Weather prediction method and device and training method and device of weather prediction model

Abstract

The invention discloses a weather prediction method, weather prediction is carried out through a weather prediction model, the weather prediction model comprises a space-time variation auto-encoder, the space-time variation auto-encoder comprises an encoder and a decoder, and the weather prediction method comprises the following steps: obtaining current weather data containing current weather condition information; encoding the current weather data in a hidden space through an encoder to obtain current weather hidden space data; the current weather hidden space data is data after the feature dimension is increased and the space-time dimension is reduced; performing weather prediction on the hidden space to obtain predicted weather hidden space data; and decoding the predicted weather hidden space data through a decoder to obtain predicted weather data. According to the method, the redundancy of the current weather data can be reduced, the key meteorological features are extracted, and the accuracy of weather prediction is improved. The invention further discloses a training method of the weather prediction model, a weather prediction device, a training device of the weather prediction model, electronic equipment and a readable storage medium.

Description

[0001] This application claims priority to the Chinese patent application filed with the China Patent Office on August 27, 2024, with application number 202411183511.1 and application name “Weather Forecasting Method and Device and Training Method and Device for Weather Forecasting Model”, the entire contents of which are incorporated by reference in this application. Technical Field

[0002] The present invention relates to the field of weather forecasting technology, and in particular to a weather forecasting method, a weather forecasting model training method, a weather forecasting device, a weather forecasting model training device, an electronic device and a computer-readable storage medium. Background Art

[0003] Weather forecasting is indispensable in modern society and is of vital importance to people's daily life, social activities, agricultural production, transportation, disaster prevention and other fields. Understanding the possible weather conditions in the future through weather forecasting is conducive to people's arrangement and adjustment of social activities, ensuring the safe operation of transportation industries such as aviation and shipping, providing more scientific planting and harvesting guidance for agricultural production, and helping to issue disaster warnings in a timely manner, guiding the public to take defensive measures and reduce losses caused by natural disasters. Therefore, it is necessary to make timely and accurate predictions and forecasts of the weather.

[0004] In the prior art, when making weather forecasts, it is necessary to rely on meteorological spatiotemporal data. Usually, the weather conditions at the next moment are predicted based on the current weather data conditions, and the current weather data obtained is usually directly input into the relevant prediction model, and the prediction is started directly without any other processing of the current weather data. Meteorological spatiotemporal data includes atmospheric variables and ground layer variables. Among them, atmospheric variables include, for example, temperature, humidity, wind speed, etc., and ground layer variables include, for example, precipitation, air pressure, etc. However, meteorological spatiotemporal data has a very high spatial resolution and has relatively large data redundancy. Therefore, when making predictions directly through current weather data, there may be a problem of slow prediction speed due to too much redundant data, and there may also be problems of inaccurate prediction results and large errors. Summary of the invention

[0005] The present invention proposes a weather forecasting method to solve the above problems.

[0006] In a first aspect, an embodiment of the present invention discloses a weather prediction method, which performs weather prediction through a weather prediction model, wherein the weather prediction model includes a spatiotemporal variational autoencoder, the spatiotemporal variational autoencoder includes an encoder and a decoder, and the weather prediction method includes:

[0007] Get current weather data containing information about current weather conditions;

[0008] The current weather data is encoded in the latent space by an encoder, and the current weather data is converted into the current weather latent space data; the current weather latent space data is the data after adding feature dimensions to the current weather data and reducing the time dimension and space dimension;

[0009] Perform weather forecast on the current weather latent space data in latent space to obtain predicted weather latent space data;

[0010] The predicted weather latent space data is decoded by a decoder and converted into predicted weather data; the predicted weather data is the data after the predicted weather latent space data has its feature dimensions reduced and the time dimension and space dimension added, and the predicted weather data is the result of the predicted weather conditions corresponding to the current weather data.

[0011] By adopting the above technical scheme, the weather forecasting method of the present invention can increase the characteristic latitude of the current weather data, reduce the time latitude and space latitude of the current weather data by performing dimensional processing on the current weather data, extract key features related to weather forecasting, and eliminate data redundancy, thereby speeding up the forecasting speed and improving the accuracy of weather forecasting.

[0012] According to another specific embodiment of the present invention, the encoder includes an encoding residual module, and the encoder encodes the current weather data in the latent space to convert the current weather data into the current weather latent space data, specifically including:

[0013] The current weather data is mapped from the standard data space to the latent space through the encoding residual module, so that the current weather data is converted into the current weather latent space data; wherein the standard data space is the data space corresponding to the current weather data.

[0014] According to another specific embodiment of the present invention, the coding residual module includes a plurality of coding residual convolutional layers, and the current weather data is mapped from the standard data space to the latent space through the coding residual module, so that the current weather data is converted into the current weather latent space data, specifically including:

[0015] The current weather data is dimensionally processed through multiple encoded residual convolutional layers to increase the feature dimension of the current weather data, reduce the time dimension and space dimension of the current weather data, and obtain the current weather latent space data.

[0016] According to another specific embodiment of the present invention, the decoder includes a decoding residual module, and the predicted weather latent space data is decoded by the decoder to convert the predicted weather latent space data into predicted weather data, specifically including:

[0017] The predicted weather latent space data is reconstructed from the latent space to the standard data space through the decoding residual module to obtain the predicted weather data.

[0018] According to another specific embodiment of the present invention, the decoding residual module includes a plurality of decoding residual convolutional layers, and the predicted weather latent space data is reconstructed from the latent space to the standard data space through the decoding residual module to obtain the predicted weather data, which specifically includes:

[0019] The predicted weather latent space data is dimensionally processed through multiple decoding residual convolutional layers to reduce the feature dimension of the predicted weather latent space data, increase the time dimension and space dimension of the predicted weather latent space data, and obtain the predicted weather data.

[0020] According to another specific embodiment of the present invention, the number of layers of the encoding residual convolution layer of the encoding residual module is the same as the number of layers of the decoding residual convolution layer of the decoding residual module, so that the multiples of the feature dimension increased by the current weather data are the same as the multiples of the feature dimension reduced by the predicted weather latent space data, and the multiples of the time dimension and space dimension reduced by the current weather data are the same as the multiples of the time dimension and space dimension increased by the predicted weather latent space data.

[0021] According to another specific embodiment of the present invention, the weather prediction model further includes a prediction model, which performs weather prediction on the current weather latent space data in the latent space to obtain the predicted weather latent space data, specifically including:

[0022] The prediction model uses the group attention mechanism to fuse the features of the current weather latent space data to obtain the predicted weather latent space data of the next state of weather conditions;

[0023] Among them, the group attention mechanism is used to quantify and group the weather features used for prediction contained in the current weather latent space data to obtain multiple feature groups, and redetermine the weights of the weather features used for prediction in the multiple feature groups for feature fusion.

[0024] According to another specific embodiment of the present invention, the current weather latent space data is used for weather forecasting by using a group attention mechanism through a forecasting model to obtain the forecast weather latent space data of the weather conditions in the next state, specifically including:

[0025] According to the timestamp corresponding to the current weather hidden space data, the prediction model uses the group attention mechanism to perform weather prediction on the current weather hidden space data to obtain the predicted weather hidden space data of the next state of weather conditions.

[0026] In a second aspect, an embodiment of the present invention further discloses a method for training a weather prediction model, which is used to train a weather prediction model for performing weather prediction in a weather prediction method in any of the aforementioned embodiments, wherein the weather prediction model includes a spatiotemporal variational autoencoder, the spatiotemporal variational autoencoder includes an encoder and a decoder, and the method for training the weather prediction model includes:

[0027] Get weather data samples;

[0028] The weather data samples are encoded in the latent space by an encoder to obtain weather latent space data corresponding to the weather data samples in the latent space; the weather latent space data is the data obtained by adding feature dimensions to the weather data samples and reducing time dimensions and space dimensions;

[0029] The weather hidden space data is decoded and restored by a decoder to obtain restored weather data corresponding to the weather hidden space data; the restored weather data is the data obtained by reducing the feature dimension of the weather hidden space data and adding the time dimension and the space dimension;

[0030] Determine a loss function based on the weather data sample and the restored weather data;

[0031] The parameters of the spatiotemporal variational autoencoder to be trained are updated according to the loss function, weather data samples and restored weather data until the spatiotemporal variational autoencoder to be trained meets the preset training conditions, thereby obtaining a trained spatiotemporal variational autoencoder.

[0032] By adopting the above technical scheme, the weather prediction model trained by the weather prediction model training method provided by the present invention can increase the characteristic latitude of the current weather data, reduce the time latitude and space latitude of the current weather data by dimensional processing of the current weather data, extract key features related to weather forecasting, and eliminate data redundancy at the same time, speed up the prediction speed, and improve the accuracy of weather forecasting.

[0033] According to another specific embodiment of the present invention, the loss function includes a reconstruction loss function and a KL divergence loss function, and the reconstruction loss function is a mean square error loss function and / or a cross entropy loss function.

[0034] According to another specific embodiment of the present invention, the weather prediction model further includes a prediction model, the prediction model is used to perform weather prediction in latent space, and the training method of the weather prediction model further includes:

[0035] Obtaining original weather data samples, timestamps corresponding to weather conditions contained in the original weather data samples, and weather forecast labels corresponding to the original weather data samples;

[0036] The prediction model uses a group attention mechanism to fuse the features of the original weather data samples to obtain the predicted weather data of the next state of weather conditions;

[0037] Determine a prediction loss function based on the weather prediction label and the predicted weather data;

[0038] The parameters of the prediction model to be trained are updated according to the prediction loss function, the predicted weather data and the weather prediction label until the prediction model to be trained meets the preset prediction training conditions, thereby obtaining a trained prediction model.

[0039] In a third aspect, an embodiment of the present invention further discloses a weather prediction device, which performs weather prediction through a weather prediction model, wherein the weather prediction model includes a spatiotemporal variational autoencoder, the spatiotemporal variational autoencoder includes an encoder and a decoder, and the weather prediction device includes:

[0040] A data acquisition module, used to acquire current weather data including current weather condition information;

[0041] The latent space encoding module is used to encode the current weather data in the latent space through an encoder, and convert the current weather data into the current weather latent space data; the current weather latent space data is the data after adding feature dimensions to the current weather data and reducing the time dimension and space dimension;

[0042] The latent space prediction module is used to perform weather prediction on the current weather latent space data in the latent space to obtain the predicted weather latent space data;

[0043] The latent space decoding module is used to decode the predicted weather latent space data through a decoder, and convert the predicted weather latent space data into predicted weather data; the predicted weather data is the data after the predicted weather latent space data reduces the feature dimension and increases the time dimension and space dimension, and the predicted weather data is the result of the predicted weather conditions corresponding to the current weather data.

[0044] By adopting the above technical solution, the weather forecasting device can perform dimensional processing on the current weather data, increase the characteristic latitude of the current weather data, reduce the time latitude and space latitude of the current weather data, extract key features related to weather forecasting, and eliminate data redundancy, speed up the forecasting speed, and improve the accuracy of weather forecasting.

[0045] In a fourth aspect, an embodiment of the present invention further discloses a weather prediction model training device, which is used to train a weather prediction model for performing weather prediction in a weather prediction method in any of the aforementioned embodiments, wherein the weather prediction model includes a spatiotemporal variational autoencoder, the spatiotemporal variational autoencoder includes an encoder and a decoder, and the weather prediction model training device includes:

[0046] A training data acquisition module is used to obtain weather data samples;

[0047] The encoding module is used to encode the weather data sample in the latent space through the encoder to obtain the weather latent space data corresponding to the weather data sample in the latent space; the weather latent space data is the data obtained by adding feature dimensions to the weather data sample and reducing the time dimension and space dimension;

[0048] A decoding module is used to decode and restore the weather hidden spatial data through a decoder to obtain restored weather data corresponding to the weather hidden spatial data; the restored weather data is the data after the weather hidden spatial data has its feature dimension reduced and its time dimension and space dimension increased;

[0049] A loss function determination module, used to determine a loss function based on weather data samples and restored weather data;

[0050] The training module is used to update the parameters of the spatiotemporal variational autoencoder to be trained according to the loss function, weather data samples and restored weather data, until the spatiotemporal variational autoencoder to be trained meets the preset training conditions, thereby obtaining a trained spatiotemporal variational autoencoder.

[0051] By adopting the above technical solution, the weather prediction model trained by the weather prediction model training device can perform dimensional processing on the current weather data, increase the characteristic latitude of the current weather data, reduce the time latitude and spatial latitude of the current weather data, extract key features related to weather forecasting, and eliminate data redundancy, speed up the prediction speed, and improve the accuracy of weather forecasting.

[0052] In a fifth aspect, an embodiment of the present invention further discloses an electronic device, comprising a processor and a memory, wherein the memory stores at least one instruction, and when the at least one instruction is executed by the processor, implements the weather prediction method in any of the aforementioned embodiments, and / or implements the training method of the weather prediction model in any of the aforementioned embodiments.

[0053] By adopting the above technical solution, the electronic device enables the trained weather prediction model to perform dimensional processing on the current weather data, increase the feature latitude of the current weather data, reduce the time latitude and space latitude of the current weather data, extract key features related to weather forecasting, and eliminate data redundancy, speed up the prediction speed, and improve the accuracy of weather forecasting.

[0054] In a sixth aspect, an embodiment of the present invention further discloses a computer-readable storage medium, in which at least one instruction is stored, and when the at least one instruction is executed, the weather prediction method in any of the aforementioned embodiments is implemented, and / or the training method of the weather prediction model in any of the aforementioned embodiments is implemented.

[0055] By adopting the above technical solution, the computer-readable storage medium enables the trained weather forecast model to perform dimensional processing on the current weather data, increase the feature latitude of the current weather data, reduce the time latitude and space latitude of the current weather data, extract key features related to weather forecasting, and eliminate data redundancy, speed up the forecasting speed, and improve the accuracy of weather forecasting. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 The process of the weather forecasting method in the embodiment of the present invention is shown as follows Figure 1 ;

[0057] Figure 2 The process of the weather forecasting method in the embodiment of the present invention is shown as follows Figure 2 ;

[0058] Figure 3 The process of the weather forecasting method in the embodiment of the present invention is shown as follows Figure 3 ;

[0059] Figure 4 The process of the weather forecasting method in the embodiment of the present invention is shown as follows Figure 4 ;

[0060] Figure 5 The process of the weather forecasting method in the embodiment of the present invention is shown as follows Figure 5 ;

[0061] Figure 6 The process of the weather forecasting method in the embodiment of the present invention is shown as follows Figure 6 ;

[0062] Figure 7 The process of the weather forecasting method in the embodiment of the present invention is shown as follows Figure 7 ;

[0063] Figure 8 The process of the training method of the weather prediction model in the embodiment of the present invention is shown as follows Figure 1 ;

[0064] Fig. 9 The process of the training method of the weather prediction model in the embodiment of the present invention is shown as follows Figure 2 ;

[0065] Fig.10 A schematic diagram showing the structure of a weather forecasting device in an embodiment of the present invention is shown;

[0066] Fig.11 A schematic diagram showing the structure of a training device for a weather prediction model according to an embodiment of the present invention is shown;

[0067] Fig.12 A schematic diagram showing the structure of an electronic device in an embodiment of the present invention is shown;

[0068] Fig.13 A schematic diagram showing the structure of the spatiotemporal variational autoencoder part of the weather prediction model in an embodiment of the present invention is shown;

[0069] Fig.14 A schematic diagram showing the structure of the prediction model part of the weather prediction model in an embodiment of the present invention. DETAILED DESCRIPTION

[0070] The following specific embodiments illustrate the implementation of the present invention, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. Although the description of the present invention will be introduced in conjunction with the preferred embodiment, this does not mean that the features of this invention are limited to this implementation. On the contrary, the purpose of introducing the invention in conjunction with the implementation is to cover other options or modifications that may extend based on the claims of the present invention. In order to provide a deep understanding of the present invention, the following description will include many specific details. The present invention can also be implemented without using these details. In addition, in order to avoid confusion or blurring the focus of the present invention, some specific details will be omitted in the description. It should be noted that the embodiments of the present invention and the features in the embodiments can be combined with each other without conflict.

[0071] It should be noted that in this specification, similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, it does not need to be further defined and explained in the subsequent drawings.

[0072] The terms “first”, “second”, etc. are only used for distinguishing descriptions and should not be understood as indicating or implying relative importance.

[0073] In the description of this embodiment, it is also necessary to explain that, unless otherwise clearly specified and limited, the terms "set", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in this embodiment can be understood according to specific circumstances.

[0074] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0075] After careful study of the prior art, the inventor found that the prior art usually directly predicts the current weather data obtained without doing other processing before prediction. However, the weather data has a large redundancy, which is not conducive to quickly and accurately predicting the weather. The inventor hopes to solve the above problem.

[0076] First, reference Figure 1 The embodiment of the present invention provides a weather prediction method, which performs weather prediction through a weather prediction model, wherein the weather prediction model includes a spatiotemporal variational autoencoder (VAE), and the spatiotemporal variational autoencoder includes an encoder and a decoder. The weather prediction method includes the following steps:

[0077] S1: Obtain current weather data including current weather condition information.

[0078] Among them, the current weather data may include weather variables, such as the temperature of the current pressure layer, the humidity of the current pressure layer, the wind speed of the current pressure layer, the air pressure of the current pressure layer, the temperature of the current ground layer, the precipitation of the current ground layer and the air pressure of the current ground layer.

[0079] S2: Encode the current weather data in the latent space through the encoder and convert the current weather data into the current weather latent space data.

[0080] Among them, the current weather latent space data is the data after adding feature dimensions to the current weather data and reducing the time dimension and space dimension. The latent space refers to the potential representation space to which the encoder maps the input data. In this space, the data is converted into a compressed, usually low-dimensional form, which can capture the main features and structure of the original input data. That is, in this embodiment, the weather data of the day is encoded in the latent space by the encoder, specifically referring to converting the current weather data into a low-dimensional form in the latent space. By re-encoding and converting the current weather data to the latent space, it is beneficial to reduce the spatiotemporal dimensions of the current weather data, increase the feature dimensions of the current weather data, and extract and retain weather features that are more relevant to weather forecasts while eliminating redundant data, which can not only reduce the amount of data, but also ensure the accuracy of subsequent weather forecasts.

[0081] It should be noted that in some possible embodiments provided by the present invention, the current weather latent space data is an abstract feature data in the latent space, which is data coupled with multiple weather characteristics, wherein the weather characteristics may include, for example, temperature, wind speed, pressure, atmospheric characteristics, etc.

[0082] S3: Perform weather prediction on the current weather latent space data in the latent space to obtain predicted weather latent space data.

[0083] Among them, a pre-trained prediction model can be used in the latent space to perform weather prediction on the current weather latent space data.

[0084] S4: Decode the predicted weather latent space data through a decoder to convert the predicted weather latent space data into predicted weather data.

[0085] The predicted weather data is the data obtained by reducing the feature dimension of the predicted weather latent spatial data and increasing the time dimension and the space dimension. The predicted weather data is the result of the predicted weather conditions corresponding to the current weather data.

[0086] The decoder of the spatiotemporal variational autoencoder can reconstruct and map the data in the latent space back to the original data space. By using the decoder to convert the forecast weather latent space data into forecast weather data, the time dimension, space dimension and feature dimension of the forecast weather data are restored, so that the weather conditions at the forecast time can be directly observed by reading the forecast weather data.

[0087] By adopting the above technical scheme, the weather forecasting method of the present invention can increase the characteristic latitude of the current weather data, reduce the time latitude and space latitude of the current weather data by performing dimensional processing on the current weather data, extract key features related to weather forecasting, and eliminate data redundancy, thereby speeding up the forecasting speed and improving the accuracy of weather forecasting.

[0088] Furthermore, the encoder includes a coding residual module, referring to Figure 2 , the above step S2 specifically includes the following steps:

[0089] S21: Mapping the current weather data from the standard data space to the latent space through the encoding residual module, so that the current weather data is converted into the current weather latent space data.

[0090] The standard data space is the data space corresponding to the current weather data. The encoding residual module (i.e., Resblock) helps solve the problem of gradient vanishing or gradient exploding during the training process of the spatiotemporal variational autoencoder, helps promote the conversion of the current weather data to the current weather latent space data, and improves the performance of the spatiotemporal variational autoencoder.

[0091] Furthermore, the coded residual module includes multiple coded residual convolutional layers, referring to Figure 3 The above step S21 specifically includes the following steps:

[0092] S211: Perform dimension processing on the current weather data through multiple coded residual convolution layers to increase the feature dimension of the current weather data, reduce the time dimension and space dimension of the current weather data, and obtain the current weather latent space data.

[0093] Among them, the more the number of coded residual convolution layers, the more times the feature dimension of the current weather data is increased, and the more times the time dimension and space dimension of the current weather data are reduced. Furthermore, the multiples of the time dimension and space dimension of the current weather data reduced by multiple coded residual convolution layers can be 4 times, 8 times or 16 times. The specific number of coded residual convolution layers and the multiples of the time dimension and space dimension reduced for the current weather data can be determined according to the actual needs of weather forecasting in different situations. By using multiple coded residual convolution layers to gradually reduce the time dimension and space dimension, and gradually increase the feature dimension, it can not only effectively reduce the time and space dimensions of the current weather data and reduce data redundancy, but also grasp important weather characteristics and prevent the loss of important feature data.

[0094] Furthermore, the decoder includes a decoding residual module, referring to Figure 4 , the above step S4 specifically includes the following steps:

[0095] S41: Reconstruct the predicted weather latent space data from the latent space to the standard data space through the decoding residual module to obtain the predicted weather data.

[0096] The decoding residual module of the decoder helps to solve the problem of gradient vanishing or gradient exploding during the training process of the spatiotemporal variational autoencoder, improve the expressiveness of the entire model, enhance the ability to reconstruct data, improve the quality of the output predicted weather data, reduce overfitting, and improve the accuracy of weather forecasts.

[0097] Furthermore, the decoding residual module includes multiple decoding residual convolutional layers, referring to Figure 5 The above step S41 specifically includes the following steps:

[0098] S411: Dimension processing is performed on the predicted weather latent space data through multiple decoding residual convolution layers to reduce the feature dimension of the predicted weather latent space data, increase the time dimension and space dimension of the predicted weather latent space data, and obtain the predicted weather data.

[0099] Among them, the number of layers of the encoding residual convolution layer of the encoding residual module is the same as the number of layers of the decoding residual convolution layer of the decoding residual module, so that the multiples of the feature dimension increased by the current weather data are the same as the multiples of the feature dimension reduced by the predicted weather latent space data, and the multiples of the time dimension and space dimension reduced by the current weather data are the same as the multiples of the time dimension and space dimension increased by the predicted weather latent space data. More specifically, the number of decoding residual convolution layers is the same as the number of encoding residual convolution layers, so as to restore the predicted weather latent space data to the same feature dimension, time dimension and space dimension as the current weather data. By using multiple decoding residual convolution layers to gradually increase the time dimension and space dimension, and gradually reduce the feature dimension, it is beneficial to ensure that the reconstructed output predicted weather data retains more weather feature details and ensure the quality of the predicted weather data.

[0100] Furthermore, the weather prediction model also includes a prediction model, reference Figure 6 , the above step S3 specifically includes the following steps:

[0101] S31: The prediction model uses a group attention mechanism to fuse the features of the current weather latent space data to obtain the predicted weather latent space data of the next weather state.

[0102] Fig.14 The structural diagram of the prediction model used in the present invention is shown. The entire prediction model performs data processing and weather prediction in latent space. The prediction model used in the present invention is the Weather Transformer model. By combining the spatiotemporal variational autoencoder with the Weather Transformer model, it is beneficial to further improve the accuracy of weather prediction while reducing the redundancy of meteorological spatiotemporal data, ensuring the stability and generalization ability of the entire weather prediction model, and making weather prediction more accurate.

[0103] The Weather Transformer model includes a Transformer module (i.e., Transformer Block). More specifically, the Weather Transformer model may include d Transformer modules (d is a positive integer), and the specific number of Transformer modules can be selected according to the actual needs of weather forecasting. The Transformer module includes a normalization layer AdaLN, a group attention mechanism module QGA, and a fully connected layer MLP.

[0104] The grouped attention mechanism module QGA uses the grouped attention mechanism for image processing. Specifically, the grouped attention mechanism is used to quantify and group the weather features used for prediction contained in the current weather latent space data to obtain multiple feature groups, and re-weight the weather features used for prediction in multiple feature groups for feature fusion. By applying the grouped attention mechanism in the Transformer module, it helps to improve computational efficiency, enhance the expressiveness of the model, reduce computational complexity, and improve the accuracy of prediction.

[0105] The fully connected layer MLP can integrate and convert the prediction results of other layers in the Transformer module to obtain the predicted weather latent space data.

[0106] Further, refer to Figure 7 The above step S31 specifically includes the following steps:

[0107] S311: According to the timestamp corresponding to the current weather hidden space data, the current weather hidden space data is predicted by the prediction model using the group attention mechanism to obtain the predicted weather hidden space data of the next weather state.

[0108] Specifically, the normalization layer AdaLN is used to receive the input timestamp. The timestamp is the timestamp of the current weather latent spatial data input to the prediction model, not the time to be predicted. The timestamp of the current weather latent spatial data can be, for example, the hour of the day or the day of the year (doy, which is the abbreviation of day of year). Since the weather in a year usually shows periodic changes, the input timestamp will contain periodic information of weather changes. By combining the current weather latent spatial data and its corresponding timestamp (i.e., the current time point when the current weather latent spatial data is collected), it helps to further improve the accuracy of the model's weather forecast.

[0109] Since meteorological spatiotemporal data usually contain meteorological layer variables and ground layer variables, the spatial resolution is usually very high and has relatively large redundancy. In order to solve the above problems, the present invention designs a spatiotemporal variational autoencoder for meteorological spatiotemporal data for the first time, which effectively reduces the spatiotemporal dimension of the data and extracts important meteorological element features while maintaining a low reconstruction error. After encoding by the spatiotemporal variational autoencoder, weather forecasts can be performed directly on the latent space. Since the feature expression ability of the latent space is stronger, it is beneficial to improve the accuracy of weather forecasts. In addition, after the time dimension and space dimension of the input weather data are reduced, an attention module (such as a global attention mechanism, a group attention mechanism, etc.) can be used to increase the number of parameters of the network model, guide the further optimization and improvement of large models in the meteorological field, improve the prediction performance of the model, and further improve the accuracy of the prediction.

[0110] Second, reference Figure 8 The embodiment of the present invention further provides a method for training a weather prediction model, which is used to train a weather prediction model used for weather prediction in the weather prediction method in any of the aforementioned embodiments, wherein the weather prediction model includes a spatiotemporal variational autoencoder (VAE), and the spatiotemporal variational autoencoder includes an encoder and a decoder. The method for training the weather prediction model includes the following steps:

[0111] S01: Get weather data samples.

[0112] Exemplarily, the weather data sample may be selected from the ERA5 dataset. The ERA5 dataset is a global climate reanalysis dataset developed and maintained by the European Centre for Medium-Range Weather Forecasts (ECMWF). The ERA5 dataset provides high-resolution global data of various meteorological elements, including atmospheric temperature, air pressure, wind speed, precipitation, etc.

[0113] S02: Encode the weather data samples in the latent space through an encoder to obtain the weather latent space data corresponding to the weather data samples in the latent space.

[0114] The weather latent spatial data is the data obtained by adding feature dimensions to the weather data samples and reducing the time dimension and the space dimension. The encoder is the encoder used in the above-mentioned weather prediction method embodiment.

[0115] S03: Decode and restore the weather hidden spatial data through a decoder to obtain restored weather data corresponding to the weather hidden spatial data.

[0116] The restored weather data is the data obtained by reducing the feature dimension of the weather latent spatial data and increasing the time dimension and the space dimension. The decoder is the decoder used in the above-mentioned weather forecasting method embodiment.

[0117] S04: Determine a loss function based on the weather data sample and the restored weather data.

[0118] Among them, the loss function includes a reconstruction loss function and a KL divergence loss function, and the reconstruction loss function can be a mean square error loss function and / or a cross entropy loss function. By adopting the reconstruction loss function, it is beneficial to measure the difference between the output restored weather data and the weather data sample, and it is beneficial to ensure that the spatiotemporal variational autoencoder retains the information of the input weather data sample as much as possible during the encoding and decoding process. By adopting the KL divergence loss function, it can be ensured that the distribution in the latent space is close to the prior distribution (such as the standard normal distribution), which is beneficial to ensure the continuity and smoothness of the latent space, help prevent overfitting, and promote the effect of data reconstruction. By combining the reconstruction loss function and the KL divergence loss function, the training process can be stabilized and the generalization ability of the model can be guaranteed.

[0119] S05: Update the parameters of the spatiotemporal variational autoencoder to be trained according to the loss function, weather data samples and restored weather data, until the spatiotemporal variational autoencoder to be trained meets the preset training conditions, thereby obtaining a trained spatiotemporal variational autoencoder.

[0120] The preset training condition may be reaching a preset training time or the number of iterations during the training process, or the loss function may converge and become stable.

[0121] By adopting the above technical scheme, the weather prediction model trained by the weather prediction model training method provided by the present invention can increase the characteristic latitude of the current weather data, reduce the time latitude and space latitude of the current weather data by dimensional processing of the current weather data, extract key features related to weather forecasting, and eliminate data redundancy at the same time, speed up the prediction speed, and improve the accuracy of weather forecasting.

[0122] refer to Fig.13 , Fig.13 FIG. 1 is a schematic diagram showing the structure of the spatiotemporal variational autoencoder used in the above embodiments of the present invention. The input weather data sample is first encoded into the latent space by the encoder of the spatiotemporal variational autoencoder, and then restored and reconstructed by the decoder of the spatiotemporal variational autoencoder to obtain the restored weather data. t-k …X t (initial input) represents the input weather data sample (multiple samples can be input to train the spatiotemporal variational autoencoder), T*C in In *H*W, T represents time (corresponding to the time dimension), C in represents the feature channel (corresponding to the feature dimension), H*W represents the space (corresponding to the spatial dimension). Latent represents the latent space, h t Represents weather hidden spatial data, corresponding to T f *C f *H f *W f is the dimension of the weather latent space data corresponding to the latent space (correspondingly, T f Indicates time, corresponding to the time dimension; C f Represents the feature channel, corresponding to the feature dimension; H f *W f represents space, corresponding to the spatial dimension), where C f Greater than C in , T f Less than T, H f *W f Less than H*W. After the weather hidden space data is restored by the decoder, the output is the restored weather data (i.e. Fig.13 The X shown on the right t-k…X t (Output result), it is expected that the input weather data sample and the output restored weather data are as consistent as possible, and the best case is that they are completely consistent), the T*C*H*W of the restored weather data is the same as the T*C*H*W of the weather data sample in *H*W are the same (i.e., the time dimension, spatial dimension, and feature dimension of the restored weather data and the weather data sample are the same, respectively).

[0123] In the above embodiment, the weather prediction model also includes a prediction model, and the prediction model is used to perform weather prediction in the latent space. Fig. 9 , the training method of the weather prediction model also includes the following steps:

[0124] S06: Obtain original weather data samples, timestamps corresponding to weather conditions contained in the original weather data samples, and weather forecast labels corresponding to the original weather data samples.

[0125] Exemplarily, the original weather data samples, the corresponding timestamps, and the weather forecast labels can all be selected from the ERA5 dataset.

[0126] S07: The prediction model uses a group attention mechanism to fuse the features of the original weather data samples to obtain the predicted weather data of the next weather condition.

[0127] The prediction model is the prediction model used in the above weather prediction method embodiment. Fig.14 ,h t is the original weather data sample, h t+1 For forecasting weather data.

[0128] S08: Determine a prediction loss function according to the weather prediction label and the predicted weather data.

[0129] Among them, illustratively, the prediction loss function may be, for example, a cross entropy loss or a MAE loss, or other loss functions designed by oneself.

[0130] S09: Update the parameters of the prediction model to be trained according to the prediction loss function, the predicted weather data and the weather prediction label, until the prediction model to be trained meets the preset prediction training conditions, thereby obtaining a trained prediction model.

[0131] The predicted preset training condition may be, for example, reaching a preset training time or the number of iterations in the training process, or the loss function converges and tends to be stable.

[0132] Thirdly, reference Fig.10The embodiment of the present invention further provides a weather prediction device 1, which performs weather prediction through a weather prediction model, wherein the weather prediction model includes a spatiotemporal variational autoencoder, and the spatiotemporal variational autoencoder includes an encoder and a decoder. The weather prediction device 1 includes a data acquisition module 11, a latent space encoding module 12, a latent space prediction module 13, and a latent space decoding module 14.

[0133] Among them, the data acquisition module 11 is connected to the latent space encoding module 12, the latent space encoding module 12 is connected to the latent space prediction module 13, and the latent space prediction module 13 is connected to the latent space decoding module 14.

[0134] The data acquisition module 11 is used to acquire current weather data including current weather condition information.

[0135] The latent space encoding module 12 is used to encode the current weather data in the latent space through an encoder, and convert the current weather data into the current weather latent space data; the current weather latent space data is the data after adding feature dimensions to the current weather data and reducing the time dimension and space dimension.

[0136] The latent space prediction module 13 is used to perform weather prediction on the current weather latent space data in the latent space to obtain predicted weather latent space data.

[0137] The latent space decoding module 14 is used to decode the predicted weather latent space data through a decoder, and convert the predicted weather latent space data into predicted weather data; the predicted weather data is the data after the predicted weather latent space data reduces the feature dimension and adds the time dimension and space dimension, and the predicted weather data is the result of the predicted weather conditions corresponding to the current weather data.

[0138] By adopting the above technical solution, the weather prediction device 1 can increase the characteristic latitude of the current weather data, reduce the time latitude and space latitude of the current weather data by dimensional processing of the current weather data, extract key features related to weather forecasting, and eliminate data redundancy, speed up the prediction speed, and improve the accuracy of weather forecasting.

[0139] Furthermore, the latent space encoding module 12 can also be used to perform the aforementioned steps S21 and S211. The latent space prediction module 13 can also be used to perform the aforementioned steps S31 and S311. The latent space decoding module 14 can also be used to perform the aforementioned steps S41 and S411.

[0140] Fourthly, reference Fig.11An embodiment of the present invention further provides a weather prediction model training device 2, which is used to train a weather prediction model used for weather prediction in a weather prediction method in any of the aforementioned embodiments, wherein the weather prediction model includes a spatiotemporal variational autoencoder, and the spatiotemporal variational autoencoder includes an encoder and a decoder. The weather prediction model training device 2 includes a training data acquisition module 21, an encoding module 22, a decoding module 23, a loss function determination module 24 and a training module 25.

[0141] Among them, the training data acquisition module 21 is connected to the encoding module 22, the encoding module 22 is connected to the decoding module 23, the decoding module 23 is connected to the loss function determination module 24, and the loss function determination module 24 is connected to the training module 25.

[0142] The training data acquisition module 21 is used to acquire weather data samples.

[0143] The encoding module 22 is used to encode the weather data samples in the latent space through an encoder to obtain the weather latent space data corresponding to the weather data samples in the latent space; the weather latent space data is the data after adding feature dimensions to the weather data samples and reducing time dimensions and space dimensions.

[0144] The decoding module 23 is used to decode and restore the weather hidden spatial data through a decoder to obtain restored weather data corresponding to the weather hidden spatial data; the restored weather data is the data after the weather hidden spatial data has its feature dimensions reduced and its time dimension and space dimension added.

[0145] The loss function determination module 24 is used to determine the loss function according to the weather data samples and the restored weather data.

[0146] The training module 25 is used to update the parameters of the spatiotemporal variational autoencoder to be trained according to the loss function, weather data samples and restored weather data until the spatiotemporal variational autoencoder to be trained meets the preset training conditions, thereby obtaining a trained spatiotemporal variational autoencoder.

[0147] By adopting the above technical scheme, the weather prediction model trained by the weather prediction model training device 2 can increase the characteristic latitude of the current weather data, reduce the time latitude and space latitude of the current weather data by dimensional processing of the current weather data, extract key features related to weather forecasting, and eliminate data redundancy, speed up the prediction speed, and improve the accuracy of weather forecasting.

[0148] Furthermore, the weather prediction model also includes a prediction model, which is used to perform weather prediction in latent space, and the training device 2 of the weather prediction model can also include a prediction model data acquisition module, a prediction model processing module, a prediction loss function determination module and a prediction model training module. Among them, the prediction model data acquisition module is connected to the prediction model processing module, the prediction model processing module is connected to the prediction loss function determination module, and the prediction loss function determination module is connected to the prediction model training module.

[0149] The prediction model data acquisition module is used to obtain original weather data samples, timestamps corresponding to the weather conditions contained in the original weather data samples, and weather prediction labels corresponding to the original weather data samples.

[0150] The prediction model processing module is used to perform feature fusion on the original weather data samples through the prediction model using the group attention mechanism to obtain the predicted weather data of the next weather condition.

[0151] The prediction loss function determination module is used to determine the prediction loss function according to the weather prediction label and the predicted weather data.

[0152] The prediction model training module is used to update the parameters of the prediction model to be trained according to the prediction loss function, predicted weather data and weather prediction labels until the prediction model to be trained meets the prediction preset training conditions, thereby obtaining a trained prediction model.

[0153] Fifth, reference Fig.12 The embodiment of the present invention further provides an electronic device 3, including a processor 32 and a memory 31, wherein the memory 31 stores at least one instruction, and when the at least one instruction is executed by the processor 32, the weather prediction method in any of the aforementioned embodiments is implemented, and / or the training method of the weather prediction model in any of the aforementioned embodiments is implemented. The memory 31 may include, for example, a system memory, a fixed non-volatile storage medium, etc. The system memory may store, for example, an operating system, an application program, a boot loader, and other programs.

[0154] By adopting the above technical solution, the electronic device 3 enables the trained weather prediction model to perform dimensional processing on the current weather data, increase the feature latitude of the current weather data, reduce the time latitude and space latitude of the current weather data, extract key features related to weather forecasting, and eliminate data redundancy, speed up the prediction speed, and improve the accuracy of weather forecasting.

[0155] In a sixth aspect, an embodiment of the present invention further provides a computer-readable storage medium, in which at least one instruction is stored, and when the at least one instruction is executed, the weather forecasting method in any of the aforementioned embodiments is implemented, and / or the training method of the weather forecasting model in any of the aforementioned embodiments is implemented.

[0156] By adopting the above technical solution, the computer-readable storage medium enables the trained weather forecast model to perform dimensional processing on the current weather data, increase the feature latitude of the current weather data, reduce the time latitude and space latitude of the current weather data, extract key features related to weather forecasting, and eliminate data redundancy, speed up the forecasting speed, and improve the accuracy of weather forecasting.

[0157] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable non-transient storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0158] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0159] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0160] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0161] Although the present invention has been illustrated and described with reference to certain preferred embodiments of the present invention, it should be understood by those skilled in the art that the above is a further detailed description of the present invention in conjunction with specific embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. Those skilled in the art may make various changes in form and details, including making several simple deductions or substitutions, without departing from the spirit and scope of the present invention.

Claims

1. A weather forecasting method, characterized in that: Weather prediction is performed by a weather prediction model, wherein the weather prediction model includes a spatiotemporal variational autoencoder, the spatiotemporal variational autoencoder includes an encoder and a decoder, and the weather prediction method includes: Get current weather data containing information about current weather conditions; The encoder encodes the current weather data in a latent space, and converts the current weather data into current weather latent space data; the current weather latent space data is the data obtained by adding feature dimensions to the current weather data and reducing time dimensions and space dimensions; Performing weather forecasting on the current weather latent space data in the latent space to obtain forecasted weather latent space data; The predicted weather latent space data is decoded by the decoder to convert the predicted weather latent space data into predicted weather data; the predicted weather data is the data after the predicted weather latent space data has its feature dimensions reduced and the time dimension and space dimension added, and the predicted weather data is the result of the predicted weather conditions corresponding to the current weather data.

2. The weather forecasting method according to claim 1, characterized in that: The encoder includes an encoding residual module, and the encoder encodes the current weather data in the latent space to convert the current weather data into current weather latent space data, specifically including: The current weather data is mapped from the standard data space to the latent space through the coding residual module, so that the current weather data is converted into the current weather latent space data; wherein the standard data space is the data space corresponding to the current weather data.

3. The weather forecasting method according to claim 2, characterized in that: The coding residual module includes a plurality of coding residual convolution layers, and the current weather data is mapped from the standard data space to the latent space through the coding residual module so as to convert the current weather data into the current weather latent space data, specifically including: The current weather data is dimensionally processed through the multiple coded residual convolutional layers to increase the feature dimension of the current weather data, reduce the time dimension and space dimension of the current weather data, and obtain the current weather latent space data.

4. The weather forecasting method according to claim 3, characterized in that: The decoder includes a decoding residual module, and the decoding of the predicted weather latent space data by the decoder to convert the predicted weather latent space data into predicted weather data specifically includes: The predicted weather latent space data is reconstructed from the latent space to the standard data space through the decoding residual module to obtain the predicted weather data.

5. The weather forecasting method according to claim 4, characterized in that: The decoding residual module includes a plurality of decoding residual convolutional layers, and the decoding residual module is used to reconstruct the predicted weather latent space data from the latent space to the standard data space to obtain the predicted weather data, specifically including: The predicted weather latent space data is dimensionally processed through the multiple decoding residual convolutional layers to reduce the feature dimension of the predicted weather latent space data, increase the time dimension and space dimension of the predicted weather latent space data, and obtain the predicted weather data.

6. The weather forecasting method according to claim 5, characterized in that: The number of layers of the encoding residual convolution layer of the encoding residual module is the same as the number of layers of the decoding residual convolution layer of the decoding residual module, so that the multiples of the feature dimension increased by the current weather data are the same as the multiples of the feature dimension reduced by the predicted weather latent space data, and the multiples of the time dimension and space dimension reduced by the current weather data are the same as the multiples of the time dimension and space dimension increased by the predicted weather latent space data.

7. The weather forecasting method according to claim 1, characterized in that: The weather prediction model also includes a prediction model, which performs weather prediction on the current weather hidden space data in the hidden space to obtain predicted weather hidden space data, specifically including: The prediction model uses a group attention mechanism to perform feature fusion on the current weather latent space data to obtain the predicted weather latent space data of the weather conditions in the next state; Among them, the group attention mechanism is used to quantify and group the weather features used for prediction contained in the current weather latent space data to obtain multiple feature groups, and redetermine the weights of the weather features used for prediction in the multiple feature groups to perform feature fusion.

8. The weather forecasting method according to claim 7, characterized in that: The method of performing weather forecasting on the current weather latent space data by using the forecasting model and adopting the group attention mechanism to obtain the forecast weather latent space data of the weather condition in the next state specifically includes: According to the timestamp corresponding to the current weather hidden space data, the prediction model adopts a group attention mechanism to perform weather prediction on the current weather hidden space data to obtain the predicted weather hidden space data of the weather conditions in the next state.

9. A method for training a weather prediction model, characterized in that: Used for training a weather prediction model for performing weather prediction in the weather prediction method according to any one of claims 1 to 8, the weather prediction model comprising a spatiotemporal variational autoencoder, the spatiotemporal variational autoencoder comprising an encoder and a decoder, and the training method of the weather prediction model comprising: Get weather data samples; The weather data sample is encoded in the latent space by the encoder to obtain weather latent space data corresponding to the weather data sample in the latent space; the weather latent space data is the data obtained by adding feature dimensions to the weather data sample and reducing time dimensions and space dimensions; The weather hidden spatial data is decoded and restored by the decoder to obtain restored weather data corresponding to the weather hidden spatial data; the restored weather data is the data obtained by reducing the feature dimension of the weather hidden spatial data and adding the time dimension and the space dimension; Determine a loss function according to the weather data sample and the restored weather data; The parameters of the spatiotemporal variational autoencoder to be trained are updated according to the loss function, the weather data samples and the restored weather data until the spatiotemporal variational autoencoder to be trained meets preset training conditions, thereby obtaining the trained spatiotemporal variational autoencoder.

10. The weather prediction model training method according to claim 9, characterized in that: The loss function includes a reconstruction loss function and a KL divergence loss function, and the reconstruction loss function is a mean square error loss function and / or a cross entropy loss function.

11. The method for training a weather prediction model according to claim 9, wherein: The weather prediction model further includes a prediction model, and the prediction model is used to perform weather prediction on the latent space. The training method of the weather prediction model further includes: Obtaining an original weather data sample, a timestamp corresponding to the weather conditions contained in the original weather data sample, and a weather forecast label corresponding to the original weather data sample; The prediction model uses a group attention mechanism to perform feature fusion on the original weather data samples to obtain predicted weather data of the weather conditions in the next state; Determine a prediction loss function according to the weather prediction label and the predicted weather data; The parameters of the prediction model to be trained are updated according to the prediction loss function, the predicted weather data and the weather prediction label until the prediction model to be trained meets the prediction preset training conditions, thereby obtaining the trained prediction model.

12. A weather forecasting device, characterized in that: Weather prediction is performed by a weather prediction model, wherein the weather prediction model includes a spatiotemporal variational autoencoder, the spatiotemporal variational autoencoder includes an encoder and a decoder, and the weather prediction device includes: A data acquisition module, used to acquire current weather data including current weather condition information; A latent space encoding module, used for encoding the current weather data in a latent space through the encoder, and converting the current weather data into current weather latent space data; the current weather latent space data is the data obtained by adding feature dimensions to the current weather data and reducing time dimensions and space dimensions; A latent space prediction module, used for performing weather prediction on the current weather latent space data in the latent space to obtain predicted weather latent space data; The latent space decoding module is used to decode the predicted weather latent space data through the decoder to convert the predicted weather latent space data into predicted weather data; the predicted weather data is the data after the predicted weather latent space data has its feature dimensions reduced and the time dimension and space dimension added, and the predicted weather data is the result of the predicted weather conditions corresponding to the current weather data.

13. A weather forecast model training device, characterized in that: Used for training a weather prediction model for performing weather prediction in a weather prediction method according to any one of claims 1 to 8, the weather prediction model comprising a spatiotemporal variational autoencoder, the spatiotemporal variational autoencoder comprising an encoder and a decoder, the training device of the weather prediction model comprising: A training data acquisition module is used to obtain weather data samples; An encoding module, used for encoding the weather data sample in a latent space through the encoder to obtain weather latent space data corresponding to the weather data sample in the latent space; the weather latent space data is the data obtained by adding feature dimensions to the weather data sample and reducing time dimensions and space dimensions; A decoding module, used for decoding and restoring the weather hidden spatial data through the decoder to obtain restored weather data corresponding to the weather hidden spatial data; the restored weather data is the data obtained by reducing the feature dimension of the weather hidden spatial data and adding the time dimension and the space dimension; A loss function determination module, used to determine a loss function according to the weather data sample and the restored weather data; A training module is used to update the parameters of the spatiotemporal variational autoencoder to be trained according to the loss function, the weather data sample and the restored weather data, until the spatiotemporal variational autoencoder to be trained meets the preset training conditions, thereby obtaining the trained spatiotemporal variational autoencoder.

14. An electronic device, characterized in that: The electronic device includes a processor and a memory, wherein the memory stores at least one instruction, and when the at least one instruction is executed by the processor, it implements the weather prediction method as described in any one of claims 1-8, and / or the training method of the weather prediction model as described in any one of claims 9-11.

15. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores at least one instruction, and when the at least one instruction is executed, it implements the weather forecasting method as described in any one of claims 1-8 and / or the training method of the weather forecasting model as described in any one of claims 9-11.

Images