing the number of input members (up to 50) under its memory limit (16G). The experiment starts with a learning rate of 0.0001. When the validation metric stopmetric remains unchanged for 5 consecutive epochs, the learning rate is reduced to 70% of the value before the minimum learning rate is 2.5E -6 . When the metric remains at the minimum learning rate for 6 epochs, the training is stopped, and the maximum number of epochs does not exceed 200. The experiment uses Pytorch 1.9.1 as the training framework. In this embodiment, the division is: no [0.0, 0.1) mm day -1 , light [0.1, 10.1) mm day -1 , medium [10.1, 25.1) mm day -1 , heavy [25.1, 50.1) mm day -1 , and heavy [50.1, 999) mm day -1Five categories, i.e. L = 5. To verify the performance of the weather forecast post-processing method of the ground meteorological station ensemble forecast in this embodiment, this embodiment mainly compares the disordered zoom graph convolution network (DZGCN) and various existing methods, and the disordered zoom graph convolution network (DZGCN) in this embodiment has 3 information extraction blocks. The kernel size is set to 5, and the stride in the zoom module is 3. The disordered zoom graph convolution network (DZGCN) in this embodiment is trained on a single node of an HPC (high performance computing) cluster containing 4 Tesla V100 using the Adam optimizer, with each GPU maximizing the number of input members (up to 50) under its memory limit (16G). The experiment starts with a learning rate of 0.0001. When the validation indicator stopmetric remains unchanged for 5 consecutive epochs, the learning rate is reduced to 70% of the value before the minimum learning rate is 2.5E -6 When the indicator remains at the minimum learning rate for 6 epochs, the training is stopped, and the maximum number of epochs does not exceed 200. The experiment uses Pytorch1.9.1 as the training framework. In this embodiment, the calculation function expression of the validation indicator stopmetric is:

[0115]

[0116]

[0117]

[0118] In the above formula, ∈ is a constant, L is the number of precipitation intensity categories (in this work, we have no [0.0, 0.1) mm·day -1 , light [0.1, 10.1) mm·day -1 , medium [10.1, 25.1) mm·day -1 , heavy [25.1, 50.1) mm·day -1 and heavy [50.1, 999) mm·day -1 five categories, i.e. L = 5), n(i,j) represents the number of predictions in category i with observed values in category j, N F (i) represents the total number of predictions in category i, N O (j) represents the total number of observed values in category j, and N is the total number of predictions. The final performance test result is shown in Table 2.

[0119] Table 2: Prediction performance results of various methods.

[0120] MAE RMSE HSS ETS 00 ]] ETS 10 ]]> ETS 20 ]] ETS 50 ]]> EM 3.205 7.508 0.387 0.3101 0.207 0.153 0.0498 BP 3.242 9.013 0.474 0.4319 0.176 0.155 0.1467 PM 5.06 11.115 0.221 0.1897 0.0534 0.0459 0.0196 WEM 3.215 8.098 0.102 0.0458 0.190 0.112 0.0188 ASTGCN 2.896 9.318 0.507 0.479 0.115 0.136 0.168 KipfGCN 2.754 9.038 0.519 0.494 0.181 0.084 0.0188 GraphWaveNet 2.815 9.467 0.507 0.482 0.118 0.102 0.0815 DZGCN 2.825 8.967 0.534 0.503 0.144 0.131 0.115

[0121] In the above table, the DZGCN in the first column is the out-of-order zooming graph convolutional network DZGCN of the present embodiment, and the other rows are existing methods for comparison, where EM is the ensemble mean method, that is, taking the mean of all ensemble prediction members as the prediction. PM is the probability matching method, which integrates spatial distribution information and is usually more accurate than the EM method. BP is the BestPercentile method proposed by Dai et al., 2016, which not only considers spatial distribution but also further considers historical experience. ASTGCN is a method for traffic flow prediction using an attention mechanism-based spatial graph attention convolutional network and a time graph attention convolutional network. KipfGCN is a Chebyshev graph convolutional network based on residual connection. GraphWaveNet is a classic algorithm of spatiotemporal graph neural network, which captures temporal relationships through temporal dilated convolution and multi-layer temporal superposition structure. The performance results of the test are all known indicators, including MAE, RMSE, HSS, ETS 00 , ETS 10 , ETS 20 , and ETS 50 , where MAE refers to the mean absolute error, RMSE refers to the root mean square error, HSS refers to the Heidke skill score, and ETS 00 , ETS 10 , ETS 20 , and ETS 50 are all fairness threat scores, and the calculation function expression is:

[0122]

[0123]

[0124] where i refers to the precipitation intensity category. ETSXX represents the ETS under the XX precipitation threshold, for example, ETS00 represents the prediction hit of no precipitation event ([0.0, 0.1) mm·day -1 ); ETS50 represents the precipitation of heavy rain event (≥50.1 mm·day -1 ), and so on.

[0125] From Table 2, among all the indicators, MAE (Mean Absolute Error) and RMSE are the average precipitation forecast performance of each station in the test set. It can be seen that first, the MAE of the intelligent prediction algorithm is improved compared with the traditional method, but the optimal value of RMSE appears in the EM algorithm. According to the characteristics of RMSE that will amplify the error, it can be known that the large error in the EM algorithm is smaller than other methods, which is consistent with the characteristics of the average method that weakens the influence of extreme value. Another method that performs well in the RMSE indicator is WEM, which is also based on historical average, and different weights are selected for different deviation values for correction. Our method ranks third in MAE and RMSE, and achieves the optimal intelligent prediction algorithm in RMSE. In MAE, our method is slightly inferior to GraphWaveNet, and significantly lower than KipfGCN. This is because GraphWaveNet and KipfGCN highly depend on Chebyshev graph convolution operation, which is based on neighborhood message passing, and can be regarded as a neighborhood node smoothing operation, so after smoothing the large precipitation in the deviation position, the average effect is better. From the hit rate, only the disordered zoom convolutional neural network (DZGCN) method is slightly higher than 0.5 in ETS00, and the performance of other intelligent prediction algorithms is 1.39% to 4.7% worse than it. The statistical method performs differently in this indicator, with the best being the BP algorithm, but it still has a significant gap compared with the intelligent prediction algorithm. In the medium precipitation hit (ETS10), the algorithm that is not sensitive to extreme value still performs best: since light and moderate precipitation is the most common type of precipitation event and the high skewness distribution of precipitation, the expected value of precipitation often falls within the range of moderate precipitation. Therefore, the estimated value of precipitation often falls within the range of moderate precipitation. Therefore, the closer the model prediction value is to the expected value, the higher the hit rate of medium precipitation forecast. EM is the method with the closest result to the center of the precipitation value, and KipfGCN is the method closest to the mean value among intelligent algorithms. The most ideal method in heavy rain forecast (ETS20) is the BP method, which is the best method in current business forecast, surpassing all intelligent prediction algorithms. It is worth noting that the performance of intelligent prediction algorithms in heavy rain forecast, first of all, ASTGCN shows amazing heavy rain fitting ability. Although this is accompanied by false positives of precipitation events (low ETS00 performance), its ability to capture special values is worth mentioning. Combined with its structural characteristics, we believe that the attention mechanism integrated in the spatial convolution plays an important role. Through network learning, the attention mechanism gives more weight to special points of outliers, that is, the spatio-temporal attention layer can pay more attention to extreme precipitation. In addition, unlike KipfGCN, which uses MAE as the loss function, GraphWaveNet, which also uses precipitation weighted Huber loss, surprisingly fails to have more improvement in these two indicators.With its network structure, we think that its temporal convolution layer fills too many 0 values when performing arbitrary dilated casual convolution in the time dimension: because it is designed based on the concept of spatiotemporal graph, the padding scheme is designed to make the current time slice only refer to historical information. This will undoubtedly affect the prediction of extreme precipitation. Our method ranks third or fourth in the rainstorm hit, far behind other methods. From the perspective of the comprehensive forecast skill (HSS), the disordered zoom-in graph convolutional network (DZGCN) method achieves the best among all methods, because it has no major defects in category prediction: EM, WEM, KipfGCN, and GraphWaveNet are all insensitive to strong precipitation hits to varying degrees; BP and ASTGCN have obvious false negatives for small precipitation. Overall, the disordered zoom-in graph convolutional network (DZGCN) is a relatively balanced data-driven post-processing ensemble prediction method for precipitation.

[0126] To further study the impact of different modules on the disordered zoom-in graph convolutional network (DZGCN), we designed two different ablation models. The base is a disordered zoom-in graph convolutional network (DZGCN) with 3 layers of information extraction blocks, kernel size of 5, and stride of 3; maeloss refers to replacing the precipitation weighted loss function in base with the MAE function; outtoin refers to replacing the zoom-out module ZO in base with the zoom-in module ZI. From the Figure 6 It can be seen from Table 4 that the performance improvement brought by the precipitation weighted loss mainly reflects in the hit skill, especially the rainstorm hit skill: the ETS50 index of maeloss is much lower than that of the base model. For the accuracy of precipitation values, directly using MAE as the loss function can obtain better results: the maeloss model significantly exceeds the disordered zoom-in graph convolutional network (DZGCN) based model in MAE and RMAE indicators. However, outtoin has a certain degree of decline in each indicator. This is because, on the one hand, the direct weighted average of all members is not conducive to the extraction of potential information among ensemble prediction members; on the other hand, the parameter quantity of the zoom-in module ZI is significantly reduced compared with the zoom-out module ZO, and the network is more difficult to learn rich information.

[0127] Figure 7The different post-processing models show different performances on the spatial distribution of different precipitation indicators, where the same indicator uses the same value range, providing different information from the general results. First, it can be seen that the performance of intelligent precipitation prediction is generally better than that of traditional statistical precipitation post-processing in MAE and RMSE, and the performance of statistical methods is generally weaker at sites with large differences along the coast. Second, in terms of precipitation classification hits, the performance of BP is similar to that of intelligent prediction algorithms, but there is still a gap. It is worth mentioning that ASTGCN does not significantly outperform the disordered zoom-in convolutional network (DZGCN) and GraphWaveNet method in the single-station average ETS50 hit rate, and GraphWaveNet performs slightly better visually than ASTGCN, which is due to the single-station average hit rate being more sensitive to spatial bias than the average indicator in the experimental area, and the size of the false alarm directly penalizes the indicator. Our method still maintains its stable performance, indicating that its high hit rate for heavy precipitation is reliable. Figure 7 For the spatial average of the indicators, from left to right: MAE, RMSE, HSS, ETS 50 ; from top to bottom are DZGCN, GraphWaveNet, ASTGCN, EM, BP, WEM.

[0128] Figure 8The cases (a) is a typical heavy rainfall caused by large-scale convective process, with obvious spatial fractal distribution characteristics of precipitation band and intensity. From the ensemble mean (EM method), it can be seen that the original output of the ensemble prediction is relatively accurate in predicting the center position of the precipitation, but the prediction of the precipitation magnitude is higher. However, most of the intelligent prediction algorithms tend to smooth the extreme value, so the GraphWave and the disordered zoom graph convolutional network (DZGCN) show higher prediction ability than the traditional method. The precipitation event in case (b) is concentrated near the estuary. Due to the complex natural and human environment in this area, the distribution of precipitation events in ground truth observation is relatively discrete: the sites observing the same kind of precipitation event are far apart. This spatial distribution is difficult to accurately predict by global numerical weather prediction model, because it is affected by the process in a short time and very small range. As can be seen from the figure, the ensemble prediction itself is still relatively reliable in predicting its precipitation center, but in terms of overall hit rate, the disordered zoom graph convolutional network (DZGCN) method is superior to other prediction algorithms. In case (c), the precipitation event in the experimental area is affected by two precipitation centers: one is located in the northwest (upper right corner of the figure), and the other is located in the southeast sea area (lower right corner of the figure), forming a unique V-shaped precipitation band. All intelligent prediction methods successfully captured this spatial feature, but in terms of single station accuracy, there is still a certain distance from the actual complex observation. Case (d) still shows the spatial distribution of precipitation affected by two large-scale air masses. Combined with the previous cases (a) and (c), it is proved that the intelligent prediction algorithm has better fitting ability than the traditional statistical method in the case of precipitation caused by large-scale process. They are more accurate for the occurrence of heavy rainfall and no rainfall events. However, referring to cases (b) and (c), when facing local different spatial precipitation distribution, all neural network-based methods including disordered zoom graph convolutional network (DZGCN) are slightly inferior in fitting local differences.

[0129] In summary, the method of the embodiment proposes an irregular position precipitation fitting based on a graph neural network, which extracts the internal spatial information of the ensemble prediction members and the potential correlation between the members using the disordered graph spatio-temporal network structure, and uses other five surface variables to post-process the ensemble precipitation prediction. This method follows the basic framework of GraphWave, but breaks its serial structure in the time dimension. Instead, multiple similar information extraction blocks with the same input are connected in parallel, and an upsampling module and a downsampling module are used instead of the original time convolution layer to extract the correlation between the ensemble members. In terms of spatial feature extraction, this method follows the operation of the existing spatio-temporal graph convolution network, slices the time series, and then uses Chebyshev graph convolution to extract irregular spatial information. In addition, we also propose a precipitation weighted Huber loss function as a guide for network training. The experiment uses the TEGGE ensemble prediction dataset commonly used in business prediction and the real ground monitoring station dataset after quality inspection. The data range is [10.25 °N-35 °N, 107 °E-122.75 °E] with a resolution of 0.01 °, which is much smaller than the 0.25 ° resolution of the ensemble prediction. Experiments have proved that, with the assistance of multiple elements, the disordered zoom-in graph convolution network (DZGCN) method proposed in this embodiment can basically achieve the fitting of actual precipitation: in terms of precipitation accuracy and hit rate of different precipitation categories, compared with four statistical precipitation, commonly used post-processing methods in operation and other three deep learning methods based on graph convolution network or graph spatio-temporal convolution network, this method obtains the best overall hierarchical prediction skill score, and in terms of RMSE, it exceeds other deep learning methods.

[0130] In addition, the embodiment also provides a weather forecast post-processing system for ensemble prediction of ground meteorological stations, which comprises a microprocessor and a memory connected to each other, and the microprocessor is programmed or configured to execute the weather forecast post-processing method for ensemble prediction of ground meteorological stations. In addition, the embodiment also provides a computer readable storage medium, wherein a computer program for programming or configuring a microprocessor to execute the weather forecast post-processing method for ensemble prediction of ground meteorological stations is stored in the computer readable storage medium.

[0131] Those skilled in the art will appreciate that embodiments of the application can be readily used as a method, apparatus, or computer program product. Accordingly, the application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the application can take the form of a computer program product on one or more computer readable storage media (including, but not limited to, disk memory, CD-ROMs, optical storage devices, etc.) embodying computer readable program code. The application is described in reference to the flow diagrams and / or block diagrams of the methods, apparatus (systems) and computer program products according to embodiments of the application illustrated in the attached drawings, which flow diagrams and / or block diagrams Figure 1 one or more functions specified in the flow diagram and / or block diagram. These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the function specified in the flow diagram and / or block diagram. Figure 1 one or more functions specified in the flow diagram and / or block diagram. These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the function specified in the flow diagram and / or block diagram. Figure 1 one or more functions specified in the flow diagram and / or block diagram. These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the function specified in the flow diagram and / or block diagram. Figure 1 one or more functions specified in the flow diagram and / or block diagram. These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the function specified in the flow diagram and / or block diagram. Figure 1 one or more functions specified in the flow diagram and / or block diagram. These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the function specified in the flow diagram and / or block diagram. Figure 1 one or more functions specified in the flow diagram and / or block diagram. These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the function specified in the flow diagram and / or block diagram.

[0132] The above description is only preferred embodiments of the application, the protection scope of the application is not limited to the above-mentioned embodiments, any technical scheme falling within the idea of the application shall fall within the protection scope of the application. It should be noted that, for ordinary skilled in the art, some improvements and refinements without departing from the principle of the application, these improvements and refinements shall be considered as the protection scope of the application.

Claims

1. A method of post-processing weather forecasts of a collection of ground weather stations ensemble forecasts, characterized in that, Comprise: Collective forecasting of ground weather stations Linear coding input disordered zoom graph convolution network DZGCN, in the disordered zoom graph convolution network DZGCN, a plurality of information extraction blocks with different network parameters are used to extract information respectively, then the features are extracted by the graph convolution layer, and then the precipitation of the ground weather station is predicted according to all the extracted features by using the prediction block. When the plurality of information extraction blocks with different network parameters respectively extract information, the information extraction block extracting information comprises: using an amplification module ZI as a gate to extract irregular position distribution space feature information of linear coding and global to local correlation information of ensemble member dimension, using a reduction module ZO as a filter to extract linear coding space features and local to global association information of member dimension, using a gating unit to integrate the extracted global to local correlation information and local to global association information, then extracting features of the integrated information through a graph convolution layer, and finally outputting after residual connection with information extracted by the input information extraction block and batch normalization; The amplification module ZI includes a plurality of graph convolution layers, and any i-th l The function expression of the i-th graph convolution layer is: ， In the above formula, is the global-to-local correlation information of the output of the first layer of the upscaling module ZI, l ReLU is an activation function, is a convolution kernel with a member dimension of 1x1, is a standard convolution operation, represents the global-to-local correlation information of the output of the first layer of the upscaling module ZI, l- 1layer output of the upscaling module ZI, is the number of layers of the upscaling module ZI, and the input of the first layer of the upscaling module ZI is where Linear is linear encoding, is a collective forecast of the ground meteorological station; The zoom-out module ZO comprises a plurality of graph convolution layers, and any i-th l The function expression of the i-th graph convolution layer is: ， ， In the above formula, is the first layer output information of the downsizing module ZO, l is the first layer output information of the downsizing module ZO, is the first layer output information of the downsizing module ZO, is the first layer output information of the downsizing module ZO, is the first layer output information of the downsizing module ZO, is the first layer output information of the downsizing module ZO, is the first layer output information of the downsizing module ZO, is the first layer output information of the downsizing module ZO, l is the first layer output information of the downsizing module ZO, is the first layer output information of the downsizing module ZO, l is the first layer output information of the downsizing module ZO, is the first layer output information of the downsizing module ZO, is the first layer output information of the downsizing module ZO, l is the first layer output information of the downsizing module ZO, is the first layer output information of the downsizing module ZO, is the first layer output information of the downsizing module ZO, l is the first layer output information of the downsizing module ZO, is the first layer output information of the downsizing module ZO, is the first layer output information of the downsizing module ZO, is the first layer output information of the downsizing module ZO, l is the first layer output information of the downsizing module ZO, is the first layer output information of the downsizing module ZO, is the first layer output information of the downsizing module ZO, is the first layer output information of the downsizing module ZO, is the first layer output information of the downsizing module ZO, 2. The post-processing method of the nowcast of a weather forecast of a collection of ground weather stations according to claim 1, characterized in that, The function expression for integrating the global to local correlation information and the local to global association information by the gating unit is: ， In the above formula, is the integrated information, g is the output activation function, is the global-to-local correlation information output by the first l layer of the amplification module ZI, is the element product, sigmoid is the activation function, is the local-to-global correlation information output by the zoom-out module ZO.

3. The post-processing method of the nowcast of a weather forecast of a collection of ground weather stations according to claim 1, characterized in that, The function expression for predicting the precipitation of the ground meteorological station according to all the extracted features by the prediction block is: ， In the above formula, is the predicted precipitation at the ground weather station, is a fully connected network, and mean is an averaging operation, is an intermediate variable, and has: ， ， ， In the above formula, is a full connection network, is an intermediate variable, Stack is a stacking operation, ~ respectively are features obtained by performing supplementary feature extraction on the information output by the first to B-th information extraction blocks through the graph convolution layer, is a feature obtained by performing supplementary feature extraction on the information output by the arbitrary b-th information extraction block through the graph convolution layer, and ReLU is an activation function, is a graph convolution operation of the graph convolution layer, is a convolution kernel of the graph convolution layer, is information output by the b-th information extraction block, and B is the number of information extraction blocks.

4. The post-processing method of the nowcast of a weather forecast of a collection of ground weather stations according to claim 1, characterized in that, The graph convolution layer adopts Chebyshev polynomials as a convolution kernel, and any i-th l The input feature of the layer graph convolution layer is: ， In the above formula, respectively, the input information member dimension of the first l layer graph convolution layer, N is the number of graph nodes, the input feature number of the first l layer graph convolution layer, the member dimension size of the input information of the first l layer graph convolution layer; and the function expression for generating any m-th slice of the first l layer graph convolution layer is:​ ， In the above formula, is the first l layer graph convolutional layer output feature information of the m slice, ReLU is an activation function, is the first l layer graph convolutional layer of the graph convolutional operation of the convolutional kernel of the first l layer graph convolutional layer, is the first l layer graph convolutional layer input feature information of the m slice; Any of l The function expression of the output features generated by the layer graph convolutional layer is: ， In the above formula, is the first l layer generates an output feature, Concat is a connection operation, is the first l layer of the graph convolution layer M l slices.​ 5. The post-processing method of the nowcast of a weather forecast of a collection of ground weather stations according to claim 1, characterized in that, The loss function used by the disordered zooming graph convolutional network DZGCN during training is: ， In the above formula, is a loss function, is a predicted precipitation at a ground weather station, is a true value of the precipitation at the ground weather station, is a weight vector, is a Huber loss, and has: ， ， in the above formula, w i is the i-th component of the weight vector, max is the maximum value, min is the minimum value, is the i-th component of the true value of the precipitation at the ground weather station is the i-th component of the true value of the precipitation at the ground weather station i is the i-th component of the true value of the precipitation at the ground weather station is a function defined by the argument, , , , , , is a hyperparameter, , , , , , ∈ [0, 1], is the i-th component of the true value of the precipitation at the ground weather station is the i-th component of the true value of the precipitation at the ground weather station i is the i-th component of the true value of the precipitation at the ground weather station 6. A weather forecast post-processing system for a ground weather station ensemble forecast, comprising a microprocessor and a memory interconnected, characterized in that, The microprocessor is programmed or configured to perform the weather forecast post-processing method of the ground meteorological station ensemble prediction according to any one of claims 1-5.

7. A computer-readable storage medium having stored therein a computer program, characterized in that, The computer program is used for being programmed or configured by the microprocessor to perform the weather forecast post-processing method of the ground meteorological station ensemble prediction according to any one of claims 1-5.

Images