Extreme weather prediction model training method and extreme weather prediction method and device

Abstract

The invention discloses a training method of an extreme weather prediction model, the extreme weather prediction model comprises a diffusion model, the diffusion model comprises a forward diffusion module and a backward diffusion module, and the training method comprises the following steps: determining a preliminary prediction weather sample used for training and a prediction label corresponding to the preliminary prediction weather sample; training a diffusion model through the preliminary prediction weather sample and the prediction label: adding noise data to the preliminary prediction weather sample through a forward diffusion module to obtain noise weather fusion data; restoring the noise weather fusion data to a predicted weather restoration sample through a back diffusion module to obtain predicted noise data; determining a loss function according to the noise data and the predicted noise data; and training a diffusion model according to the loss function, the prediction weather restoration sample and the prediction label to obtain a trained extreme weather prediction model. And the trained model is more accurate in prediction. The invention further discloses a training device, an extreme weather prediction method and device, 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 202411183518.3 and application name “Extreme Weather Forecast Model Training Method and Extreme Weather Forecast Method and Device”, the entire contents of which are incorporated by reference in this application. Technical Field

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

[0003] With the impact of global climate change, the frequency and intensity of extreme weather events are increasing, which has brought serious impacts on human society and the natural environment. Extreme weather events include extreme rainfall, extreme strong winds and other situations. Extreme rainfall will have a negative impact on the gross regional product (GRP). Studies have shown that an increase in the number of rainy days and extreme daily rainfall will lead to a significant decline in the global macroeconomic growth rate. Extreme strong winds may affect the normal operation of wind turbines, hinder wind power generation, and may also cause house damage and affect transportation. In order to reduce the losses caused by extreme weather to human social activities and enable people to take protective measures in advance, it is necessary to accurately forecast extreme weather or provide early warnings for extreme weather.

[0004] Due to the mixing effect in the atmospheric system, different gases and particles in the atmosphere will mix through physical and dynamic processes. These mixing processes have an important impact on weather and climate. Therefore, accurate prediction of extreme weather is a very difficult task, especially for extreme precipitation (i.e. extreme rainfall) and extreme strong winds.

[0005] In the prior art, for weather forecasting, the numerical weather prediction (NWP) model is usually used to predict various weather parameters. The numerical weather prediction (NWP) model is a model that uses mathematical equations and numerical methods to simulate the state of the atmosphere and predict future weather. The NWP model is based on physics and atmospheric dynamics. However, due to the variable distribution of extreme weather events in time and space, with high spatiotemporal variability, extreme weather is also related to multi-scale factors, such as small-scale raindrops and large-scale entire weather systems. The NWP model is difficult to cover everything, and the attention to details is not enough, and the prediction of extreme weather is still not accurate enough.

[0006] In the prior art, there are also methods for predicting the weather through artificial intelligence machine learning. However, these weather forecast models based on artificial intelligence machine learning are usually modeled in an autoregressive manner and optimized through the RMSE loss function. When these models are used for weather forecasting, they are prone to unrealistic smooth prediction problems, such as ignoring the fluctuations of rapid weather changes, lacking a grasp of weather details, and failing to capture some extreme weather events. As the forecast time increases, the problem of smooth prediction will become more obvious, leading to an underestimation of the intensity of extreme weather events and inaccurate extreme weather forecasts. Summary of the invention

[0007] The present invention proposes a training method for an extreme weather prediction model to solve the above problems.

[0008] In a first aspect, an embodiment of the present invention discloses a training method for an extreme weather prediction model, the extreme weather prediction model is used to predict extreme weather, the extreme weather prediction model includes a diffusion model, the diffusion model includes a forward diffusion module and a reverse diffusion module, and the training method for the extreme weather prediction model includes:

[0009] Determine preliminary forecast weather samples for training the extreme weather forecast model and forecast labels corresponding to the preliminary forecast weather samples;

[0010] The diffusion model is trained using preliminary weather forecast samples and forecast labels, including:

[0011] Adding noise data to the preliminary forecast weather samples through the forward diffusion module, so that the preliminary forecast weather samples become noise weather fusion data;

[0012] The noise weather fusion data is restored to the predicted weather restoration sample through the reverse diffusion module to obtain the predicted noise data;

[0013] Determine the loss function based on the added noise data and the predicted noise data;

[0014] The diffusion model is trained according to the loss function, predicted weather restoration samples and predicted labels to update the parameters of the diffusion model until the preset training conditions are met, thereby obtaining a trained diffusion model and thus a trained extreme weather prediction model.

[0015] By adopting the above technical scheme, the training method of the extreme weather prediction model adds noise to the preliminary predicted weather samples so that the preliminary predicted weather samples become noise weather fusion data, and then obtains the loss function related to the noise, trains the diffusion model so that the diffusion model can restore the predicted weather results from the noise weather fusion data, and makes the weather results predicted by the trained extreme weather prediction model have more weather condition details, thereby improving the weather forecasting skills of the model and improving the accuracy of the model's weather forecast.

[0016] According to another specific embodiment of the present invention, adding noise data to the preliminary forecast weather sample through the forward diffusion module so that the preliminary forecast weather sample becomes noise weather fusion data, specifically including:

[0017] The forward diffusion module uses the Markov chain principle to add noise unit data to the preliminary forecast weather samples multiple times, so that the preliminary forecast weather samples become noise weather fusion data;

[0018] Determine noise data based on preliminary forecast weather samples and noise weather fusion data;

[0019] The noise unit data is Gaussian noise, and the noise data is Gaussian noise distribution data.

[0020] According to another specific embodiment of the present invention, the loss function is MSE loss.

[0021] In a second aspect, an embodiment of the present invention further discloses a training device for an extreme weather prediction model, the extreme weather prediction model is used to predict extreme weather, the extreme weather prediction model includes a diffusion model, the diffusion model includes a forward diffusion module and a reverse diffusion module, and the training device for the extreme weather prediction model includes:

[0022] A training preparation module, used to determine preliminary forecast weather samples for training an extreme weather forecast model and forecast labels corresponding to the preliminary forecast weather samples;

[0023] The training module is used to train the diffusion model through preliminary predicted weather samples and predicted labels, specifically including: adding noise data to the preliminary predicted weather samples through the forward diffusion module to make the preliminary predicted weather samples become noise weather fusion data; restoring the noise weather fusion data to the predicted weather restoration samples through the reverse diffusion module to obtain predicted noise data; determining the loss function according to the added noise data and the predicted noise data; training the diffusion model according to the loss function, the predicted weather restoration samples and the predicted labels to update the parameters of the diffusion model until the preset training conditions are met, thereby obtaining a trained diffusion model, and thus obtaining a trained extreme weather prediction model.

[0024] By adopting the above technical scheme, the training device of the extreme weather prediction model adds noise to the preliminary predicted weather samples so that the preliminary predicted weather samples become noise weather fusion data, and then obtains the loss function related to the noise, and trains the diffusion model so that the diffusion model can restore the predicted weather results from the noise weather fusion data, and makes the weather results predicted by the trained extreme weather prediction model have more weather condition details, thereby improving the weather forecasting skills of the model and improving the accuracy of the model's weather forecast.

[0025] In a third aspect, an embodiment of the present invention further discloses an extreme weather prediction method, which predicts extreme weather through an extreme weather prediction model, the extreme weather prediction model is trained by the extreme weather prediction model training method of any of the aforementioned embodiments, the extreme weather prediction model includes a prediction model and a diffusion model, and the extreme weather prediction method includes:

[0026] Determine current weather variable data based on current weather conditions;

[0027] According to the current weather variable data, the weather conditions at the next preset time are predicted through the prediction model to obtain preliminary predicted weather data;

[0028] The preliminary forecast weather data is processed for detail recovery through the diffusion model to obtain the final forecast weather data; the final forecast weather data is the prediction result of the extreme weather forecast model for the extreme weather at the next preset time.

[0029] By adopting the above technical solution, the extreme weather prediction method predicts extreme weather through the trained extreme weather prediction model, and the obtained extreme weather prediction results have more details and the prediction of extreme weather is more accurate.

[0030] According to another specific embodiment of the present invention, the current weather variable data includes: 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.

[0031] According to another specific embodiment of the present invention, the prediction model is selected from one of the following: FuXi model, Pangu-Weather model, Graphcast model.

[0032] According to another specific embodiment of the present invention, extreme weather includes: extreme precipitation, extreme wind speed, extreme temperature and tropical cyclone;

[0033] Among them, the standard for extreme precipitation is determined according to the classification standard of 24-hour precipitation in the Chinese national standard (GB / T 28592-2012); the standard for extreme wind speed is determined according to the standard of Beaufort wind scale; the standard for extreme temperature is determined according to the percentile threshold in the ERA-Interim surface air temperature reanalysis dataset; the extreme degree of tropical cyclones is determined according to the difference between the predicted value predicted by the extreme weather prediction model and the actually observed value.

[0034] In a fourth aspect, an embodiment of the present invention further discloses an extreme weather prediction device, characterized in that extreme weather is predicted by an extreme weather prediction model, the extreme weather prediction model is trained by the extreme weather prediction model training method according to any one of claims 1 to 3, the extreme weather prediction model includes a prediction model and a diffusion model, and the extreme weather prediction device includes:

[0035] The current weather acquisition module is used to determine the current weather variable data according to the current weather conditions;

[0036] The preliminary weather forecast module is used to predict the weather conditions at the next moment through the forecast model according to the current weather variable data to obtain preliminary forecast weather data;

[0037] The final weather forecast module is used to perform detail recovery processing on the preliminary forecast weather data through the diffusion model to obtain the final forecast weather data; the final forecast weather data is the result of the extreme weather forecast model's extreme weather forecast.

[0038] By adopting the above technical solution, the extreme weather prediction device predicts extreme weather through the trained extreme weather prediction model, and the obtained extreme weather prediction results have more details and the prediction of extreme weather is more accurate.

[0039] 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 training method of the extreme weather prediction model in any of the aforementioned embodiments, and / or implements the extreme weather prediction method in any of the aforementioned embodiments.

[0040] By adopting the above technical solution, the electronic device adds noise to the preliminary predicted weather samples to make the preliminary predicted weather samples become noise weather fusion data, and then obtains the loss function related to the noise, and trains the diffusion model so that the diffusion model can restore the predicted weather results from the noise weather fusion data, and make the predicted weather results have more weather condition details, thereby improving the weather forecasting skills of the model. The trained extreme weather prediction model will make more accurate predictions when making extreme weather forecasts.

[0041] 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. When the at least one instruction is executed, it implements the training method of the extreme weather prediction model in any of the aforementioned embodiments, and / or implements the extreme weather prediction method in any of the aforementioned embodiments.

[0042] By adopting the above technical solution, a computer-readable storage medium adds noise to the preliminary predicted weather samples so that the preliminary predicted weather samples become noise weather fusion data, and then obtains a loss function related to noise, and trains a diffusion model so that the diffusion model can restore the predicted weather results from the noise weather fusion data, and make the predicted weather results have more weather condition details, thereby improving the weather forecasting skills of the model, and the trained extreme weather prediction model predicts more accurately when making extreme weather forecasts. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 The process of the training method of the extreme weather prediction model in the embodiment of the present invention is shown as follows Figure 1 ;

[0044] Figure 2 The process of the training method of the extreme weather prediction model in the embodiment of the present invention is shown as follows Figure 2 ;

[0045] Figure 3 The process of the training method of the extreme weather prediction model in the embodiment of the present invention is shown as follows Figure 3 ;

[0046] Figure 4 A flow chart showing an extreme weather prediction method according to an embodiment of the present invention;

[0047] Figure 5 A schematic diagram showing the structure of a training device for an extreme weather prediction model according to an embodiment of the present invention;

[0048] Figure 6 A schematic diagram showing the structure of an extreme weather prediction device in an embodiment of the present invention is shown;

[0049] Figure 7 A schematic diagram showing the structure of an electronic device in an embodiment of the present invention is shown;

[0050] Figure 8 A schematic diagram showing an extreme weather prediction model in an embodiment of the present invention;

[0051] Fig. 9 A schematic diagram showing the comparison of the extreme weather prediction model of the present invention with other prediction models in terms of CSI indicators;

[0052] Fig.10A schematic diagram showing the comparison of the extreme weather prediction model of the present invention with other prediction models in terms of SEDI indicators;

[0053] Fig.11 A schematic diagram showing a comparison between the extreme weather prediction model of the present invention and other prediction models in terms of prediction of extreme total precipitation (TP);

[0054] Fig.12 A schematic diagram showing the comparison between the extreme weather prediction model of the present invention and other prediction models in the prediction of 10-meter wind speed (WS10);

[0055] Fig.13 A schematic diagram showing a comparison between the extreme weather prediction model of the present invention and other prediction models in the prediction of tropical cyclones using the IBTrACS data set;

[0056] Fig.14 A schematic diagram showing a comparison between the extreme weather prediction model of the present invention and other prediction models in predicting tropical cyclones using the ERA5 dataset. DETAILED DESCRIPTION

[0057] 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.

[0058] 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.

[0059] The terms “first”, “second”, etc. are only used for distinguishing descriptions and should not be understood as indicating or implying relative importance.

[0060] 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.

[0061] 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.

[0062] In the prior art, when forecasting extreme weather, the weather forecast model is prone to not paying enough attention to weather details, resulting in a smooth forecast problem, which leads to inaccurate forecasts of extreme weather. The inventor hopes to solve the above problem.

[0063] First, refer to Figure 1 The present invention provides a training method for an extreme weather prediction model. The extreme weather prediction model is used to predict extreme weather. The extreme weather prediction model includes a diffusion model, and the diffusion model includes a forward diffusion module and a reverse diffusion module. The training method for the extreme weather prediction model specifically includes the following steps:

[0064] S1: Determine preliminary forecast weather samples for training the extreme weather prediction model and the forecast labels corresponding to the preliminary forecast weather samples.

[0065] The preliminary forecast weather sample may be actual weather condition data at a selected specific moment, and the forecast label corresponding to the preliminary forecast weather sample may be forecast weather condition data expected to be obtained for the actual weather condition data at the next moment of the specific moment.

[0066] Further, in this embodiment, the dataset of preliminary forecast weather samples and forecast labels for training the extreme weather prediction model can be selected from the ERA5 dataset, which has a spatial resolution of 0.25° and a temporal resolution of 6 hours. Exemplarily, the dataset of preliminary forecast weather samples and forecast labels for training the extreme weather prediction model can be data from 2002 to 2017 in the ERA5 dataset. Exemplarily, when testing after the model training is completed, data from other years in the ERA5 dataset can be used as a test set, for example, data from 2018 can be used as a test set.

[0067] S2: Train the diffusion model with preliminary forecast weather samples and forecast labels.

[0068] refer to Figure 2In the above embodiment, step S2 specifically includes the following steps:

[0069] S21: adding noise data to the preliminary forecast weather samples through the forward diffusion module, so that the preliminary forecast weather samples become noise weather fusion data.

[0070] Among them, the noise weather fusion data can look like pure noise data. By converting the preliminary forecast weather samples into noise weather fusion data, it is helpful to establish the mapping of the model from noise to weather-related data, which helps the subsequent diffusion model to accurately learn how to restore from noise to weather, and improve the accuracy of the restored weather data.

[0071] S22: The noise weather fusion data is restored to the predicted weather restoration sample through the reverse diffusion module to obtain the predicted noise data.

[0072] The predicted noise data is the data obtained by learning and fitting the diffusion model.

[0073] S23: Determine a loss function based on the added noise data and the predicted noise data.

[0074] In this embodiment, the loss function is MSE loss.

[0075] S24: Train the diffusion model according to the loss function, the predicted weather restoration samples and the predicted labels to update the parameters of the diffusion model until the preset training conditions are met, thereby obtaining a trained diffusion model and thus obtaining a trained extreme weather prediction model.

[0076] The preset training condition may be a preset number of training iterations.

[0077] By adopting the above technical scheme, the training method of the extreme weather prediction model adds noise to the preliminary predicted weather samples so that the preliminary predicted weather samples become noise weather fusion data, and then obtains the loss function related to the noise, and trains the diffusion model so that the diffusion model can restore the predicted weather results from the noise weather fusion data, and make the predicted weather results have more weather conditions details, thereby improving the weather forecasting skills of the model and improving the accuracy of weather forecasting.

[0078] Further, refer to Figure 3 , step S21 specifically includes the following steps:

[0079] S211: Using the forward diffusion module, the Markov chain principle is used to add noise unit data to the preliminary forecast weather sample for multiple times, so that the preliminary forecast weather sample becomes noise weather fusion data.

[0080] S212: Determine noise data based on the preliminary forecast weather sample and the noise weather fusion data.

[0081] Wherein, the noise unit data is Gaussian noise, and the noise data is Gaussian noise distribution data. For example, Figure 8 N noise This is the noise data.

[0082] Specifically, the action process of the diffusion model includes a forward diffusion process and a reverse diffusion process. The forward diffusion process is executed by a forward diffusion module, and the reverse diffusion process is executed by a reverse diffusion module.

[0083] For example, when data samples are randomly drawn from a given real data distribution X0~q(X0), the forward diffusion process can be run as a Markov chain. Among them, Markov Chain is a statistical random process. The principle of Markov Chain is that the next state of a system depends only on the current state and has nothing to do with the previous state. In the forward diffusion process, Gaussian noise is gradually added to the preliminary forecast weather sample according to a predetermined variance plan. For example, a total of T steps are taken to add Gaussian noise to the preliminary forecast weather sample (T is a positive integer), then the variance plan can be: β1, β2, ..., βT, that is, in the process of adding Gaussian noise, the first step is to add Gaussian noise to the preliminary forecast weather sample according to the variance of β1 to obtain data X1; the second step is to continue to add Gaussian noise according to the variance of β2 on the basis of data X1 (that is, the preliminary forecast weather sample data obtained after adding Gaussian noise in the first step) to obtain data X2, and so on; finally, after adding Gaussian noise according to the variance of βT in the Tth step, data XT is obtained. In the above process, β1<β2<…<βT also needs to be satisfied.

[0084] In the above process, the variance plan is crucial for the effective operation of the forward diffusion process, and can affect the quality of the generated prediction results and the convergence behavior of the diffusion model. Since weather-related data has very complex variation patterns in time and space, there is usually a certain correlation between data of different time steps or space steps. By designing a variance plan and gradually adding Gaussian noise according to the variance plan, the diffusion model can better capture the correlation between different weather data, learn the complex structure and complex patterns of weather data, and improve the accuracy of subsequent weather forecasts.

[0085] In addition, by gradually adding noise variance and increasing the noise variance added at each step, it is also helpful to avoid the gradient explosion problem during the diffusion model training process. If Gaussian noise is directly added to the preliminary weather forecast sample with a larger variance, the preliminary weather forecast sample may very quickly become a form that looks like pure noise data, making it difficult for the diffusion model to accurately learn the complex laws and structures of weather data, and it is difficult to learn the effective information related to weather forecast in weather data, resulting in inaccurate weather forecast results of the trained model.

[0086] Furthermore, in the above forward diffusion process, the Markov chain principle is used, and the result after adding noise according to the variance plan at each step can be determined by the following formula:

[0087]

[0088] Among them, X t represents the data after adding noise in the current step, X t-1 represents the data obtained after adding noise in the previous step of the current step, β t represents the variance corresponding to the addition of Gaussian noise in the tth step of the variance plan, I represents unit variance, N represents Gaussian distribution, q(X t |X t-1 ) indicates that from X t-1 To X t The state transition probability distribution.

[0089] When the noise amplitude added at each step is small enough (i.e., β t is small enough), and the total number of steps T is large enough, the result X t It will be very close to the isotropic Gaussian distribution, which will help the diffusion model to learn more accurately and obtain the data samples of q(X0) in the subsequent back diffusion process.

[0090] By adopting the Markov chain principle in the process of training the diffusion model, since the result of the current step only depends on the result obtained in the previous step of the current step, and the consistency of data distribution can be guaranteed (that is, the distribution of the data obtained after adding noise is still consistent with the distribution of the real data), it helps to simplify model training, speed up training, and ensure the accuracy of the prediction results of the trained model.

[0091] Then, the diffusion model predicts the reverse distribution q(X t-1 |X t ) is used for learning. At this time, samples can be taken from N(0, I) (i.e., XT), and the reverse diffusion process is run to obtain data samples of q(X0). In the reverse diffusion process, the parameterized model (q(X0) is learned. t-1 |X t ) is approximated by a parameterized model), that is, a model that restores the original clear image from the noisy weather data. More specifically, the reverse diffusion process is to gradually reduce the noise of the noisy weather fusion data from the final noisy weather fusion data obtained in the forward diffusion process. Since q(X t-1 |X t ) is a compound Gaussian distribution, so for sufficiently small β t , parameterized model p θIt also conforms to the Gaussian distribution, and the parameterization process only designs the mean and variance. The specific formula is as follows:

[0092] p θ (X t-1 |X T )=N(X t-1 ;μ θ (X t ,t),∑ θ (X t ,t)) (Formula 2)

[0093] Among them, p θ (X t-1 |X T ) represents the parameterized model, X t represents the data after adding noise in the current step, X t-1 represents the data obtained after adding noise in the previous step of the current step, N represents Gaussian distribution, μ θ (X t ,t) represents the mean, ∑ θ (X t ,t) represents the variance.

[0094] Through the above-mentioned reverse diffusion process, the parameterized model is learned and the noise is gradually subtracted, so that the noise weather fusion data is gradually restored to the predicted weather restoration sample, which helps to gradually restore the image details and image clarity of the weather-related data in the noise weather fusion data. Furthermore, based on the noise weather fusion data and the predicted weather restoration sample, the predicted noise data can be calculated. The specific formula involved in the reverse diffusion process of gradually subtracting noise is as follows:

[0095]

[0096] Among them, t ranges from 1 to T (T is the number of steps used to gradually add noise to the noise weather fusion data during the forward diffusion process), X t ~q(X t |X0) is obtained by gradually adding Gaussian noise ε to the preliminary forecast weather samples, ε θ is the predicted noise data. The above formula 3 is the loss function that can be used in the diffusion model training process, which is the MSE loss.

[0097] Furthermore, the extreme weather prediction model also includes a prediction model (for example, a prediction model that has been pre-trained). In other embodiments, the diffusion model can be trained according to all weather variable conditions of the prediction model. During the training process, the prediction model predicts the weather prediction result X in real time. k, the more the time required for prediction is advanced, the lower the prediction accuracy will be. At this time, the diffusion model can be further trained based on the time step k of the prediction model as a condition, and the posterior probability shown in the following formula can be obtained accordingly:

[0098]

[0099] Among them, Y k For X k The corresponding true label, t ranges from 1 to T (T is the number of steps taken to gradually add noise to the noisy weather fusion data during the forward diffusion process).

[0100]

[0101] The diffusion model can directly apply the above X k To make predictions and play a denoising role, the above MSE loss function L can be used for training.

[0102] Furthermore, the loss function used to train the diffusion model can also be a latitude-weighted RMSE loss function. By increasing the latitude weighting, it is more conducive to focusing on the weather conditions at latitudes that are more critical to weather forecasting, reducing prediction errors and improving the accuracy of weather forecasts.

[0103] By adopting the technical solutions of the above-mentioned embodiments, the current weather conditions are first predicted through the prediction model. After obtaining the preliminary prediction results, the preliminary prediction results may have problems such as smooth prediction and inaccurate prediction. The trained diffusion model is then used to further optimize the preliminary prediction results, perform diffusion noise reduction, introduce more weather details, and the introduced details are consistent with the data distribution, which is conducive to further improving the accuracy of weather forecasts.

[0104] Further, continue to refer to Figure 8 In other embodiments, other conditions (Condition) may be introduced when training the diffusion model, such as data related to the time step (k) used in the prediction, data related to the land sea mask (Land seamask), and data related to the terrain (Orography). By introducing the above conditions during training, the diffusion model can more accurately learn the laws of data related to weather prediction such as the atmosphere and terrain, and the trained diffusion model can more accurately optimize and restore the predicted weather results, further improving the accuracy of weather prediction.

[0105] In the above embodiments, only the diffusion model is trained and its parameters are updated, while the prediction model does not participate in the training, and the parameters of the prediction model remain unchanged. For example, the preset training condition can be 120,000 iterations of training for 24 hours, and is performed on a cluster equipped with 8 Nvidia A100 GPUs, with a batch size of 1 for each GPU, and can be optimized using the AdamW optimizer. The relevant parameters can be, for example, β1=0.9 and β2=0.95, and the initial learning rate is 2.5×10 -5 , the weight decay coefficient is 0.1. Furthermore, the Scheduled DropPath with a dropout rate of 0.2 can be used to alleviate the overfitting situation.

[0106] Second, reference Figure 5 The present invention provides a training device 1 for an extreme weather prediction model, wherein the extreme weather prediction model is used to predict extreme weather, and the extreme weather prediction model includes a diffusion model, and the diffusion model includes a forward diffusion module and a reverse diffusion module. The training device 1 for the extreme weather prediction model includes a training preparation module 11 and a training module 12. The training preparation module 11 and the training module 12 are connected.

[0107] The training preparation module 11 is used to determine preliminary predicted weather samples for training the extreme weather prediction model and prediction labels corresponding to the preliminary predicted weather samples.

[0108] The training module 12 is used to train the diffusion model through preliminary predicted weather samples and predicted labels, specifically including: adding noise data to the preliminary predicted weather samples through the forward diffusion module to make the preliminary predicted weather samples become noise weather fusion data; restoring the noise weather fusion data to the predicted weather restoration samples through the reverse diffusion module to obtain predicted noise data; determining the loss function based on the added noise data and the predicted noise data; training the diffusion model based on the loss function, the predicted weather restoration samples and the predicted labels to update the parameters of the diffusion model until the preset training conditions are met, thereby obtaining a trained diffusion model, and thus obtaining a trained extreme weather prediction model.

[0109] By adopting the above technical scheme, the training device 1 of the extreme weather prediction model adds noise to the preliminary predicted weather samples so that the preliminary predicted weather samples become noise weather fusion data, and then obtains the loss function related to the noise, and trains the diffusion model so that the diffusion model can restore the predicted weather results from the noise weather fusion data, and make the weather results predicted by the trained extreme weather prediction model have more weather condition details, thereby improving the weather forecasting skills of the model and improving the accuracy of the model's weather forecast.

[0110] Furthermore, the training module 12 may also be used to execute steps S21, S22, S23, S24, S211 and S212 in the aforementioned embodiments.

[0111] Thirdly, reference Figure 4 The present invention provides an extreme weather prediction method, which predicts extreme weather through an extreme weather prediction model, the extreme weather prediction model is trained by the extreme weather prediction model training method in any of the above embodiments, and the extreme weather prediction model includes a prediction model and a diffusion model. The extreme weather prediction method specifically includes the following steps:

[0112] S01: Determine current weather variable data according to current weather conditions.

[0113] Among them, the current variable data may include 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. By understanding the temperature, humidity, wind speed and air pressure of the current weather in the pressure layer, and understanding the temperature, precipitation and air pressure of the current ground layer, it is helpful to understand the weather details of the current weather system in various aspects, which helps to improve the accuracy of subsequent prediction results.

[0114] S02: Predict the weather conditions at the next preset time using a prediction model based on the current weather variable data to obtain preliminary predicted weather data.

[0115] The preset time may be a time that is a certain time away from the current time, or may be a time that is expected to be predicted. For example, if the current time is 00:00, the next preset time may be 06:00 on the same day as the current time, or 03:00 on the next day after the current time. The next preset time may be determined based on the actual predicted demand. More specifically, refer to Figure 8 , Figure 8 Medium X t The superscript t indicates the current time, for example, 00 o'clock. Figure 8 Medium X t+1 The superscript t+1 indicates the next preset time, for example, 06:00. Figure 8 X in t+1 This is the preliminary forecast weather data.

[0116] Furthermore, X t That is, the current weather variable data, which may include the atmospheric state variables of the current weather input into the model, including the temperature, humidity, wind and pressure of the barometric layer, as well as the temperature, pressure, precipitation conditions of the ground layer, etc.

[0117] S03: Performing detailed recovery processing on the preliminary forecast weather data through the diffusion model to obtain the final forecast weather data; the final forecast weather data is the forecast result of the extreme weather forecast model on the extreme weather at the next preset time.

[0118] Continue to refer Figure 8 , Figure 8 In For the final forecast weather data.

[0119] The forecast results of the extreme weather prediction model are corrected by the diffusion model, that is, the preliminary predicted weather data of the extreme weather prediction model are processed for detail recovery. In the present invention, the diffusion model is a generative model based on DDPM (diffusion probability model), which can transform the original complex data into multiple relatively simple Gaussian distributions, and can generate many details in the iterative denoising process, so that the preliminary predicted weather data can have more details (that is, the details of the predicted weather variables), for example, the spatial structure of the temperature distribution or wind speed distribution can be made clearer and richer.

[0120] By adopting the above technical solution, the extreme weather prediction method predicts extreme weather through the trained extreme weather prediction model, and the obtained extreme weather prediction results have more details and the prediction of extreme weather is more accurate.

[0121] Furthermore, the prediction model of the extreme weather prediction model can be selected from one of the following: FuXi model, Pangu-Weather model, and Graphcast model.

[0122] Among them, the Pangu-Weather model is Huawei Cloud's Pangu meteorological model, which can provide global weather forecasts in seconds. Its meteorological forecast results can include potential, humidity, wind speed, temperature, sea level pressure and other aspects.

[0123] The Graphcast model is a data-driven global weather forecast model developed by researchers from DeepMind and Google. It is a global medium-term weather forecast based on graph neural networks. It can provide medium-term forecasts of key global weather indicators with a resolution of 0.25°, which is equivalent to a spatial resolution of about 25 km x 25 km near the equator and a global grid of 721 x 1440 pixels.

[0124] Preferably, the prediction model is a FuXi model. Specifically, the FuXi model is an autoregressive model that can predict the weather conditions at the next moment based on the weather conditions input at the current moment. Exemplarily, the input weather conditions can include variables of the atmosphere and the ground layer.

[0125] In this embodiment, the FuXi model includes a cube embedding layer, a U-Transformer layer, and a fully connected layer. Among them, the cube embedding layer can be used to reduce the spatial dimension and time dimension of the input data, which helps to speed up the training process of the FuXi model. The U-Transformer layer includes a downsampling layer, an upsampling layer, and an intermediate layer. Among them, the intermediate layer includes a plurality of Swin Transformer V2 modules, for example, 48, and the Swin Transformer V2 module can be used to expand capacity and resolution. The upsampling layer can be used to process the embedded data of the embedding layer. The fully connected layer can be used to generate the final weather forecast results.

[0126] In the above embodiments, the predicted extreme weather may include extreme precipitation, extreme wind speed, extreme temperature and tropical cyclone. Among them, the standard of extreme precipitation is determined according to the classification standard of 24-hour precipitation in the Chinese national standard (GB / T 28592-2012); the standard of extreme wind speed is determined according to the standard of Beaufort wind scale; the standard of extreme temperature is determined according to the percentile threshold in the ERA-Interim surface temperature reanalysis data set; the extreme degree of tropical cyclone is determined according to the difference between the predicted value predicted by the extreme weather prediction model and the actually observed value.

[0127] Exemplarily, the extreme weather prediction model of the present invention is trained for five predicted surface variables, including 2-meter temperature (T2M), 10-meter u wind component (U10), 10-meter v wind component (V10), mean sea level pressure MSL and extreme total precipitation TP, and the forecasting performance of the extreme weather prediction model is tested. Taking the FuXi model used in the extreme weather prediction model of the present invention as an example, the forecasting performance of the extreme weather prediction model (i.e., FuXi-Extreme), the FuXi model and the ECMWF HRES model (hereinafter referred to as the HRES model) are compared.

[0128] The forecast performance tested includes the forecast performance of extreme total precipitation (TP), the forecast performance of 10-meter wind speed (WS10), and the forecast performance of 2-meter temperature (T2M). The dataset used is the ERA reanalysis dataset. Among them, 10-meter wind speed (WS10) refers to the wind speed measured at a standard height of 10 meters above the ground. The height of 10 meters is a standard reference height commonly used for weather forecasting. At this height, the wind speed is less directly affected by ground obstacles (such as buildings, trees, etc.), and can better reflect the movement of the atmosphere. 2-meter temperature (T2M) refers to the temperature measured at a standard height of 2 meters above the ground. This height is close to the general height of human activities and is less affected by surface characteristics (such as soil temperature, vegetation cover, etc.), and can better represent the actual temperature conditions in the lower layers of the atmosphere.

[0129] In addition, the present invention also evaluates the forecast performance of tropical cyclones (TCs) for extreme weather prediction models, and the data set used is the International Best Typhoon Track Dataset (IBTrACS). Each track in the IBTrACS data set represents a 6-hour time series, which represents the position of the tropical cyclone TC eye in latitude and longitude coordinates, as well as other relevant characteristics of the tropical cyclone TC at that specific time and location. When the FuXi model, the extreme weather prediction model, and the HRES model detect a cyclone at the same time, all tropical cyclone tracks are evaluated to ensure that all models use the same set of events for evaluation. In addition to the IBTrACS data set, the ECMWF TC tracking method is applied to the ERA5 data set to extract tropical cyclone TC tracks and intensities for use in tropical cyclone TC forecast evaluation.

[0130] The evaluation indicators for weather forecasts include the critical success index (CSI) and the symmetric extreme dependence index (SEDI). The CSI index can quantify the success of the forecast hit rate. The specific formula is as follows:

[0131]

[0132] Among them, TP represents true positive (number of hits), FP represents false positive (number of false positives), and FN represents false negative (number of missed reports). The range of CSI is 0 to 1. The closer the CSI is to 1, the more accurate the prediction.

[0133] Among them, in order to calculate the CSI indicators of TP and WS10, specific thresholds are required. For the extreme total precipitation accumulated for 6 hours, TP, the 24-hour precipitation value in the Chinese national standard (GB / T 28592-2012) was referenced and divided by 4 to obtain the thresholds of 2.5, 6.25, 25, 62.5 and 70 mm for 6-hour precipitation. For wind speed WS10, the Beaufort wind scale was used, and the wind speed thresholds were set at 13.9, 17.2, 20.8, 24.5, 28.5 and 32.7 m / s.

[0134] The SEDI index is mainly used to evaluate the synchronous occurrence of extreme weather events in different locations or times. The specific formula is as follows:

[0135]

[0136] Among them, TP represents true positive (number of hits), FP represents false positive (number of false alarms), FN represents false negative (number of missed alarms), TN represents true negative, F represents false alarm rate, and H represents hit rate. SEDI values ​​range from -1 to 1. When SEDI is 0, it means that the two extreme weather events are independent of each other; when SEDI is 1, it means that the two extreme weather events are completely synchronized; when SEDI is -1, it means that the two extreme weather events are completely staggered.

[0137] Forecasts of tropical cyclones (TCs) are evaluated in terms of track and intensity. Track error is quantified using the mean absolute error (MAE), measuring the distance between the actual observed and predicted TC center positions. The intensity assessment involves evaluating the maximum WS10 near the TC center and the MSL of the TC center. The intensity forecast is evaluated using the root mean square error (RMSE).

[0138] refer to Fig. 9 , the WS10 and 6-hour cumulative TP of the HRES model, the FuXi model and the extreme weather prediction model of the present invention in the 5-day forecast were evaluated using the test data of the ERA5 dataset in 2018. It can be seen from the figure that at lower thresholds, such as the 2.5 mm threshold for TP, the CSI gap between the FuXi model and the extreme weather prediction model is relatively small, while for more extreme values, the difference increases significantly. In addition, the CSI values ​​of the FuXi model and the extreme weather prediction model are very close at the beginning of the forecast, but as the forecast time increases, the extreme weather prediction model significantly outperforms the FuXi model. Overall, the extreme weather prediction model has the highest CSI scores in WS10 and TP. In terms of WS10, the extreme weather prediction model is also superior to the FuXi model and the HRES model.

[0139] refer to Fig.10, shows the SEDI index scores of T2M, TP and WS10 for the HRES model, FuXi model and extreme weather prediction model based on different percentiles as a function of the forecast lead time for the 5-day forecast. As the percentile increases, the SEDI score decreases, indicating that the prediction of more extreme events is more challenging. Among the three evaluated variables, TP showed the lowest score, while T2M showed the highest score, and the SEDI score of T2M remained above 0.6 throughout the 5-day forecast period. As can be seen from the figure, the extreme weather prediction model achieved the highest scores in all three variables and forecast performance, and has superior forecasting capabilities for extreme T2M, WS10 and TP in the 5-day forecast.

[0140] Typhoon Rumbia occurred from August 15 to 21, 2018. Fig.11 and Fig.12 shows the forecast initialized at 12UTC on August 15, 2018, Fig.11 Showing the situation of TP, Fig.12 The case of WS10 is shown. As can be seen from the figure, the extreme weather prediction model can generate forecasts with more details, especially in the 60-hour and 120-hour forecasts.

[0141] The forecast assessment of tropical cyclones (TCs) includes the assessment of track and intensity predictions.

[0142] Fig.13 The statistical comparison of the HRES model, FuXi model and the extreme weather prediction model in the 5-day forecast is shown, which is an evaluation of the track and intensity forecast using the IBTrACS dataset. In terms of TC track forecast, when the MAE is smaller, it indicates better forecasting ability. The forecasting ability of the FuXi model and the extreme weather prediction model is comparable, and there is no significant difference in the track error, both of which are superior to the HRES model. In terms of TC intensity forecast, the extreme weather prediction model performs better than the FuXi model in terms of WS10. As for the MSL forecast, the difference between the FuXi model and the extreme weather prediction model is negligible. The HRES model achieved the best performance in TC intensity forecast, maintaining the lowest RMSE value throughout the forecast period.

[0143] Fig.14 The statistical comparison of the HRES model, FuXi model and extreme weather prediction model in the 5-day forecast is shown, which is an evaluation of the track and intensity forecast using the ERA5 dataset. In terms of TC track forecast, the FuXi model and the extreme weather prediction model are superior to the HRES model. In terms of TC intensity forecast, the FuXi model and the extreme weather prediction model have smaller RMSE values ​​than the HRES model, and the FuXi model and the extreme weather prediction model are superior to the HRES model.

[0144] In general, according to the above experimental data, taking into account the forecasts of various extreme weather conditions, the weather forecast results of the extreme weather forecast model in the above embodiments of the present invention are more accurate. In the present invention, the detailed features of the weather forecast results generated by the forecast model are enhanced by training the diffusion model, and the weather forecasting skills of the model are improved. The trained extreme weather forecast model has more accurate forecast results when performing extreme weather forecasts.

[0145] Fourthly, reference Figure 6 The present invention provides an extreme weather prediction device 2, which predicts extreme weather through an extreme weather prediction model, the extreme weather prediction model is trained by any of the extreme weather prediction model training methods in the above embodiments, and the extreme weather prediction model includes a prediction model and a diffusion model. The extreme weather prediction device 2 includes a current weather acquisition module 21, a preliminary weather prediction module 22, and a final weather prediction module 23.

[0146] Among them, the current weather acquisition module 21 is connected to the preliminary weather prediction module 22, and the preliminary weather prediction module 22 is connected to the final weather prediction module 23.

[0147] The current weather acquisition module 21 is used to determine the current weather variable data according to the current weather conditions.

[0148] The preliminary weather forecast module 22 is used to predict the weather conditions at the next moment through a forecasting model according to the current weather variable data to obtain preliminary forecast weather data.

[0149] The final weather prediction module 23 is used to perform detail recovery processing on the preliminary predicted weather data through a diffusion model to obtain final predicted weather data; the final predicted weather data is the result of the extreme weather prediction model predicting the extreme weather.

[0150] By adopting the above technical solution, the extreme weather prediction device 2 predicts extreme weather through the trained extreme weather prediction model, and the obtained extreme weather prediction results have more details and the prediction of extreme weather is more accurate.

[0151] Furthermore, the current weather variable data includes: 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. By obtaining the weather variable data of various aspects of the current weather, various details of the current weather conditions can be obtained, which helps the extreme weather prediction model to more accurately predict the weather at the next moment.

[0152] Furthermore, the prediction model in the extreme weather prediction model can be selected from one of the following: FuXi model, Pangu-Weather model, and Graphcast model.

[0153] The extreme weather predicted by the extreme weather prediction model can include extreme precipitation, extreme wind speed, extreme temperature and tropical cyclones. Among them, the standard of extreme precipitation is determined according to the classification standard of 24-hour precipitation in the Chinese national standard (GB / T 28592-2012); the standard of extreme wind speed is determined according to the standard of Beaufort wind scale; the standard of extreme temperature is determined according to the percentile threshold in the ERA-Interim surface temperature reanalysis data set; the extreme degree of tropical cyclones is determined according to the difference between the predicted value predicted by the extreme weather prediction model and the actual observed value.

[0154] Fifth, reference Figure 7 The embodiment of the present invention further discloses 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 training method of the extreme weather prediction model in any of the aforementioned embodiments is implemented, and / or the extreme weather prediction method 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.

[0155] In this embodiment, the electronic device 3 adds noise to the preliminary predicted weather samples to make the preliminary predicted weather samples become noise weather fusion data, and then obtains a loss function related to noise, trains a diffusion model so that the diffusion model can restore the predicted weather results from the noise weather fusion data, and makes the predicted weather results have more weather condition details, thereby improving the weather forecasting skills of the model, and the trained extreme weather prediction model predicts more accurately when performing extreme weather forecasts.

[0156] 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 training method of the extreme weather prediction model in any of the aforementioned embodiments is implemented, and / or the extreme weather prediction method in any of the aforementioned embodiments is implemented.

[0157] By adopting the above technical solution, a computer-readable storage medium adds noise to the preliminary predicted weather samples so that the preliminary predicted weather samples become noise weather fusion data, and then obtains a loss function related to noise, and trains a diffusion model so that the diffusion model can restore the predicted weather results from the noise weather fusion data, and make the predicted weather results have more weather condition details, thereby improving the weather forecasting skills of the model, and the trained extreme weather prediction model predicts more accurately when making extreme weather forecasts.

[0158] 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.

[0159] 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.

[0160] 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.

[0161] 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.

[0162] 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 training method for an extreme weather prediction model, characterized in that: The extreme weather prediction model is used to predict extreme weather. The extreme weather prediction model includes a diffusion model. The diffusion model includes a forward diffusion module and a reverse diffusion module. The training method of the extreme weather prediction model includes: Determining preliminary predicted weather samples for training the extreme weather prediction model and prediction labels corresponding to the preliminary predicted weather samples; The diffusion model is trained by using the preliminary forecast weather samples and the forecast labels, specifically including: adding noise data to the preliminary forecast weather sample through the forward diffusion module, so that the preliminary forecast weather sample becomes noise weather fusion data; The noise weather fusion data is restored to the predicted weather restoration sample by the reverse diffusion module to obtain the predicted noise data; Determine a loss function according to the added noise data and the predicted noise data; The diffusion model is trained according to the loss function, the predicted weather restoration sample and the predicted label to update the parameters of the diffusion model until the preset training conditions are met, thereby obtaining the trained diffusion model and thus obtaining the trained extreme weather prediction model.

2. The extreme weather prediction model training method according to claim 1, characterized in that: The adding noise data to the preliminary forecast weather sample by the forward diffusion module so that the preliminary forecast weather sample becomes noise weather fusion data specifically includes: By using the forward diffusion module, the noise unit data is added to the preliminary forecast weather sample multiple times by using the Markov chain principle, so that the preliminary forecast weather sample becomes noise weather fusion data; Determine the noise data according to the preliminary forecast weather sample and the noise weather fusion data; The noise unit data is Gaussian noise, and the noise data is Gaussian noise distribution data.

3. The extreme weather prediction model training method according to claim 1, characterized in that: The loss function is MSE loss.

4. A training device for an extreme weather prediction model, characterized in that: The extreme weather prediction model is used to predict extreme weather. The extreme weather prediction model includes a diffusion model. The diffusion model includes a forward diffusion module and a reverse diffusion module. The training device of the extreme weather prediction model includes: A training preparation module, used to determine preliminary predicted weather samples for training the extreme weather prediction model and prediction labels corresponding to the preliminary predicted weather samples; A training module is used to train the diffusion model through the preliminary forecast weather samples and the forecast labels, specifically including: adding noise data to the preliminary forecast weather samples through the forward diffusion module to make the preliminary forecast weather samples become noise weather fusion data; restoring the noise weather fusion data to the forecast weather restoration samples through the reverse diffusion module to obtain forecast noise data; determining a loss function based on the added noise data and the forecast noise data; training the diffusion model based on the loss function, the forecast weather restoration samples and the forecast labels to update the parameters of the diffusion model until the preset training conditions are met, thereby obtaining the trained diffusion model, and thus obtaining the trained extreme weather prediction model.

5. An extreme weather prediction method, characterized in that: The extreme weather is predicted by an extreme weather prediction model, wherein the extreme weather prediction model is trained by the extreme weather prediction model training method according to any one of claims 1 to 3, wherein the extreme weather prediction model includes a prediction model and a diffusion model, and the extreme weather prediction method includes: Determine current weather variable data based on current weather conditions; Predicting the weather conditions at the next preset time using the prediction model according to the current weather variable data to obtain preliminary predicted weather data; The preliminary forecast weather data is processed for detail recovery through the diffusion model to obtain final forecast weather data; the final forecast weather data is the forecast result of the extreme weather forecast model on the extreme weather at the next preset time.

6. The extreme weather prediction method according to claim 5, characterized in that: The current weather variable data include: 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.

7. The extreme weather prediction method according to claim 5, characterized in that: The prediction model is selected from one of the following: FuXi model, Pangu-Weather model, and Graphcast model.

8. The extreme weather prediction method according to claim 5, characterized in that: The extreme weather includes: extreme precipitation, extreme wind speed, extreme temperature and tropical cyclones; Among them, the standard of extreme precipitation is determined according to the classification standard of 24-hour precipitation in the Chinese national standard (GB / T 28592-2012); the standard of extreme wind speed is determined according to the standard of Beaufort wind scale; the standard of extreme temperature is determined according to the percentile threshold in the ERA-Interim surface air temperature reanalysis data set; the extreme degree of the tropical cyclone is determined according to the difference between the predicted value predicted by the extreme weather prediction model and the actually observed value.

9. An extreme weather prediction device, characterized in that: The extreme weather is predicted by an extreme weather prediction model, wherein the extreme weather prediction model is trained by the extreme weather prediction model training method according to any one of claims 1 to 3, wherein the extreme weather prediction model includes a prediction model and a diffusion model, and the extreme weather prediction device includes: The current weather acquisition module is used to determine the current weather variable data according to the current weather conditions; A preliminary weather forecast module, used to predict the weather conditions at the next moment through the forecast model according to the current weather variable data, and obtain preliminary forecast weather data; The final weather forecast module is used to perform detail recovery processing on the preliminary forecast weather data through the diffusion model to obtain final forecast weather data; the final forecast weather data is the result of the extreme weather forecast model's prediction of the extreme weather.

10. 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, implements the training method of the extreme weather prediction model as described in any one of claims 1-3, and / or implements the extreme weather prediction method as described in claims 5-8.

11. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores at least one instruction, which, when executed, implements the training method for the extreme weather prediction model as described in any one of claims 1-3, and / or implements the extreme weather prediction method as described in claims 5-8.

Images