A sandstorm weather prediction method, system, device and medium based on generative AI

Abstract

The present invention discloses a sandstorm weather prediction method, system, device, and medium based on generative AI, relating to the technical field of sandstorm prediction. The method comprises: constructing an MM-DDPM model, and using the MM-DDPM model to jointly encode the correlation between meteorological elements and sandstorm labels in a sandstorm dataset to generate sandstorm enhancement samples; using the sandstorm enhancement samples to perform parameter tuning on a pre-trained meteorological large model to obtain a trained sandstorm prediction model, and using the sandstorm prediction model to perform weather prediction to obtain a sandstorm prediction result; the sandstorm prediction result includes the occurrence time, duration, and spatial distribution of the sandstorm event; the pre-trained meteorological large model is constructed based on the Fengwu model and the sandstorm fitting layer. The present invention can break through the limitations of existing technologies in data scarcity and imbalance by generating reliable sandstorm data samples, thereby improving the accuracy and stability of sandstorm predictions.

Description

Technical Field

[0001] The present invention relates to the technical field of sandstorm prediction technology, and in particular to a sandstorm weather prediction method, system, equipment and medium based on generative AI. Background Art

[0002] In the field of sandstorm weather prediction, existing methods are mainly divided into numerical prediction methods based on physical models and statistical prediction methods based on machine learning. However, these methods have significant limitations when dealing with sparse and unbalanced data samples.

[0003] Numerical prediction methods based on physical models are currently a commonly used technology in meteorological forecasting, mainly relying on the simulation of atmospheric physical processes. Common physical models such as the WRF (Weather Research and Forecasting) model simulate the movement and evolution of the atmosphere by solving a series of complex atmospheric dynamics and thermodynamic equations. Although such methods can theoretically capture the physical essence of weather processes, they have the following shortcomings: (1) High computational cost: Numerical prediction models are complex and require a large amount of computing resources, especially in extreme weather events such as sandstorms that involve multiple atmospheric processes. The computational cost is high and the time requirements are strict; (2) Accuracy depends on model parameters: The accuracy of numerical models depends on the initial conditions and the setting of model parameters. These parameters are often difficult to accurately obtain or adjust, resulting in insufficient accuracy in the prediction of local extreme weather events (such as sandstorms).

[0004] In recent years, with the rise of data-driven technology, artificial intelligence (AI) methods represented by machine learning have been widely used in weather forecasting, such as large meteorological models such as PanguWeather, Fengwu, Fuxi, and Graphcast. By using historical meteorological data to train models, machine learning methods can identify patterns in data and predict future weather. However, existing machine learning methods still face many challenges in sandstorm prediction: (1) Strong data dependence: Machine learning models are highly dependent on a large amount of training data, but the occurrence of sandstorm events is scarce, and data samples are usually small and unbalanced, making it difficult for the model to fully learn the laws of sandstorm occurrence, resulting in overfitting or unstable prediction problems; (2) Lack of physical information fusion: Most machine learning models only focus on the statistical patterns of the data itself, fail to effectively combine atmospheric physical processes, and find it difficult to capture the complex meteorological background when sandstorms occur, limiting their prediction accuracy.

[0005] In recent years, generative AI technologies such as Generative Adversarial Network (GAN) and Variational Autoencoder (VAE) have performed well in solving the problem of data imbalance and can enhance the diversity of datasets by generating virtual samples. These technologies have achieved good results in some fields, but they still have the following shortcomings in sandstorm weather forecasting: (1) Single-modal generation: Existing generative AI methods usually only generate a single type of data and fail to consider the joint generation of multiple meteorological factors at the same time. This makes the match between the generated virtual data and the real meteorological conditions not high, affecting the reliability of the generated data; (2) Insufficient physical consistency: Traditional generative models mostly ignore the physical constraints in meteorological data. The generated virtual data may not conform to the actual physical laws, which limits their application in meteorological forecasting.

[0006] Given the limitations of these existing technologies, how to generate sandstorm data samples that can match meteorological conditions has become a major challenge in existing technologies. Summary of the Invention

[0007] The purpose of the present invention is to provide a sandstorm weather prediction method, system, equipment and medium based on generative AI, which can break through the limitations of existing technologies in data scarcity and imbalance by generating reliable sandstorm data samples, and improve the accuracy and stability of sandstorm prediction.

[0008] To achieve the above object, the present invention provides the following solutions:

[0009] A sandstorm weather prediction method based on generative AI, including:

[0010] Obtain sandstorm datasets from historical data;

[0011] Constructing an MM-DDPM model, and using the MM-DDPM model to jointly encode the correlation between meteorological elements and sandstorm labels in the sandstorm dataset to generate sandstorm enhanced samples; the sandstorm enhanced samples include meteorological element data and sandstorm data with physical correlation; the MM-DDPM model is constructed based on a joint encoder and a joint loss;

[0012] The sandstorm enhancement samples are used to tune the parameters of the pre-trained meteorological model to obtain a trained sandstorm prediction model, and the sandstorm prediction model is used to perform weather forecasts to obtain sandstorm prediction results; the sandstorm prediction results include the occurrence time, duration and spatial distribution of the sandstorm event; the pre-trained meteorological model is constructed based on the Fengwu model and the sandstorm fitting layer.

[0013] Optionally, the historical data adopts the sandstorm dataset LSDSSIMR data constructed by FY-4A satellite impact and meteorological reanalysis data, and ERA5 meteorological data.

[0014] Optionally, the specific process of generating the sandstorm enhancement sample is as follows:

[0015] The meteorological element data X w and sandstorm data X d By combining the encoder E(X d , X w ) is converted into a shared latent variable space representation Z, and the data of the two modalities are mapped to the same latent variable space through the joint encoder, so as to achieve modeling and preserve the relationship between the two modalities during the diffusion process; in the forward diffusion process, noise is gradually added to the latent variable Z so that its distribution approaches the standard Gaussian distribution Specifically expressed as:

[0016]

[0017] Where Z t is the latent variable representation of the diffusion process at time t, where t belongs to the integer range [0, T]; the reverse generation process starts from the latent variable Z after adding the maximum Gaussian noise T First, we gradually denoise and restore the initial joint latent variable representation Z0, thereby generating sandstorm data related to meteorological conditions. The conditional probability distribution generated in reverse is:

[0018]

[0019] Where μ θ and Σ θ are the mean and covariance matrices respectively, depending on the current latent variable representation Z t and weather data X w The final latent variable Z0 generated is used to restore the meteorological elements through the decoder D and sandstorm data

[0020] Optionally, the joint loss is expressed as:

[0021]

[0022] Among them, the first term represents the reconstruction loss in the latent variable space, the second term and the third term represent the reconstruction loss of meteorological data and sandstorm data respectively, and λ1 and λ2 are both loss balance coefficients. represents the expected value of the latent variable reconstruction loss under the conditions of joint time t, latent variable space Z0, and noise ∈, and Respectively represent the meteorological data Xw and sandstorm data X d The expectation of the reconstruction error.

[0023] Optionally, the specific process of performing weather forecasting using the sandstorm prediction model is as follows:

[0024] In the sandstorm prediction model, the input data is mapped to the joint latent variable space H through a layer of Transformer shared encoder, which is expressed as:

[0025]

[0026] in, Indicates that the Fengwu model is The output result is the input. Subsequently, a fully connected fitting layer is used to map future meteorological factors to the probability of sandstorm occurrence, and the probability of sandstorm occurrence is output through the Sigmoid activation function:

[0027] Output=σ(W2·ReLU(W1·H+b1)+b2)

[0028] Among them, W1, W2 and b1, b2 are the weights and bias of the network, σ is the Sigmoid activation function, and the output Output represents the probability of sandstorm occurrence. Since sandstorm data is binary data, the binary cross entropy loss function is used to optimize the model. The expression of the loss function is:

[0029]

[0030] Among them, y i For real sandstorm labels, The probability of sandstorm occurrence is predicted by the model. In the sandstorm prediction model, the main part of the meteorological model is frozen and no longer participates in training. Only the parameters are tuned on the sandstorm fitting layer to approximate the sandstorm data. The enhanced samples generated by the MM-DDPM model are used to solve the data imbalance problem. During the parameter tuning process, the enhanced samples and the real sandstorm data are input into the model together, so that the fitting layer can more accurately capture the pattern of sandstorm occurrence. After the parameter tuning is completed, the sandstorm prediction model is used to predict the occurrence of sandstorms in the future time step. Among them, the input meteorological element X T+1:T+K , the model outputs the probability of sandstorms occurring in the future as:

[0031]

[0032] in, is the probability of sandstorm occurrence in the next K time steps, M DustNet The final sandstorm prediction model.

[0033] The present invention also provides a sandstorm weather prediction system based on generative AI, comprising:

[0034] Data acquisition module, used to obtain sandstorm datasets from historical data;

[0035] a sample enhancement module for constructing an MM-DDPM model and using the MM-DDPM model to jointly encode the correlation between meteorological elements and sandstorm labels in the sandstorm dataset to generate sandstorm-enhanced samples; the sandstorm-enhanced samples include meteorological element data and sandstorm data with physical correlation; the MM-DDPM model is constructed based on a joint encoder and a joint loss;

[0036] The weather forecast module is used to use the sandstorm enhancement samples to tune the parameters of the pre-trained meteorological model to obtain a trained sandstorm prediction model, and use the sandstorm prediction model to perform weather forecasts to obtain sandstorm prediction results; the sandstorm prediction results include the occurrence time, duration and spatial distribution of the sandstorm event; the pre-trained meteorological model is constructed based on the Fengwu model and the sandstorm fitting layer.

[0037] The present invention also provides an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute the above-mentioned sandstorm weather prediction method based on generative AI.

[0038] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-mentioned sandstorm weather prediction method based on generative AI.

[0039] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0040] The present invention discloses a sandstorm weather forecasting method, system, device, and medium based on generative AI. The method includes constructing a MM-DDPM model and using it to jointly encode the correlation between meteorological elements and sandstorm labels in a sandstorm dataset to generate enhanced sandstorm samples. The enhanced sandstorm samples are used to optimize the parameters of a pre-trained meteorological model to obtain a trained sandstorm prediction model. The sandstorm prediction model is then used to perform weather forecasts to obtain sandstorm prediction results. The sandstorm prediction results include the occurrence time, duration, and spatial distribution of sandstorm events. The pre-trained meteorological model is constructed based on the Fengwu model and a sandstorm fitting layer. By generating reliable sandstorm data samples, the present invention can overcome the limitations of existing technologies in data scarcity and imbalance, thereby improving the accuracy and stability of sandstorm forecasts. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0042] Figure 1 Schematic diagram of the flow of the sandstorm weather prediction method based on generative AI of the present invention;

[0043] Figure 2 This is a schematic diagram of sample balancing based on the DDPM model in this embodiment;

[0044] Figure 3 Schematic diagram of the multimodal denoising diffusion probability model MM-DDPM model in this embodiment;

[0045] Figure 4 This is a schematic diagram of data flow during DustNet model training in this embodiment;

[0046] Figure 5 This is a comparison chart of the sandstorm prediction effect and the actual observation result on May 31, 2022 in this embodiment; among them, part (a) is the actual observation chart; part (b) is the prediction result chart. DETAILED DESCRIPTION

[0047] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0048] The purpose of the present invention is to provide a sandstorm weather prediction method, system, equipment and medium based on generative AI, which can break through the limitations of existing technologies in data scarcity and imbalance by generating reliable sandstorm data samples, and improve the accuracy and stability of sandstorm prediction.

[0049] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0050] The present invention provides a sandstorm weather prediction method based on generative AI, comprising:

[0051] Step 100: Obtain a sandstorm dataset from historical data.

[0052] Step 200: Construct an MM-DDPM model, and use the MM-DDPM model to jointly encode the correlation between meteorological elements and sandstorm labels in the sandstorm dataset to generate sandstorm enhanced samples; the sandstorm enhanced samples include meteorological element data and sandstorm data with physical correlation; the MM-DDPM model is constructed based on a joint encoder and a joint loss.

[0053] Step 300: Parameters of a pre-trained meteorological model are tuned using the sandstorm enhancement sample to obtain a trained sandstorm prediction model, and the sandstorm prediction model is used to perform weather forecasting to obtain a sandstorm prediction result; the sandstorm prediction result includes the occurrence time, duration, and spatial distribution of the sandstorm event; the pre-trained meteorological model is constructed based on the Fengwu model and the sandstorm fitting layer.

[0054] As can be seen, in this embodiment, first, to address the current situation of sparse and unevenly distributed sandstorm weather dataset samples, a multimodal diffusion model is proposed to generate virtual sandstorm data samples, solving the sample imbalance problem and providing more diverse training data for the model. Secondly, by introducing an existing high-precision pre-trained meteorological large model and utilizing transfer learning technology, a sandstorm fitting layer is added to the pre-trained model to construct a DustNet model that deeply correlates meteorological data with sandstorm events. This model fully utilizes meteorological conditions to gain insight into future sandstorm occurrences, improving the model's accuracy and robustness in predicting sandstorm weather. Extreme weather events such as sandstorms have a serious impact on human production and life. The implementation of this invention will make a significant contribution to reducing disaster losses and protecting people's lives and property.

[0055] Based on the above scheme, the following embodiments are provided.

[0056] like Figure 1 As shown, this embodiment combines a multimodal diffusion model with a large meteorological model to generate sandstorm samples that match real meteorological conditions, thereby improving the model's ability to predict sandstorm weather.

[0057] First, the study was conducted using the LSDSSIMR data, a sandstorm dataset constructed based on the FY-4A satellite impact and meteorological reanalysis data, and ERA5 meteorological data. The LSDSSIMR dataset covers most of the areas where sandstorms occur most frequently, with a spatial resolution of 4km and an image size of 640×1280. The data was collected at intervals of 15 minutes from March to May every year from 2020 to 2022, providing a large amount of research data for sandstorm research. Although LSDSSIMR provides rich research data, sandstorms are still extremely rare events. Therefore, in this embodiment, the sample balancing method is first explored to expand the meteorological elements and the number of sandstorm labels for sandstorms based on the generative AI-diffusion model, such as Figure 2 As shown in Figure 2. The DDPM model is a type of diffusion model. Through forward diffusion and reverse denoising, the model fully learns the distribution of samples, then generates new samples from the noise with the same distribution as the samples, achieving sample balance from the perspective of data enhancement. However, in the research task of sandstorm prediction based on meteorological input, simply generating sandstorm labeled samples still cannot achieve an effective sample process. Therefore, it is necessary to develop a generation method that simultaneously generates meteorological element and sandstorm label data and ensures that there is a realistic correlation between them.

[0058] Based on this, the present invention first constructs a Multi-Model Denoising Diffusion Probabilistic Model (MM-DDPM model), which jointly encodes the correlation between meteorological elements and sandstorm labels, and simultaneously generates meteorological element data and sandstorm data with physical correlation, thereby achieving sample balance of the sandstorm data set. The MM-DDPM model not only retains the ability of the DDPM model to gradually denoise and generate data, but also introduces a multimodal information processing module into the model structure, so that the generated sandstorm data samples are consistent with the meteorological data. Compared with the conventional DDPM model, the MM-DDPM model takes into account more interdependencies between modes during the generation process. For example, the specific structure comparison Figure 3 The core idea of ​​the MM-DDPM model is to represent the joint distribution of meteorological elements and sandstorm data as part of the diffusion process based on conditional generation, and to generate sandstorm data that matches the meteorological conditions through reverse denoising. The implementation process is as follows:

[0059] First, the meteorological element data X w and sandstorm data X d By combining the encoder E(X d , X w) is converted into a shared latent variable space representation Z. This joint encoder ensures that the data of the two modalities are mapped to the same latent variable space, so that their mutual relationship during the diffusion process can be modeled and preserved. During the forward diffusion process, the latent variable Z is gradually added with noise so that its distribution approaches the standard Gaussian distribution.

[0060]

[0061] Where Z t is the latent variable representation of the diffusion process at time t, where t belongs to the integer range [0, T]; the reverse generation process starts from the latent variable Z after adding the maximum Gaussian noise T First, we gradually denoise and restore the initial joint latent variable representation Z0, thereby generating sandstorm data related to meteorological conditions. The conditional probability distribution generated in reverse is:

[0062]

[0063] Where μ θ and Σ θ are the mean and covariance matrices respectively, depending on the current latent variable representation Z t and weather data X w The generated final latent variable representation Z0 can be restored to the meteorological elements through the decoder D and sandstorm data

[0064]

[0065] During training, the loss function of the MM-DDPM model is designed to simultaneously reconstruct meteorological data and dust storm data to ensure that the generated results are physically consistent:

[0066]

[0067] Among them, the first term represents the reconstruction loss in the latent variable space, the second and third terms represent the reconstruction losses of meteorological data and sandstorm data respectively, and λ1 and λ2 are the balance coefficients of the loss. represents the expected value of the latent variable reconstruction loss under the conditions of joint time t, latent variable space Z0, and noise ∈, and Respectively represent the meteorological data X w and sandstorm data X d By minimizing the loss function L and optimizing the model parameters θ, we can ensure that the generated meteorological and sandstorm data are physically consistent, and finally obtain the generated meteorological elements. and sandstorm data Through this joint encoding and diffusion mechanism, the MM-DDPM model can learn the complex correlations between meteorological elements and sandstorm data in the latent variable space, and simultaneously generate physically related data samples, thereby balancing the sandstorm dataset and improving the model's generalization ability and prediction accuracy.

[0068] After generating a virtual sandstorm dataset, meteorological elements are used to approximate sandstorm label values ​​based on a pre-trained meteorological model (taking Fengwu as an example). Since generating sandstorm data directly from future meteorological data obtained from the Fengwu model will lead to error accumulation, a combination of transfer learning and fine-tuning is adopted. By introducing a sandstorm fitting module, the model output is adjusted to model the relationship between meteorological data and sandstorm events. Finally, the sandstorm weather prediction model DustNet is obtained through transfer learning and fine-tuning based on the meteorological model and sample balanced sandstorm data, as shown in the figure. Figure 4 shown.

[0069] First, in DustNet, future meteorological factors predicted by large meteorological models cannot be directly used for sandstorm prediction. Therefore, a sandstorm fitting layer is introduced to approximate the probability of sandstorm occurrence. To better map meteorological factors to sandstorm data, the model uses a layer of Transformer shared encoder to map the input data to a joint latent variable space H, which is expressed as:

[0070]

[0071] in Indicates Fengwu is the output result of the input. Subsequently, a fully connected fitting layer is used to map future meteorological factors to the probability of sandstorm occurrence, and the probability of sandstorm occurrence is output through the Sigmoid activation function:

[0072] Output=σ(W2·ReLU(W1·H+b1)+b2)

[0073] Where W1, W2 and b1, b2 are the network weights and biases, σ is the Sigmoid activation function, and the output represents the probability of a sandstorm. Since the sandstorm data is binary data, the binary cross entropy loss function is used to optimize the model. The loss function is expressed as:

[0074]

[0075] Among them, y i For real sandstorm labels, The probability of sandstorm occurrence predicted by the model. In DustNet, the main part of the meteorological model is frozen, that is, it no longer participates in training. Only fine-tuning is performed on the sandstorm fitting layer to approximate the sandstorm data. The enhanced samples generated by the multimodal diffusion model can alleviate the data imbalance problem. During the fine-tuning process, the enhanced samples and the real sandstorm data are input into the model together, so that the fitting layer can more accurately capture the pattern of sandstorm occurrence. After fine-tuning is completed, the DustNet model can predict the occurrence of sandstorms in the future time step. Specifically, the meteorological element X is input. T+1 :T+K, the model outputs the probability of sandstorms occurring in the future:

[0076]

[0077] in, is the probability of sandstorm occurrence in the next K time steps, M DustNet This is the final DustNet model.

[0078] The technical solution in this embodiment uses the MM-DDPM model to generate virtual sandstorm data samples, addressing the data imbalance issue. It then constructs the DustNet sandstorm prediction model to accurately predict future sandstorms. This method not only improves the model's predictive capabilities but also expands the application scope of existing large-scale meteorological models. It has broad application prospects in extreme weather forecasting and is expected to provide a scientific basis and technical support for scientific research and practical applications in fields such as meteorological forecasting and disaster prevention.

[0079] Therefore, the present invention innovatively solves the problem of scarce and unevenly distributed sandstorm data samples. By introducing a multimodal diffusion model (MM-DDPM model), the present invention can generate virtual sandstorm data samples that are consistent with actual physical conditions, significantly improve the diversity of the data set, enhance the training effect of the model, and thus effectively improve the accuracy and generalization ability of sandstorm predictions. In addition, the present invention also freezes the main part of the pre-trained meteorological model, adds a fitting layer specifically for fitting sandstorm data, and combines it with transfer learning technology to accurately approximate future meteorological elements to the occurrence of sandstorms. This process effectively avoids the phenomenon of gradual accumulation of errors, while greatly improving the accuracy of predictions and the robustness of the model.

[0080] Specifically, the MM-DDPM model generates virtual dust storm data to compensate for the scarcity of original dataset samples. Figure 5The sandstorm prediction results shown intuitively demonstrate the effectiveness and accuracy of the present invention. After the generated enhanced samples were used for model training, the model's prediction results were highly consistent with the actual occurrence of sandstorms. The model not only successfully captured the occurrence time of sandstorm events, but was also able to accurately predict their duration and spatial distribution, demonstrating extremely high practical value. Compared with traditional methods, the present invention avoids the problem of gradual error accumulation during the prediction process by adding a fitting layer on the basis of a pre-trained atmospheric model and combining it with fine-tuning technology, thereby ensuring the stability and reliability of sandstorm predictions. While improving the prediction accuracy, the present invention also greatly enhances the model's ability to cope with complex meteorological conditions.

[0081] Furthermore, the present invention innovatively proposes a sandstorm sample generation method based on a multimodal denoising probability diffusion model (MM-DDPM model), which can effectively generate physically consistent virtual sandstorm data. This model jointly encodes meteorological elements and sandstorm label data to generate virtual samples that match actual meteorological conditions, thereby compensating for the sparse and uneven distribution of sandstorm datasets. This innovation lies in increasing the diversity of the dataset through the generated virtual data, providing more sample support for the training of sandstorm prediction models.

[0082] In addition, the present invention proposes a method based on transfer learning, which introduces a pre-trained meteorological large model (taking FengWu as an example) to construct a sandstorm meteorological large model DustNet. By freezing the main part of the meteorological large model and adding a fitting layer specifically for fitting sandstorm events, future meteorological elements are deeply associated with the occurrence of sandstorms. The fitting layer approaches future meteorological forecast data to sandstorm data through nonlinear mapping of meteorological elements, effectively avoiding the error accumulation problem and enhancing the prediction accuracy of the model. This design effectively shortens the model training time and improves the specificity for sandstorm prediction through transfer learning technology. By using the enhanced samples generated by the MM-DDPM model, the model can achieve sample balance during the training process, further improving the generalization ability of the prediction model. The fine-tuning process targets the binarization characteristics of sandstorm data and combines the generated data for label fitting, thereby ensuring the accuracy and reliability of the prediction results.

[0083] The DustNet sandstorm prediction model proposed in this paper is based on a pre-trained large-scale meteorological model and incorporates a fitting layer. Through a multi-layered network structure, it establishes the complex relationship between meteorological factors and sandstorm events, achieving high prediction accuracy and robustness. By inputting meteorological data over a specific time span, the DustNet model can predict the time and location of future sandstorm events, significantly improving the ability to predict extreme weather events.

[0084] By combining key technologies and innovative points, the present invention effectively solves problems such as uneven sandstorm data and low prediction accuracy, and has broad application prospects in the fields of meteorological forecasting and disaster prevention.

[0085] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0086] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A sandstorm weather prediction method based on generative AI, characterized in that: include: Obtain sandstorm datasets from historical data; Constructing an MM-DDPM model, and using the MM-DDPM model to jointly encode the correlation between meteorological elements and sandstorm labels in the sandstorm dataset to generate sandstorm enhanced samples; the sandstorm enhanced samples include meteorological element data and sandstorm data with physical correlation; the MM-DDPM model is constructed based on a joint encoder and a joint loss; Parameters of a pre-trained meteorological model are tuned using the sandstorm enhancement samples to obtain a trained sandstorm prediction model, and weather forecasting is performed using the sandstorm prediction model to obtain a sandstorm prediction result; the sandstorm prediction result includes the occurrence time, duration, and spatial distribution of the sandstorm event; the pre-trained meteorological model is constructed based on the Fengwu model and the sandstorm fitting layer; The specific generation process of the sandstorm enhanced sample is as follows: The meteorological element data X w and sandstorm data X d By combining the encoder E(X d , X w ) is converted into a shared latent variable space representation Z, and the data of the two modalities are mapped to the same latent variable space through the joint encoder, so as to achieve modeling and preserve the relationship between the two modalities during the diffusion process; in the forward diffusion process, noise is gradually added to the latent variable Z so that its distribution approaches the standard Gaussian distribution Specifically expressed as: Where Z t is the latent variable representation of the diffusion process at time t, where t belongs to the integer range [0, T]; the reverse generation process starts from the latent variable Z after adding the maximum Gaussian noise T First, we gradually denoise and restore the initial joint latent variable representation Z0, thereby generating sandstorm data related to meteorological conditions. The conditional probability distribution generated in reverse is: Where, and are the mean and covariance matrices respectively, depending on the current latent variable representation Z t and weather data X w The final latent variable Z0 generated is used to restore the meteorological elements through the decoder D and sandstorm data The specific process of weather forecasting using the sandstorm prediction model is as follows: In the sandstorm prediction model, the input data is mapped to the joint latent variable space H through a layer of Transformer shared encoder, which is expressed as: in, Indicates that the Fengwu model is The output result is the input. Subsequently, a fully connected fitting layer is used to map future meteorological factors to the probability of sandstorm occurrence, and the probability of sandstorm occurrence is output through the Sigmoid activation function: Output=σ(W2·ReLU(W1·H+b1)+b2) Among them, W1, W2 and b1, b2 are the weights and bias of the network, σ is the Sigmoid activation function, and the output Output represents the probability of sandstorm occurrence. Since sandstorm data is binary data, the binary cross entropy loss function is used to optimize the model. The expression of the loss function is: Among them, y i For real sandstorm labels, The probability of sandstorm occurrence is predicted by the model. In the sandstorm prediction model, the main part of the meteorological model is frozen and no longer participates in training. Only the parameters are tuned on the sandstorm fitting layer to approximate the sandstorm data. The enhanced samples generated by the MM-DDPM model are used to solve the data imbalance problem. During the parameter tuning process, the enhanced samples and the real sandstorm data are input into the model together, so that the fitting layer can more accurately capture the pattern of sandstorm occurrence. After the parameter tuning is completed, the sandstorm prediction model is used to predict the occurrence of sandstorms in the future time step. Among them, the input meteorological element X T+1:T+K , the model outputs the probability of sandstorms occurring in the future as: in, is the probability of sandstorm occurrence in the next K time steps, M DustNet The final sandstorm prediction model.

2. The sandstorm weather prediction method based on generative AI according to claim 1 is characterized in that: The historical data are LSDSSIMR data of sandstorm dataset constructed by FY-4A satellite impact and meteorological reanalysis data, and ERA5 meteorological data.

3. The sandstorm weather prediction method based on generative AI according to claim 1 is characterized in that: The joint loss is expressed as: Among them, the first term represents the reconstruction loss in the latent variable space, the second term and the third term represent the reconstruction loss of meteorological data and sandstorm data respectively, and λ1 and λ2 are both loss balance coefficients. represents the expected value of the latent variable reconstruction loss under the conditions of joint time t, latent variable space Z0, and noise ∈, and Respectively represent the meteorological data X w and sandstorm data X d The expectation of the reconstruction error.

4. A sandstorm weather forecasting system based on generative AI, applied to the method according to any one of claims 1 to 3, characterized in that: include: Data acquisition module, used to obtain sandstorm datasets from historical data; a sample enhancement module for constructing an MM-DDPM model and using the MM-DDPM model to jointly encode the correlation between meteorological elements and sandstorm labels in the sandstorm dataset to generate sandstorm-enhanced samples; the sandstorm-enhanced samples include meteorological element data and sandstorm data with physical correlation; the MM-DDPM model is constructed based on a joint encoder and a joint loss; The weather forecast module is used to use the sandstorm enhancement samples to tune the parameters of the pre-trained meteorological model to obtain a trained sandstorm prediction model, and use the sandstorm prediction model to perform weather forecasts to obtain sandstorm prediction results; the sandstorm prediction results include the occurrence time, duration and spatial distribution of the sandstorm event; the pre-trained meteorological model is constructed based on the Fengwu model and the sandstorm fitting layer.

5. An electronic device, characterized in that: It includes a memory and a processor, the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform the sandstorm weather prediction method based on generative AI according to any one of claims 1-3.

6. A computer-readable storage medium, characterized in that It stores a computer program, which, when executed by a processor, implements the sandstorm weather prediction method based on generative AI as described in any one of claims 1 to 3.

Images