Global cyclone forecasting method combining physical mechanism and data driving

Abstract

The invention provides a global cyclone forecasting method combining a physical mechanism and data driving, and relates to the field of cyclone forecasting. The invention provides a potential diffusion model architecture of physical inspiration, deep fusion of a Transform codec, physical constraint explicit modeling and a multi-modal information fusion technology. The architecture realizes technical innovation in the following key links: integrating historical tropical cyclone data, historical environmental field data and future environmental field forecast data, and providing comprehensive environmental information support for prediction; in a core modeling link, input is mapped to a high-dimensional submerged space for denoising and then decoded back to a data space to generate a prediction sequence, and particularly, a physical inspired gating feature interaction mechanism is embedded in a decoding stage so as to explicitly model the physical constraint relationship among the trajectory, the wind speed and the air pressure. According to the scheme, the stability and physical consistency of tropical cyclone prediction on the long-term time scale are effectively improved, and the method has remarkable technical advantages and application value.

Description

Technical Field

[0001] This invention relates to the field of cyclone forecasting, specifically to a global cyclone forecasting method that combines physical mechanisms with data-driven approaches. Background Technology

[0002] Cyclone forecasting is a core research direction in the field of meteorological disaster early warning. Its core objective is to accurately predict key attributes of cyclones, such as their trajectory (latitude and longitude), central pressure, and maximum sustained wind speed, so as to provide support for disaster prevention and mitigation.

[0003] Among related technologies, the paper (Zhang, S., Mu, P., Huang, C., Zhang, J., & Bai, C. (2025). TC-Diffuser: Bi-Condition Multi-Modal Diffusion for Tropical CycloneForecasting. Proceedings of the AAAI Conference on Artificial Intelligence, 39(1). https: / / doi.org / 10.1609 / aaai.v39i1.32099) discloses a bi-conditional multimodal diffusion model (TC-Diffuser), mainly used for joint forecasting of the trajectory and intensity of global tropical cyclones. The core advantage of this scheme is that it is the first to combine the probability generation capability of diffusion models with multimodal fusion technology, and models the evolution process of tropical cyclones through "bi-conditional constraints" (historical observation data and environmental field data), which solves the problem that traditional deep learning models are difficult to handle uncertainty; at the same time, the integration of multimodal data improves the richness of feature representation, and the stability of short-term forecasts (such as 6-24 hours) is better than some traditional models.

[0004] However, although the above schemes utilize the probability generation capabilities of the diffusion framework to handle uncertainty, they cannot fundamentally solve the problem of physical consistency of forecast results by relying solely on historical data and environmental fields for dual-condition constraints, thus limiting long-term forecast performance. Summary of the Invention

[0005] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a global cyclone forecasting method that combines physical mechanisms with data-driven approaches, thus solving the problem of physical consistency in global cyclone forecasting results.

[0006] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: A global cyclone forecasting method that combines physical mechanisms with data-driven approaches includes: Acquire and preprocess historical tropical cyclone data, historical environmental field data, and future environmental field forecast data; among which, historical tropical cyclone data includes historical trajectory, wind speed, and air pressure data; Under the physical-inspired latent diffusion model architecture, historical tropical cyclone data, environmental field data obtained by splicing historical environmental field data and future environmental field forecast data, and the current denoised time step are vectorized respectively. The three types of vectorized features are aligned and fused and used as input to the Transformer encoder. A unified context vector is output through a multi-head self-attention mechanism. Starting from pure Gaussian noise and the maximum denoising time step, the current noisy latent variable, the current denoising time step, and the unified context vector are used as inputs to the physics-inspired decoder. The output of the previous noisy latent variable is updated in reverse to gradually generate the denoised latent variable. The physics-inspired decoder adopts a Transformer decoder with an integrated physics-inspired gating attention module. The physics-inspired gating attention module is used to explicitly model the physical constraints between task-specific features such as trajectory, wind speed, and air pressure. The denoised latent variables are used as input to the convolutional decoder to predict the trajectory, wind speed, and pressure sequence of future tropical cyclones.

[0007] Preferably, historical tropical cyclone data is vectorized using a GRU time-series encoder; A sliding window Transformer encoder is used to vectorize historical environmental field data and future environmental field forecast data.

[0008] Preferably, the physical heuristic decoder is composed of several stacked decoding blocks with identical structures, each decoding block integrating a physically heuristic gated attention module with identical structure; in the process of reverse updating the output of a noisy latent variable, each layer of decoding blocks is used for: The current denoising time step is encoded into a temporal embedding vector and merged with the output of the previous decoding block to obtain time-aware latent variables; wherein, the first decoding block merges the temporal embedding vector with the current noisy latent variable output from the previous denoising step; Self-attention feature modeling of time-perception latent variables is performed using a multi-head self-attention mechanism; By using a multi-head cross-attention mechanism to interact with self-attention features and a unified context vector, fused features are obtained. By using a physics-inspired gating attention module, the physical constraints between task-specific features such as trajectory, wind speed, and air pressure are explicitly modeled to obtain physically consistent unified features. The unified features are updated through a feedforward network and output to the next layer of the decoding block; the last layer of the decoding block outputs the previous noisy latent variable.

[0009] Preferably, the step of explicitly modeling the physical constraints between task-specific features such as trajectory, wind speed, and air pressure using a physics-inspired gating attention module to obtain physically consistent unified features includes: The fused features are decoupled into task-specific features corresponding to trajectory, wind speed and air pressure respectively by task projection mapping; By using a cross-task attention mechanism, specific features of each task can interact with each other, and the interaction features of specific features of each task can be obtained respectively. A gating adjustment mechanism is introduced to adaptively balance the contribution ratio of the currently selected task-specific features and their interaction features in order to update the currently selected task-specific features. Reintegrate the updated task-specific features to obtain a physically consistent unified feature.

[0010] Preferably, a joint optimization strategy combining the diffusion noise prediction objective and the result reconstruction objective is used to train the potential diffusion model architecture.

[0011] Preferably, a learnable uncertainty weighting mechanism is used to dynamically balance the losses corresponding to the diffusion noise prediction target and the result reconstruction target; wherein, the uncertainty weighting mechanism refers to dynamically adjusting the weights of the two types of losses by introducing trainable parameters for diffusion loss uncertainty and reconstruction loss uncertainty respectively.

[0012] A global cyclone forecasting system that combines physical mechanisms with data-driven approaches includes: The acquisition and preprocessing module is used to acquire and preprocess historical tropical cyclone data, historical environmental field data, and future environmental field forecast data; among which, historical tropical cyclone data includes historical trajectory, wind speed, and air pressure data; The encoding module is used to vectorize historical tropical cyclone data, environmental field data obtained by splicing historical environmental field data and future environmental field forecast data, and the current denoising time step under the physical-inspired latent diffusion model architecture. The three types of vectorized features are aligned and fused and used as input to the Transformer encoder. A unified context vector is output through a multi-head self-attention mechanism. The denoising module starts from pure Gaussian noise and the maximum denoising time step, and uses the current noisy latent variable, the current denoising time step, and the unified context vector as input to the physics-inspired decoder. It then updates the output of the previous noisy latent variable in reverse order to gradually generate denoised latent variables. The physics-inspired decoder uses a Transformer decoder with an integrated physics-inspired gating attention module, which is used to explicitly model the physical constraints between task-specific features such as trajectory, wind speed, and air pressure. The prediction module is used to take the denoised latent variables as input to the convolutional decoder to predict the trajectory, wind speed, and pressure sequence of future tropical cyclones.

[0013] A storage medium storing a computer program, wherein the computer program causes a computer to perform the global cyclone forecasting method described above.

[0014] An electronic device, comprising: One or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing cyclone forecasting as described above.

[0015] (III) Beneficial Effects This invention provides a global cyclone forecasting method that combines physical mechanisms with data-driven approaches. Compared with existing technologies, it has the following advantages: This invention proposes a physically inspired latent diffusion model architecture that deeply integrates Transformer encoder-decoder, physically constrained explicit modeling, and multimodal information fusion techniques. This architecture achieves technological innovation in the following key aspects: integrating historical tropical cyclone data, historical environmental field data, and future environmental field forecast data to provide comprehensive environmental information support for prediction; in the core modeling stage, the input is first mapped to a high-dimensional latent space for denoising, and then decoded back into the data space to generate the prediction sequence. In particular, a physically inspired gated feature interaction mechanism is embedded in the decoding stage to explicitly model the physical constraint relationships between trajectory, wind speed, and air pressure. This solution effectively improves the stability and physical consistency of tropical cyclone prediction over long-term timescales, demonstrating significant technical advantages and application value. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A block diagram illustrating a global cyclone forecasting method combining physical mechanisms and data-driven approaches, provided as an embodiment of the present invention. Figure 2 A flowchart illustrating a global cyclone forecasting method combining physical mechanisms and data-driven approaches, provided as an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a physical heuristic decoder provided in an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] The Phys-Diff model architecture provided by this invention is an integrated technical solution that deeply integrates Transformer codec, physical constraint explicit modeling, and multimodal information fusion technology. Overall, this technical solution is a cyclone prediction method based on latent space diffusion generation. By introducing multimodal condition constraints and a physically inspired consistency modeling mechanism, it achieves the collaborative prediction of the future trajectory (such as latitude and longitude) and intensity elements (such as central pressure and maximum sustained wind speed) of tropical cyclones.

[0020] The key innovations of this invention are as follows: 1) We designed a Physics-Inspired Gated Attention (PIGA) module. Through a four-step process of “decomposition-interaction-gating-fusion”, we decoupled the features of tropical cyclones into three task-specific streams: trajectory, wind speed, and air pressure. Then, we used cross-task attention to capture the physical dependencies between attributes, which effectively enhanced the physical consistency of the prediction and solved the problem of the lack of physical correlation caused by the independent modeling of tropical cyclone attributes in existing deep learning methods.

[0021] 2) Introduce a multimodal data fusion strategy to integrate historical tropical cyclone data, ERA5 historical environmental reanalysis data, and FengWu future environmental forecast data. Combine this with time step embedding to generate a unified context vector, providing comprehensive environmental information support for forecasting and improving the effectiveness of long-term forecasts.

[0022] 3) Construct a physics-inspired latent diffusion model architecture. Based on the Transformer encoder and decoder, the input is first mapped to a high-dimensional latent space for diffusion and denoising, and then decoded back to the data space to generate predictions. This breaks through the limitations of traditional diffusion models that directly process raw data, enhances the model's ability to represent complex cyclone data distributions, and ensures prediction reliability.

[0023] 4) An uncertainty-weighted dual-loss training strategy is proposed, which combines diffusion loss (used to measure the mean square error between predicted noise and real noise) and reconstruction loss (used to measure the error between predicted sequence and true value). By dynamically balancing the weights of the two losses through learnable parameters, the model training is stabilized and the overall prediction accuracy is further improved.

[0024] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0025] Example 1: like Figure 1 As shown, this embodiment of the invention provides a global cyclone forecasting method that combines physical mechanisms with data-driven approaches, including: S1. Acquire and preprocess historical tropical cyclone data, historical environmental field data, and future environmental field forecast data; among which, historical tropical cyclone data includes historical trajectory, wind speed, and air pressure data; S2. Under the physical-inspired latent diffusion model architecture, historical tropical cyclone data, environmental field data obtained by splicing historical environmental field data and future environmental field forecast data, and the current denoising time step are vectorized respectively. The three types of vectorized features are aligned and fused and used as the input of the Transformer encoder. A unified context vector is output through a multi-head self-attention mechanism. S3. Starting from pure Gaussian noise and the maximum denoising time step, the current noisy latent variable, the current denoising time step, and the unified context vector are used as inputs to the physics-inspired decoder. The output of the previous noisy latent variable is updated in reverse to gradually generate the denoised latent variable. The physics-inspired decoder adopts a Transformer decoder with an integrated physics-inspired gating attention module. The physics-inspired gating attention module is used to explicitly model the physical constraints between task-specific features such as trajectory, wind speed, and air pressure. S4. Use the denoised latent variables as input to the convolutional decoder to predict the trajectory, wind speed, and pressure sequence of future tropical cyclones.

[0026] This solution effectively improves the stability and physical consistency of tropical cyclone forecasts over long timescales, demonstrating significant technical advantages and application value.

[0027] It should be noted that the present invention does not directly perform regression modeling in the original physical quantity space, but instead constructs a physics-inspired latent diffusion model architecture. The predicted target is first mapped to the latent space, and a diffusion generation process involving progressive noise addition and denoising is performed within the latent space. Specifically: like Figure 2 As shown, Figure 2 A flowchart of a global cyclone forecasting method that combines physical mechanisms with data-driven approaches is disclosed.

[0028] See Figure 2 During the training phase, a convolutional encoder is set up. The future tropical cyclone sequence (true value) x0 is mapped to the initial latent variable z0 = E(x0). The diffusion model defines a noise scheduling sequence of length T. Let αt =1-β t , Where β t Let α be the noise injection rate at the t-th diffusion time step; t α s Let be the signal retention coefficients at the t-th and s-th diffusion time steps, respectively; is the cumulative signal retention coefficient at the t-th diffusion time step.

[0029] During forward diffusion (noise addition), the model gradually disrupts the latent variable structure using a fixed Gaussian noise injection strategy, causing it to... Evolved to approximately isotropic pure noise The forward process satisfies a Markov chain. , Let be the noisy latent variable at the t-th diffusion time step. Let be the probability density function of a multivariate Gaussian distribution. It is the mean vector. Let I be the covariance matrix and I be the identity matrix. The whole represents the initial latent variables. by Linear decay, with added variance of Gaussian noise. Equivalently, the result of a single step of noise addition can be represented in the form of a single sample: The true noise ε follows a standard multivariate Gaussian distribution with a mean of 0 and a covariance of identity matrix I. Used to control the signal retention ratio Used to control noise intensity. It is understandable that as the diffusion time step t increases, The effective information gradually diminishes, eventually... This provides a "learnable, unified starting point" for reverse generation.

[0030] In the reverse denoising process, the model explicitly introduces physical constraints. It's important to emphasize that physical consistency isn't just post-processed at the final output, but rather persists throughout each reverse time step: during each denoising step... The computation is performed jointly by the Transformer decoder and the PIGA gated interaction. For a noise prediction network with parameter θ, Let be the noisy latent variable at the t-th denoising time step. During the training phase During the inference phase, the sampled data is pure Gaussian noise. Specifically, a unified context vector c (the acquisition process is detailed below) is first injected into the denoised features through cross-attention. Then, within the block, task-specific features of trajectory / wind speed / air pressure are decomposed, and cross-task attention interactions and gating fusion are performed to ensure that the physical coupling relationship is effectively implemented. The noise estimation stage is encoded, thus directly affecting the noisy latent variables at the (t-1)th denoising time step. The update direction ensures that physical consistency is continuously strengthened during the gradual generation process. Based on the noise prediction results, the reverse update adopts the following standard DDPM (Denoising Diffusion Probabilistic Models) form: The random noise w follows a standard multivariate Gaussian distribution with a mean of 0 and a covariance of identity matrix I. When t=1, w=0 is taken to obtain a deterministic output; σ t This is a preset or fixed backsampling variance (used to preserve the generated uncertainty). The physical meaning of this iteration is: the network first estimates the noise added in the forward process, and then adjusts it according to the scheduling coefficients. Gradually stripped away, making Compare To make it "cleaner", iteratively obtain the denoised latent variables. Ultimately, the convolutional encoder is used to predict future tropical cyclone sequences. .

[0031] In particular, during the training phase, this embodiment of the invention employs a joint optimization strategy that combines the diffusion noise prediction objective with the result reconstruction objective.

[0032] Firstly, regarding the core objective of diffusion denoising, the diffusion loss is defined as the difference between the actual noise ε and the network predicted noise. The error (e.g., mean squared error MSE) is calculated, and the expectation is performed on time step t, initial latent variable z0, and true noise ε during training: in, For the spread loss, For mathematical expectation, It is a 2-norm.

[0033] Secondly, to ensure consistency between the final generated results and the actual future state, a reconstruction loss is defined as a direct measure of the predicted future tropical cyclone sequence. Error (e.g., mean square error, MSE) between the future tropical cyclone sequence (true value) x0 and the actual value: in, To rebuild the losses.

[0034] Furthermore, to avoid the two types of objectives hindering each other during training and to improve the stability of multi-task training, a learnable uncertainty weighting mechanism is introduced to dynamically balance the two losses. The overall objective function is: in, and These are the learnable parameters corresponding to diffusion loss and reconstruction loss, respectively, used to characterize the uncertainties of diffusion denoising task and reconstruction task and achieve adaptive trade-off.

[0035] Understandably, the aforementioned joint optimization method ensures both the correctness of the stepwise denoising process and a high degree of consistency between the final prediction and the true future state. Furthermore, it dynamically balances different loss terms using learnable uncertainty weights to improve the stability of multi-task training and overall prediction performance. This method guarantees both the correctness of the stepwise denoising process and a high degree of consistency between the final prediction and the true future state.

[0036] Furthermore, it should be noted that steps S1 to S4 of the above technical solution are the inference stages of the physics-inspired latent diffusion model architecture, that is, they only correspond to the reverse denoising stage and do not involve the forward diffusion (noise addition) process. Based on the above detailed explanation, the following will describe each step of the solution in detail: In step S1, historical tropical cyclone data, historical environmental field data, and future environmental field forecast data are acquired and preprocessed; among them, historical tropical cyclone data includes historical trajectory, wind speed, and air pressure data.

[0037] This step focuses on addressing data heterogeneity by transforming data from different sources and types into standardized input. Technically, this involves: First, three types of core data are integrated: historical tropical cyclone data are derived from the IBTrACS (International Climate Stewardship Best Track Archive) dataset from 1980 to 2022, which includes trajectories (latitude and longitude), maximum sustained wind speeds, and central pressures; environmental field data are extracted from historical environmental field data of ERA5 (European Centre for Medium-Range Weather Forecasts Reanalysis Data 5) and future environmental field prediction data from the FengWu model.

[0038] Furthermore, all data were cropped to a 10° radius region centered on the tropical cyclone center (final dimensions 69×80×80). For standardization, the trajectory coordinates were normalized relative to the initial time coordinates to eliminate the influence of the starting position; wind speed, air pressure, and other environmental variables were standardized using the mean and standard deviation of the training set to eliminate dimensional differences.

[0039] Compared with related technologies, the embodiments of the present invention break through the limitations of traditional single data standardization, design a special normalization scheme for trajectory coordinates, and integrate historical and future dual environmental data to improve the model's adaptability to tropical cyclones in different sea areas, laying a solid foundation for multimodal fusion.

[0040] In step S2, under the physical-inspired latent diffusion model architecture, historical tropical cyclone data, environmental field data obtained by splicing historical environmental field data and future environmental field forecast data, and the current denoised time step are vectorized respectively. The three types of vectorized features are aligned and fused and used as input to the Transformer encoder. A unified context vector is output through a multi-head self-attention mechanism.

[0041] This step focuses on addressing the problem of insufficient multimodal fusion. It utilizes a three-tiered architecture—GRU temporal coding, Swin Transformer spatial coding, and Transformer encoder—to construct the encoder portion of the denoising network, achieving deep fusion of multi-source data. Specifically: First, a GRU (Gated Recurrent Unit) temporal encoder processes tropical cyclone sequences from multiple historical moments, outputting feature vectors that capture the temporal evolution patterns. Then, a Swing Transformer encoder processes the spliced ​​data of historical and future environmental fields, extracting spatial feature tokens. Simultaneously, the denoising time step is mapped to a high-dimensional time step embedding to encode diffusion stage information. Finally, the three types of features are spliced ​​and input into a standard Transformer encoder, achieving deep interaction through a self-attention mechanism, outputting a unified context vector c, which serves as a conditional input to guide the update of latent variables in each subsequent denoising time step.

[0042] See Figure 2 The Transformer encoder consists of stacked coding blocks with identical structures. Each coding block sequentially includes diffuse embedding, layer normalization, multiple attention, and a feedforward network. Since a standard Transformer encoder is used, its specific processing steps will not be described here.

[0043] Unlike related technologies that only integrate historical observations and single environmental data and use simple splicing or weighted fusion methods, the embodiments of this invention incorporate future environmental field data for the first time. Through a dedicated time-space-cross-modal interactive architecture, it ensures the effective fusion of multimodal information, so that the denoising process not only depends on the evolution law of the historical cyclone itself, but also makes full use of the influence of the environmental background on the development path and intensity changes of the cyclone, thereby significantly improving the reliability of medium and long-term forecasts.

[0044] In step S3, starting from pure Gaussian noise and the maximum denoising time step, the current noisy latent variable, the current denoising time step, and the unified context vector are used as inputs to the physics-inspired decoder. The output of the previous noisy latent variable is updated in reverse to gradually generate the denoised latent variable. The physics-inspired decoder adopts a Transformer decoder with an integrated physics-inspired gating attention module. The physics-inspired gating attention module is used to explicitly model the physical constraint relationship between task-specific features such as trajectory, wind speed, and air pressure.

[0045] This step aims to address the disconnect between uncertainty modeling and physical laws. Through a latent diffusion process of "forward noise addition and backward denoising," it combines physical constraints to achieve a balance between probabilistic prediction and physical consistency. Specifically, this step corresponds to performing a backward denoising process in the latent space.

[0046] See Figure 3 The physics-inspired decoder consists of several stacked decoding blocks with identical structures, each of which integrates a physics-inspired gated attention module with the same structure. Each decoding block sequentially includes diffuse embedding, layer normalization, multi-head self-attention, layer normalization, multi-head cross-attention, a physics-inspired gated attention module, a feedforward network, and layer normalization. The physics-inspired gated attention module sequentially includes task projection mapping, cross-task attention, and a gating adjustment mechanism.

[0047] and Figure 3 Corresponding to the decoder structure shown, each decoding block includes steps such as explicit modeling of the diffusion time step, self-attention feature modeling, conditional cross-attention fusion, and feedforward network update, which are used to refine the denoising results layer by layer in the latent space.

[0048] Specifically, in the process of reverse updating the output of the previous noisy latent variable, each decoding block is used for: S10. Encode the current denoising time step into a time embedding vector and merge it with the output of the previous layer decoding block to obtain time-aware latent variables; wherein, the first layer decoding block merges the time embedding vector with the current noisy latent variables output from the previous denoising step.

[0049] S20. Self-attention feature modeling of time-perceived latent variables is performed using a multi-head self-attention mechanism.

[0050] S30. The self-attention features and the unified context vector are interacted through a multi-head cross-attention mechanism to obtain fused features.

[0051] S40. Through the physics-inspired gating attention module, the physical constraint relationship between the trajectory, wind speed and air pressure of the task-specific features is explicitly modeled to obtain physically consistent unified features.

[0052] S50. Update the unified features through the feedforward network and output them to the next layer of the decoding block; wherein, the last layer of the decoding block outputs the previous noisy latent variable.

[0053] Thus, by stacking multiple layers of decoding blocks, latent variables gradually approach the true future cyclone state at different semantic levels and different denoising time steps.

[0054] See Figure 3 Furthermore, a physically inspired gated attention mechanism (PIGA module) is introduced into the decoder to explicitly characterize the physical relationships between multiple predictive elements of tropical cyclones through a four-step process of decomposition-interaction-gating-fusion. This mechanism performs structured processing of the current latent space features in each decoding block, ensuring that the model continuously satisfies physical consistency constraints during generation. Specifically, S40 mentioned above includes: S401. The fused features are decoupled into task-specific features corresponding to trajectory, wind speed and air pressure respectively through task projection mapping.

[0055] S402. Through a cross-task attention mechanism, the specific features of each task interact with each other, and the interaction features of the specific features of each task are obtained respectively.

[0056] S403. Introduce a gating adjustment mechanism to adaptively balance the contribution ratio of the currently selected task-specific features and their interaction features in order to update the currently selected task-specific features.

[0057] S404. Reintegrate the updated task-specific features to obtain a physically consistent unified feature.

[0058] It should be noted that after executing S401, operations S402~S403 are executed symmetrically three times inside the Physics-Inspired Gated Attention (PIGA) module to update the trajectory, wind speed, and air pressure features respectively. Then, in S404, the updated trajectory, wind speed, and air pressure features are fused to obtain a physically consistent unified feature.

[0059] In step S4, the denoised latent variables are used as input to the convolutional decoder to predict the trajectory, wind speed, and pressure sequence of future tropical cyclones.

[0060] This step uses a convolutional decoder to restore the denoised latent variables to the data space and output the prediction results, thereby completing the generative prediction of the trajectory (latitude and longitude) and intensity elements (wind speed and air pressure) sequence of future tropical cyclones. This latent space diffusion generation mechanism improves modeling stability while ensuring expressive power and can naturally characterize the uncertainties in the prediction process.

[0061] This concludes the complete process of the global cyclone forecasting method that combines physical mechanisms with data-driven approaches, as described in this embodiment of the invention.

[0062] Example 2: This invention provides a global cyclone forecasting system that combines physical mechanisms with data-driven approaches, comprising: The acquisition and preprocessing module is used to acquire and preprocess historical tropical cyclone data, historical environmental field data, and future environmental field forecast data; among which, historical tropical cyclone data includes historical trajectory, wind speed, and air pressure data; The encoding module is used to vectorize historical tropical cyclone data, environmental field data obtained by splicing historical environmental field data and future environmental field forecast data, and the current diffusion time step under the physical-inspired potential diffusion model architecture. The three types of vectorized features are aligned and fused and used as input to the Transformer encoder. A unified context vector is output through a multi-head self-attention mechanism. The denoising module starts from pure Gaussian noise and the maximum denoising time step, and uses the current noisy latent variable, the current denoising time step, and the unified context vector as input to the physics-inspired decoder. It then updates the output of the previous noisy latent variable in reverse order to gradually generate denoised latent variables. The physics-inspired decoder uses a Transformer decoder with an integrated physics-inspired gating attention module, which is used to explicitly model the physical constraints between task-specific features such as trajectory, wind speed, and air pressure. The prediction module is used to take the denoised latent variables as input to the convolutional decoder to predict the trajectory, wind speed, and pressure sequence of future tropical cyclones.

[0063] Example 3: This invention provides a storage medium storing a computer program that causes a computer to execute the global cyclone forecasting method as described in Embodiment 1.

[0064] Example 4: This invention provides an electronic device, comprising: One or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing cyclone forecasting as described above.

[0065] It is understood that the global cyclone forecasting system, storage medium and electronic device that combine physical mechanisms and data-driven approaches provided in the embodiments of the present invention correspond to the global cyclone forecasting method that combines physical mechanisms and data-driven approaches provided in the embodiments of the present invention. The explanations, examples and beneficial effects of the relevant contents can be referred to the corresponding parts of the method, and will not be repeated here.

[0066] In summary, compared with existing technologies, it has the following beneficial effects: 1. The embodiments of the present invention introduce a physics-inspired gated attention module to explicitly model the physical dependencies between attributes such as tropical cyclone trajectory, air pressure and wind speed (such as the negative correlation between air pressure and wind speed, and the trajectory being driven by the environmental field), realize feature decoupling and structured representation in the latent space, effectively improve the physical consistency of forecasts, and significantly reduce the accumulation of errors in long-term forecasts.

[0067] 2. The embodiments of the present invention fully integrate multimodal data, including historical tropical cyclone data, ERA5 historical environmental reanalysis data, and FengWu future environmental forecast data. Through the Transformer encoder-decoder architecture, it realizes efficient interaction of multi-source information, enhances the ability of features to capture environmental evolution, and supports the accuracy of long-term forecasts.

[0068] 3. In this embodiment of the invention, physical constraints are embedded in the potential diffusion model to ensure that the prediction results conform to the principles of meteorological dynamics. At the same time, an adaptive multi-task loss balancing mechanism is adopted to stabilize training and optimize the overall forecast accuracy.

[0069] 4. The embodiments of the present invention combine the probability diffusion framework with physical constraints, and achieve reliable uncertainty modeling through uncertainty weighted loss, while ensuring that the probability prediction results conform to physical laws and improving the reliability of decision-making.

[0070] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0071] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A global cyclone forecasting method combining physical mechanisms and data-driven approaches, characterized in that, include: Acquire and preprocess historical tropical cyclone data, historical environmental field data, and future environmental field forecast data; among which, historical tropical cyclone data includes historical trajectory, wind speed, and air pressure data; Under the physical-inspired latent diffusion model architecture, historical tropical cyclone data, spliced ​​historical environmental field data and future environmental field forecast data, and the current denoised time step are vectorized respectively. The three types of vectorized features are aligned and fused and used as input to the Transformer encoder. A unified context vector is output through a multi-head self-attention mechanism. Starting from pure Gaussian noise and the maximum denoising time step, the current noisy latent variable, the current denoising time step, and the unified context vector are used as inputs to the physics-inspired decoder. The output of the previous noisy latent variable is updated in reverse to gradually generate the denoised latent variable. The physics-inspired decoder adopts a Transformer decoder with an integrated physics-inspired gating attention module. The physics-inspired gating attention module is used to explicitly model the physical constraints between task-specific features such as trajectory, wind speed, and air pressure. The denoised latent variables are used as input to the convolutional decoder to predict the trajectory, wind speed, and pressure sequence of future tropical cyclones.

2. The global cyclone forecasting method as described in claim 1, characterized in that, Historical tropical cyclone data were vectorized using a GRU time-series encoder; A sliding window Transformer encoder is used to vectorize historical environmental field data and future environmental field forecast data.

3. The global cyclone forecasting method as described in claim 1, characterized in that, The physical heuristic decoder is composed of several stacked decoding blocks with identical structures, each decoding block integrating a physically heuristic gated attention module with identical structure; in the process of reverse updating the output of a noisy latent variable, each layer of decoding blocks is used for: The current denoising time step is encoded into a temporal embedding vector and merged with the output of the previous decoding block to obtain time-aware latent variables; wherein, the first decoding block merges the temporal embedding vector with the current noisy latent variable output from the previous denoising step; Self-attention feature modeling of time-perception latent variables is performed using a multi-head self-attention mechanism; By using a multi-head cross-attention mechanism to interact with self-attention features and a unified context vector, fused features are obtained. By using a physics-inspired gating attention module, the physical constraints between task-specific features such as trajectory, wind speed, and air pressure are explicitly modeled to obtain physically consistent unified features. The unified features are updated through a feedforward network and output to the next layer of the decoding block; the last layer of the decoding block outputs the previous noisy latent variable.

4. The global cyclone forecasting method as described in claim 3, characterized in that, The method involves explicitly modeling the physical constraints between task-specific features such as trajectory, wind speed, and air pressure using a physics-inspired gating attention module to obtain physically consistent unified features, including: The fused features are decoupled into task-specific features corresponding to trajectory, wind speed and air pressure respectively by task projection mapping; By using a cross-task attention mechanism, specific features of each task can interact with each other, and the interaction features of specific features of each task can be obtained respectively. A gating adjustment mechanism is introduced to adaptively balance the contribution ratio of the currently selected task-specific features and their interaction features in order to update the currently selected task-specific features. Reintegrate the updated task-specific features to obtain a physically consistent unified feature.

5. The global cyclone forecasting method as described in claim 1, characterized in that, A joint optimization strategy combining the diffusion noise prediction objective and the result reconstruction objective is adopted to train the potential diffusion model architecture.

6. The global cyclone forecasting method as described in claim 5, characterized in that, A learnable uncertainty weighting mechanism is used to dynamically balance the losses corresponding to the diffusion noise prediction target and the result reconstruction target; wherein, the uncertainty weighting mechanism refers to dynamically adjusting the weights of the two types of losses by introducing trainable parameters for diffusion loss uncertainty and reconstruction loss uncertainty respectively.

7. A global cyclone forecasting system combining physical mechanisms and data-driven approaches, characterized in that, include: The acquisition and preprocessing module is used to acquire and preprocess historical tropical cyclone data, historical environmental field data, and future environmental field forecast data; among which, historical tropical cyclone data includes historical trajectory, wind speed, and air pressure data; The encoding module is used to vectorize historical tropical cyclone data, environmental field data obtained by splicing historical environmental field data and future environmental field forecast data, and the current denoising time step under the physical-inspired latent diffusion model architecture. The three types of vectorized features are aligned and fused and used as input to the Transformer encoder. A unified context vector is output through a multi-head self-attention mechanism. The denoising module starts from pure Gaussian noise and the maximum denoising time step, and uses the current noisy latent variable, the current denoising time step, and the unified context vector as input to the physics-inspired decoder. It then updates the output of the previous noisy latent variable in reverse order to gradually generate denoised latent variables. The physics-inspired decoder uses a Transformer decoder with an integrated physics-inspired gating attention module, which is used to explicitly model the physical constraints between task-specific features such as trajectory, wind speed, and air pressure. The prediction module is used to take the denoised latent variables as input to the convolutional decoder to predict the trajectory, wind speed, and pressure sequence of future tropical cyclones.

8. A storage medium, characterized in that, It stores a computer program, wherein the computer program causes the computer to perform the global cyclone forecasting method as described in any one of claims 1 to 6.

9. An electronic device, characterized in that, include: One or more processors; Memory; And one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing the cyclone forecasting method as described in any one of claims 1 to 6.

Images