A storm surge water level increase prediction method and device, electronic equipment and storage medium

Abstract

The application discloses a storm surge water level increase prediction method and device, electronic equipment and a storage medium, relates to the technical field of seawater monitoring, and comprises the following steps: acquiring multi-source observation data of a target sea area, wherein the multi-source observation data comprises satellite remote sensing data, nearshore monitoring data, meteorological model data and basin data; performing space-time alignment and abnormality processing on the multi-source observation data to generate standardized space-time grid data; performing hybrid prediction modeling on the standardized space-time grid data through a physical data double-driving modeling layer to obtain a target probability water level increase field; and dynamically correcting the target probability water level increase field by using real-time observation data to obtain a storm surge water level increase prediction value. The application combines a physical mechanism model and a probability generation model, overcomes the limitation of a single model, effectively captures the nonlinear evolution process of a storm surge, and adapts to complex coastal topography and variable weather conditions through a space-time double-driving architecture, thereby reducing the risk of missing reports or false reports.

Description

Technical Field

[0001] This application relates to the field of seawater monitoring technology, and in particular to a method, device, electronic equipment, and storage medium for predicting storm surge. Background Technology

[0002] Storm surge is an abnormal rise in sea level caused by extreme weather events such as typhoons and extratropical cyclones. The resulting seawater intrusion and coastal erosion seriously threaten the economy and lives of coastal areas. Against the backdrop of global climate change, extreme weather events are becoming increasingly frequent. Traditional forecasting methods rely primarily on empirical statistical methods and numerical models, but these have significant limitations: empirical formulas depend on linear relationships in historical data, making it difficult to capture dynamic changes in typhoon paths; while numerical models (such as FVCOM and ADCIRC) can simulate multi-physics coupling processes, they require high computing power and are computationally time-consuming (15-30 minutes per simulation), making it difficult to meet the need for minute-level early warnings. Summary of the Invention

[0003] In view of this, embodiments of this application provide a storm surge prediction method, apparatus, electronic device, and storage medium to improve the accuracy of storm surge prediction.

[0004] One aspect of this application provides a storm surge prediction method, the method comprising the following steps:

[0005] Acquire multi-source observation data of the target sea area, including satellite remote sensing data, nearshore monitoring data, meteorological model data, and watershed data;

[0006] The multi-source observation data is spatiotemporally aligned and anomaly processed to generate standardized spatiotemporal grid data;

[0007] By using a dual-drive modeling layer based on physical data, hybrid prediction modeling is performed on the standardized spatiotemporal grid data to obtain the target probability-increasing water field.

[0008] The target probability flood field is dynamically corrected using real-time observation data to obtain the storm surge flood prediction value.

[0009] In some embodiments, the step of spatiotemporal alignment of the multi-source observation data includes the following steps:

[0010] Cubic spline interpolation is used to unify the multi-source observation data to a set time step and a set spatial resolution.

[0011] Sensor noise in the multi-source observation data is removed using the isolated forest algorithm, and missing values ​​in the multi-source observation data are filled using a graph convolutional network.

[0012] In some embodiments, the step of performing hybrid prediction modeling on the standardized spatiotemporal grid data through a dual-drive physical data modeling layer to obtain the target probabilistic water-increasing field includes the following steps:

[0013] Data on storm surge rise processes from offshore to nearshore areas, based on physical mechanisms;

[0014] Based on the storm surge data, a target probability surge field is generated using a probability diffusion model, and the uncertainty of the target probability surge field is quantified.

[0015] In some embodiments, the numerical model based on physical mechanisms simulates storm surge data from the open sea to the nearshore, including the following steps:

[0016] The four-layer nested model of offshore-nearshore-estuary-watershed is used as the numerical model, and the storm surge data from offshore to nearshore are simulated based on the four-layer nested model.

[0017] in:

[0018] The outer ocean layer of the four-layer nested model uses the FVCOM model to simulate ocean circulation exceeding a preset scale at a set resolution;

[0019] The nearshore layer of the four-layer nested model is embedded with shallow water equation residual constraints through a physical information neural network; wherein, the physical information neural network is encrypted to a set size.

[0020] The FVCOM model is used to output the offshore water level field as the initial condition of the physical information neural network, and the nearshore water increase predicted by the physical information neural network is used to feed back to the FVCOM boundary conditions.

[0021] In some embodiments, the step of constructing the probability diffusion model includes the following steps:

[0022] The diffusion process is pre-trained on ERA5 reanalysis data to construct the latent space;

[0023] The probability distribution of the water-increasing field is generated through a reverse diffusion process;

[0024] Monte Carlo Dropout is used to quantify uncertainty and output the confidence interval of the predicted value.

[0025] In some embodiments, the step of dynamically correcting the target probability storm surge field using real-time observation data to obtain the storm surge prediction value includes the following steps:

[0026] Based on the ensemble Kalman filter assimilation framework, the state vector of the target probability-enhancing water field is updated every set time interval; wherein, the state vector includes water level, flow velocity, and salinity;

[0027] The background error covariance matrix of the target probability-increased water field is calculated using the Schur product approximation.

[0028] The target probabilistic storm surge field is subjected to edge-cloud collaborative incremental learning through a federated learning framework to obtain the storm surge prediction value.

[0029] When the ADWIN algorithm detects that the concept drift index of the target probability storm surge field is greater than a set threshold, it returns to the step of updating the state vector of the target probability storm surge field every set time interval based on the ensemble Kalman filter assimilation framework until the concept drift index is less than the set threshold, thereby obtaining the final storm surge prediction value.

[0030] In some embodiments, the method further includes the following steps:

[0031] Based on the storm surge rise prediction values, a flood risk heat map and emergency decision-making instructions are generated;

[0032] The steps for generating the flood risk heat map include the following:

[0033] Coastline segmentation was performed using the U-Net++ model, and terrain parameters were extracted using the sea-land segmentation V1.1 dataset.

[0034] The storm surge rise predictions are overlaid with LiDAR topographic data to calculate the inundation depth for different return periods;

[0035] The Dijkstra algorithm was used to optimize evacuation routes and generate the flood risk heat map that includes multimodal early warning information.

[0036] Another aspect of this application embodiment provides a storm surge prediction device, the device comprising:

[0037] The data acquisition unit is used to acquire multi-source observation data of the target sea area, including satellite remote sensing data, nearshore monitoring data, meteorological model data, and watershed data.

[0038] The data preprocessing unit is used to perform spatiotemporal alignment and anomaly processing on the multi-source observation data to generate standardized spatiotemporal grid data;

[0039] The water increase field modeling unit is used to perform hybrid prediction modeling on the standardized spatiotemporal grid data through a dual-drive modeling layer of physical data to obtain the target probability water increase field.

[0040] The storm surge prediction unit is used to dynamically correct the target probabilistic storm surge field using real-time observation data to obtain the storm surge prediction value.

[0041] Another aspect of this application embodiment provides an electronic device, including a processor and a memory;

[0042] The memory is used to store programs;

[0043] The processor executes the program to implement any of the methods described above.

[0044] Another aspect of this application provides a computer-readable storage medium storing a program that is executed by a processor to implement the method described in any of the above embodiments.

[0045] This application includes at least the following beneficial effects:

[0046] This application acquires multi-source observation data of the target sea area, including satellite remote sensing data, nearshore monitoring data, meteorological model data, and watershed data. It performs spatiotemporal alignment and anomaly processing on the multi-source observation data to generate standardized spatiotemporal grid data. A dual-drive modeling layer based on physical data is used to perform hybrid prediction modeling on the standardized spatiotemporal grid data to obtain the target probabilistic storm surge field. Real-time observation data is used to dynamically correct the target probabilistic storm surge field to obtain the predicted storm surge value. This application combines a physical mechanism model and a probabilistic generation model to overcome the limitations of a single model and effectively capture the nonlinear evolution process of storm surges. A closed-loop feedback mechanism continuously optimizes the prediction results, adapting to extreme scenarios such as sudden changes in typhoon paths. It achieves seamless integration from data acquisition and hybrid modeling to decision output, enhancing system robustness. Physical constraints ensure that the prediction results conform to the laws of fluid dynamics, improving the credibility of disaster decision-making. The spatiotemporal dual-drive architecture adapts to complex coastal topography and variable meteorological conditions, reducing the risk of missed or false alarms. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 A flowchart illustrating a storm surge prediction method provided in this application embodiment;

[0049] Figure 2 This is a structural block diagram of a storm surge prediction device provided in an embodiment of this application. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0051] Before providing a detailed description of the embodiments of this application, some related technologies involved in the embodiments of this application will be described first, as follows:

[0052] Expanding the dimensions of data acquisition:

[0053] Initially relying solely on tide gauge data (as described above), the system now integrates multi-source data from meteorological satellites, buoys, and radar, improving spatiotemporal resolution to the hourly level. For example, the system integrates 0.25°×0.25° gridded meteorological data with 1 / 12° high-precision sea surface height data, eliminating noise through difference correction, trend analysis, and normalization, and extracting key features (such as typhoon center pressure and maximum wind speed radius).

[0054] Model architecture evolution path:

[0055] Empirical statistical methods: Based on empirical formulas for pressure-water increase (such as harmonic analysis), but neglecting the dynamic changes in typhoon paths, resulting in larger prediction errors. Numerical models: Hydrodynamic models such as FVCOM and ADCIRC require input typhoon path point data and have a computational domain resolution of 5km, but struggle to characterize the impact of complex nearshore topography. Machine learning: Models such as LSTM and CNN handle nonlinear temporal features, but are highly dependent on feature engineering.

[0056] Technological convergence trend:

[0057] In recent years, hybrid approaches combining physical models and data-driven methods have emerged, such as combining the FETSCM model with inverse distance weighted interpolation. However, the efficiency of multi-model collaboration is low. An improved approach uses VMD decomposition to extract low-frequency trend components in the 0.01-0.1Hz range and combines this with a TCN network to capture long-term dependencies. Experiments show a reduction in prediction error.

[0058] Some related technical solutions are implemented as follows:

[0059] ① Numerical simulation scheme:

[0060] Technical features: Based on the Navier-Stokes equations, coupled with an atmosphere-ocean-wave module. For example, the FVCOM model of the National Marine Forecasting Center requires input typhoon path point data, and the computational domain resolution is 5km.

[0061] Limitations: Each simulation takes a long time, making it difficult to meet the needs of real-time early warning, and it is sensitive to initial field errors.

[0062] ②Statistical learning plan:

[0063] Typical model: Multiple linear regression: Establishing a linear relationship between water increase and air pressure and wind speed.

[0064] Random forest model: handles non-linear relationships, but is highly dependent on feature engineering.

[0065] Optimization direction: Transform typhoon path points into multi-dimensional temporal features through temporal feature extraction algorithms, and combine them with graph neural networks to capture spatial correlations, thereby improving prediction accuracy.

[0066] ③ Hybrid intelligent solution:

[0067] A comprehensive system architecture integrating multiple artificial intelligence technologies (such as machine learning, deep learning, expert systems, and fuzzy logic) aims to improve overall intelligence through technological complementarity and synergistic optimization. Its core lies in integrating the advantages of different intelligent technologies, such as combining the nonlinear modeling capabilities of neural networks with the logical reasoning capabilities of knowledge engineering, to form an intelligent decision-making system that integrates multimodal data fusion and multi-algorithm collaboration. This solution typically comprises three layers: a data layer (multi-source heterogeneous data acquisition and preprocessing), an algorithm layer (hybrid model training and optimization), and an application layer (intelligent decision-making and execution), achieving cross-domain technology integration through modular design.

[0068] Problems with related technologies:

[0069] (1) Insufficient dynamic modeling capability:

[0070] Traditional numerical models (such as FVCOM and ADCIRC) rely on physical equations, but they struggle to capture nonlinear processes such as abrupt changes in typhoon path and pressure, leading to a surge in errors under extreme weather conditions. Empirical statistical methods (such as harmonic analysis) only establish static linear relationships and cannot adapt to the spatiotemporal heterogeneity of typhoon intensity and path, resulting in generally high 24-hour prediction errors.

[0071] (2) Inefficient data utilization:

[0072] Single models (such as LSTM) only utilize temporal features, neglecting the multi-dimensional correlations such as the spatial distribution of typhoon paths and ocean topography, resulting in insufficient feature utilization. The fusion depth of multi-source data (meteorology, hydrology, satellite remote sensing) is insufficient, failing to achieve cross-modal feature interaction, leading to the loss of key physical mechanisms (such as storm surge-wave coupling).

[0073] (3) Real-time performance and computational bottlenecks:

[0074] Pure numerical models require 15-30 minutes for a single simulation, which is insufficient to meet the needs of minute-level early warning. While deep learning models (such as CNNs) accelerate inference, they have a large number of parameters, making deployment on edge devices difficult. When historical data is insufficient (<10 years), machine learning models suffer from severe overfitting and reduced generalization ability.

[0075] Reference Figure 1 This application provides a storm surge prediction method, specifically including the following steps S100~S130:

[0076] S100: Acquire multi-source observation data of the target sea area, including satellite remote sensing data, nearshore monitoring data, meteorological model data, and watershed data;

[0077] S110: Perform spatiotemporal alignment and anomaly processing on the multi-source observation data to generate standardized spatiotemporal grid data;

[0078] S120: The standardized spatiotemporal grid data is mixed and predicted to be modeled by a dual-drive modeling layer of physical data to obtain the target probability water increase field;

[0079] S130: The target probability flood field is dynamically corrected using real-time observation data to obtain the storm surge flood prediction value.

[0080] Optionally, the step of spatiotemporally aligning the multi-source observation data includes the following steps:

[0081] Cubic spline interpolation is used to unify the multi-source observation data to a set time step and a set spatial resolution.

[0082] Sensor noise in the multi-source observation data is removed using the isolated forest algorithm, and missing values ​​in the multi-source observation data are filled using a graph convolutional network.

[0083] Optionally, the step of performing hybrid prediction modeling on the standardized spatiotemporal grid data through a dual-drive physical data modeling layer to obtain the target probabilistic water-increasing field includes the following steps:

[0084] Data on storm surge rise processes from offshore to nearshore areas, based on physical mechanisms;

[0085] Based on the storm surge data, a target probability surge field is generated using a probability diffusion model, and the uncertainty of the target probability surge field is quantified.

[0086] Optionally, the numerical model based on physical mechanisms simulates storm surge data from the open sea to the nearshore, including the following steps:

[0087] The four-layer nested model of offshore-nearshore-estuary-watershed is used as the numerical model, and the storm surge data from offshore to nearshore are simulated based on the four-layer nested model.

[0088] in:

[0089] The outer ocean layer of the four-layer nested model uses the FVCOM model to simulate ocean circulation exceeding a preset scale at a set resolution;

[0090] The nearshore layer of the four-layer nested model is embedded with shallow water equation residual constraints through a physical information neural network; wherein, the physical information neural network is encrypted to a set size.

[0091] The FVCOM model is used to output the offshore water level field as the initial condition of the physical information neural network, and the nearshore water increase predicted by the physical information neural network is used to feed back to the FVCOM boundary conditions.

[0092] Optionally, the step of constructing the probability diffusion model includes the following steps:

[0093] The diffusion process is pre-trained on ERA5 reanalysis data to construct the latent space;

[0094] The probability distribution of the water-increasing field is generated through a reverse diffusion process;

[0095] Monte Carlo Dropout is used to quantify uncertainty and output the confidence interval of the predicted value.

[0096] Optionally, the step of dynamically correcting the target probability storm surge field using real-time observation data to obtain the storm surge prediction value includes the following steps:

[0097] Based on the ensemble Kalman filter assimilation framework, the state vector of the target probability-enhancing water field is updated every set time interval; wherein, the state vector includes water level, flow velocity, and salinity;

[0098] The background error covariance matrix of the target probability-increased water field is calculated using the Schur product approximation.

[0099] The target probabilistic storm surge field is subjected to edge-cloud collaborative incremental learning through a federated learning framework to obtain the storm surge prediction value.

[0100] When the ADWIN algorithm detects that the concept drift index of the target probability storm surge field is greater than a set threshold, it returns to the step of updating the state vector of the target probability storm surge field every set time interval based on the ensemble Kalman filter assimilation framework until the concept drift index is less than the set threshold, thereby obtaining the final storm surge prediction value.

[0101] Optionally, the method further includes the following steps:

[0102] Based on the storm surge rise prediction values, a flood risk heat map and emergency decision-making instructions are generated;

[0103] The steps for generating the flood risk heat map include the following:

[0104] Coastline segmentation was performed using the U-Net++ model, and terrain parameters were extracted using the sea-land segmentation V1.1 dataset.

[0105] The storm surge rise predictions are overlaid with LiDAR topographic data to calculate the inundation depth for different return periods;

[0106] The Dijkstra algorithm was used to optimize evacuation routes and generate the flood risk heat map that includes multimodal early warning information.

[0107] The following section will provide a detailed introduction and explanation of the solutions in the embodiments of this application, using specific application examples.

[0108] This embodiment proposes a spatiotemporal dual-driven hybrid storm surge prediction scheme that integrates physical mechanisms and data-driven approaches to achieve a closed-loop process of "multi-source data → feature extraction → dynamic modeling → real-time correction".

[0109] Specifically, this embodiment includes the following technical solutions:

[0110] (1) Overall system architecture design:

[0111] The framework adopts a four-layer nested structure: a multi-source observation layer, a data fusion layer, a physical data dual-driven modeling layer, and a dynamic correction and decision output layer, realizing multi-scale data fusion and multi-modal model collaboration. The components of each layer are as follows:

[0112] ① Multi-source observation layer:

[0113] Satellite remote sensing: SWOT satellites acquire 0.5°×0.5° sea surface height anomalies (SSHA), which are then combined with GNSS-R to retrieve sea surface wind speed and tide levels. Nearshore monitoring: Coastal-based high-frequency ground wave radar (HFR) retrieves wind, wave, and current fields with a resolution of 100 meters. Watershed data: The WRF meteorological model outputs 1km gridded precipitation / wind fields, and the SWAT model provides runoff forecasts.

[0114] ② Data Fusion Layer:

[0115] Spatiotemporal alignment: Cubic spline interpolation is used to unify heterogeneous data to a time step of 0.1h and a spatial resolution of 50m.

[0116] Anomaly detection: Sensor noise is removed using the isolated forest algorithm, and missing values ​​are filled in using a graph convolutional network (GCN).

[0117] ③ Dual-drive modeling layer based on physical data:

[0118] The four-layer nested model of offshore-nearshore-estuary-watershed, the generative assimilation system constructs a potential space based on the probabilistic diffusion model (PDM) and generates a high-probability water-increasing field through a reverse diffusion process.

[0119] ④ Decision-making level:

[0120] Dynamic correction module: Utilizes the LSTM-GM model to update boundary conditions in real time. Risk visualization: Integrates the sea-land segmentation-coastline dataset (sea-landsegmentationV1.1) to generate a heatmap of inundation risk.

[0121] (2) Module 1: Multi-source observation layer and implementation process:

[0122] The multi-source observation data layer is the "sensory neural network" of the storm surge prediction system. Through high-precision, multi-dimensional data acquisition and fusion, it provides a spatiotemporally continuous observational foundation for the upper-layer model. Its core architecture consists of three parts: a collaborative acquisition system, intelligent preprocessing, and a dynamic fusion engine. The specific implementation process is as follows:

[0123] ① Multi-source data collaborative acquisition system:

[0124] (a) Satellite remote sensing data acquisition:

[0125] Sensor configuration: SWOT satellite: acquires 0.5°×0.5° sea surface height anomaly (SSHA), covering the globe but with a revisit period of 12 hours.

[0126] GNSS-R satellite: Inverts sea surface wind speed (accuracy ±0.3m / s) and significant wave height, calculated through the phase delay of reflected signals.

[0127] Data acquisition strategy: A dual-satellite collaborative observation mode is adopted, with SWOT providing a large-scale sea surface height field and GNSS-R supplementing nearshore wind field details.

[0128] Data download frequency: SWOT generates global grid data every 3 hours, and GNSS-R updates local areas every 15 minutes.

[0129] (b) Deployment of nearshore monitoring network:

[0130] High-frequency ground wave radar (HFR): Deployed at coastal base stations (distance ≤ 50km), it uses dual-band (HF+VHF) to achieve flow field inversion with a resolution of 0.1m. The sampling rate is 1Hz, the detection range is within 200km, and multipath interference is suppressed through beamforming technology.

[0131] Upgrade of tide gauge stations: Traditional tide gauge stations are equipped with pressure sensors (accuracy ±1cm) and combined with BeiDou positioning to eliminate land subsidence errors.

[0132] Data transmission: Minute-level uploads are achieved via a 5G private network, supporting resume download and data encryption.

[0133] (c) Meteorological-hydrological data fusion:

[0134] Meteorological model data: WRF model outputs 1km gridded wind field (updated every 6 hours), ECMWF provides typhoon track probability forecast (every 2 hours).

[0135] Data preprocessing: Kriging interpolation is used to fill in the pattern blind spots, and the error covariance matrix is ​​dynamically updated.

[0136] Watershed monitoring data: SWAT model outputs runoff forecasts, which are combined with LiDAR topographic data (5m resolution) to calculate the inflow to the sea.

[0137] An ADCP (Acoustic Doppler Current Profiler) was deployed in the estuary area to monitor the intensity of freshwater input in real time.

[0138] ② Intelligent preprocessing of heterogeneous data:

[0139] (a) Spatiotemporal alignment and format standardization:

[0140] Time synchronization: The clocks of each node are unified based on the NTP protocol, and the SWOT data is interpolated to a 1-minute time step using cubic spline interpolation.

[0141] Spatial resampling: SWOT data is projected onto the UTM coordinate system and aligned with the HFR radar grid.

[0142] Data format conversion: SWOTLevel-2 data (HDF5) → NetCDF4, HFR raw data (binary) → GeoTIFF.

[0143] Metadata standardization: Adopt ISO19115 to describe attributes such as data source, coordinate system, and precision.

[0144] (b) Anomaly detection and repair technology:

[0145] Statistical anomaly detection: Outliers are identified based on the isolated forest algorithm (contamination=0.01), such as when the water level change value at the tide gauge station is greater than 3σ, a correction is triggered.

[0146] Radar flow field anomalies: Power spectral density analysis was used to eliminate false flow velocities caused by ocean wave interference.

[0147] Interpolation strategy: Short-term missing (<1h): KNN algorithm (k=5) fills in the missing data based on neighboring site data.

[0148] Long-term missing values ​​(>1h): Introduce the SWAN wave model to generate alternative values, combined with terrain similarity weighting.

[0149] (c) Noise suppression and enhancement:

[0150] Wavelet denoising: The db4 wavelet basis is used for HFR radar data, and the VisuShrink threshold rule is used to improve the signal-to-noise ratio.

[0151] Spatiotemporal filtering: Spatial dimension: Gaussian kernel (σ=500m) to smooth local fluctuations.

[0152] Time dimension: Moving average window (30 minutes) suppresses high-frequency noise.

[0153] ③ Multimodal feature extraction and enhancement:

[0154] (a) Feature Engineering Engine:

[0155] Typhoon path coding: VisionTransformer is used to convert the path point sequence into a 128-dimensional embedding vector to capture dynamic features such as curvature and turning angle.

[0156] Input: latitude and longitude of the typhoon center, wind speed, and air pressure (time resolution 15 minutes).

[0157] Quantification of topographic impact: Based on U-Net++, extract coastline morphological parameters (bay mouth width, slope) and calculate topographic amplification factor (such as bay resonance effect).

[0158] (b) Cross-modal alignment and enhancement:

[0159] Contrastive learning framework: Construct triplet loss to bring similar disaster scenarios closer together (such as "typhoon + high tide") and push away outlier samples.

[0160] Data augmentation: CycleGAN converts clear-sky SWOT images into simulated sea level height anomalies under the influence of typhoons, expanding the training set.

[0161] Knowledge graph construction: Establish an entity relationship graph of "typhoon-wave-flood level" and embed the TransE algorithm to enhance semantic association.

[0162] ④ Dynamic Fusion Engine:

[0163] (a) Multi-source data collaborative fusion:

[0164] Adaptive weight allocation: The weights between observation points are calculated based on the spatial correlation index, with higher weights for closer points.

[0165] Dynamic adjustment: When a sensor fails, the weight of neighboring nodes is automatically increased.

[0166] Error covariance propagation: Constructing the optimal interpolation operator to integrate background field error and observation error.

[0167] (b) Edge-cloud collaborative processing:

[0168] Edge node deployment: The JetsonOrin platform runs a lightweight preprocessing module (VMD decomposition + feature extraction) with a latency of <150ms.

[0169] Supports offline mode: caches data when communication is interrupted and automatically retransmits it after communication is restored.

[0170] Cloud-based fusion center: The A100 GPU cluster performs global data assimilation, supporting parallel computing of 100+ nodes (single fusion time <3 minutes).

[0171] (3) Module 2: Multi-source data deep fusion framework and implementation process:

[0172] ① Data acquisition and heterogeneous integration:

[0173] (a) Multi-source data acquisition system:

[0174] Satellite remote sensing data: SWOT satellites acquire 0.5°×0.5° sea surface height anomalies (SSHA), and GNSS-R retrieves sea surface wind speed and tide levels, covering the globe.

[0175] Nearshore monitoring data: Shore-based high-frequency ground wave radar (HFR): 0.1m resolution wind, wave and current field, detection range within 200km.

[0176] Tide gauge stations: minute-level water level observation, spatial distribution density 50 km² / station.

[0177] Meteorological model data: WRF outputs 1km gridded precipitation / wind field (updated every 6 hours), ECMWF provides typhoon track probability forecast.

[0178] Watershed data: SWAT model runoff forecast (15-minute step) combined with LiDAR topographic data (5-meter resolution).

[0179] (b) Heterogeneous data integration technology:

[0180] Spatiotemporal alignment: Cubic spline interpolation was used to unify the SWOT (12h) and HFR (1s) data to a time step of 0.1h.

[0181] Spatial resampling: SWOT data is aligned with the shore-based radar data grid (100m×100m) after projection transformation (UTM).

[0182] Standardized formats: Satellite data (HDF5) → NetCDF4, Radar data (binary) → GeoTIFF, Tide data (CSV) → Parquet.

[0183] Unified metadata description: Adopting the ISO19115 standard, defining attributes such as data source, coordinate system, and precision.

[0184] ② Data preprocessing and quality improvement:

[0185] (a) Anomaly detection and repair:

[0186] Statistical anomaly detection: Outliers are identified based on the isolated forest algorithm, such as triggering corrections when the water level change value at the tide gauge station exceeds 3σ.

[0187] Interpolation strategy: Short-term missing data (<1h): Fill in the missing data based on neighboring site data using the KNN algorithm (k=5).

[0188] Long-term missing (>1h): Introduce physical models (such as the SWAN wave model) to generate alternative values.

[0189] (b) Noise suppression:

[0190] Wavelet denoising: The db4 wavelet basis is used for HFR radar data, and the threshold rule is VisuShrink.

[0191] Spatiotemporal filtering: Construct a spatial Gaussian kernel (σ=500m) and a time moving average (window=30min) to suppress random fluctuations.

[0192] (c) Data augmentation:

[0193] Adversarial generation: CycleGAN is used to convert clear-sky SWOT images into simulated sea level height anomalies under the influence of typhoons, thus expanding the training set.

[0194] ③ Multimodal feature extraction and enhancement:

[0195] (a) Feature Engineering Engine:

[0196] Temporal features: Typhoon path encoding: The ViT model transforms the path point sequence into a 128-dimensional embedding vector to capture dynamic features such as curvature and turning angle.

[0197] Meteorological factor extraction: The GRU network processes WRF wind field data and outputs 6-dimensional time-varying features (wind speed, wind direction, air pressure, etc.).

[0198] Spatial Features: U-Net++ Segmentation of Coastline Morphology: Input LiDAR topographic data, output multiple morphological parameters such as bay mouth width and water depth gradient.

[0199] Attention mechanism: The Transformer module calculates spatial attention weights to highlight key areas (such as estuaries and bays).

[0200] (b) Cross-modal alignment:

[0201] Comparative learning: Construct triplet loss to bring similar disaster scenarios closer together (such as "typhoon + high tide") and push away outlier samples.

[0202] Knowledge Graph: Establish an entity relationship graph of "typhoon-wave-flood level" and embed the TransE algorithm to enhance semantic association.

[0203] ④ Spatiotemporal decomposition and hybrid modeling:

[0204] (a) Spatiotemporal decomposition module:

[0205] VMD-TS decomposition: The water-enhancing sequence is decomposed into 5 modal components (IMF), retaining the low-frequency trend term of 0.01-0.1Hz.

[0206] A Time Scale Attention (TSA) module is introduced to dynamically allocate the weights of each IMF (Softmax normalization).

[0207] Frequency domain feature enhancement: Wavelet packet decomposition is performed on the high-frequency IMF (>0.5Hz) to extract energy features in the 0.1-0.3Hz range.

[0208] (b) Hybrid prediction model architecture:

[0209] Physical guidance channel: PINN embedded shallow water equation: loss function includes equation residuals.

[0210] Outer sea layer FVCOM simulation: computational domain resolution 1km, coupled with an atmosphere-ocean module (WRF-FVCOM).

[0211] Data-driven channel: Improved LSTM-GM model: Gating mechanism controls the length of historical dependencies (maximum backtracking 72 hours).

[0212] Graph Convolutional Networks (GCNs): Model spatial correlations between sites, with adjacency matrices based on geographical distance and hydraulic connections.

[0213] (c) Generative assimilation system:

[0214] Probabilistic Diffusion Model (PDM): The diffusion process is pre-trained on ERA5 reanalysis data, and the reverse process generates a high-probability water-filling field. The latent space has 128 dimensions, and the denoising network adopts a U-Net architecture (4 downsampling modules + skip connections).

[0215] Uncertainty quantification: Monte Carlo Dropout forward propagation randomly masks 20% of neurons, outputting mean and variance (μ±2σ confidence interval).

[0216] ⑤ Dynamic fusion and real-time correction:

[0217] (a) Edge-cloud collaborative computing:

[0218] Edge nodes: NVIDIA JetsonOrin deployment lightweight module (VMD decomposition + feature extraction).

[0219] Cloud Center: A100 GPU cluster trains hybrid models, supporting parallel computing of 100+ nodes (6 hours per training session).

[0220] (b) Online incremental learning:

[0221] Federated learning framework: Each node uploads gradients to the cloud, and the FedAvg algorithm updates the global model.

[0222] Concept drift detection: The ADWIN algorithm is used to monitor changes in data distribution. When the drift index is greater than 0.25, model retraining is triggered.

[0223] (c) Dynamic feedback mechanism:

[0224] EnKF assimilation: State vector: xk=[η,u,S]T (water level, flow velocity, salinity).

[0225] Background error covariance: Schur product approximation (B≈HΣHT).

[0226] Real-time correction rate: The model is updated every 5 minutes with the latest observation data, and the error correction rate is >90%.

[0227] ⑥ System Deployment and Optimization:

[0228] (a) Model compression and acceleration:

[0229] Knowledge distillation: Teacher model (FVCOM-PINN) guides student model.

[0230] Quantization inference: FP32 → INT8 quantization reduces memory usage, and Jetson Nano supports 10FPS real-time inference.

[0231] (b) Visualization and decision support:

[0232] 3D flooding simulation: The UE5 engine renders risk heatmaps and supports dynamic loading of LOD (Level of Detail).

[0233] Emergency response generation: Dijkstra's algorithm optimizes evacuation routes, combined with OSM road network data.

[0234] (4) Module 3: Physical Data Dual-Driven Modeling Layer:

[0235] The dual-drive modeling layer based on physical data integrates physical mechanisms with data-driven technologies to construct a hybrid prediction system that combines interpretability and high accuracy. Its core architecture consists of three main modules: physical constraint modeling, data-driven modeling, and a hybrid optimization mechanism, achieving a closed-loop process from multi-source data to dynamic prediction.

[0236] ① Physical constraint modeling: Neural networks embedding physical laws:

[0237] (a) Bidirectional coupling between PINN (Physical Information Neural Network) and FVCOM:

[0238] Architecture Design:

[0239] Outer sea layer: The FVCOM model was used to simulate large-scale ocean circulation with a grid resolution of 1 km, covering the area from the open sea to the nearshore.

[0240] Nearshore layer: Topographic-fluid interactions are captured using the PINN network (an improvement based on U-Net++), with the mesh refined to 500m.

[0241] Coupling interface: The FVCOM outputs the offshore water potential field as the initial condition for PINN, and the nearshore water level rise predicted by PINN is fed back to the FVCOM boundary conditions.

[0242] Implementation steps:

[0243] Data preprocessing: The sea surface height anomaly retrieved from SWOT satellites is fused with GNSS-R wind speed data to generate the initial field of the outer sea layer.

[0244] FVCOM Simulation: Based on the wind field output from the WRF meteorological model, FVCOM is driven to calculate the superposition effect of offshore tides and storm surges.

[0245] PINN Training: Using the FVCOM output as the supervision signal, and embedding shallow water equation residual constraints, PINN is trained to predict nearshore water level rise.

[0246] (b) Physically constrained data-driven model:

[0247] Feature engineering:

[0248] Typhoon path coding: VisionTransformer is used to convert the typhoon path point sequence into a spatiotemporal feature vector to capture dynamic features such as path curvature and turning angle.

[0249] Quantification of topographic impact: Based on U-Net++, coastline morphological parameters (such as bay mouth width and slope) are extracted, and the topographic amplification factor is calculated.

[0250] ② Data-driven modeling: Generative assimilation and multi-scale decomposition:

[0251] (a) Probabilistic diffusion model (PDM) constructs the latent space:

[0252] Implementation process:

[0253] Pre-training phase: The diffusion process is trained on ERA5 reanalysis data to learn the probability distribution characteristics of storm surge levels.

[0254] Potential spatial sampling: High-probability water accumulation fields are generated through backdiffusion to capture long-range dependencies of extreme events.

[0255] Uncertainty quantification: Monte Carlo Dropout randomly masks neurons and outputs the predicted mean and variance (confidence interval).

[0256] (b) Spatiotemporal decomposition and feature enhancement:

[0257] VMD-TS decomposition:

[0258] Decomposition strategy: Employ fully adaptive noise ensemble empirical mode decomposition (CEEMDAN) to decompose the water-increasing sequence into low-frequency trend terms (IMF1-3) and high-frequency fluctuation terms (IMF4-5).

[0259] Feature selection: Based on the energy percentage threshold (>80%), the dominant mode is retained and random noise is removed.

[0260] Spatiotemporal attention mechanism:

[0261] Cross-modal interaction: Construct a multimodal attention matrix for typhoon path, ocean waves, and wind speed, and dynamically allocate feature weights.

[0262] Temporal modeling: TCN (Temporal Convolutional Network) is used to capture the lag effect of typhoon path change on water level increase.

[0263] (c) Deep learning flood mapping:

[0264] Dataset Construction:

[0265] STURM-Flood dataset: integrates 200,000 flood inundation images, annotating the inundation depth and extent.

[0266] Coastline segmentation: Based on the sea-landsegmentationV1.1 dataset, coastline morphological features are extracted.

[0267] Model architecture:

[0268] U-Net++ Improvement: Introducing a Spatial Attention Module (SAM) to enhance the ability to identify small-scale flooded areas.

[0269] Loss function: joint cross-entropy, Dice loss and SSIM loss, to optimize boundary clarity and overall accuracy.

[0270] ③ Hybrid optimization mechanism: dynamic fusion and real-time correction:

[0271] (a) Dual-channel dynamic fusion

[0272] Physical channel: The long-term trend field is output based on the FVCOM-PINN coupling model to ensure physical interpretability.

[0273] Data Channel: Improved LSTM-GM model to handle short-term fluctuations, and gating mechanism to control the length of historical dependencies.

[0274] Fusion strategy:

[0275] Uncertainty-aware weighting: Dynamically adjust the weights of physical and data channels based on the variance of the Monte Carlo Dropout output.

[0276] Spatiotemporal alignment: Aligning the spatial scales of data with different resolutions using the Transformer module.

[0277] (b) Real-time correction and incremental learning:

[0278] EnKF assimilation framework:

[0279] State vector: contains multiple physical quantities such as water level, flow velocity, and salinity, and is updated every 15 minutes.

[0280] Background error covariance: The Schur product approximation is used to balance computational efficiency and accuracy.

[0281] Federated learning optimization:

[0282] Edge-cloud collaboration: Lightweight models are deployed at each monitoring site, and gradients are aggregated in the cloud to update global parameters.

[0283] Concept drift detection: Based on the ADWIN algorithm, monitor changes in data distribution and trigger model retraining.

[0284] (5) Module 4: Dynamic Correction and Decision Output Layer:

[0285] The dynamic correction and decision output layer is the system's "intelligent brain," achieving a closed-loop process from prediction to action through real-time data-driven correction, multi-dimensional risk quantification, and intelligent decision generation. Its core architecture consists of three parts: a real-time correction engine, a risk visualization platform, and an emergency response system. The specific implementation process is as follows:

[0286] ① Real-time correction engine:

[0287] (a) Dynamic assimilation of multi-source data:

[0288] Implementation steps:

[0289] Data Acquisition: Integrates real-time data streams from SWOT satellite (sea surface height), GNSS-R (wind speed), HFR radar (flow field), and tide gauge stations (water level), with a time resolution down to the second level.

[0290] Spatiotemporal alignment: Cubic spline interpolation is used to unify heterogeneous data to a 100m grid and a 1-minute time step.

[0291] Anomaly detection: Sensor noise is removed using the isolated forest algorithm, and missing values ​​are filled in using a graph convolutional network (GCN).

[0292] Assimilation update: The model state variables (water level, flow velocity) are adjusted using ensemble Kalman filtering (EnKF), and the background error covariance is approximated using the Schur product.

[0293] (b) Model adaptive correction:

[0294] Implementation steps:

[0295] Error analysis: Compare the predicted water level with the measured value at the tide gauge station to calculate the spatiotemporal error distribution (such as root mean square error RMSE).

[0296] Parameter adjustment: Based on Bayesian optimization, adjust the physical constraint weights of PINN (such as the weight λ of the shallow water equation residual term) to optimize prediction bias.

[0297] Boundary feedback: The corrected offshore field (such as storm surge) is used as the boundary condition of FVCOM to drive the dynamic update of the nearshore model.

[0298] ② Risk visualization platform:

[0299] (a) Multi-dimensional risk quantification:

[0300] Implementation steps:

[0301] Flooding area mapping: Based on the STURM-Flood dataset (200,000 images), U-Net++ was used to segment the flooded area and output a high-precision mask.

[0302] Depth heatmap generation: Combining topographic elevation and flood prediction, the inundation depth distribution is calculated for different return periods (10 years / 100 years).

[0303] Disaster chain analysis: Overlaying wave rise (based on the SWAN model) and storm surge (based on the SWAVE model) to assess the risk of levee overtopping.

[0304] (b) Multimodal information fusion:

[0305] Implementation steps:

[0306] Text generation: Based on the GPT-4 model, disaster descriptions are automatically generated (such as "The water level at Chiwan Station in the Pearl River Estuary is expected to increase by 2.1m, exceeding the warning level by 1.5m").

[0307] Video synthesis: Combining heat maps, flooding animations, and voice broadcasts to generate short video early warning products that are suitable for social media dissemination.

[0308] Interactive query: Allows users to click on any grid point to view in-depth information such as historical comparisons and probability distributions.

[0309] ③ Emergency Response System:

[0310] (a) Intelligent evacuation decision-making:

[0311] Implementation steps:

[0312] Route planning: Optimize evacuation routes based on Dijkstra's algorithm and dynamically avoid congested road sections (real-time traffic data from Gaode API).

[0313] Resource scheduling: Generate a priority map for the deployment of supplies (sandbags, lifeboats) based on the population density of the affected population (from census data).

[0314] Multi-department collaboration: By connecting to the emergency management platform via API, instructions are automatically pushed to fire, medical, and transportation departments.

[0315] (b) Real-time early warning issuance:

[0316] Implementation steps:

[0317] Multi-channel push notification: Simultaneous release via SMS (12379 early warning platform), radio (emergency broadcasting system), and electronic road signs (VMS).

[0318] Precise targeting: Based on user location (mobile phone GPS) and occupation (fisherman, port worker), customized risk avoidance guides are pushed.

[0319] Feedback mechanism: Collect flooding data reported by the public (via a crowdsourcing app) for dynamic model correction.

[0320] In summary, this embodiment includes the following key technical points:

[0321] 1. Spatiotemporal alignment method for multi-source heterogeneous data:

[0322] A spatiotemporal alignment method for multi-source heterogeneous data based on dynamic interpolation and feature fusion is proposed. This method unifies satellite remote sensing data and near-shore radar data to a time step of 0.1 hours through cubic spline interpolation algorithm, achieves spatial alignment by using UTM projection and adaptive grid resampling technology, and constructs a spatiotemporal benchmark framework for multi-source data by combining metadata standardization description (ISO19115). This solves the problem of prediction error accumulation caused by spatiotemporal resolution mismatch in traditional methods.

[0323] 2. Anomaly data repair method based on adversarial generation and physical constraints:

[0324] An anomalous data repair method that integrates Generative Adversarial Networks (GANs) and physical equation constraints is proposed. CycleGAN is used to convert clear-sky satellite imagery into simulated sea surface height anomalies under the influence of typhoons. Combined with shallow water equation residual constraints, alternative values ​​that conform to physical laws are generated, thus overcoming the bottleneck of traditional interpolation methods in data loss repair under extreme weather conditions.

[0325] 3. Multi-scale decomposition and cross-modal attention fusion method:

[0326] A feature enhancement method based on fully adaptive CEEMDAN decomposition and cross-modal attention mechanism is proposed. The flooding sequence is decomposed into low-frequency trend terms (IMF1-3) and high-frequency fluctuation terms (IMF4-5). The feature weights of multi-source data such as typhoon path, waves, and wind speed are dynamically allocated through cross-modal attention matrix to improve the identification accuracy of small-scale flooded areas.

[0327] 4. Physical-Data Dual-Driven Hybrid Modeling Architecture:

[0328] A hybrid modeling method that bidirectionally couples a Physical Information Neural Network (PINN) with a Probabilistic Diffusion Model (PDM) is proposed. The PINN embeds shallow water equations to constrain the dynamic processes of the outer sea layer, while the PDM is used to pre-train a latent space generator on ERA5 data to generate a high-probability water accumulation field and quantify the uncertainty (confidence interval μ±2σ), thus solving the problem of lack of interpretability of pure data models.

[0329] 5. Federated learning-driven dynamic weight allocation method:

[0330] A dynamic weight allocation method based on federated learning is proposed. Each monitoring station uploads local correction parameters to the cloud, and the global model parameters are aggregated through the FedAvg algorithm. The fusion weights of the physical and data channels are dynamically adjusted by combining the variance output of Monte Carlo Dropout, thereby reducing prediction errors.

[0331] 6. Enhanced emergency decision-making methods integrating virtual and real elements:

[0332] A virtual-real fusion emergency decision-making method based on UE5 engine and AR technology is proposed. It constructs a three-dimensional virtual coastline scene (LOD dynamic loading), identifies the real coastline through the mobile phone camera and overlays a virtual flood line, and optimizes the evacuation route by combining Dijkstra algorithm (dynamically avoiding congested road sections), thereby improving evacuation efficiency.

[0333] 7. Real-time correction system with edge-cloud collaboration:

[0334] A real-time correction system that combines edge computing and cloud collaboration deploys a lightweight EnKF module on Jetson edge nodes, updates only the state variables of the area near the typhoon path, and approximates the background error covariance through the Schur product, supporting real-time early warning one hour before the typhoon eyewall makes landfall.

[0335] 8. Knowledge graph-guided typhoon path coding method:

[0336] A typhoon path encoding method based on knowledge graph and VisionTransformer is proposed. The method constructs an entity relationship graph of "typhoon-wave-flood level" (embedded with TransE algorithm) and uses VisionTransformer to transform the path point sequence into a 128-dimensional spatiotemporal feature vector to capture dynamic features such as path curvature and turning angle, thereby improving the accuracy of path prediction.

[0337] 9. Simulation method for tidal river sections adapted to grid nesting:

[0338] A simulation method for tidal river sections based on LiDAR terrain and adaptive mesh nesting is proposed. The method uses OceanMesh2D to generate a three-layer nested mesh and combines it with the SWAN wave model to calculate wave run-up, thus solving the problem of large simulation deviations in the vertical mixing process of traditional models under complex terrain.

[0339] 10. Methods for issuing early warning information regarding blockchain-based evidence storage:

[0340] A blockchain-based method for storing and disseminating early warning information through multiple channels stores early warning information (text, heatmaps, evacuation routes) on the blockchain via smart contracts, verifies data integrity using zero-knowledge proofs, and connects to terminals such as emergency broadcasting and electronic road signs via API, supporting adaptive push notifications in multiple languages.

[0341] Reference Figure 2 This application provides a storm surge prediction device, comprising:

[0342] The data acquisition unit is used to acquire multi-source observation data of the target sea area, including satellite remote sensing data, nearshore monitoring data, meteorological model data, and watershed data.

[0343] The data preprocessing unit is used to perform spatiotemporal alignment and anomaly processing on the multi-source observation data to generate standardized spatiotemporal grid data;

[0344] The water increase field modeling unit is used to perform hybrid prediction modeling on the standardized spatiotemporal grid data through a dual-drive modeling layer of physical data to obtain the target probability water increase field.

[0345] The storm surge prediction unit is used to dynamically correct the target probabilistic storm surge field using real-time observation data to obtain the storm surge prediction value.

[0346] It is understood that the content of the above method embodiments is applicable to the present device embodiments. The specific functions implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0347] In some alternative embodiments, the functions / operations mentioned in the block diagrams may not occur in the order shown in the operation diagrams. For example, depending on the functions / operations involved, two consecutively shown blocks may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order. Furthermore, the embodiments presented and described in the flowcharts of this application are provided by way of example to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logic flows presented herein. Alternative embodiments are contemplated in which the order of various operations is changed and sub-operations described as part of a larger operation are executed independently.

[0348] Furthermore, although this application is described in the context of functional modules, it should be understood that, unless otherwise stated, one or more of the described functions and / or features may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in a separate physical device or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding this application. Rather, given the properties, functions, and internal relationships of the various functional modules in the apparatus disclosed herein, the actual implementation of the module will be understood within the scope of conventional technology for an engineer. Therefore, those skilled in the art can implement the application set forth in the claims using ordinary techniques without excessive experimentation. It is also understood that the specific concepts disclosed are merely illustrative and not intended to limit the scope of this application, which is determined by the full scope of the appended claims and their equivalents.

[0349] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0350] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0351] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0352] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0353] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," 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 this application. In this specification, the 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.

[0354] Although embodiments of this application have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the claims and their equivalents.

[0355] The above is a detailed description of the preferred embodiments of this application, but this application is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of this application, and these equivalent modifications or substitutions are all included within the scope defined by the claims of this application.

Claims

1. A method for predicting storm surge, characterized in that, The method includes the following steps: Acquire multi-source observation data of the target sea area, including satellite remote sensing data, nearshore monitoring data, meteorological model data, and watershed data; The multi-source observation data is spatiotemporally aligned and anomaly processed to generate standardized spatiotemporal grid data; By using a dual-drive modeling layer based on physical data, hybrid prediction modeling is performed on the standardized spatiotemporal grid data to obtain the target probability-increasing water field. The target probability flood field is dynamically corrected using real-time observation data to obtain the storm surge flood prediction value; The process of performing hybrid prediction modeling on the standardized spatiotemporal grid data through a dual-driven physical data modeling layer to obtain the target probabilistic water-increasing field includes the following steps: Data on storm surge rise processes from offshore to nearshore areas, based on physical mechanisms; Based on the storm surge data, a target probabilistic storm surge field is generated using a probability diffusion model, and the uncertainty of the target probabilistic storm surge field is quantified. The steps for constructing the probability diffusion model include the following: The diffusion process is pre-trained on ERA5 reanalysis data to construct the latent space; The probability distribution of the water-increasing field is generated through a reverse diffusion process; Monte Carlo Dropout is used to quantify uncertainty and output the confidence interval of the predicted value.

2. The storm surge prediction method according to claim 1, characterized in that, The steps for spatiotemporal alignment of the multi-source observation data include the following: Cubic spline interpolation is used to unify the multi-source observation data to a set time step and a set spatial resolution. Sensor noise in the multi-source observation data is removed using the isolated forest algorithm, and missing values ​​in the multi-source observation data are filled using a graph convolutional network.

3. The storm surge prediction method according to claim 1, characterized in that, The numerical model based on physical mechanisms simulates storm surge data from the open sea to the nearshore, including the following steps: The four-layer nested model of offshore-nearshore-estuary-watershed is used as the numerical model, and the storm surge data from offshore to nearshore are simulated based on the four-layer nested model. in: The outer ocean layer of the four-layer nested model uses the FVCOM model to simulate ocean circulation exceeding a preset scale at a set resolution; The nearshore layer of the four-layer nested model is embedded with shallow water equation residual constraints through a physical information neural network; wherein, the physical information neural network is encrypted to a set size. The FVCOM model is used to output the offshore water level field as the initial condition of the physical information neural network, and the nearshore water increase predicted by the physical information neural network is used to feed back to the FVCOM boundary conditions.

4. The storm surge prediction method according to claim 1, characterized in that, The process of dynamically correcting the target probability storm surge field using real-time observation data to obtain the storm surge prediction value includes the following steps: Based on the ensemble Kalman filter assimilation framework, the state vector of the target probability-enhancing water field is updated every set time interval; wherein, the state vector includes water level, flow velocity, and salinity; The background error covariance matrix of the target probability-increased water field is calculated using the Schur product approximation. The target probabilistic storm surge field is subjected to edge-cloud collaborative incremental learning through a federated learning framework to obtain the storm surge prediction value. When the ADWIN algorithm detects that the concept drift index of the target probability storm surge field is greater than a set threshold, it returns to the step of updating the state vector of the target probability storm surge field every set time interval based on the ensemble Kalman filter assimilation framework until the concept drift index is less than the set threshold, thereby obtaining the final storm surge prediction value.

5. A storm surge prediction method according to any one of claims 1 to 4, characterized in that, The method further includes the following steps: Based on the storm surge rise prediction values, a flood risk heat map and emergency decision-making instructions are generated; The steps for generating the flood risk heat map include the following: Coastline segmentation was performed using the U-Net++ model, and terrain parameters were extracted using the sea-land segmentation V1.1 dataset. The storm surge rise predictions are overlaid with LiDAR topographic data to calculate the inundation depth for different return periods; The Dijkstra algorithm was used to optimize evacuation routes and generate the flood risk heat map that includes multimodal early warning information.

6. A storm surge prediction device, characterized in that, The device is used to implement the storm surge prediction method as described in claim 1, and the device includes: The data acquisition unit is used to acquire multi-source observation data of the target sea area, including satellite remote sensing data, nearshore monitoring data, meteorological model data, and watershed data. The data preprocessing unit is used to perform spatiotemporal alignment and anomaly processing on the multi-source observation data to generate standardized spatiotemporal grid data; The water increase field modeling unit is used to perform hybrid prediction modeling on the standardized spatiotemporal grid data through a dual-drive modeling layer of physical data to obtain the target probability water increase field. The storm surge prediction unit is used to dynamically correct the target probabilistic storm surge field using real-time observation data to obtain the storm surge prediction value.

7. An electronic device, characterized in that, The electronic device includes a processor and a memory; The memory is used to store programs; The processor executes the program to implement the method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The storage medium stores a program that is executed by a processor to implement the method as described in any one of claims 1 to 5.

Images