Training method, using method and related device of high-resolution weather forecast model

Abstract

The embodiment of the invention discloses a training method, a use method and a related device of a high-resolution weather forecast model. The training method comprises the following steps: acquiring a pre-collected training sample data set; inputting the first forecast data of the low-resolution rainfall forecast in the training sample data set, the atmospheric precipitable water amount data obtained by the Beidou data inversion and the surface meteorological observation data into a preset deep learning model for high-resolution weather forecast model training to obtain second forecast data of the high-resolution weather forecast; determining whether the deep learning model converges or not by using the second forecast data, a first label in the training sample data set and a preset loss function; and if so, taking the converged deep learning model as a high-resolution weather forecast model. Through the above mode, high-resolution weather forecast based on deep learning and fusing Beidou and meteorological data is realized, the forecast precision is improved, and the accuracy of short temporary rainfall forecast is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of meteorological forecasting, and in particular to a training method, a use method and a related device of a high-resolution weather forecast model. Background Art

[0002] Current weather forecasts mainly rely on numerical weather forecast models, which usually have low spatial resolution and find it difficult to accurately predict extreme weather phenomena such as local short-term heavy rainfall.

[0003] In order to improve the forecast accuracy, some studies have also tried to improve the forecast effect by fusing multi-source data (such as satellite, radar, ground observation, etc.). For example, Assiri (2022) used a random forest regression model to fuse satellite and precipitation observation data to improve the accuracy of precipitation prediction, but the accuracy of short-term precipitation forecast still needs to be improved.

[0004] How to achieve high-resolution weather forecasting is still an urgent problem to be solved. Summary of the invention

[0005] The main purpose of the present invention is to provide a training method, a use method and related devices for a high-resolution weather forecast model, which can solve the problem that the accuracy of short-term precipitation forecast in the prior art still needs to be improved.

[0006] To achieve the above object, the first aspect of the present invention provides a training method for a high-resolution weather forecast model, the training method comprising:

[0007] Acquire a pre-collected training sample data set, wherein the training sample data set includes at least training sample data and actual precipitation labels of the training sample data, wherein the training sample data includes at least first forecast data of low-resolution precipitation forecasts of multiple weather precipitation models, atmospheric precipitable water data obtained by inversion of Beidou data, and surface meteorological observation data, and the actual precipitation label includes a first label of high-resolution actual precipitation data;

[0008] Inputting the first forecast data, atmospheric precipitable water data and surface meteorological observation data into a preset deep learning model to perform model training for high-resolution weather forecast, thereby obtaining second forecast data for high-resolution weather forecast;

[0009] Determining whether the deep learning model converges using the second forecast data, the first label, and a preset loss function;

[0010] If the deep learning model converges, the converged deep learning model is used as the high-resolution weather forecast model.

[0011] In a feasible implementation, the deep learning model includes a first model for data optimization, a second model for data fusion, and a third model for downscaling that are connected in sequence, and the actual precipitation label also includes a second label of low-resolution actual precipitation data. Then, the first forecast data, atmospheric precipitable water data, and surface meteorological observation data are input into a preset deep learning model for model training of high-resolution weather forecasting, and the first forecast result is obtained, including:

[0012] Inputting the first forecast data, atmospheric precipitable water data and surface meteorological observation data into a first model for data optimization model training to obtain optimized forecast data of low-resolution precipitation forecasts of various weather precipitation models;

[0013] Determining whether the first model converges using the first loss value between the optimized forecast data and the second label;

[0014] If the first loss value indicates that the first model has not converged, the model parameters of the first model are updated using the first loss value to obtain an updated first model, and the step of inputting the first forecast data, atmospheric precipitable water data, and surface meteorological observation data into the first model for data optimization model training is returned to obtain optimized forecast data of low-resolution precipitation forecasts of various weather precipitation models;

[0015] If the first loss value indicates that the first model has converged, a converged first model is obtained, and the accuracy information of each weather forecast model and the optimized forecast data are used for weighted fusion to obtain first fused forecast data for low-resolution precipitation forecast, and the optimized forecast data corresponding to each weather precipitation model is input into the second model for data fusion model training to obtain second fused forecast data for low-resolution precipitation forecast;

[0016] Determining whether the second model converges by using a second loss value between the first fused forecast data and the second fused forecast data;

[0017] If the second loss value indicates that the second model has not converged, the model parameters of the second model are updated using the second loss value to obtain an updated second model, and the step of inputting the optimized forecast data corresponding to each weather precipitation model into the second model for data fusion model training to obtain second fused forecast data of low-resolution precipitation forecast is returned to be executed;

[0018] If the second loss value indicates that the second model has converged, a converged second model is obtained, and each of the second fused forecast data is input into the third model for downscaled model training to obtain second forecast data for high-resolution precipitation forecast.

[0019] In a feasible implementation, the using the second forecast data, the first label and the preset loss function to determine whether the deep learning model converges includes:

[0020] Determining a third loss value of the third model based on the second forecast data, the first label and the loss function;

[0021] If the third loss value is not lower than the preset loss threshold, it is determined that the third model has not converged, and the model parameters of the third model are updated using the third loss value to obtain an updated third model, and the step of inputting each of the second fused forecast data into the third model for downscaling model training to obtain the second forecast data of high-resolution precipitation forecast is returned to execute;

[0022] If the third loss value is lower than a preset loss threshold, it is determined that the third model converges, and the converged deep learning model includes a converged first model, a second model, and a third model connected in sequence.

[0023] In a feasible implementation, the training method further includes:

[0024] The first forecast data, atmospheric precipitable water data and surface meteorological observation data are subjected to data preprocessing to obtain the preprocessed first forecast data, atmospheric precipitable water data and surface meteorological observation data, the training sample data include the preprocessed first forecast data, atmospheric precipitable water data and surface meteorological observation data, and the preprocessing includes at least normalization processing and interpolation processing.

[0025] To achieve the above object, the second aspect of the present invention provides a method for using a high-resolution weather forecast model, the method comprising:

[0026] Obtain current forecast data of low-resolution precipitation forecasts from multiple weather and precipitation models, current atmospheric precipitable water data obtained by Beidou data inversion, and current surface meteorological observation data;

[0027] The current forecast data, the current atmospheric precipitable water data and the current surface meteorological observation data are input into a high-resolution weather forecast model for high-resolution weather forecasting to obtain third forecast data for the high-resolution weather forecast. The high-resolution weather forecast model is trained using the training method shown in the first aspect and any feasible implementation method.

[0028] To achieve the above object, the third aspect of the present invention provides a training device for a high-resolution weather forecast model, the training device comprising:

[0029] Data acquisition module: used to acquire a pre-collected training sample data set, wherein the training sample data set includes at least training sample data and actual precipitation labels of the training sample data, wherein the training sample data includes at least first forecast data of low-resolution precipitation forecasts of multiple weather precipitation models, atmospheric precipitable water data obtained by inversion of Beidou data, and surface meteorological observation data, and the actual precipitation label includes a first label of high-resolution actual precipitation data;

[0030] Model training module: used for inputting the first forecast data, atmospheric precipitable water data and surface meteorological observation data into a preset deep learning model to perform model training for high-resolution weather forecast, and obtain second forecast data for high-resolution weather forecast;

[0031] Convergence judgment module: used to determine whether the deep learning model converges by using the second forecast data, the first label and a preset loss function;

[0032] Model determination module: used to use the converged deep learning model as the high-resolution weather forecast model if the deep learning model converges.

[0033] In a feasible implementation, the deep learning model includes a first model for data optimization, a second model for data fusion, and a third model for downscaling that are connected in sequence, and the actual precipitation label also includes a second label of low-resolution actual precipitation data, and the training device also includes:

[0034] Data optimization module: used for inputting the first forecast data, atmospheric precipitable water data and surface meteorological observation data into the first model for data optimization model training, and obtaining optimized forecast data of low-resolution precipitation forecasts of various weather precipitation models; using the first loss value between the optimized forecast data and the second label to determine whether the first model has converged; if the first loss value indicates that the first model has not converged, using the first loss value to update the model parameters of the first model to obtain an updated first model, and returning to execute the step of inputting the first forecast data, atmospheric precipitable water data and surface meteorological observation data into the first model for data optimization model training, and obtaining optimized forecast data of low-resolution precipitation forecasts of various weather precipitation models;

[0035] Data fusion module: used for obtaining a converged first model if the first loss value indicates that the first model has converged, and performing weighted fusion using the accuracy information of each weather forecast model and the optimized forecast data to obtain first fused forecast data for low-resolution precipitation forecast, and inputting the optimized forecast data corresponding to each weather and precipitation model into the second model for data fusion model training to obtain second fused forecast data for low-resolution precipitation forecast; using a second loss value between the first fused forecast data and the second fused forecast data to determine whether the second model has converged; if the second loss value indicates that the second model has not converged, updating the model parameters of the second model using the second loss value to obtain an updated second model, and returning to execute the step of inputting the optimized forecast data corresponding to each weather and precipitation model into the second model for data fusion model training to obtain second fused forecast data for low-resolution precipitation forecast;

[0036] Data downscaling module: used for obtaining a converged second model if the second loss value indicates that the second model has converged, and inputting each of the second fused forecast data into the third model for downscaling model training to obtain second forecast data for high-resolution precipitation forecast.

[0037] To achieve the above-mentioned object, the fourth aspect of the present invention provides a device for using a high-resolution weather forecast model, the device comprising:

[0038] Data acquisition module: used to obtain the current forecast data of low-resolution precipitation forecasts of multiple weather precipitation models, the current atmospheric precipitable water data obtained by Beidou data inversion, and the current surface meteorological observation data;

[0039] Weather forecast module: used to input the current forecast data, current atmospheric precipitable water data and current surface meteorological observation data into a high-resolution weather forecast model for high-resolution weather forecast, and obtain third forecast data for high-resolution weather forecast. The high-resolution weather forecast model is trained using the training method shown in the first aspect and any feasible implementation method.

[0040] To achieve the above-mentioned purpose, the fifth aspect of the present invention provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor executes the steps of the training method shown in the first aspect and any feasible implementation method or the using method shown in the second aspect.

[0041] To achieve the above-mentioned purpose, the sixth aspect of the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the training method shown in the first aspect and any feasible implementation method or the using method shown in the second aspect.

[0042] The embodiments of the present invention have the following beneficial effects:

[0043] The present invention provides a training method for a high-resolution weather forecast model, the training method comprising: obtaining a pre-collected training sample data set, the training sample data set at least comprising training sample data and actual precipitation labels of the training sample data, the training sample data at least comprising first forecast data of low-resolution precipitation forecasts of multiple weather precipitation models, atmospheric precipitable water data obtained by inverting Beidou data, and surface meteorological observation data, the actual precipitation label comprising a first label of high-resolution actual precipitation data; inputting the first forecast data, atmospheric precipitable water data, and surface meteorological observation data into a preset deep learning model to perform model training for high-resolution weather forecast, and obtaining second forecast data for high-resolution weather forecast; using the second forecast data, the first label, and a preset loss function to determine whether the deep learning model converges; if the deep learning model converges, using the converged deep learning model as the high-resolution weather forecast model. Through the above method, high-resolution weather forecast based on deep learning integration of Beidou and meteorological data, the first forecast data of low-resolution precipitation forecasts of multiple weather precipitation models, atmospheric precipitable water data obtained by inverting Beidou data, surface meteorological observation data, and other multi-source data are efficiently integrated to achieve high-resolution weather forecast, improve forecast accuracy, and improve the accuracy of short-term precipitation forecast. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0045] in:

[0046] Figure 1 A flowchart of a training method for a high-resolution weather forecast model in an embodiment of the present invention;

[0047] Figure 2 A model architecture diagram of a high-resolution weather forecast model in an embodiment of the present invention;

[0048] Figure 3A flowchart of a method for using a high-resolution weather forecast model in an embodiment of the present invention;

[0049] Figure 4 A structural block diagram of a training device for a high-resolution weather forecast model in an embodiment of the present invention

[0050] Figure 5 A structural block diagram of a device for using a high-resolution weather forecast model in an embodiment of the present invention;

[0051] Figure 6 4 is a structural block diagram of a computer device in an embodiment of the present invention. DETAILED DESCRIPTION

[0052] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0053] See also Figure 1 , Figure 1 The flowchart of a training method of a high-resolution weather forecast model in an embodiment of the present invention is shown in FIG. 1 . The method can be applied to a terminal or a server. The terminal can be a desktop terminal or a mobile terminal. The mobile terminal can be at least one of a mobile phone, a tablet computer, a laptop computer, etc. The server can be implemented by an independent server or a server cluster composed of multiple servers. This embodiment is described by applying to a terminal as an example. Figure 1 The training methods shown include:

[0054] 101. Acquire a pre-collected training sample data set, wherein the training sample data set includes at least training sample data and actual precipitation labels of the training sample data, wherein the training sample data includes at least first forecast data of low-resolution precipitation forecasts of multiple weather precipitation models, atmospheric precipitable water data obtained by inverting Beidou data, and surface meteorological observation data, and the actual precipitation label includes a first label of high-resolution actual precipitation data;

[0055] It is understandable that the weather forecast may be a forecast of weather conditions such as temperature, wind speed or precipitation. In this application, the examples are mainly centered around precipitation conditions, which may be short-term precipitation conditions. In order to improve the high-resolution forecast of short-term precipitation forecasts, this application collects a large amount of historical precipitation forecast data and actual precipitation data in advance, wherein the precipitation forecast data may be the first forecast data P of low-resolution weather forecasts of multiple weather precipitation models i. i, the actual precipitation data can be high-resolution actual precipitation data, and the weather precipitation model includes but is not limited to the weather precipitation models of the meteorological centers of various countries such as the European Center for Mesoscale Weather Forecasts (ECMWF), the China Meteorological Administration (CMA), the National Center for Environmental Prediction (NCEP) of the United States and the Japan Meteorological Agency (JMA). Then, the actual precipitation label is marked for the first forecast data, and the actual precipitation label can be the high-resolution actual precipitation data collected above, for example, the low-resolution first forecast data at the nth moment is marked with the high-resolution actual precipitation data at the nth moment, wherein n can be a certain time node in history, and the time interval between specific time nodes can be adaptively selected according to the actual situation, and is not limited here.

[0056] Furthermore, in order to improve the forecast accuracy, the present application also pre-collects the atmospheric precipitable water data and surface meteorological observation data obtained by inverting Beidou data as training sample data to participate in the training of the high-resolution weather forecast model. Specifically, the atmospheric precipitable water data and surface meteorological observation data at the nth moment can be used to optimize the first forecast data at the nth moment, so as to realize high-resolution weather forecast with multi-source data fusion.

[0057] Through the above training sample data, the relationship between historical low-resolution precipitation forecast data and historical high-resolution actual precipitation data can be learned during the model training process, so as to achieve high-resolution weather forecast by relying on low-resolution precipitation forecast data.

[0058] In a feasible implementation, in order to improve the model training efficiency, the collected original training sample data will also be preprocessed, that is, the training method also includes: preprocessing the first forecast data, atmospheric precipitation data and surface meteorological observation data to obtain the preprocessed first forecast data, atmospheric precipitation data and surface meteorological observation data, the training sample data includes the preprocessed first forecast data, atmospheric precipitation data and surface meteorological observation data, and the preprocessing includes at least normalization and interpolation. Specifically, low-resolution precipitation forecast data P of multiple weather precipitation models are obtained. i , atmospheric precipitable water data (PrecipitableWaterVapor, PWV) obtained by Beidou data inversion Beidou) and surface meteorological observation data (surface meteorological observation data include but are not limited to temperature T, air pressure P, humidity H, wind speed W, wind direction V). Through data normalization, ensure that the data from different data sources have the same numerical range, thereby improving the training effect of the model. Then, use the Kriging interpolation method to unify all data to the same spatial resolution and time step to ensure data consistency. This step is to ensure that the subsequent deep learning model can efficiently process and optimize the forecast data. Finally, collect actual low-resolution and high-resolution precipitation data as training labels.

[0059] 102. Input the first forecast data, atmospheric precipitable water data, and surface meteorological observation data into a preset deep learning model to perform model training for high-resolution weather forecast, thereby obtaining second forecast data for high-resolution weather forecast;

[0060] It should be noted that the deep learning model may include a first model for data optimization, a second model for data fusion, and a third model for downscaling, which are connected in sequence. The first model may be a dual-channel convolutional neural network, the second model may be a self-attention learning network, and the third model may be a downscaling network. Figure 2 , Figure 2 is a model architecture diagram of a high-resolution weather forecast model in an embodiment of the present invention, such as Figure 2 The model architecture shown in the figure is composed of three deep learning networks, namely, a two-channel convolutional neural network for data optimization (referred to as data optimization two-channel network), a self-attention learning network for data fusion (referred to as data fusion network), and a downscaling network for data downscaling (referred to as downscaling network). Figure 2 The data optimized dual-channel network shown may include an input layer, a convolution layer, a fusion layer, two fully connected layers, and an output layer connected in sequence, wherein the input layer is dual-channel, and the convolution layer is also dual-channel, wherein the training sample data is the input of the input layer, and the output layer is connected to the input layer of the data fusion network; Figure 2 The data fusion network shown may include an input layer connected in sequence, several feature extraction layers connected in series, a self-attention mechanism layer, two fully connected layers, and an output layer, wherein the output layer is connected to the input layer of the downscaling network; Figure 2 The downscaling network shown may include an input layer, a high-frequency feature extraction enhancement layer, a feature combination layer, an upsampling layer, a refinement layer, and an output layer connected in sequence, wherein the output layer is used to output a high-resolution precipitation forecast.

[0061] Furthermore, the actual precipitation label includes not only the first label of the actual precipitation data with high resolution but also the second label of the actual precipitation data with low resolution, then step 102 includes the following steps A01 to A07:

[0062] A01, inputting the first forecast data, atmospheric precipitable water data and surface meteorological observation data into a first model for data optimization model training, and obtaining optimized forecast data of low-resolution precipitation forecasts of various weather precipitation models;

[0063] A02, using the first loss value between the optimized forecast data and the second label to determine whether the first model converges;

[0064] A03. If the first loss value indicates that the first model has not converged, the model parameters of the first model are updated using the first loss value to obtain an updated first model, and the step of inputting the first forecast data, atmospheric precipitable water data, and surface meteorological observation data into the first model for data optimization model training is returned to obtain optimized forecast data of low-resolution precipitation forecasts of various weather precipitation models;

[0065] A04. If the first loss value indicates that the first model converges, a converged first model is obtained, and the accuracy information of each weather forecast model and the optimized forecast data are used for weighted fusion to obtain first fused forecast data of low-resolution precipitation forecast, and the optimized forecast data corresponding to each weather precipitation model is input into the second model for data fusion model training to obtain second fused forecast data of low-resolution precipitation forecast;

[0066] It should be noted that after data preprocessing, the low-resolution precipitation forecast data P of multiple weather precipitation models are input. i , Precipitable Water Vapor (PWV) data obtained from Beidou data inversion Beidou ) and surface meteorological observation data (including temperature T, air pressure P, humidity H, wind speed W, wind direction V), and use deep learning models and actual observed meteorological data to optimize precipitation forecast data, that is, use atmospheric precipitable water data and surface meteorological observation data to optimize the low-resolution precipitation forecast data P of each weather precipitation model. i .

[0067] Specifically, a two-channel convolutional neural network is constructed with the following structure:

[0068] Input layer: Channel 1: Receive low-resolution precipitation forecast data P i ; Channel 2: Beidou water vapor data PWV Beidou And surface meteorological observation data (including temperature T, air pressure P, humidity H, wind speed W, wind direction V).

[0069] Convolutional layer: Channel 1 and channel 2 are both set with 4 convolutional layers (32 filters per layer, convolution kernel size 3x3) to extract spatial features in the data. The convolution operation formula (1) is:

[0070] C=φ(W*P i +b) #(1)

[0071] Among them, φ is the activation function, using ReLU; W is the convolution kernel, b is the bias, and * represents the convolution operation. The convolution layer captures the local features of the data by applying filters, which helps to extract the complex patterns in the weather forecast data. Channel 1 obtains feature F forecast,i , channel 2 gets feature F obs .

[0072] Fusion layer: A fully connected layer (64 neurons) that combines the output features F of the two channels forecast,i and F obs Connect and generate comprehensive features F combined .

[0073] Fully connected layers: Through 2 fully connected layers (128 neurons in each layer).

[0074] Output layer: a fully connected layer (1 neuron) that outputs optimized low-resolution precipitation forecast data (i.e. the optimized forecast data mentioned above) and accuracy information.

[0075] In order to train the model, the actual low-resolution precipitation data O is used as the label (i.e., the second label mentioned above), and the mean square error (MSE) loss function is used to measure the gap between the model prediction and the actual observation. This gap is the first loss value, and the formula is as follows (2):

[0076]

[0077] The back propagation algorithm is used to optimize parameters and the Adam optimizer is used to minimize the loss function, thereby optimizing the model parameters and ultimately generating more accurate forecast data.

[0078] After the model training is completed, the optimization results of each weather precipitation forecast model i are calculated The error between the predicted and actual precipitation data O is used to obtain the accuracy information MSE of each weather precipitation forecast model i. i .

[0079] Finally, the optimized forecast data of low-resolution precipitation forecast of multiple weather precipitation models i are obtained. And its precision information MSE iEnter the next step, which includes optimizing the forecast data corresponding to each weather precipitation model The second model is input to perform data fusion model training to obtain the second fusion forecast data P of low-resolution precipitation forecast. fusion , and the accuracy information MSE of each weather forecast model i i The optimized forecast data are weighted and fused to obtain the first fused forecast data of low-resolution precipitation forecast, and the first fused forecast data and the second fused forecast data P are directly weighted. fusion The loss between them is used to train the second model.

[0080] A05. Determine whether the second model converges by using a second loss value between the first fused forecast data and the second fused forecast data;

[0081] A06. If the second loss value indicates that the second model has not converged, the model parameters of the second model are updated using the second loss value to obtain an updated second model, and the step of inputting the optimized forecast data corresponding to each weather precipitation model into the second model for data fusion model training to obtain second fused forecast data of low-resolution precipitation forecast is returned to be executed;

[0082] A07. If the second loss value indicates that the second model converges, a converged second model is obtained, and the second fused forecast data is input into the third model for downscaled model training to obtain second forecast data for high-resolution precipitation forecast.

[0083] It should be noted that the forecast is optimized using multiple weather precipitation models obtained in the previous step. And its precision information MSE i , build a deep learning model to weightedly fuse multiple optimized forecast data. The deep learning model here is a self-attention learning network. The specific structure is as follows:

[0084] Input layer: receiving various optimized precipitation forecast data and its precision information MSE i .

[0085] Feature extraction layer: Four convolutional layers (64 filters per layer, convolution kernel size 3x3) are used to extract the features of each optimized forecast data. The formula is as follows (3):

[0086]

[0087] Self-attention mechanism layer: Generate a dynamic weight matrix for each data source through the self-attention mechanism. Automatically adjust the weights of different data sources based on the local characteristics of the input features. This fine-grained local feature processing will make it easier to capture local heavy rainfall phenomena than existing methods:

[0088] (1) The input feature tensor F i Mapped to the query vector (Q), key vector (K), and value vector (V) in the self-attention mechanism. Generate a dynamic weight matrix through the following steps:

[0089] For each feature map F i , using independent weight matrices W Q , W K , and W V , use Xavier to initialize the weight matrix and assign parameters, and then use the weight matrix to i Mapped to low-dimensional query, key and value space, as shown in formula (4):

[0090] Q i =W Q ·F i ,K i =W K ·F i ,V i =W V ·F i #(4)

[0091] Q i , K i and V i The dimension is the feature map size × dk, where dk is the feature dimension after dimensionality reduction, which is set to 32 here.

[0092] (2) Perform a dot product operation on the query and key of all feature maps to calculate the similarity between different feature maps:

[0093] First, for each feature map Q i and K i Calculate the similarity score as shown in formula (5):

[0094]

[0095] In formula (5), the softmax operation ensures that the similarity scores of each feature map are normalized to form a probability distribution. i,j is a weight matrix with size of input and size of forecast data graphs, representing the feature graph F i and F jSimilarity in feature space. A high value in the attention weight matrix indicates that the local features of these feature maps at a specific location have high similarity.

[0096] (3) Use the weights to sum the value vectors. For each query position, the weight matrix α i,j , which is used to weight the feature map F j The value vector V j , as shown in formula (6):

[0097]

[0098] (4) All F' i Further combination to generate the final fusion feature F fusion , as shown in formula (7):

[0099] F fusion =concat(F'1,F'2,…,F' n )#(7)

[0100] Concat is a commonly used operation, which refers to splicing (connecting) two or more tensors in a specific dimension. fusion Go to the next level.

[0101] Fully connected layer: F fusion It passes through 2 fully connected layers (64 neurons in each layer) followed by a ReLU activation function.

[0102] Output layer: The fused features are finally passed through a fully connected layer (1 neuron) to generate the final fused prediction data P fusion .

[0103] Among them, direct weighted fusion is used (the weight here is the accuracy MSE obtained during data optimization i The model is trained with weather precipitation forecast data (determined by the model) as labels. In this way, the model can effectively fuse all input features into a complete precipitation forecast result and ensure that the output reflects the weighted information of local weather information and multi-source data, thereby achieving higher-precision precipitation forecasts. The loss function uses the mean square error (MSE), which is used to calculate the loss between the predicted value and the label (i.e., the second loss value mentioned above). Parameter optimization is performed through the back-propagation algorithm, using the Adam optimizer.

[0104] Finally, the fused high-precision precipitation forecast data P is obtained. fusion Entering the next step, the next step includes converting the second fused forecast data P fusionThe third model is input to perform downscaling model training to obtain the second forecast data P of high-resolution precipitation forecast. HR .

[0105] 103. Determine whether the deep learning model converges by using the second forecast data, the first label and a preset loss function;

[0106] In a feasible implementation, the using the second forecast data, the first label and the preset loss function to determine whether the deep learning model converges includes:

[0107] B01. Determine a third loss value of the third model based on the second forecast data, the first label and the loss function;

[0108] B02. If the third loss value is not lower than a preset loss threshold, it is determined that the third model has not converged, and the model parameters of the third model are updated using the third loss value to obtain an updated third model, and the step of inputting each of the second fused forecast data into the third model for downscaling model training to obtain the second forecast data of high-resolution precipitation forecast is returned to be executed;

[0109] B03. If the third loss value is lower than a preset loss threshold, it is determined that the third model converges, and the converged deep learning model includes a converged first model, a second model, and a third model connected in sequence.

[0110] It should be noted that the fused high-precision precipitation forecast data P obtained in the previous step is used fusion ,In order to improve the resolution of forecast data, a downscaling deep learning network is constructed again. The steps are as follows:

[0111] Input layer: receives the fused low-resolution forecast data P fusion .

[0112] High-frequency feature extraction enhancement layer: First, the input low-resolution data is decomposed into high-frequency and low-frequency components by wavelet decomposition: the high-frequency component retains the detail information and is processed by independent convolution; the low-frequency component retains the overall trend and is processed by the standard convolution layer. This can significantly enhance the model's ability to capture small-scale phenomena such as local heavy rainfall and extreme weather.

[0113] (1) P fusion When performing wavelet decomposition, the data is decomposed into four subbands:

[0114] Low-frequency component LL: The data is filtered horizontally and vertically through a low-pass filter to obtain the overall structural information and indicate the main trend of the data.

[0115] High frequency components LH, HL, HH: These components represent the detail information in the horizontal and vertical directions respectively.

[0116] LH captures low-frequency horizontal and high-frequency vertical information.

[0117] HL captures high-frequency information horizontally and low-frequency information vertically.

[0118] HH captures high-frequency information in both horizontal and vertical directions.

[0119] (2) Extract high-frequency and low-frequency components:

[0120] Low frequency component P low : The LL subband is used as the low-frequency component to represent the main trends and spatial structure characteristics of the input.

[0121] High frequency component P high : It consists of LH, HL and HH sub-bands, representing details and local change information.

[0122] Then LH, HL and HH are concatenated to form a high-frequency feature tensor containing multiple directional information. The formula is as follows (8):

[0123] P high =concat(LH,HL,HH) #(8)

[0124] Furthermore, the high frequency component P high After three convolution layers (the first convolution layer: 32 filters, the convolution kernel size is 3×3, the second convolution layer: 64 filters, the convolution kernel size is 3×3, the third convolution layer: 128 filters, the convolution kernel size is 3×3), the detail features are enhanced; the low-frequency component P low Then it is processed through two layers of convolution (the first layer of convolution: 32 filters, the convolution kernel size is 3×3, the second layer of convolution: 64 filters, the convolution kernel size is 3×3) to extract large-scale features. The processed components enter the next layer.

[0125] Feature combination layer: The processed high-frequency and low-frequency components are recombined and the final fusion feature F is generated through the fully connected layer. enhanced .

[0126] Upsampling layer: After two layers of deconvolution layers (128 filters per layer, convolution kernel size 4x4, stride 2), each layer is followed by a ReLU activation function to improve the spatial resolution and obtain F upsampled , the formula is:

[0127] F upsampled =φ(W T *F enhanced +b) #(9)

[0128] Where W T is the transposed convolution kernel, b is the bias, and * represents the deconvolution operation. In this step, the resolution of the input feature map is increased according to the formula (10):

[0129]

[0130] Among them, InputSize is the input size, Stride is the stride, Kernel Size is the convolution kernel size, Padding and Output Padding are both 0, so the original resolution n*w will become (4n+6)*(4w+6), thus the resolution becomes higher. Therefore, when using high-resolution actual precipitation data as labels, it is important to keep the resolution of the precipitation forecast data with higher resolution consistent (using the Kriging interpolation method).

[0131] Refinement layer: After 4 convolutional layers (64 filters per layer, convolution kernel size 3x3), each convolution layer is followed by a ReLU activation function to further refine the features and obtain F refined .

[0132] Output layer: After the convolution layer (1 filter, convolution kernel size 3x3), the linear Sigmoid activation function is used to ensure the appropriate output range, and finally a high-resolution weather precipitation forecast P is obtained. HR .

[0133] The model training uses high-resolution actual precipitation data as labels (i.e., the first label mentioned above), and uses the mean square error (MSE) to calculate the error between the predicted value and the true value (i.e., the third loss value mentioned above). The Adam optimizer is used to adjust the network parameters through the back propagation algorithm to minimize the loss function. Among them, the preset loss threshold can be set based on the above principle.

[0134] 104. If the deep learning model converges, the converged deep learning model is used as the high-resolution weather forecast model.

[0135] It can be understood that the deep learning model is composed of data optimization, data fusion and data downscaling models. The convergence of the deep learning model means that the above-mentioned first model, second model and third model all converge. The converged deep learning model includes the converged first model, second model and third model. Using the converged deep learning model as the high-resolution weather forecast model includes using the converged first model, second model and third model as the high-resolution weather forecast model.

[0136] Through the above method, a high-resolution weather forecasting method based on deep learning that integrates Beidou and meteorological data is realized, which is characterized by using measured meteorological data to optimize low-resolution precipitation forecasts, and the technology used for downscaling processing to improve resolution after fusion, wherein the three network architectures and data preprocessing designs are data preprocessing module, data optimization module, data fusion module, and data downscaling module. Among them, the data fusion module at least includes the use of a dynamic weight generation mechanism based on the self-attention mechanism, which automatically adjusts the weights according to the local characteristics of different data sources to achieve efficient multi-source data fusion. Among them, the downscaling module at least includes: high-frequency feature extraction enhancement layer: using wavelet transform and other technologies to decompose the data into high-frequency and low-frequency components, and then performing convolution processing on them separately and then combining them. It can be understood that the extension of the above-mentioned high-resolution precipitation forecasting method can be used for high-resolution forecasting in weather forecast factors. In summary, the high-resolution precipitation forecasting method can bring the following beneficial effects:

[0137] 1) Improve the accuracy of short-term precipitation forecast: The existing weather forecast model has deficiencies in short-term precipitation forecast, mainly because the spatial resolution of the forecast data is low, making it difficult to accurately capture local heavy precipitation phenomena. The present invention introduces Beidou meteorological and ground meteorological data, and uses deep learning technology to optimize precipitation forecast data, improve forecast accuracy, and improve the accuracy of short-term precipitation forecast.

[0138] 2) Achieve efficient fusion of multi-source data: The effective fusion of multi-source meteorological data is an important way to improve forecast accuracy. Existing methods have difficulties in dealing with the accuracy and consistency of different data sources. The present invention constructs a multi-source data fusion model based on deep learning. The model uses a feature-based dynamic weight generation mechanism to dynamically adjust the fusion weights of different data sources according to real-time observation data to achieve efficient fusion and improve the quality of forecast data.

[0139] 3) Improve the spatial resolution of forecast data: High-resolution weather forecasts are of great significance for disaster prevention and mitigation, especially in areas with complex terrain and changeable climate. However, existing downscaling methods have limited effects. The present invention introduces a high-frequency feature enhancement module into the fused forecast data using a deep learning model to improve the resolution of the forecast data and capture small-scale weather phenomena such as local heavy rainfall.

[0140] The present invention provides a training method for a high-resolution weather forecast model, the training method comprising: obtaining a pre-collected training sample data set, the training sample data set at least comprising training sample data and actual precipitation labels of the training sample data, the training sample data at least comprising first forecast data of low-resolution precipitation forecasts of multiple weather precipitation models, atmospheric precipitable water data obtained by inverting Beidou data, and surface meteorological observation data, the actual precipitation label comprising a first label of high-resolution actual precipitation data; inputting the first forecast data, atmospheric precipitable water data, and surface meteorological observation data into a preset deep learning model to perform model training for high-resolution weather forecast, and obtaining second forecast data for high-resolution weather forecast; using the second forecast data, the first label, and a preset loss function to determine whether the deep learning model converges; if the deep learning model converges, using the converged deep learning model as the high-resolution weather forecast model. Through the above method, high-resolution weather forecast based on deep learning integration of Beidou and meteorological data, the first forecast data of low-resolution precipitation forecasts of multiple weather precipitation models, atmospheric precipitable water data obtained by inverting Beidou data, surface meteorological observation data, and other multi-source data are efficiently integrated to achieve high-resolution weather forecast, improve forecast accuracy, and improve the accuracy of short-term precipitation forecast.

[0141] See also Figure 3 , Figure 3 The flowchart of a method for using a high-resolution weather forecast model in an embodiment of the present invention is shown in FIG. 1 . The method can be applied to a terminal or a server. The terminal can be a desktop terminal or a mobile terminal. The mobile terminal can be at least one of a mobile phone, a tablet computer, a laptop computer, etc. The server can be implemented by an independent server or a server cluster composed of multiple servers. This embodiment is described by applying to a terminal as an example. Figure 3 The method of use shown includes the following steps:

[0142] 301. Obtain current forecast data of low-resolution precipitation forecasts of multiple weather precipitation models, current atmospheric precipitable water data obtained by inversion of Beidou data, and current surface meteorological observation data;

[0143] 302. Input the current forecast data, current atmospheric precipitable water data and current surface meteorological observation data into a high-resolution weather forecast model to perform high-resolution weather forecast, thereby obtaining third forecast data of the high-resolution weather forecast;

[0144] The high-resolution weather forecast model is based on Figure 1It is obtained by training with the training method shown. It can be understood that after the training of the high-resolution weather forecast model is completed, the high-resolution weather forecast model can be used to realize the high-resolution weather forecast of real-time rolling forecast. Specifically, the current forecast data of the low-resolution precipitation forecast of multiple weather precipitation models, the current atmospheric precipitable water data obtained by Beidou data inversion and the current surface meteorological observation data are collected in real time, and the current forecast data, the current atmospheric precipitable water data and the current surface meteorological observation data are input into the high-resolution weather forecast model for high-resolution weather forecast, and the third forecast data of the high-resolution weather forecast is obtained. The third forecast data is used to reflect the high-resolution weather forecast data corresponding to the current forecast data.

[0145] Among them, in order to achieve more accurate real-time rolling forecasts, the latest observation data (actual precipitation data) and forecast data (precipitation forecast data) are continuously introduced. The specific steps are as follows:

[0146] Get the latest observation data regularly: Get the latest weather observation information through Beidou satellite data and ground-based meteorological observation data.

[0147] Real-time optimization of model parameters: Update the parameters of deep learning models using the latest observations, ensuring that the model is always up to date.

[0148] Generate the latest forecast data: Use the optimized model to generate the latest high-resolution weather forecast data to ensure the real-time and accuracy of the forecast data.

[0149] The present invention provides a method for using a high-resolution weather forecast model, the method comprising: obtaining current forecast data of low-resolution precipitation forecasts of multiple weather precipitation models, current atmospheric precipitable water data obtained by inverting Beidou data, and current surface meteorological observation data; inputting the current forecast data, current atmospheric precipitable water data, and current surface meteorological observation data into the high-resolution weather forecast model for high-resolution weather forecasting, and obtaining third forecast data of the high-resolution weather forecast; wherein the high-resolution weather forecast model is adopted as follows Figure 1 The training method shown in the figure is used for training. Through the above method, high-resolution weather forecast based on deep learning integrating Beidou and meteorological data is realized, and high-resolution weather forecast with efficient fusion of multi-source data such as current forecast data of low-resolution precipitation forecast of multiple weather precipitation models, current atmospheric precipitable water data obtained by Beidou data inversion and current surface meteorological observation data is realized, which improves the forecast accuracy and the accuracy of short-term precipitation forecast.

[0150] See also Figure 4 , Figure 4 FIG. 1 is a structural block diagram of a training device for a high-resolution weather forecast model in an embodiment of the present invention. Figure 4 The training device shown comprises:

[0151] Data acquisition module 401: used to acquire a pre-collected training sample data set, wherein the training sample data set includes at least training sample data and actual precipitation labels of the training sample data, wherein the training sample data includes at least first forecast data of low-resolution precipitation forecasts of multiple weather precipitation models, atmospheric precipitable water data obtained by inversion of Beidou data, and surface meteorological observation data, and the actual precipitation label includes a first label of high-resolution actual precipitation data;

[0152] Model training module 402: used for inputting the first forecast data, atmospheric precipitable water data and surface meteorological observation data into a preset deep learning model to perform model training for high-resolution weather forecast, so as to obtain second forecast data for high-resolution weather forecast;

[0153] Convergence judgment module 403: used to determine whether the deep learning model converges by using the second forecast data, the first label and the preset loss function;

[0154] Model determination module 404: used to use the converged deep learning model as the high-resolution weather forecast model if the deep learning model converges.

[0155] It should be noted that if Figure 4 The functions of each module in the training device shown are Figure 1 The contents of each step in the training method shown are similar, so we will not repeat them here to avoid repetition. For details, please refer to Figure 1 The content of each step in the training method is shown.

[0156] The present invention provides a training device for a high-resolution weather forecast model, the training device comprising: a data acquisition module: used to acquire a pre-collected training sample data set, the training sample data set at least comprising training sample data and actual precipitation labels of the training sample data, the training sample data at least comprising first forecast data of low-resolution precipitation forecasts of multiple weather precipitation models, atmospheric precipitable water data obtained by inverting Beidou data and surface meteorological observation data, the actual precipitation label comprising a first label of high-resolution actual precipitation data; a model training module: used to input the first forecast data, atmospheric precipitable water data and surface meteorological observation data into a preset deep learning model to perform model training for high-resolution weather forecast, and obtain second forecast data for high-resolution weather forecast; a convergence judgment module: used to determine whether the deep learning model converges by using the second forecast data, the first label and a preset loss function; a model determination module: used to use the converged deep learning model as the high-resolution weather forecast model if the deep learning model converges. Through the above-mentioned training device, high-resolution weather forecast based on deep learning that integrates Beidou and meteorological data, the first forecast data of low-resolution precipitation forecasts of multiple weather precipitation models, atmospheric precipitable water data obtained by inverting Beidou data, surface meteorological observation data, and other multi-source data are efficiently integrated to achieve high-resolution weather forecast, thereby improving forecast accuracy and the accuracy of short-term precipitation forecasts.

[0157] In a feasible implementation, the deep learning model includes a first model for data optimization, a second model for data fusion, and a third model for downscaling that are connected in sequence, and the actual precipitation label also includes a second label of low-resolution actual precipitation data, and the training device also includes:

[0158] Data optimization module: used for inputting the first forecast data, atmospheric precipitable water data and surface meteorological observation data into the first model for data optimization model training, and obtaining optimized forecast data of low-resolution precipitation forecasts of various weather precipitation models; using the first loss value between the optimized forecast data and the second label to determine whether the first model has converged; if the first loss value indicates that the first model has not converged, using the first loss value to update the model parameters of the first model to obtain an updated first model, and returning to execute the step of inputting the first forecast data, atmospheric precipitable water data and surface meteorological observation data into the first model for data optimization model training, and obtaining optimized forecast data of low-resolution precipitation forecasts of various weather precipitation models;

[0159] Data fusion module: used for obtaining a converged first model if the first loss value indicates that the first model has converged, and performing weighted fusion using the accuracy information of each weather forecast model and the optimized forecast data to obtain first fused forecast data for low-resolution precipitation forecast, and inputting the optimized forecast data corresponding to each weather and precipitation model into the second model for model training of data fusion to obtain second fused forecast data for low-resolution precipitation forecast; using a second loss value between the first fused forecast data and the second fused forecast data to determine whether the second model has converged; if the second loss value indicates that the second model has not converged, updating the model parameters of the second model using the second loss value to obtain an updated second model, and returning to execute the step of inputting the optimized forecast data corresponding to each weather and precipitation model into the second model for model training of data fusion to obtain second fused forecast data for low-resolution precipitation forecast;

[0160] Data downscaling module: used for obtaining a converged second model if the second loss value indicates that the second model has converged, and inputting each of the second fused forecast data into the third model for downscaling model training to obtain second forecast data for high-resolution precipitation forecast.

[0161] Furthermore, the use of the above modules can bring the following beneficial effects:

[0162] 1) Improve the accuracy of short-term precipitation forecast: The existing weather forecast model has deficiencies in short-term precipitation forecast, mainly because the spatial resolution of the forecast data is low, making it difficult to accurately capture local heavy precipitation phenomena. The present invention introduces Beidou meteorological and ground meteorological data, and uses deep learning technology to optimize precipitation forecast data, improve forecast accuracy, and improve the accuracy of short-term precipitation forecast.

[0163] 2) Achieve efficient fusion of multi-source data: The effective fusion of multi-source meteorological data is an important way to improve forecast accuracy. Existing methods have difficulties in dealing with the accuracy and consistency of different data sources. The present invention constructs a multi-source data fusion model based on deep learning. The model uses a feature-based dynamic weight generation mechanism to dynamically adjust the fusion weights of different data sources according to real-time observation data to achieve efficient fusion and improve the quality of forecast data.

[0164] 3) Improve the spatial resolution of forecast data: High-resolution weather forecasts are of great significance for disaster prevention and mitigation, especially in areas with complex terrain and changeable climate. However, existing downscaling methods have limited effects. The present invention introduces a high-frequency feature enhancement module into the fused forecast data using a deep learning model to improve the resolution of the forecast data and capture small-scale weather phenomena such as local heavy rainfall.

[0165] See also Figure 5 , Figure 5 FIG. 1 is a structural block diagram of a device for using a high-resolution weather forecast model in an embodiment of the present invention. Figure 5 The device shown for use includes:

[0166] Data acquisition module 501: used to obtain current forecast data of low-resolution precipitation forecasts of multiple weather precipitation models, current atmospheric precipitable water data obtained by Beidou data inversion, and current surface meteorological observation data;

[0167] Weather forecast module 502: used to input the current forecast data, current atmospheric precipitation data and current surface meteorological observation data into a high-resolution weather forecast model to perform high-resolution weather forecast, and obtain third forecast data of high-resolution weather forecast. The high-resolution weather forecast model is based on Figure 1 The training method shown is used for training.

[0168] It should be noted that if Figure 5 The device shown is used with Figure 3 The contents of each step in the usage method shown are similar, so they are not described here to avoid repetition. For details, please refer to Figure 3 The contents of each step in the usage method shown.

[0169] The present invention provides a device for using a high-resolution weather forecast model, the device comprising: a data acquisition module for acquiring current forecast data of low-resolution precipitation forecasts of multiple weather precipitation models, current atmospheric precipitable water data obtained by inverting Beidou data, and current surface meteorological observation data; a weather forecast module for inputting the current forecast data, current atmospheric precipitable water data, and current surface meteorological observation data into a high-resolution weather forecast model for high-resolution weather forecasting, and obtaining third forecast data of the high-resolution weather forecast, wherein the high-resolution weather forecast model is a model that adopts the following method: Figure 1 The training method shown in the figure is used for training. Through the above method, high-resolution weather forecast based on deep learning integrating Beidou and meteorological data is realized, and high-resolution weather forecast with efficient fusion of multi-source data such as current forecast data of low-resolution precipitation forecast of multiple weather precipitation models, current atmospheric precipitable water data obtained by Beidou data inversion and current surface meteorological observation data is realized, which improves the forecast accuracy and the accuracy of short-term precipitation forecast.

[0170] Figure 6 FIG. 1 shows an internal structure diagram of a computer device in an embodiment. The computer device may be a terminal or a server. Figure 6As shown, the computer device includes a processor, a memory and a network interface connected via a system bus. The memory includes a non-volatile storage medium and an internal memory. The non-volatile storage medium of the computer device stores an operating system and may also store a computer program. When the computer program is executed by the processor, the processor can implement the above method. The internal memory may also store a computer program. When the computer program is executed by the processor, the processor can implement the above method. Those skilled in the art can understand that Figure 6 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0171] In one embodiment, a computer device is provided, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the following steps: Figure 1 Or the steps of the method shown in 3.

[0172] In one embodiment, a computer-readable storage medium is provided, which stores a computer program. When the computer program is executed by a processor, the processor performs the following steps: Figure 1 Or the steps of the method shown in 3.

[0173] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0174] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0175] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A method for training a high-resolution weather forecast model, characterized in that: The training method comprises: Acquire a pre-collected training sample data set, wherein the training sample data set includes at least training sample data and actual precipitation labels of the training sample data, wherein the training sample data includes at least first forecast data of low-resolution precipitation forecasts of multiple weather precipitation models, atmospheric precipitable water data obtained by inversion of Beidou data, and surface meteorological observation data, and the actual precipitation label includes a first label of high-resolution actual precipitation data; Inputting the first forecast data, atmospheric precipitable water data and surface meteorological observation data into a preset deep learning model to perform model training for high-resolution weather forecast, thereby obtaining second forecast data for high-resolution weather forecast; Determining whether the deep learning model converges using the second forecast data, the first label, and a preset loss function; If the deep learning model converges, the converged deep learning model is used as the high-resolution weather forecast model.

2. The training method according to claim 1, characterized in that: The deep learning model includes a first model for data optimization, a second model for data fusion, and a third model for downscaling that are connected in sequence, and the actual precipitation label also includes a second label of low-resolution actual precipitation data. Then, the first forecast data, atmospheric precipitable water data, and surface meteorological observation data are input into a preset deep learning model for model training of high-resolution weather forecast, and the first forecast result obtained includes: Inputting the first forecast data, atmospheric precipitable water data and surface meteorological observation data into a first model for data optimization model training to obtain optimized forecast data of low-resolution precipitation forecasts of various weather precipitation models; Determining whether the first model converges using the first loss value between the optimized forecast data and the second label; If the first loss value indicates that the first model has not converged, the model parameters of the first model are updated using the first loss value to obtain an updated first model, and the step of inputting the first forecast data, atmospheric precipitable water data, and surface meteorological observation data into the first model for data optimization model training is returned to obtain optimized forecast data of low-resolution precipitation forecasts of various weather precipitation models; If the first loss value indicates that the first model has converged, a converged first model is obtained, and the accuracy information of each weather forecast model and the optimized forecast data are used for weighted fusion to obtain first fused forecast data for low-resolution precipitation forecast, and the optimized forecast data corresponding to each weather precipitation model is input into the second model for data fusion model training to obtain second fused forecast data for low-resolution precipitation forecast; Determining whether the second model converges by using a second loss value between the first fused forecast data and the second fused forecast data; If the second loss value indicates that the second model has not converged, the model parameters of the second model are updated using the second loss value to obtain an updated second model, and the step of inputting the optimized forecast data corresponding to each weather precipitation model into the second model for data fusion model training to obtain second fused forecast data of low-resolution precipitation forecast is returned to be executed; If the second loss value indicates that the second model has converged, a converged second model is obtained, and each of the second fused forecast data is input into the third model for downscaled model training to obtain second forecast data for high-resolution precipitation forecast.

3. The training method according to claim 2, characterized in that: The using the second forecast data, the first label and the preset loss function to determine whether the deep learning model converges includes: Determining a third loss value of the third model based on the second forecast data, the first label and the loss function; If the third loss value is not lower than the preset loss threshold, it is determined that the third model has not converged, and the model parameters of the third model are updated using the third loss value to obtain an updated third model, and the step of inputting each of the second fused forecast data into the third model for downscaling model training to obtain the second forecast data of high-resolution precipitation forecast is returned to execute; If the third loss value is lower than a preset loss threshold, it is determined that the third model converges, and the converged deep learning model includes a converged first model, a second model, and a third model connected in sequence.

4. The training method according to claim 1, characterized in that: The training method further comprises: The first forecast data, atmospheric precipitable water data and surface meteorological observation data are subjected to data preprocessing to obtain the preprocessed first forecast data, atmospheric precipitable water data and surface meteorological observation data, the training sample data include the preprocessed first forecast data, atmospheric precipitable water data and surface meteorological observation data, and the preprocessing includes at least normalization processing and interpolation processing.

5. A method for using a high-resolution weather forecast model, characterized in that: The method of use includes: Obtain current forecast data of low-resolution precipitation forecasts from multiple weather precipitation models, current atmospheric precipitable water data obtained by Beidou data inversion, and current surface meteorological observation data; The current forecast data, the current atmospheric precipitable water data and the current surface meteorological observation data are input into a high-resolution weather forecast model for high-resolution weather forecasting to obtain third forecast data for the high-resolution weather forecast. The high-resolution weather forecast model is trained using the training method as shown in any one of claims 1 to 4.

6. A training device for a high-resolution weather forecast model, characterized in that: The training device comprises: Data acquisition module: used to acquire a pre-collected training sample data set, wherein the training sample data set includes at least training sample data and actual precipitation labels of the training sample data, wherein the training sample data includes at least first forecast data of low-resolution precipitation forecasts of multiple weather precipitation models, atmospheric precipitable water data obtained by inversion of Beidou data, and surface meteorological observation data, and the actual precipitation label includes a first label of high-resolution actual precipitation data; Model training module: used for inputting the first forecast data, atmospheric precipitable water data and surface meteorological observation data into a preset deep learning model to perform model training for high-resolution weather forecast, and obtain second forecast data for high-resolution weather forecast; Convergence judgment module: used to determine whether the deep learning model converges by using the second forecast data, the first label and a preset loss function; Model determination module: used to use the converged deep learning model as the high-resolution weather forecast model if the deep learning model converges.

7. The training device according to claim 6, characterized in that: The deep learning model includes a first model for data optimization, a second model for data fusion, and a third model for downscaling that are connected in sequence, and the actual precipitation label also includes a second label of low-resolution actual precipitation data, and the training device also includes: Data optimization module: used for inputting the first forecast data, atmospheric precipitable water data and surface meteorological observation data into the first model for data optimization model training, and obtaining optimized forecast data of low-resolution precipitation forecasts of various weather precipitation models; using the first loss value between the optimized forecast data and the second label to determine whether the first model has converged; if the first loss value indicates that the first model has not converged, using the first loss value to update the model parameters of the first model to obtain an updated first model, and returning to execute the step of inputting the first forecast data, atmospheric precipitable water data and surface meteorological observation data into the first model for data optimization model training, and obtaining optimized forecast data of low-resolution precipitation forecasts of various weather precipitation models; Data fusion module: used for obtaining a converged first model if the first loss value indicates that the first model has converged, and performing weighted fusion using the accuracy information of each weather forecast model and the optimized forecast data to obtain first fused forecast data for low-resolution precipitation forecast, and inputting the optimized forecast data corresponding to each weather and precipitation model into the second model for model training of data fusion to obtain second fused forecast data for low-resolution precipitation forecast; using a second loss value between the first fused forecast data and the second fused forecast data to determine whether the second model has converged; if the second loss value indicates that the second model has not converged, updating the model parameters of the second model using the second loss value to obtain an updated second model, and returning to execute the step of inputting the optimized forecast data corresponding to each weather and precipitation model into the second model for model training of data fusion to obtain second fused forecast data for low-resolution precipitation forecast; Data downscaling module: used for obtaining a converged second model if the second loss value indicates that the second model has converged, and inputting each of the second fused forecast data into the third model for downscaling model training to obtain second forecast data for high-resolution precipitation forecast.

8. A device for using a high-resolution weather forecast model, characterized in that: The use device comprises: Data acquisition module: used to obtain the current forecast data of low-resolution precipitation forecasts of multiple weather precipitation models, the current atmospheric precipitable water data obtained by Beidou data inversion, and the current surface meteorological observation data; Weather forecast module: used for inputting the current forecast data, current atmospheric precipitable water data and current surface meteorological observation data into a high-resolution weather forecast model for high-resolution weather forecast, and obtaining third forecast data of the high-resolution weather forecast, wherein the high-resolution weather forecast model is trained using the training method as shown in any one of claims 1 to 4.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the processor is caused to perform the steps of the training method according to any one of claims 1 to 4 or the using method according to claim 5 .

10. A computer device comprising a memory and a processor, characterized in that: The memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the training method according to any one of claims 1 to 4 or the using method according to claim 5.

Images