A marine disaster risk prevention and control early warning system

Abstract

The application relates to the technical field of disaster early warning, in particular to a marine disaster risk prevention and control early warning system. The system comprises a data fusion module, a risk prediction module, a risk assessment module and an early warning release module. The data fusion module is used for collecting data in real time from various channels and performing standardized preprocessing and fusion. The risk prediction module is used for predicting storm surge risks by using a space-time graph neural network component, a physical information neural network component and a hierarchical deep neural network component contained in the fused data. The risk assessment module is used for receiving generated storm surge risk prediction data, combining vulnerability information of disaster-bearing bodies, assessing disaster risk severity, and determining corresponding risk levels. The early warning release module is used for generating corresponding early warning information based on the determined risk levels and releasing the early warning information through multiple channels. The system converts high-fidelity prediction results into operable risk assessment, automatically generates and releases storm surge early warning, and overcomes the limitations of the prior art in storm surge prediction accuracy, physical authenticity and long-term stability.

Description

Technical Field

[0001] This invention relates to the field of disaster early warning technology, specifically a marine disaster risk prevention and early warning system. Background Technology

[0002] Storm surge, an abnormal rise and fall of sea level caused by severe weather systems such as tropical or extratropical cyclones, is one of the most destructive marine hazards in coastal areas. Against the backdrop of frequent extreme weather events due to global climate change, the intensity and frequency of storm surges are on the rise, posing an increasingly serious threat to coastal populations, critical infrastructure (such as ports and offshore platforms), and economic activities.

[0003] Currently, storm surge disaster risk prevention and control faces multiple challenges. In particular, existing early warning systems typically have the following deficiencies:

[0004] The predictive models are outdated and struggle to capture complex dynamics: Existing technologies largely rely on traditional data mining techniques and expert systems. These methods have limited capabilities in handling storm surges, a dynamic process involving complex air-sea interactions, coastline geometry, and seabed topography, characterized by nonlinearity and strong spatiotemporal coupling. Furthermore, they are often "black box" models, lacking adherence to physical laws, which casts doubt on the reliability of the prediction results.

[0005] Accumulated errors in long-term forecasts lead to insufficient timeliness of early warnings: Both traditional numerical weather prediction models and purely data-driven deep learning models face the problem of rapid accumulation of prediction errors as the time step increases when forecasting long-term storm surge processes. This problem severely limits the lead time and accuracy of early warnings.

[0006] Due to limitations in computational power and algorithms, traditional models and solutions cannot effectively handle the complex physical processes of storm surges. Currently, risk prevention and control efforts are mainly driven by the government at a macro level, such as conducting risk assessments at the county and city level. However, such macro-level assessments lack the refined, high-resolution predictions of inundation extent and depth required for specific ports, critical infrastructure, or coastal communities, making it difficult to meet the precise disaster prevention needs at both the engineering and community levels. Summary of the Invention

[0007] To address the aforementioned deficiencies in the existing technology, this invention provides a marine disaster risk prevention and early warning system for conducting high-precision, long-term, and refined storm surge risk assessment and early warning.

[0008] One aspect of this invention provides a marine disaster risk prevention and early warning system, which runs on computer hardware (including processor and storage device) and includes a data fusion module, a risk prediction module, a risk assessment module and an early warning release module;

[0009] The data fusion module is used to collect meteorological data, marine observation data, seabed topography data, and remote sensing data for obtaining storm parameters, and to perform standardized preprocessing and fusion.

[0010] The risk prediction module is connected to the data fusion module and is used to predict storm surge risk based on the fused data, utilizing the spatiotemporal graph neural network component, physical information neural network component, and hierarchical deep neural network component contained therein.

[0011] The spatiotemporal graph neural network is used to simulate the nearshore storm surge inundation process, the physical information neural network is used to ensure the physical consistency of storm surge prediction results, and the hierarchical deep neural network is used to make long-term stable predictions of large-scale storm surges.

[0012] The risk assessment module is used to receive storm surge prediction data generated by the risk prediction module, and combine it with the vulnerability information of the disaster-bearing body to assess the severity of the disaster risk and determine the corresponding risk level.

[0013] The early warning release module is used to generate corresponding early warning information based on the risk level generated by the risk assessment module, and release it through multiple channels.

[0014] Preferably, the risk prediction module further includes a convolutional autoencoder, which is used to perform dimensionality reduction processing on the spatial prediction data.

[0015] Preferably, the hierarchical deep neural network is used for long-term stable prediction of large-scale storm surges, specifically including:

[0016] Using the encoder portion of the convolutional autoencoder, high-dimensional historical storm surge spatial field data is compressed into low-dimensional latent vectors that can capture its main spatial modes.

[0017] A set of parallel, hierarchical deep neural networks, each corresponding to a different prediction time step, are trained to learn the mapping relationship from key storm parameters to the low-dimensional latent vector.

[0018] In the prediction phase, the network with the largest time step is first used to directly predict the potential vector at the farthest moment based on the current storm parameters. Then, the network with smaller time step is used step by step to take the results of the larger-scale prediction as input and interpolate and refine the prediction at intermediate moments, thereby generating a complete, high-temporal-resolution future potential vector time series.

[0019] Using the decoder part of the convolutional autoencoder, the future potential vector time series is reconstructed frame by frame into a high-dimensional, dynamically evolving storm surge spatial field prediction result.

[0020] Preferably, the spatiotemporal graph neural network is used to simulate the nearshore storm surge inundation process, specifically including:

[0021] The target sea area is abstracted as a graph structure, where nodes represent key geographical locations in space, and the set of edges represents the physical relationships between the nodes;

[0022] By utilizing graph convolutional network layers to perform operations on the graph structure, spatial dependencies are learned by aggregating information from neighboring nodes.

[0023] The output of the graph convolutional network layer is input into the recurrent neural network layer to capture the dynamic evolution of time series data at each node and generate a forecast of the storm surge inundation process.

[0024] Preferably, the physical information neural network is used to ensure the physical consistency of storm surge prediction results. Specifically, this is achieved by integrating the physical control equations into the loss function used for training the spatiotemporal graph neural network model. The integration formula is as follows:

[0025] ;

[0026] In the formula, The integrated loss function, For the model's data loss, For physical loss, This is a weighting factor used to balance data-driven errors and physics-driven errors.

[0027] Preferably, the physical control equation is the shallow water equation.

[0028] Another aspect of the present invention provides a method for marine disaster risk prevention and early warning, comprising the following steps:

[0029] The data fusion module collects and integrates meteorological data, marine observation data, seabed topography data, and remote sensing data used to obtain storm parameters to generate standardized spatiotemporal data.

[0030] The risk prediction module is used to predict the evolution of future storm surges. Specifically, it uses a spatiotemporal graph neural network component to simulate nearshore inundation, a physical information neural network component to ensure physical consistency, and a hierarchical deep neural network component to stabilize long-term predictions.

[0031] Based on the storm surge prediction results and combined with the vulnerability information of the disaster-bearing body, a quantitative risk assessment is conducted using the risk assessment module.

[0032] Based on the risk assessment results, the early warning release module generates and releases early warning information.

[0033] Compared with the prior art, the advantages of this invention are:

[0034] Spatiotemporal graph neural networks can abstract complex geographical entities such as coastlines, islands, and ports, along with their hydrodynamic relationships, into graph structures, thereby efficiently simulating the propagation and inundation processes of storm surges in irregular spaces. Combined with multi-source data fusion capabilities, the system can assess storm surges and their resulting secondary disaster effects, providing a holistic risk view.

[0035] By introducing a physical information neural network component, known physical laws (such as shallow water equations) are directly integrated into the loss function of the model training. This system not only relies on data-driven approaches but also follows the dynamic principles of storm surges, improving the model's generalization ability in sparse data regions and the physical authenticity of the prediction results. This avoids predictions that violate physical common sense that may occur with purely data-driven models.

[0036] By first effectively reducing the dimensionality of the high-dimensional storm surge spatial field, and then using a hierarchical network for stable prediction at multiple time scales, this system can achieve stable and accurate forecasts over a longer period, thus gaining valuable time windows for disaster prevention and mitigation decision-making. Attached Figure Description

[0037] Figure 1 This is a system architecture diagram of a marine disaster risk prevention and early warning system proposed in this invention. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0039] The system described in this invention is deployed on a distributed computing architecture, which includes, but is not limited to: a data acquisition server cluster for data access, a high-performance computing cluster for model training and inference, a spatiotemporal database for storing historical and real-time data, and a publishing server for issuing early warning information. The system's software layer consists of the following modules: a data fusion module, a risk prediction module, a risk assessment module, and an early warning publishing module.

[0040] The data fusion module is used to collect meteorological data, marine observation data, seabed topography data, and remote sensing data for obtaining storm parameters, and to perform standardized preprocessing and fusion.

[0041] This module retrieves data from global and regional data centers through standardized data access services (such as APIs) or customized data parsing scripts. Key data includes:

[0042] Ocean and meteorological observation data: real-time and historical physical oceanographic and meteorological parameters, such as tide level, wave height, wind speed and direction, atmospheric pressure, etc.

[0043] Satellite remote sensing data: high-resolution multispectral imagery and synthetic aperture radar (SAR) data are used for large-scale monitoring of storm parameters, sea surface conditions, and the extent of coastal inundation after a disaster.

[0044] Geophysical data: High-precision global seabed topography data used in storm surge modeling is a key static factor influencing storm surge rise and propagation.

[0045] Numerical weather forecast data: Wind and pressure field forecast products from global models, which are the core dynamic fields driving storm surge models.

[0046] The collected raw data undergoes standardized preprocessing and fusion processes to ensure data quality and usability, including data cleaning, unit unification, and spatiotemporal coordinate normalization.

[0047] For data from different sources and with varying spatiotemporal resolutions, a fusion algorithm is employed to generate a unified data product. For example, for on-site sensor data from buoys and tide gauge stations, the system uses a time series analysis model for processing; this data serves as the "ground truth" for model training and validation. For satellite altimetry and SAR data, the system uses deep learning models such as CNN and U-Net for analysis to obtain key information such as sea level height, storm parameters, and inundation extent.

[0048] In addition, numerical weather prediction data, which is the core driving force of the storm surge model, is integrated. This data is optimized through data assimilation techniques and XGBoost-based correction methods to provide the model with more accurate forcing fields such as wind and pressure. For the seabed topography data, which serves as the static geographic input to the model, the system employs techniques such as kriging interpolation or spline interpolation to support the subsequent construction of spatiotemporal maps.

[0049] Finally, the module outputs a spatiotemporally aligned, uniformly formatted high-dimensional data cube, which serves as the input to the risk prediction module.

[0050] The risk prediction module is connected to the data fusion module. Based on the fused data, it uses the spatiotemporal graph neural network component, physical information neural network component and hierarchical deep neural network component to predict storm surge risk.

[0051] The spatiotemporal graph neural network is used to simulate the nearshore storm surge inundation process, the physical information neural network is used to ensure the physical consistency of storm surge prediction results, and the hierarchical deep neural network is used to make long-term stable predictions of large-scale storm surges.

[0052] Spatiotemporal graph neural networks simulate nearshore storm surge inundation processes, specifically including:

[0053] Step 1: Graph Construction: First, abstract the target nearshore waters into a graph structure. In this system, the node set V represents key locations in space, such as tide gauge stations, ports, or coastal grid points. The edge set E represents the hydrodynamic connectivity or spatial proximity between these nodes. The characteristics of each node are time-series data (such as water level and flow velocity) from the data fusion module.

[0054] Step 2, Spatial Feature Extraction: Graph Convolutional Network (GCN) layers are used to perform computations on the constructed graph. GCN updates the representation of the central node by aggregating information from neighboring nodes, enabling the model to learn how storm surges propagate in complex geographical environments.

[0055] Step 3: Temporal Feature Extraction: The output of the GCN layer is input into a recurrent neural network layer (such as LSTM or GRU) to capture the dynamic evolution of parameters such as water level and flow velocity at each node.

[0056] Traditional grid-based models (such as CNNs) suffer from poor geometric adaptability and low computational efficiency when dealing with irregular coastlines, complex seabed topography, and non-uniformly distributed observation stations. This module, however, uses a spatiotemporal graph neural network to perform short- to medium-term (e.g., 6-24 hours) refined storm surge inundation forecasts for nearshore areas with complex topological connections.

[0057] The physical information neural network component is used to ensure the physical consistency of storm surge prediction results, specifically including:

[0058] Loss function design: First, determine the loss function. Since the physical information neural network component is used to modify the loss function of the neural network, an additional physical loss term is added. This term represents the residual of the model prediction result to the physical control equation (such as the shallow water equation) describing the storm surge.

[0059] Loss function calculation: The total loss function of the system is defined as the weighted sum of data loss and physical loss.

[0060] ;

[0061] in, For the total loss function, Let be the mean square error between the predicted and observed values. Let be the mean square value of the physical equation residuals at randomly sampled points in the spatiotemporal domain. This is a weighting parameter ranging from 0 to 1, used to balance data-driven and physics-driven errors. This parameter determines the importance of the physics loss term relative to the data loss term: when... When the value approaches 1, model training focuses more on following the laws of physics; when... When the value approaches 0, the focus shifts more towards fitting the observed data.

[0062] The spatiotemporal partial derivatives required to calculate the residuals of the physical equations are obtained through the automatic differentiation function built into the deep learning framework.

[0063] Purely data-driven AI models, lacking an understanding of fluid dynamics principles, may produce predictions that violate physical laws. This module utilizes a physical information neural network component to force the storm surge prediction field generated by the spatiotemporal graph neural network to conform to fluid dynamic laws, thereby improving the model's generalization ability and reliability.

[0064] Hierarchical deep neural networks are used for long-term stable prediction of large-scale storm surges, specifically including:

[0065] First, a convolutional autoencoder is trained using historical storm surge data. The encoder part of the convolutional autoencoder learns how to compress the high-dimensional storm surge spatial field into a low-dimensional latent vector, which captures the main modes of the entire spatial field. The decoder then learns how to reconstruct the complete spatial field from this low-dimensional vector.

[0066] Next, a hierarchical deep neural network system consisting of multiple parallel sub-networks is trained, with each sub-network specifically responsible for predicting the evolution of the aforementioned low-dimensional latent vector at different time scales.

[0067] When performing a long-term forecast, the sub-network with the largest time step is first used to directly predict the potential vector at the furthest time based on the current storm parameters (such as central pressure, maximum wind speed radius, and location). Then, the forecasts are progressively scaled down, using the prediction results at larger time scales to constrain and guide the predictions at smaller time scales, until the complete sequence of potential vectors with the highest required time resolution is finally generated.

[0068] Finally, the predicted latent vector time series is fed frame by frame into the decoder of the convolutional autoencoder to reconstruct the complete process of the high-dimensional storm surge spatial field evolution over time.

[0069] For large-scale storm surge spatial fields, using a single-step iterative approach for long-term prediction amplifies even small errors at each step and propagates to the next, eventually leading to an explosive accumulation of errors that renders long-term predictions completely invalid. This module utilizes a hierarchical deep neural network and a convolutional autoencoder to address the problem of long-term, stable prediction of large-scale storm surges. Its hierarchical, multi-scale prediction mechanism can suppress the stepwise accumulation of errors.

[0070] This module combines the aforementioned spatiotemporal graph neural network, physical information neural network components, and hierarchical deep neural network to form a multi-scale storm surge prediction scheme. First, the hierarchical deep neural network generates a large-scale, long-term forecast of future storm surge evolution. Then, in high-risk nearshore areas of particular concern, the system uses this large-scale prediction result as boundary conditions to drive a spatiotemporal graph neural network model constrained by the physical information neural network, performing higher-resolution, refined inundation predictions that take into account the influence of local complex terrain.

[0071] The risk assessment module receives storm surge prediction data generated by the risk prediction module, and combines it with the vulnerability information of the disaster-bearing body to assess the severity of the disaster risk and determine the corresponding risk level.

[0072] The system maintains a geospatial database that stores the location, type, and vulnerability parameters of key disaster-bearing bodies (such as port facilities, power plants, residential areas, and transportation arteries), such as the inundation depth-loss curve of buildings.

[0073] This module performs spatial overlay analysis on the storm surge inundation intensity field output by the risk prediction module and the vulnerability map of the disaster-bearing body.

[0074] By using a rule-based or machine learning-based logic engine, the expected loss or functional failure probability of each disaster-bearing entity under the predicted scenario is calculated, thereby obtaining its risk level.

[0075] The module ultimately outputs a dynamic risk map, which indicates the storm surge risk level of different regions and disaster-bearing bodies at different points in the future.

[0076] The early warning release module is used to generate corresponding early warning information based on the risk level generated by the risk assessment module, and release it through multiple channels.

[0077] When the storm surge risk level of a certain area or disaster-bearing body calculated by the risk assessment module exceeds the preset threshold, the early warning issuance process is triggered.

[0078] This module generates early warning messages and pushes them to the publishing server, which then distributes them to different channels, including national early warning systems, mobile communication networks, and broadcast television systems.

[0079] Example 1:

[0080] This embodiment uses a simulated typhoon event as an example to illustrate the workflow of the solution in detail.

[0081] S1, Data Fusion.

[0082] After the system starts, the data fusion module automatically obtains the following data through the API interface:

[0083] We obtain the latest numerical weather forecast data from the government meteorological department to get the typhoon's path, central pressure, and wind field forecast for the next 72 hours.

[0084] Real-time satellite cloud images are obtained from satellite data centers, and key parameters such as the typhoon's current location and maximum wind speed radius are extracted using image analysis algorithms.

[0085] Real-time tide level, wave height, and air pressure data from multiple tide gauge stations and buoys at the port are obtained as initial conditions and validation benchmarks for model prediction.

[0086] The system utilizes its built-in global seabed topography database to obtain high-precision water depth data for the port and adjacent sea areas.

[0087] This module cleans, unifies, and aligns the above data in time and space to generate a standardized data cube for use by the risk prediction module.

[0088] S2, Risk Prediction.

[0089] The risk prediction module receives the fused data and performs multi-scale predictions in the following order:

[0090] Large-scale, long-term forecasting: First, the latest typhoon parameters (central pressure, maximum wind speed radius, latitude and longitude, etc.) are input into a pre-trained hierarchical deep neural network. This hierarchical deep neural network uses its hierarchical network structure to predict the evolution of storm surge water levels along the coast where the port is located over the next 72 hours.

[0091] Nearshore Refined Prediction: Next, the spatiotemporal graph neural network is activated, using the large-scale prediction results generated by the hierarchical deep neural network as its open boundary conditions. This spatiotemporal graph neural network utilizes a pre-constructed geographic graph structure (where tide gauges, critical infrastructure, and coastal grid points are nodes, and water flow channels are edges) to perform high-resolution simulations of storm surge propagation and reflection within the port and the inundation process in low-lying areas over the next 24 hours.

[0092] Physical consistency constraint: At each step of the computation in the spatiotemporal graph neural network, the physical information neural network component substitutes the output of the spatiotemporal graph neural network (predicted water level and flow velocity) into the shallow water equation and calculates its residual. This residual is added as a penalty term to the loss function of the spatiotemporal graph neural network, thereby forcing the model's predictions to conform to the basic laws of fluid dynamics, ensuring the physical accuracy of the predictions, especially in regions where observational data is sparse.

[0093] S3, Dynamic Risk Assessment.

[0094] The risk assessment module receives a high-resolution inundation prediction map (containing the inundation range and water depth at various future times) from a spatiotemporal graph neural network. This module then performs spatial overlay analysis on this map against a pre-stored critical infrastructure GIS database. The system automatically calculates that within the next 18 hours, several subway station entrances near the port will be submerged by water levels exceeding 1 meter, and the risk level is rated as "high".

[0095] S4. Warning issued.

[0096] The risk assessment results immediately trigger the early warning release module. This module automatically generates an early warning message based on a preset template. This message is immediately pushed to the release server and simultaneously distributed to various channels, including the national early warning system, mobile communication networks, and broadcast television systems.

[0097] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0098] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A marine disaster risk prevention and early warning system, characterized in that, It includes a data fusion module, a risk prediction module, a risk assessment module, and an early warning release module; The data fusion module is used to collect meteorological data, marine observation data, seabed topography data, and remote sensing data for obtaining storm parameters, and to perform standardized preprocessing and fusion. The risk prediction module is connected to the data fusion module and is used to predict storm surge risk based on the fused data, utilizing the spatiotemporal graph neural network component, physical information neural network component, and hierarchical deep neural network component contained therein. The spatiotemporal graph neural network is used to simulate the nearshore storm surge inundation process, the physical information neural network is used to ensure the physical consistency of storm surge prediction results, and the hierarchical deep neural network is used to make long-term stable predictions of large-scale storm surges. The physical information neural network is used to ensure the physical consistency of storm surge prediction results. Specifically, this is achieved by integrating the physical control equations into the loss function used for training the spatiotemporal graph neural network model. The integration formula is as follows: ； In the formula, The integrated loss function, For the model's data loss, L physics For physical loss, This is a weighting factor used to balance data-driven errors and physics-driven errors; The hierarchical deep neural network is used for long-term stable prediction of large-scale storm surges, specifically including: By utilizing the encoder part of a convolutional autoencoder, high-dimensional historical storm surge spatial field data is compressed into low-dimensional latent vectors that can capture its main spatial modes. A set of parallel, hierarchical deep neural networks, each corresponding to a different prediction time step, are trained to learn the mapping relationship from key storm parameters to the low-dimensional latent vector. In the prediction phase, the network with the largest time step is first used to directly predict the potential vector at the farthest moment based on the current storm parameters. Then, the network with smaller time step is used step by step to take the results of the larger-scale prediction as input and interpolate and refine the prediction at intermediate moments, thereby generating a complete, high-temporal-resolution future potential vector time series. Using the decoder part of the convolutional autoencoder, the future potential vector time series is reconstructed frame by frame into a high-dimensional, dynamically evolving storm surge spatial field prediction result; The risk assessment module is used to receive storm surge space field prediction data generated by the risk prediction module, and combine it with the vulnerability information of the disaster-bearing body to assess the severity of disaster risk and determine the corresponding risk level. The early warning release module is used to generate corresponding early warning information based on the risk level determined by the risk assessment module, and release it through multiple channels.

2. The marine disaster risk prevention and early warning system according to claim 1, characterized in that, The risk prediction module also includes a convolutional autoencoder, which is used to perform dimensionality reduction processing on the spatial prediction data.

3. The marine disaster risk prevention and early warning system according to claim 1, characterized in that, The spatiotemporal graph neural network is used to simulate nearshore storm surge inundation processes, specifically including: The target sea area is abstracted as a graph structure, where nodes represent key geographical locations in space, and the set of edges represents the physical relationships between the nodes; By utilizing graph convolutional network layers to perform operations on the graph structure, spatial dependencies are learned by aggregating information from neighboring nodes. The output of the graph convolutional network layer is input into the recurrent neural network layer to capture the dynamic evolution of time series data at each node and generate a forecast of the storm surge inundation process.

4. The marine disaster risk prevention and early warning system according to claim 1, characterized in that, The physical control equations are shallow water equations.

5. A method for marine disaster risk prevention and early warning, characterized in that, Includes the following steps: The data fusion module collects and integrates meteorological data, marine observation data, seabed topography data, and remote sensing data used to obtain storm parameters to generate standardized spatiotemporal data. The risk prediction module is used to predict the evolution of future storm surges. Specifically, it uses a spatiotemporal graph neural network component to simulate nearshore inundation, a physical information neural network component to ensure physical consistency, and a hierarchical deep neural network component to stabilize long-term predictions. The physical information neural network component ensures physical consistency by integrating the physical control equations into the loss function used for training the spatiotemporal graph neural network model. The integration formula is as follows: ； In the formula, The integrated loss function, For the model's data loss, For physical loss, This is a weighting factor used to balance data-driven errors and physics-driven errors; The hierarchical deep neural network component performs stable long-term predictions, specifically including: By utilizing the encoder part of a convolutional autoencoder, high-dimensional historical storm surge spatial field data is compressed into low-dimensional latent vectors that can capture its main spatial modes. A set of parallel, hierarchical deep neural networks, each corresponding to a different prediction time step, are trained to learn the mapping relationship from key storm parameters to the low-dimensional latent vector. In the prediction phase, the network with the largest time step is first used to directly predict the potential vector at the farthest moment based on the current storm parameters. Then, the network with smaller time step is used step by step to take the results of the larger-scale prediction as input and interpolate and refine the prediction at intermediate moments, thereby generating a complete, high-temporal-resolution future potential vector time series. Using the decoder part of the convolutional autoencoder, the future potential vector time series is reconstructed frame by frame into a high-dimensional, dynamically evolving storm surge spatial field prediction result; Based on the storm surge spatial field prediction results, combined with the vulnerability information of the disaster-bearing body, a quantitative risk assessment is conducted using the risk assessment module. Based on the risk assessment results, the early warning release module generates and releases early warning information.

Images