A physical information neural network driven typhoon scene generation method and system for an open sea island platform

Abstract

The application discloses a kind of physical information neural network driven ocean island typhoon scene generation method and system, comprising: collecting multi-source heterogeneous meteorological data and carrying out space-time alignment preprocessing;Build coarse scale space-time probability prediction model, utilize graph topology learning network to capture the spatial correlation of meteorological elements, integrate the state space model of linear complexity to efficiently process the long-time sequence dependence of typhoon evolution, and generate coarse resolution probabilistic typhoon scene through multivariate joint distribution probability model;Further build physical downscaling model, input coarse scale prediction result as condition, embed atmospheric fluid mechanics equation as physical hard constraint in loss function, and physically consistent downscaling is carried out to coarse scale scene;Finally output high-resolution typhoon scene with probability reliability and physical authenticity.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of meteorological disaster prediction, artificial intelligence and power system resilience assessment, and specifically to a method and system for generating typhoon scenarios on ocean islands driven by physical information neural networks. Background Technology

[0002] Extreme natural disasters such as typhoons pose a severe challenge to the safe and stable operation of island microgrids that rely primarily on renewable energy. Accurately characterizing the spatiotemporal evolution of typhoon disasters is a prerequisite for conducting resilience assessments and proactive defense decisions for island energy systems.

[0003] Existing typhoon scene generation methods face a dilemma in practical applications: it is difficult to balance physical realism, probabilistic reliability, and computational efficiency. Specifically, this manifests as follows: (1) Low computational efficiency of numerical models. Although traditional numerical weather prediction conforms to physical laws, solving differential equations is extremely time-consuming. Generating large-scale scene sets often takes several days, which cannot meet the minute-level real-time emergency response needs of sudden disasters on islands. (2) Physical distortion of pure data-driven models. Existing deep learning lacks fluid dynamics constraints, and the generated wind fields often exhibit artifacts that violate conservation laws, making it impossible to reproduce key structures such as the typhoon eye. In addition, observational data of distant islands is scarce, and pure data models are prone to overfitting and have poor generalization ability. (3) Edge computing power and memory bottlenecks. When processing ultra-long meteorological sequences, the computational complexity of the mainstream Transformer architecture increases quadratically, leading to memory explosion, making it impossible to deploy on island edge computing terminals with limited computing power. (4) Extreme long-tail risks are underestimated. Existing methods focus more on mean prediction and ignore the nonlinear tail correlation between variables such as wind speed and air pressure. It is difficult to capture extreme disaster scenarios at the level of black swans, resulting in insufficient stress test intensity for power system resilience assessment.

[0004] Therefore, a new method for generating typhoon scenes on distant islands is urgently needed to solve the above problems. Summary of the Invention

[0005] Purpose of the invention: To address the problems existing in the prior art, this invention provides a method and system for generating typhoon scenes on ocean islands driven by physical information neural networks.

[0006] Technical solution:

[0007] This invention proposes a method for generating typhoon scenes on distant islands driven by a physical information neural network, comprising:

[0008] Historical meteorological data and real-time observation data of the target sea area are acquired, a spatiotemporal dataset is constructed, the spatiotemporal dataset is spatiotemporally aligned and normalized, and a dynamic graph structure describing the correlation of meteorological variables in the target sea area is constructed.

[0009] A coarse-scale spatiotemporal probability prediction model is constructed, including a graph neural network for extracting spatial features, a linear state space for capturing temporal evolution, and a two-branch network for predicting joint probabilities. The dynamic graph structure is used as the input of the coarse-scale spatiotemporal probability prediction model, and the coarse-resolution typhoon prediction scenario is generated using the coarse-scale spatiotemporal probability prediction model.

[0010] A refined physical downscaling model is constructed, including a deep residual structure formed by several hidden layers. The coarse-resolution typhoon prediction scenario is taken as input, and a prediction vector is output after linear transformation by the hidden layers. The prediction vector includes the predicted values ​​of meteorological variables in the target sea area.

[0011] A training set is extracted from the coarse-resolution typhoon prediction scenario, and the refined physical downscaling model is trained in stages using a strategy of minimizing the composite loss function.

[0012] The coarse-resolution typhoon prediction scenario is input into the trained refined physical downscaling model to obtain the wind field output.

[0013] Furthermore, the nodes of the dynamic graph are spatial grid points in the spatiotemporal dataset, and the features of the nodes are real-time observation data of each meteorological variable in the spatiotemporal dataset; the Pearson correlation coefficient is calculated based on the historical meteorological data of the spatiotemporal dataset to construct the edges of the dynamic graph; the adjacency matrix of the dynamic graph structure combines a geographic adjacency matrix based on a Gaussian kernel function and a semantic adjacency matrix based on a Pearson correlation coefficient.

[0014] Furthermore, in the coarse-scale spatiotemporal probability prediction model, the graph neural network, the linear state space, and the two-branch network are cascaded sequentially, including: arranging the graph-level embedding vector sequence output by the graph neural network in temporal order to form a temporal feature sequence containing spatial topological information, which is used as the input of the linear state space; using the hidden state features output by the linear state space as the input of the two-branch network, and using the Copula function to post-process the hidden state.

[0015] Furthermore, the linear state space employs a selective state propagation mechanism to handle long-term time-series dependencies, performs zero-order preserved discretization on continuous parameters, and outputs hidden state features based on time-varying equations.

[0016] Furthermore, the dual-branch network includes an edge distribution estimation branch and a dependency structure estimation branch:

[0017] The marginal distribution estimation branch includes a first set of multilayer perceptrons, which uses the latent state features output by the state space model as the input of the first set of multilayer perceptrons to predict the marginal distribution parameters of each meteorological variable.

[0018] The dependency structure estimation branch includes a second set of multilayer perceptrons, which takes the hidden state features as input to obtain the original vector, maps the original vector to the non-zero elements of the lower triangular matrix, constructs the covariance matrix, standardizes the covariance matrix, and outputs the correlation coefficient matrix as the parameter of the Copula function.

[0019] Furthermore, the generation of the coarse-resolution typhoon forecast scenario includes:

[0020] Set the initial typhoon state vector, including meteorological variables and corresponding values ​​for the target sea area;

[0021] Using the dynamic graph structure as input, the coarse-scale spatiotemporal probability prediction model is iteratively called to predict the typhoon state vector at the current moment and output the joint probability distribution parameters of the meteorological variables at the next moment.

[0022] The predicted state is obtained by sampling through Cholesky decomposition. This predicted state is then fed back into the coarse-scale spatiotemporal probability prediction model. The prediction process is then executed recursively until a complete typhoon evolution sequence is generated.

[0023] Furthermore, the refined physical downscaling model is a conditional physical information neural network, comprising an input layer, hidden layers, and an output layer:

[0024] The typhoon evolution sequence is used as a conditional feature vector, and the spatiotemporal coordinate vector representing the target sea area is input into the input layer. The conditional feature vector and the spatiotemporal coordinate vector are then concatenated.

[0025] The hidden layer structure includes several hidden layers, with adjacent hidden layers directly connected by skip connections; the hidden layer structure is used to perform a linear transformation on the concatenated input, and the Swish function is used as the activation function;

[0026] The output layer does not have an activation function; it directly outputs the prediction vector through a linear transformation.

[0027] Furthermore, the composite loss function includes physical residual loss, data fitting loss, boundary condition loss, and anchor point regularization loss; the physical residual loss is constructed based on the two-dimensional incompressible Navier-Stokes equations and includes the continuity equation residual and the momentum equation residual.

[0028] Furthermore, the phased training includes:

[0029] In the early stages of training, pre-training is performed based on physical residual loss, data fitting loss, and boundary condition loss. The data fitting loss is given a very high weight to establish the initial manifold of the flow field. As training progresses, anchor point regularization loss is added to gradually increase the weight of physical residual loss, forcing the model output to approximate the solution space that satisfies the Navier-Stokes equations, thereby generating a high-fidelity wind field.

[0030] On the other hand, this invention also proposes a physical information neural network-driven system for generating typhoon scenes on distant islands, comprising:

[0031] The data acquisition module is used to acquire historical meteorological data and real-time observation data of the target sea area, construct a spatiotemporal dataset, perform spatiotemporal alignment and normalization processing on the spatiotemporal dataset, and construct a dynamic graph structure describing the correlation of meteorological variables in the target sea area.

[0032] The coarse-scale spatiotemporal probability prediction model includes a graph neural network for extracting spatial features, a linear state space for capturing temporal evolution, and a two-branch network for predicting joint probabilities. The dynamic graph structure is used as the input of the coarse-scale spatiotemporal probability prediction model, and the coarse-resolution typhoon prediction scenario is generated using the coarse-scale spatiotemporal probability prediction model.

[0033] The refined physical downscaling model includes a deep residual structure formed by several hidden layers. The coarse-resolution typhoon prediction scenario is taken as input, and the prediction vector is output after linear transformation by the hidden layers. The prediction vector includes the predicted values ​​of meteorological variables in the target sea area.

[0034] The training module is used to separate the training set from the coarse-resolution typhoon prediction scenario and to train the refined physical downscaling model in stages using a strategy of minimizing the composite loss function.

[0035] The wind field generation module is used to input the coarse-resolution typhoon prediction scenario into the trained refined physical downscaling model to obtain the wind field output.

[0036] Beneficial effects: Compared with the prior art, the present invention makes the following improvements:

[0037] (1) This invention effectively solves the problem of low computational efficiency of traditional numerical models by constructing a two-stage generation framework of coarse-scale probabilistic prediction and fine-scale physical downscaling. This invention avoids the direct numerical solution of complex differential equations and uses a trained deep learning model for inference, which shortens the time to generate a large-scale typhoon scene set from several days to minutes, significantly improving computational efficiency and meeting the real-time emergency decision-making needs of sudden disasters in remote islands.

[0038] (2) The refined physical downscaling model constructed in this invention embeds the Navier-Stokes equations as hard constraints into the loss function, effectively solving the problems of physical distortion and poor generalization ability in pure data-driven models. This embedded physical design forces the model output to strictly follow the laws of mass and momentum conservation, eliminating non-physical artifacts and accurately reproducing key fluid structures such as the eye of a typhoon; at the same time, physical constraints, as a strong regularization method, enable the model to maintain good generalization ability even when island observation data is scarce, avoiding overfitting.

[0039] (3) This invention uses a linear state-space model to replace the traditional Transformer architecture in the coarse-scale spatiotemporal probability prediction model, effectively solving the problem of computing power and memory bottlenecks at the edge. This design reduces the computational complexity of processing long-sequence meteorological data from quadratic O(L...) 2 The computation time was reduced to linear O(L), which significantly reduced memory usage and computational overhead, enabling the model to be deployed lightweight on edge computing terminals on islands with limited computing power, thus achieving efficient edge inference.

[0040] (4) This invention effectively solves the problem of underestimation of extreme long-tail risks by introducing a Copula function to handle the nonlinear dependencies between multiple variables and combining it with physical constraint fine-tuning. This enables the model to not only focus on mean prediction, but also to accurately characterize the tail correlation of variables such as wind speed and air pressure under extreme conditions. The generated scenario set can effectively cover extreme typhoon events at the "black swan" level, providing a higher intensity stress test sample for power system resilience assessment. Attached Figure Description

[0041] Figure 1 This is a flowchart of the method of the present invention;

[0042] Figure 2 This is a comparative analysis diagram of the algorithm of this invention with other deep learning algorithms;

[0043] Figure 3 This is an example diagram of a typical typhoon integrated scenario analysis generated by an embodiment of the present invention;

[0044] Figure 4 This is a schematic diagram of the data flow direction of wind field generation in an embodiment of the present invention;

[0045] Figure 5 It is a graph of the convergence curve of the loss function during the training process of the refined physical downscaling model;

[0046] Figure 6 This is an example diagram showing the wind field generated according to an embodiment of the present invention. Detailed Implementation

[0047] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the following embodiments are for illustrative purposes only and do not limit the scope of the invention. This embodiment was performed on a high-performance computing workstation with the following hardware configuration: Intel Xeon Gold 6248R CPU, RTX4090D GPU, and 512GB of memory. The software environment was based on the Linux Ubuntu 20.04 operating system, using PyTorch 2.7 as the deep learning framework, DGL 1.1 as the graph neural network library, and DeepXDE as the physics computation library.

[0048] A method for generating typhoon scenes driven by a physical information neural network, such as... Figure 1 The flowchart shown is for this method, which specifically includes the following steps:

[0049] S1: Multi-source data fusion and dynamic graph topology construction.

[0050] Historical meteorological data and real-time observation data of the target sea area were acquired to construct a spatiotemporal dataset. To address the issue of inconsistent spatiotemporal resolution among multi-source heterogeneous meteorological data, a spatiotemporal alignment and normalization mechanism was established.

[0051] This embodiment collects multi-source heterogeneous meteorological data for the Wanshan Islands from the Joint Typhoon Warning Center (JTWC), including typhoon center latitude and longitude, maximum wind speed, central pressure, movement speed, and direction of movement every 6 hours. Reanalysis meteorological field data for the Wanshan Islands is also collected from the ERA5 dataset, including the surface 10m wind field and sea level pressure field at the corresponding time.

[0052] This embodiment utilizes cubic spline interpolation and bilinear interpolation techniques to unify the multi-source heterogeneous meteorological data to a fixed time step and relative coordinate system. Specifically, the following operations are performed:

[0053] (1) Cubic spline interpolation was used to uniformly resample the path data and meteorological field data. A fixed time step.

[0054] (2) Perform spatial coordinate transformation to construct a relative moving coordinate system with the typhoon center as the origin, eliminate the influence of absolute latitude and longitude, and make the model focus on the internal structure of the typhoon. At the same time, the direction and speed of movement in the meteorological data are converted into changes in longitude and latitude in a rectangular coordinate system to characterize the movement characteristics of the typhoon.

[0055] (3) Data normalization: Z-Score standardization is performed on the constructed multivariate long time series dataset.

[0056]

[0057] in, This is the historical statistical average. Standard deviation, To prevent tiny amounts of division by zero, For multivariate long time series datasets, The output is for standardization processing.

[0058] Based on this, a dynamic graph structure describing the correlation of meteorological elements in the typhoon-affected area is constructed. The typhoon-affected area is the aforementioned target sea area. The nodes in the dynamic graph correspond to spatial grid points in the spatiotemporal dataset. Meteorological observation values ​​in the spatiotemporal dataset (such as changes in wind speed and air pressure over time) are input into the graph structure as dynamic features of the nodes.

[0059] Unlike traditional mapping methods that rely solely on geographic distance, this invention employs a geographic-semantic fusion strategy to construct an adjacency matrix. Specifically, it combines a geographic adjacency matrix based on a Gaussian kernel function with a semantic adjacency matrix based on the Pearson correlation coefficient, including:

[0060] (1) Construct a geographic adjacency matrix based on Gaussian kernel function :

[0061]

[0062] in, Let be the spherical distance between node i and node j. To control the standard deviation of the neighborhood width, This is the distance truncation threshold.

[0063] (2) Construct a semantic adjacency matrix based on Pearson correlation coefficient :

[0064] Based on historical meteorological time series data of each node, the Pearson correlation coefficient between nodes is calculated to capture long-distance teleconnection characteristics:

[0065]

[0066] in, For nodes With nodes The Pearson correlation coefficient between them These are historical meteorological observations for a node, with subscripts indicating the specific node number and time step. ), This represents the total length of the historical meteorological time series. A correlation threshold is set. Construct the matrix:

[0067]

[0068] (3) Geographic-semantic fusion:

[0069] The two matrices are then weighted and fused to obtain the final dynamic graph adjacency matrix. This mechanism can not only capture the physical transmission between geographically adjacent nodes, but also capture the teleconnection features of geographically distant nodes that have highly synchronized meteorological changes, laying a topological foundation for subsequent spatiotemporal feature extraction.

[0070] S2: Construct a coarse-scale spatiotemporal probability prediction (ST-GMC) model.

[0071] The coarse-scale spatiotemporal probabilistic prediction model employs an end-to-end cascaded architecture, aiming to simultaneously capture the spatial topological dependencies, long-term evolution patterns, and probabilistic dependencies among multiple variables during typhoon evolution. The overall data processing flow of the model consists of the following three cascaded modules:

[0072] Spatial feature extraction module: The model first uses a graph neural network to process the dynamic graph structure sequence input by S1 in parallel, and aggregates the discrete node features of each time step into a continuous graph-level embedding vector to encode the spatial topological information of the meteorological field.

[0073] Temporal Evolution Capture Module: This module arranges the graph-level embedding vectors in temporal order and uses them as input to the linear state-space model. It utilizes a linear complexity algorithm to capture long-range temporal dependencies and outputs the final hidden state features containing complete spatiotemporal context information. ;

[0074] Joint Probability Prediction Module: The final hidden state features are input into the joint distribution probability module, which uses a two-branch network to predict the marginal distribution parameters and inter-variable correlation parameters of meteorological variables, thereby constructing a typhoon prediction scenario containing uncertainty information.

[0075] Through the collaborative work of the three modules mentioned above, the ST-GMC model achieves an end-to-end mapping from high-dimensional discrete graph structure data to low-dimensional continuous probability distribution parameters.

[0076] Specifically, the workflow of each module in the coarse-scale spatiotemporal probability prediction model is as follows:

[0077] (2.1) Use graph neural networks to process the dynamic graph structure obtained by S1 to capture spatial topological dependencies.

[0078] At each time step, the feature representation of the current node is updated by aggregating information from neighboring nodes through graph convolution operators. By stacking multiple layers of graph convolutional networks, the model can progressively capture the spatial topological dependencies within the meteorological field from local to global perspectives, and output a graph-level embedding vector sequence that represents the current spatial state of the system.

[0079] The node features are updated using graph convolution operations to extract spatial topological features. The calculation formula is as follows:

[0080]

[0081] in, It is the first Layer in time The node feature matrix; It is an adjacency matrix with self-loops added. It is the identity matrix; yes The angle matrix; It is the first The learnable weight matrix of the layer; It is a non-linear activation function.

[0082] (2.2) To address the extremely long time-series dependence of typhoon evolution, a linear state-space model is introduced for time-series extraction.

[0083] Specifically, a cascaded connection is established between the graph neural network and the linear state-space model: the graph-level embedding vector sequence output in step (2.1) is arranged in an ordered manner according to the time dimension to form a temporal feature sequence containing spatial topological information. This sequence is used as the input signal of the linear state-space model, enabling the model to capture long-term temporal dependencies based on spatially encoded high-level features, rather than the original meteorological data.

[0084] In a preferred embodiment of the present invention, the time-series evolution capture uses the Mamba state-space model as a specific implementation scheme. The linear state-space model module is based on the discretized form of the state equation of a linear time-invariant system. In order to adapt to discretely sampled meteorological data, the zero-order hold technique is used to discretize the parameters of the continuous system.

[0085] It should be noted that the Mamba model is only a typical example of a linear state-space model. The core purpose of this type of model is to leverage its linear time complexity to address the computational bottleneck in processing long-sequence meteorological data. Therefore, those skilled in the art should understand that, apart from the Mamba model, other state-space model architectures with linear complexity or linear attention mechanism networks can be used as equivalent alternatives in the temporal evolution capture unit of this invention, provided they do not depart from the design concept of this invention.

[0086] This invention employs a selective state-space model to handle long time series. The spatial feature sequence extracted by a graph neural network (GNN) is input into the state-space module. To enable the model to adaptively adjust the state update strategy based on the input typhoon spatial features, this embodiment introduces a selection mechanism, converting the parameters in the discretization process into input functions, thereby transforming the linear time-invariant system into a linear time-varying system. The specific calculation process is as follows:

[0087] Let the vector output from step (2.1) be represented as... The vector is taken as the time of step (2.2). The input spatial feature vector is processed through a linear projection layer. , and Calculate the parameters dynamically at the current moment:

[0088]

[0089]

[0090]

[0091] in, These are the input control parameters for the current time step (used to control the degree of influence of the current input features on the hidden state of the model). The output projection parameters at the current moment (used to control the mapping relationship between the model's hidden states and the final output). The current time step is Softplus, and the activation function is Softplus. (Based on dynamically generated...) For continuous parameters Zero-order preserved discretization is performed to obtain time-varying discrete parameters. and :

[0092]

[0093]

[0094] Finally, the state-space model updates the state and outputs based on the following time-varying equations:

[0095]

[0096]

[0097] In the formula, It is in a hidden state. It is the output, the core matrix. Based on the input of the linear state space Dynamic changes. Through the above mechanism, the model can adaptively focus on key information or forget irrelevant information based on the context, thereby achieving a linear complexity of O(L) for sequence processing while maintaining a strong ability to model long-range dependencies.

[0098] Specifically, to address the issue that the model cannot utilize traditional convolution for parallel acceleration after the introduction of the selection mechanism, this embodiment further incorporates a parallel scanning algorithm. This algorithm leverages the associative property of the state update process to transform the recursive process, which originally required serial computation, into a parallelizable prefix sum computation. This improvement ensures that the computational complexity of the model remains strictly O(L) when processing long sequences, and fully utilizes the parallel computing capabilities of GPUs. Compared to the traditional Transformer architecture's quadratic complexity of O(L) as the sequence length increases, this represents a significant improvement. 2 This solution significantly improves inference efficiency and reduces memory usage, thus enabling deployment on edge devices with limited computing power.

[0099] like Figure 2 The figure shows a performance comparison between the proposed GNN-Mamba model and mainstream benchmark models (LSTM, Transformer, CNN-LSTM). The upper part of the figure shows the convergence curve of the test loss (MSE) as a function of training epochs, where the red curve represents the model of this invention. It can be seen that it has the fastest convergence speed and the lowest final error, proving the effectiveness of the cascaded architecture in feature capture. The lower part of the figure shows a comparison of the training time of each model under the same hardware environment. The model of this invention (red bar) maintains high accuracy while its training time is on par with the lightweight CNN-LSTM, significantly outperforming the computationally intensive Transformer architecture. This figure is included to visually demonstrate that the linear state-space model used in this invention successfully achieves dual optimization of accuracy and computational efficiency when processing long-sequence meteorological data, verifying the effectiveness of the linear complexity O(L) design in step (2.2).

[0100] (2.3) In order to quantify the uncertainty of the prediction results, the Copula function is used to post-process the hidden state output by the state-space model to construct the joint probability distribution of key meteorological variables of typhoon.

[0101] The final hidden state features output by the state-space model in step (2.2) As the input to this module, the strong negative correlation between the typhoon's maximum wind speed and central pressure (i.e., the meteorological data collected in S1), coupled with the nonlinear tail correlation under extreme weather conditions, means that simple independent prediction cannot capture this dependency. Therefore, this module employs a two-branch structure, specifically comprising the following two branches: a marginal distribution estimation branch, containing a first set of multilayer perceptrons (MLPs) for independently predicting the marginal distribution parameters of each meteorological variable; and a dependency structure estimation branch, containing a second set of multilayer perceptrons for predicting the Copula function parameters describing the correlation between variables.

[0102] Specifically, step (2.3) includes the following sub-steps:

[0103] (2.3.1) Utilizing the hidden state features output by the state-space model Predicting the first MLP using the first set Marginal distribution parameters of meteorological variables :

[0104]

[0105] In the formula, It is the first Learnable weights for each marginal distribution prediction head It is used to predict the first A multilayer perceptron network for the marginal distribution parameters of several meteorological variables. Parameters Used to determine the cumulative distribution function of the variable.

[0106] (2.3.2) The relevant parameters of the Copula function are predicted by the second set of multilayer perceptrons to capture the coupling relationship between wind speed and air pressure. In order to ensure the mathematical validity of the correlation coefficient matrix, this embodiment does not directly predict the correlation matrix, but constructs it by using the inverse process of Cholesky decomposition.

[0107] The specific calculation process is as follows: First, ... Input the second MLP, output the original vector z; then map z to the non-zero elements of the lower triangular matrix L:

[0108]

[0109] in The original vector output by the second group of MLPs In, mapped to the matrix of the first... Line number The element components at column positions. Using this Cholesky factor L, construct the covariance matrix. This operation is mathematically guaranteed. It must be symmetric positive semi-definite. Finally, regarding the covariance matrix... After standardization, the final correlation coefficient matrix R, which serves as the Gaussian Copula parameter, is obtained. The expression for the elements in R is:

[0110]

[0111] From hidden state The transformation process to the correlation coefficient matrix R is completely differentiable, ensuring that the entire model can be trained end-to-end through backpropagation and always produces a structurally effective probabilistic dependency model.

[0112] S3: Use the spatiotemporal dataset as input to the coarse-scale spatiotemporal probability prediction model, and use the coarse-scale spatiotemporal probability prediction model to generate a coarse-resolution typhoon prediction scenario containing uncertainty information.

[0113] This step aims to learn the complex spatiotemporal patterns of typhoon evolution from historical data and generate coarse-resolution macro-meteorological scenarios containing uncertainty information. Using the preprocessed spatiotemporal dataset constructed in S1 as input, and employing the coarse-scale spatiotemporal probabilistic prediction model (ST-GMC) in S2, a rolling time-domain prediction strategy is used to generate coarse-resolution typhoon forecast scenarios containing uncertainty information.

[0114] The specific data processing flow and parameter settings in this embodiment are as follows:

[0115] (3.1) First, in order to cover the diverse evolution paths of typhoons, this embodiment uses the Monte Carlo Sampling method to initialize the typhoon status.

[0116] Define the initial state vector of the typhoon as ,in P represents latitude and longitude. c V represents the central air pressure. max This indicates the maximum wind speed. The term represents the corresponding rate of change. A sampling space Ω is constructed based on the historical meteorological statistical characteristics of the target sea area, and an initial state vector set is provided. The formula for generating it is as follows:

[0117]

[0118] In the formula, Represents a uniform distribution sampling function. and Here, represents the lower and upper bound vectors of the state variables, derived from historical data, and N represents the total number of generated scenarios. This formula is used to randomly initialize a large-scale typhoon sample within the multidimensional state space, serving as the starting point for scenario evolution.

[0119] (3.2) Perform rolling time-domain prediction using the ST-GMC model constructed in S2. Specifically, the typhoon state vector at the current moment and the dynamic graph topology constructed in S1 are used as inputs. The ST-GMC model is iteratively called to output the joint probability distribution parameters of the meteorological variables at the next moment, and the predicted state is obtained by sampling through Cholesky decomposition. The predicted state is fed back to the model input as new historical data. The above process is recursively executed until a complete typhoon evolution sequence is generated.

[0120] like Figure 3 The figure shows a visualization of the prediction results for a typical typhoon sample (sample 69) generated using the aforementioned rolling time-domain strategy. The upper part of the figure shows the typhoon trajectory prediction, with the horizontal axis representing longitude (between 120°E and 132°E) and the vertical axis representing latitude (between 10.0°N and 13.0°N). The blue line represents the historical input trajectory, the green line represents the actual trajectory, and the red line represents the mean trajectory predicted by the model. The lower part shows the typhoon intensity (maximum wind speed) prediction, with the horizontal axis representing the time step in hours and the vertical axis representing the maximum wind speed near the typhoon's center in meters per second. The pink shaded area represents the 95% confidence interval of the model output. This demonstrates that the model of this invention can not only accurately predict the future movement trend of typhoons but also quantify the uncertainty range of the prediction through confidence intervals, and the actual values ​​all fall within these confidence intervals, proving that the probabilistic model constructed in this invention has good reliability and coverage.

[0121] (3.3) To further enhance the physical interpretability of the generated scenarios and achieve cross-scale mapping from macroscopic statistical features to microscopic physical fields, this embodiment constructs a data-physical coupled scenario reconstruction mechanism. Specifically, it includes:

[0122] (3.3.1) Smoothing of intensity evolution: Introducing an exponential moving average smoother to smooth the maximum wind speed change rate dv max Make corrections. Set the smoothing factor α = 0.4, and the calculation formula is as follows:

[0123]

[0124] Update the wind speed for the next time step using the smoothed rate of change: This mechanism forces wind speed changes to follow physical inertia, avoiding violent, non-physical oscillations.

[0125] (3.3.2) Holland wind field reconstruction: Based on the generated typhoon center location, central pressure, and maximum wind speed, the two-dimensional wind field is reconstructed using a parametric Holland vortex model. Wherein, the radius of the maximum wind speed R... max The pressure profile parameter B is dynamically calculated based on the generated latitude and intensity:

[0126]

[0127]

[0128] in The latitude of the center of the generated typhoon. air density, It is a natural constant. The ambient background air pressure is used. This step restores the low-dimensional trajectory points to two-dimensional wind field slices with physical structure.

[0129] (3.4) Repeat steps (3.2)-(3.3) above until the typhoon dissipates (the dissipation threshold is defined as V). max (<17.5m / s) or move out of the target area. Finally, a coarse-resolution typhoon scenario is generated.

[0130] S4: Construct a refined physical downscaling model based on a conditional physical neural network (TP-PINN).

[0131] This step aims to transform the coarse-scale, parameterized probabilistic scenario set generated by S3 into a high-resolution, continuous physical wind field that satisfies the laws of hydrodynamics. This embodiment constructs a Conditional Physical Information Neural Network (TP-PINN), which does not rely on traditional discrete grids but instead learns continuous function mappings to achieve wind field reconstruction at arbitrary resolutions.

[0132] Specifically, the refined physical downscaling model uses a deep residual network based on a fully connected architecture as its backbone structure, and its specific structure, internal logic, and data processing flow are as follows:

[0133] (4.1) Network input layer design and feature fusion.

[0134] The model's input vector aims to establish a correlation between the macroscopic typhoon state and microscopic spatiotemporal points. The input layer receives two parts of data: the spatiotemporal coordinate vector x. st =(x,y,t) and conditional eigenvector c.

[0135] Where x st This represents any continuous spatial location and time point within the typhoon area, belonging to the typhoon's impact area (target sea area). Here, x and y are normalized latitude and longitude coordinates, and t is the normalized time step. The conditional feature vector c is directly extracted from the coarse-scale scenario data generated by S3. For each specific spatiotemporal query point (x, y, t), its corresponding macroscopic state of the typhoon center is extracted, including the central pressure P. c Maximum wind speed V max Movement speed V moveAnd the relative distance and azimuth from the typhoon center to that point.

[0136] Finally, feature concatenation is performed to combine the spatiotemporal coordinate vector x. st The conditional eigenvector c is concatenated with the conditional eigenvector c to form a high-dimensional input tensor. This design enables the network to dynamically adjust its predictions of physical quantities at a spatiotemporal point (x,y,t) based on different typhoon intensities and locations (condition c), thus achieving "conditional" generation.

[0137] (4.2) Network hidden layer structure and core parameters.

[0138] To ensure the network has sufficient expressive power to fit complex hydrodynamic features, while avoiding the gradient vanishing problem in deep network training, this embodiment constructs a deep residual structure containing L hidden layers (L=8 in this embodiment). Specifically, each hidden layer contains N neurons (N=100 in this embodiment) and performs a linear transformation:

[0139]

[0140] in This is the weight matrix. For bias vectors, This represents the output vector of the hidden layer, with the superscript indicating the layer number. Regarding the choice of activation function, considering that residuals need to be calculated using physical equations in subsequent S5, which requires the network output to be second-order differentiable with respect to the input coordinates, this embodiment abandons the commonly used ReLU function and instead uses the Swish function. ,in These are learnable scaling parameters. Both functions possess infinitely differentiable smoothness, effectively supporting automatic differentiation calculations and ensuring accurate propagation of physical constraints.

[0141] In addition, to improve the training stability and convergence speed of the model, skip connections are introduced between every two hidden layers. This represents a non-linear activation function, and the specific formula is as follows:

[0142]

[0143] This design enables the network to learn identity mappings more easily, effectively mitigating the degradation problem of deep networks.

[0144] (4.3) Mapping of network output layer and physical quantity.

[0145] The network's output layer does not contain an activation function; it directly outputs the predicted physical quantity vector through a linear transformation. :

[0146]

[0147] Where u(x,y,t) represents the zonal wind speed component (unit: m / s) at that spatiotemporal point, v(x,y,t) represents the meridional wind speed component (unit: m / s) at that spatiotemporal point, and p(x,y,t) represents the sea level pressure (unit: hPa) at that spatiotemporal point. Through this direct mapping method, the model can output continuously changing physical field values, providing a foundation for subsequent physical constraint calculations.

[0148] (4.4) Internal logic and data flow.

[0149] like Figure 4 As shown, the model's data processing flow follows the principle of microscopic function approximation under macroscopic constraints. During forward propagation, the network receives any typhoon scenario generated by S3 as condition c and traverses the coordinates of interest (x, y, t) within the target sea area. Since the input coordinates are continuous variables, the network essentially constructs a grid-independent continuous function approximator. This means that once the model is trained, the coarse scenario generated by S3 can be resampled and the physical field reconstructed at arbitrarily high resolution according to user needs, thus breaking through the accuracy limitations of traditional discrete grids and achieving cross-scale downscaling from "coarse-scale probabilistic scenarios" to "refined high-resolution physical wind fields".

[0150] S5: Training and solving physical constraints of the TP-PINN model.

[0151] This step uses the coarse-scale, parameterized probability scenario set generated by S3 as the training dataset. Addressing the issue of physical information neural networks easily getting trapped in local optima when solving complex flow fields, this embodiment does not employ reinforcement learning. Instead, it borrows the common course learning concept from deep learning and designs a two-stage dynamic training strategy of "data preheating - physical embedding." This strategy is a mature optimization method in the field of physics-driven deep learning, and the specific process includes two stages:

[0152] Phase 1 (Data-driven pre-training): In the early stages of training, the data fitting loss is given extremely high weight (α≈0), and a large learning rate is used to quickly guide the model to fit the statistical distribution of coarse-scale scenario data and establish the initial manifold of the flow field.

[0153] The second stage (physical constraint fine-tuning): As training progresses, the weight of the physical residual loss is gradually increased, forcing the model output to approximate the solution space that satisfies the Navier-Stokes equations, and finally generating a physically consistent high-fidelity wind field.

[0154] To ensure that the generated wind field does not violate physical laws, a composite loss function containing four sub-losses is designed. Construct a composite loss function :

[0155]

[0156] In the formula, For physical residual loss, For data fitting loss, For boundary condition loss, For anchor point regularization loss, These are the corresponding weighting coefficients. Specifically, the residuals for each term are expressed as:

[0157] (1) Physical residual loss Based on the two-dimensional incompressible Navier-Stokes equations, the calculation formula is as follows:

[0158]

[0159] Among them, residual Includes the residuals of the continuity equation and the residuals of the momentum equation. For the first The spatial coordinates and time of a spatiotemporal point. The conditional feature vectors provided for the coarse-resolution typhoon prediction scenario are concatenated with the spatiotemporal coordinate vectors at the input layer as network input. The partial derivatives of the network output with respect to the input are calculated using an automatic differentiation mechanism to constrain the generated wind field to satisfy the laws of mass and momentum conservation.

[0160]

[0161] In the formula, These are the horizontal and vertical components of wind speed, respectively. For air pressure, air density, Let be the kinematic viscosity coefficient. The physical residual loss is the mean of the squared residuals mentioned above.

[0162] Physical residual loss This is the core constraint. Automatic differentiation techniques are used to calculate the partial derivatives of the network output with respect to the input, constructing the residuals of the continuity equation and the momentum equation. Minimizing this loss forces the network output to conform to the fluid dynamics mechanism.

[0163] (2) Data fitting loss This step anchors the model predictions to known observations, constraining the model's predictions at the observation points to approximate the actual reanalysis data and preventing solution space drift caused by purely physical constraints. It calculates the mean squared error (MSE) between the model's predictions and the actual values ​​at sparse, low-resolution reference data points:

[0164]

[0165] in This represents the two-dimensional wind speed component predicted by the model at the reference data point, where It is oriented east-west. It is oriented north-south. This represents the true two-dimensional wind speed component at the corresponding reference data point, given by the observation data.

[0166] (3) Boundary condition loss This item is used to apply the domain boundary. To assess the physical behavior of the model, Dirichlet boundary conditions are applied to the computational domain boundaries to ensure the rationality of the physical behavior. For the typhoon model, a far-field condition is applied, meaning that wind speed should decay to zero at the boundaries of the computational domain. This is a Dirichlet boundary condition, and its loss function is defined as the mean square error between the model's predicted and target values ​​at the boundary points.

[0167]

[0168] in, It is at the border The number of upsampled points. By minimizing this, the network is forced to learn to generate a physically plausible and relatively calm wind field in an area far from the typhoon's center.

[0169] Pre-training is performed based on the aforementioned three residuals, using only data-driven loss. Training is performed to obtain initial weights. This stage enables the model to quickly fit the data distribution. The physical residual loss from before is introduced for fine-tuning, and anchor point regularization loss is added to prevent catastrophic forgetting. :

[0170]

[0171] The loss penalty is based on the current weight. With initial weights The orthogonal distance prevents the model from forgetting data patterns, thus maintaining the ability to fit the observed data while ensuring physical authenticity.

[0172] like Figure 5 The graph shown is the convergence curve of the loss function during the training process in this embodiment. The horizontal axis represents the number of training rounds, and the vertical axis represents the loss value, reflecting the error between the predicted value and the true value during training. Through the above steps, this invention can generate a high-fidelity typhoon scenario set that combines probabilistic reliability and physical realism, which can be used to support the resilience assessment of microgrids on remote islands.

[0173] S6: The coarse-resolution typhoon prediction scenario obtained in S2 is used as the input of the physical downscaling model. The physical downscaling model trained in S5 maps the coarse-resolution typhoon prediction scenario into a high-resolution fine wind field and outputs it.

[0174] The specific output effect is as follows: Figure 6 The image shows a visualization of the physical downscaling effect of the TP-PINN model on the test samples in this embodiment. The horizontal and vertical axes of each subplot are planar grid indices, used to represent the horizontal position in physical space. The upper left and upper right subplots show the input coarse-resolution wind field (41×41 grid) and the model output high-resolution wind field (100×100 grid), respectively. The lower left subplot shows the bicubic interpolation result (100×100 grid) as a baseline comparison, which intuitively demonstrates the model's significant super-resolution reconstruction capability. Different colors in these three subplots represent different wind speeds, in m / s. The lower right subplot shows a heatmap of the difference between the TP-PINN prediction and the traditional bilinear interpolation result (100×100 grid). Different colors in the figure represent the wind speed difference between the PINN result and the interpolation result. The high-frequency texture in this part represents the nonlinear fluid details that the model learns through physical equation constraints and that traditional interpolation cannot recover.

[0175] To verify the physical consistency and generalization ability of the method of this invention, a comparative experiment was conducted between the method of this invention and a pure data-driven method under the same training set and hyperparameter configuration. The results are shown in Table 1. Table 1 aims to quantitatively demonstrate the performance trade-off advantage of the present invention after introducing physical constraints: that is, by sacrificing a small amount of statistical fitting accuracy, a significant reduction in hydrodynamic residual (PDE Loss) is achieved (physical realism is improved by about 50%).

[0176] Table 1 Comparison between the method of this invention and the data-driven method

[0177] Evaluation metrics (data fitting error) Pure data-driven PINN rate of change RMSE wind speed (m / s) 8.1264 9.7257 -10.68% MAE wind speed (m / s) 6.2423 7.966 -10.03% PDE total residuals 2115.93 1042.44 +50.73% Continuity equation residuals 12.37 7.43 +39.90% Momentum equation residual 49.097 35.3 +28.10%

[0178] This result proves that the wind field generated by this invention strictly follows the physical laws of the Navier-Stokes equations while satisfying the constraints of the observation data, thus eliminating the non-physical artifacts common in pure data-driven methods.

[0179] Furthermore, the offshore island scenario described in this invention is merely a typical application demonstration of this method and does not constitute a limitation on the application field of this invention. Based on the coarse-scale probabilistic inference-refined physical downscaling coupling architecture proposed in this invention, those skilled in the art will understand that the method of this invention is also applicable to other marine engineering scenarios with data sparsity characteristics or high physical security requirements, specifically including but not limited to:

[0180] (1) Offshore wind farm operation and maintenance: used to generate extreme wind condition maps of micro-site selection areas of wind farms, to assist in typhoon-resistant design of wind turbines and prediction of power generation;

[0181] (2) Safety assessment of offshore drilling platforms: used to simulate sudden severe sea conditions in the waters surrounding the drilling platform, providing a basis for platform structural strength verification and personnel evacuation decisions;

[0182] (3) Meteorological risk warning for coastal nuclear power plants: used to simulate the evolution characteristics of local microclimates around nuclear power plants under extreme weather conditions.

[0183] Meanwhile, without departing from the core design concept of this invention, simple replacements or modifications to the internal components of the model, such as replacing the graph neural network with a graph attention network, or replacing the Navier-Stokes equations in TP-PINN with shallow water equations to adapt to different fluid scenarios, should all be considered equivalent embodiments of this invention and included within the scope of protection of this invention.

Claims

1. A method for generating typhoon scenes on distant islands driven by a physical information neural network, characterized in that, include: Historical meteorological data and real-time observation data of the target sea area are acquired, a spatiotemporal dataset is constructed, the spatiotemporal dataset is spatiotemporally aligned and normalized, and a dynamic graph describing the correlation of meteorological variables in the target sea area is constructed. A coarse-scale spatiotemporal probability prediction model is constructed, comprising a graph neural network for extracting spatial features, a linear state space for capturing temporal evolution, and a two-branch network for predicting joint probabilities. The dynamic graph is used as input to the coarse-scale spatiotemporal probability prediction model, which generates a coarse-resolution typhoon prediction scenario. In the coarse-scale spatiotemporal probability prediction model, the graph neural network, the linear state space, and the two-branch network are cascaded sequentially. This includes: arranging the graph-level embedding vector sequence output by the graph neural network temporally to form a temporal feature sequence containing spatial topological information, which is used as input to the linear state space; and using the latent state features output by the linear state space as input to the two-branch network, and post-processing the latent state features using a Copula function. A refined physical downscaling model is constructed, including a deep residual structure formed by several hidden layers. The coarse-resolution typhoon prediction scenario is taken as input, and a prediction vector is output after linear transformation by the hidden layers. The prediction vector includes the predicted values ​​of meteorological variables in the target sea area. A training set is extracted from the coarse-resolution typhoon prediction scenario, and the refined physical downscaling model is trained in stages using a strategy of minimizing the composite loss function. The composite loss function includes physical residual loss, data fitting loss, boundary condition loss, and anchor point regularization loss. The physical residual loss is constructed based on the two-dimensional incompressible Navier-Stokes equations and includes the continuity equation residual and the momentum equation residual. The coarse-resolution typhoon prediction scenario is input into the trained refined physical downscaling model to obtain the wind field output.

2. The method for generating typhoon scenes on ocean-going islands driven by a physical information neural network according to claim 1, characterized in that, The nodes of the dynamic graph are spatial grid points in the spatiotemporal dataset, and the features of the nodes are real-time observation data of various meteorological variables in the spatiotemporal dataset. The Pearson correlation coefficient is calculated based on the historical meteorological data of the spatiotemporal dataset to construct the edges of the dynamic graph. The adjacency matrix of the dynamic graph combines a geographic adjacency matrix based on Gaussian kernel function and a semantic adjacency matrix based on Pearson correlation coefficient.

3. The method for generating typhoon scenes on ocean-going islands driven by a physical information neural network according to claim 2, characterized in that, The linear state space employs a selective state propagation mechanism to handle long-term time-series dependencies, performs zero-order preserved discretization on continuous parameters, and outputs hidden state features based on time-varying equations.

4. The method for generating typhoon scenes on ocean-going islands driven by a physical information neural network according to claim 3, characterized in that, The dual-branch network includes an edge distribution estimation branch and a dependency structure estimation branch: The marginal distribution estimation branch includes a first set of multilayer perceptrons, which uses the latent state features output from the linear state space as input to the first set of multilayer perceptrons to predict the marginal distribution parameters of each meteorological variable. The dependency structure estimation branch includes a second set of multilayer perceptrons, which takes the hidden state features as input to obtain the original vector, maps the original vector to the non-zero elements of the lower triangular matrix, constructs the covariance matrix, standardizes the covariance matrix, and outputs the correlation coefficient matrix as the parameter of the Copula function.

5. The method for generating typhoon scenes on ocean-going islands driven by a physical information neural network according to claim 4, characterized in that, The generation of coarse-resolution typhoon forecast scenarios includes: Set the initial typhoon state vector, including meteorological variables and corresponding values ​​for the target sea area; Using the dynamic graph as input, the coarse-scale spatiotemporal probability prediction model is iteratively called to predict the typhoon state vector at the current moment and output the joint probability distribution parameters of the meteorological variables at the next moment. The predicted state is obtained by sampling through Cholesky decomposition. This predicted state is then fed back into the coarse-scale spatiotemporal probability prediction model. The prediction process is then executed recursively until a complete typhoon evolution sequence is generated.

6. The method for generating typhoon scenes on distant islands driven by a physical information neural network according to claim 5, characterized in that, The refined physical downscaling model is a conditional physical information neural network, comprising an input layer, hidden layers, and an output layer: The typhoon evolution sequence is used as a conditional feature vector, and the spatiotemporal coordinate vector representing the target sea area is input into the input layer. The conditional feature vector and the spatiotemporal coordinate vector are then concatenated. The hidden layer structure includes several hidden layers, with adjacent hidden layers directly connected by skip connections; the hidden layer structure is used to perform a linear transformation on the concatenated input, and the Swish function is used as the activation function; The output layer does not have an activation function; it directly outputs the prediction vector through a linear transformation.

7. The method for generating typhoon scenes on ocean-going islands driven by a physical information neural network according to claim 6, characterized in that, The phased training includes: In the early stages of training, pre-training is performed based on physical residual loss, data fitting loss, and boundary condition loss. Initial weights are assigned to the data fitting loss to establish the initial manifold of the flow field. As training progresses, anchor point regularization loss is added, and the weight of physical residual loss is gradually increased to force the model output to approximate the solution space that satisfies the Navier-Stokes equations, thereby generating a high-fidelity wind field.

8. A physical information neural network-driven system for generating typhoon scenes on distant islands, characterized in that, include: The data acquisition module is used to acquire historical meteorological data and real-time observation data of the target sea area, construct a spatiotemporal dataset, perform spatiotemporal alignment and normalization processing on the spatiotemporal dataset, and construct a dynamic graph describing the correlation of meteorological variables in the target sea area. A coarse-scale spatiotemporal probability prediction model includes a graph neural network for extracting spatial features, a linear state space for capturing temporal evolution, and a two-branch network for predicting joint probabilities. The dynamic graph is used as input to the coarse-scale spatiotemporal probability prediction model, which generates a coarse-resolution typhoon prediction scenario. In the coarse-scale spatiotemporal probability prediction model, the graph neural network, the linear state space, and the two-branch network are cascaded sequentially. This includes: arranging the graph-level embedding vector sequence output by the graph neural network temporally to form a temporal feature sequence containing spatial topological information, which is used as input to the linear state space; and using the latent state features output by the linear state space as input to the two-branch network, and post-processing the latent state features using a Copula function. The refined physical downscaling model includes a deep residual structure formed by several hidden layers. The coarse-resolution typhoon prediction scenario is taken as input, and the prediction vector is output after linear transformation by the hidden layers. The prediction vector includes the predicted values ​​of meteorological variables in the target sea area. The training module is used to separate a training set from the coarse-resolution typhoon prediction scenario and to train the refined physical downscaling model in stages using a strategy of minimizing the composite loss function. The composite loss function includes physical residual loss, data fitting loss, boundary condition loss, and anchor point regularization loss. The physical residual loss is constructed based on the two-dimensional incompressible Navier-Stokes equations and includes the continuity equation residual and the momentum equation residual. The wind field generation module is used to input the coarse-resolution typhoon prediction scenario into the trained refined physical downscaling model to obtain the wind field output.

Images