Deep learning-based storm surge intelligent prediction method and device, and medium

Abstract

The invention provides an intelligent storm surge prediction method and device based on deep learning and a medium, and relates to the field of storm surge disaster risk assessment, and the method comprises the steps: obtaining multi-source heterogeneous data of a typhoon storm surge disaster, and carrying out the preprocessing; the multi-source heterogeneous data comprises meteorological and marine observation data and high-precision geographic information data; constructing a typhoon and storm surge prediction network; the typhoon and storm surge prediction network comprises a multi-scale spatial-temporal feature module, a physical mechanism enhancement module and a spatial-temporal sequence prediction model based on a fusion attention mechanism and a graph convolutional network; training a typhoon and storm surge prediction network through multi-source heterogeneous data; acquiring meteorological and tide level live data; and inputting the meteorological and tide level actual data into the trained typhoon and storm surge prediction network to obtain a typhoon and storm surge state prediction result. A physical mechanism is introduced into the neural network to predict the states of the typhoon and the storm surge, and the prediction precision and efficiency are remarkably improved.

Description

Technical Field

[0001] The present application relates to the field of storm surge disaster risk assessment, and in particular to a deep learning-based storm surge intelligent prediction method, device, and medium. Background Art

[0002] Storm surge prediction is of great significance to humanity. Traditional research has primarily relied on an indicator system approach. This approach selects representative indicators from three dimensions: the hazard-prone environment, hazard-causing factors, and the hazard-bearing body. Methods such as the analytic hierarchy process (AHP) and the entropy weighting method are used to determine indicator weights, and finally, a weighted summation is used to generate a comprehensive risk index. Internationally, organizations such as the UNDP, UNDRR, FEMA, and NOAA have published a series of guidelines and standards for storm surge disaster risk assessment, which are widely used in practice.

[0003] However, this type of method still has some shortcomings: first, the assessment is mostly limited to static scenarios and lacks dynamic coupling with real-time and future disaster prediction information; second, the determination of indicator quantification, classification, weights and thresholds mainly relies on expert experience and judgment, which is highly subjective and lacks uncertainty analysis; third, the physical structure and functional connections within the disaster-bearing body are not considered in detail, making it difficult to accurately characterize its vulnerability.

[0004] In recent years, both domestic and foreign countries have begun to explore the introduction of artificial intelligence technology into the field of disaster risk assessment, such as using remote sensing big data and machine learning to automatically extract disaster-prone body information, using deep learning to optimize indicator weights and thresholds, and building multi-agent models to simulate disaster scenarios.

[0005] Disadvantages of existing technology:

[0006] (1) Existing storm surge prediction methods generally have the following shortcomings: empirical statistical methods have low prediction accuracy, lack clear physical mechanisms, and are difficult to cope with complex land, sea, and air interactions; numerical simulation methods have low computational efficiency, rely heavily on high-resolution input data, and are difficult to calibrate model parameters; machine learning methods are mostly based on historical data reproduction, with limited extrapolation to future scenarios and insufficient interpretability.

[0007] (2) In terms of rapid disaster assessment, traditional methods mainly rely on static indicator systems, rarely considering the dynamic integration of forecast information and real-time monitoring data, and lacking spatiotemporal dynamic analysis of disaster situations. The selection of assessment indicator weights and thresholds is highly subjective, lacking uncertainty analysis. The vulnerability of disaster-bearing objects is not detailed enough, for example, factors such as the internal structure of buildings and the cascading effects of key infrastructure are ignored. The practicality and operability of the assessment results need to be improved.

[0008] (3) There is insufficient aggregation and integration of multi-source heterogeneous disaster big data, a lack of multidisciplinary risk knowledge maps, and a low level of intelligence in risk analysis, scenario simulation, and decision support. There is an urgent need to give full play to the enabling role of artificial intelligence to achieve refined management of storm surge disasters.

[0009] In summary, the application of artificial intelligence in disaster risk assessment is still in its infancy, and further breakthroughs are needed in data aggregation, knowledge mining, and interdisciplinary integration. Summary of the Invention

[0010] The purpose of the present invention is to provide a storm surge intelligent prediction method, device and medium based on deep learning in order to solve the problems of poor prediction accuracy and low detection efficiency in existing storm surge prediction methods.

[0011] The above-mentioned purpose of this application is achieved through the following technical solutions:

[0012] S1: Acquire and preprocess multi-source heterogeneous data on typhoon storm surge disasters; multi-source heterogeneous data include: meteorological and oceanographic observation data and high-precision geographic information data;

[0013] S2: Build a typhoon and storm surge prediction network; the typhoon and storm surge prediction network includes: a multi-scale spatiotemporal feature module, a physical mechanism enhancement module, and a spatiotemporal sequence prediction model based on a fusion attention mechanism and graph convolutional network;

[0014] S3: Training typhoon and storm surge prediction networks using multi-source heterogeneous data;

[0015] S4: Obtaining real-time meteorological and tidal data; inputting the real-time meteorological and tidal data into the trained typhoon and storm surge prediction network to obtain typhoon and storm surge status prediction results.

[0016] Optionally, step S1 includes:

[0017] Preprocessing includes: outlier detection, missing value interpolation, and detrending.

[0018] Optionally, step S2 includes:

[0019] The multi-scale spatiotemporal feature module is used to extract the spatiotemporal features of typhoons and ocean environments from multi-source heterogeneous data. The spatiotemporal features include: typhoon statistical features, environmental field spatial features, and terrain field spatial features. The specific steps are as follows:

[0020] Through the multi-scale spatiotemporal feature module, typhoon statistical characteristics are extracted with each time step as the center. Typhoon statistical characteristics include: movement distance, direction change and intensity change;

[0021] Assume that the typhoon center position at the i-th time step is (xi ,y i ), the distance to the next step is d i , the direction change is θ i , the maximum sustained wind speed change of the typhoon is Δv i , the change in the lowest central pressure is Δp i , the statistical characteristics within the time scale T can be expressed as:

[0022]

[0023] Where t is the current time step; T represents the size of the sliding window, T = 6, 12, 24 or 48; D T represents the average distance the typhoon center moves within the time scale T; d i Indicates the distance from the typhoon center in the i-th time step to the next step; Θ T represents the average value of the typhoon direction change within the time scale T; θ i Indicates the change in typhoon direction at the i-th time step; ΔV T ΔV represents the average value of the change in the maximum sustained wind speed (MSW) of the typhoon within the time scale T; T It represents the average value of the change of the typhoon's minimum central pressure (MCP) within the time scale T;

[0024] For the tide level observation sequence r=(r1,r2,…,r t ), extract the maximum water surge S within the sliding window T at each tide gauge station max and cumulative water increase S sum :

[0025] S max =max r i

[0026]

[0027] Where r i is the tide level at the i-th moment, r (MSL,i) is the astronomical mean tide level at the i-th moment;

[0028] Through the multi-scale spatiotemporal feature module, the environmental field spatial features and topographic field spatial features of meteorological and ocean observation data are extracted;

[0029] For the spatial dimension, the typhoon center position (x t ,y t ) is the origin, and the preset range from the origin center is divided into 8 sectors, each sector is 45°; let s represent the eight sectors, s = 1, 2, ..., 8;

[0030] set up They represent the average longitudinal and latitudinal wind speeds, average air pressure, average humidity, average tide level, and average sea surface temperature in the sth region at time t, respectively. The spatial characteristics of the environmental field at time t are represented by six 8-dimensional vectors as follows:

[0031]

[0032]

[0033] In the typhoon prediction area, N typhoon landing paths are divided, each path L n Use buffer (L (n,1) , L (n,r) ) indicates that l and r represent the left and right boundaries of the buffer respectively;

[0034] Extract the weighted average coastline tortuosity C within each buffer zone n and weighted mean elevation H n :

[0035]

[0036] Where n = 1, 2, ..., N; k is L n Number of internal DEM grid cells, c i and h i are the coastline tortuosity and elevation values of the i-th grid, w i is the grid area weight;

[0037] (C n ,H n ) as the spatial features of the terrain field; the spatiotemporal features are used as the input of the spatiotemporal sequence prediction model.

[0038] Optionally, step S2 further includes:

[0039] The spatial characteristics of the environmental field at time t are expressed by matrix X as follows:

[0040]

[0041] Normalize the matrix and scale the data to the [0,1] interval to obtain the normalized matrix X norm ;

[0042] To X norm Process and construct the covariance matrix C, whose element c j,k , defined as:

[0043]

[0044] Where C is a 6×6 matrix, representing the correlation between six meteorological and oceanographic physical quantities; Represents the normalized matrix X norm The mean of the elements in the jth column;

[0045] Perform eigendecomposition on the covariance matrix C to obtain the eigenvalue λ l and the corresponding eigenvector e l , meeting Ce l =λ l e l , and arrange the eigenvalues in descending order;

[0046] Select the eigenvectors corresponding to the first q eigenvalues so that the cumulative contribution rate reaches the preset threshold. The cumulative contribution rate formula is:

[0047]

[0048] The first q vectors constitute the principal component matrix E q =[e1,e 2, ...,e q ];

[0049] For multivariate time series, considering a d-dimensional continuous-time dynamic system, the information flow is calculated as follows:

[0050] dx=F(x,t)dt+B(x,t)dω

[0051] Where F=(F1,F2,...,F d ) T is any nonlinear function about x and t, ω is the standard Wiener process vector, B={b ij} is the perturbation amplitude matrix;

[0052] For a d-dimensional continuous-time dynamic system, from the variable x j to x i Information flow T j→i The calculation formula is:

[0053]

[0054] Among them, dx ji Indicates dx1...dx i-1 dx i+1 ...dx j-1 dx j+1 ...dx n , which means that the difference between x i and x j The integral differential element when integrating all variables except ; E is the mathematical expectation; ρ i =ρ i (x i) is x i The marginal probability density function, ρ ji is x j In x i Probability density function under the conditions; represents the d-2 dimensional real space;

[0055] If T j→i ≠0, then the variable x j For variable x i There is a significant causal effect, and these key variables are grouped into a key variable set V key ;

[0056] E q As a new feature with V key The variables in are used as inputs to the spatiotemporal series prediction model.

[0057] Optionally, step S2 further includes:

[0058] The spatiotemporal sequence prediction model adopts an encoder-decoder architecture;

[0059] The encoder uses two layers of stacked bidirectional gated recurrent units to encode the spatiotemporal sequences of typhoons and storm surges respectively.

[0060] Assume that the temporal and spatial characteristics of the typhoon at time i are input as The temporal and spatial characteristic input of the storm surge at time i is: The spatiotemporal characteristics of the typhoon at time i are The hidden state Storm surge characteristics The hidden state

[0061] Introduce the attention layer and calculate the attention weights of each spatial region of the typhoon and the storm surge:

[0062]

[0063] in, Indicates the size of attention; tanh is a nonlinear function; represents the normalized attention weight; It is the storm surge hidden state that aggregates spatial attention; represents the attention weight of the storm surge on the r-th typhoon spatial region at time i; is the typhoon hidden state of the r-th spatial region, r=1,2,…,8; W s ,W t , b a is the attention layer parameter;

[0064] Will and Concatenate at all time steps and input into a graph convolutional network to model the evolution of the typhoon spatial area in the time dimension:

[0065]

[0066] Among them, H (l) is the regional feature matrix of the lth layer, A is the normalized adjacency matrix, for The degree matrix of (l) is a trainable parameter; ReLU represents the activation function;

[0067] Stack L layers of graph convolutional networks to obtain the regional representation H that integrates spatiotemporal associations r ;

[0068] The decoder uses two LSTM layers, with the i-th moment As input, predict the typhoon state parameters and storm surge state at time i+1;

[0069] Typhoon status parameters include: typhoon intensity and typhoon center location;

[0070] For the prediction of typhoon intensity, the decoding process is as follows:

[0071]

[0072] in, is the hidden state of LSTM of typhoon intensity, MLP p Contains two fully connected layers and ReLU activation; p (i+1) ,v (i+1) , They represent the predicted minimum central pressure of the typhoon at time i+1, the predicted maximum sustained wind speed of the typhoon, the radius of the typhoon level 7 wind circle, and the radius of the typhoon level 10 wind circle respectively;

[0073] For the prediction of the typhoon center position, H r Introducing decoding:

[0074]

[0075] where x (i+1) ,y (i+1) Respectively represent the horizontal and vertical coordinates of the typhoon;

[0076] The water increase at site j on the shore of the storm surge and the largest water increase in the region For prediction, use a dual LSTM decoder:

[0077]

[0078]

[0079] in, represents the hidden state of the LSTM prediction model for storm surge shore segment site j at time i; Represents the hidden state of the LSTM model used to predict the maximum water increase in the entire storm surge area at time i.

[0080] Optionally, step S3 includes:

[0081] The sliding window method is used to sample historical multi-source heterogeneous data and construct training samples;

[0082] Assume that the window length is T, then a training sample formed by sampling at time t is: (X t ,Y t )

[0083] Among them, X t =[x (t-T+1) ,…,x t ] is the input sequence, To predict the target sequence, the typhoon state parameters and storm surge state at the future time T0 are included;

[0084] Select the root mean square error as the loss function:

[0085]

[0086]

[0087] L total =α1L p +α2L v +α3L r +α4L d +α5L s

[0088] Among them, p i , represents the lowest central pressure of the typhoon observed and predicted by the model at the i-th time step; v i , represents the maximum sustained wind speed of the typhoon observed and predicted by the model at the i-th time step; r i , represents the typhoon radius of observation and model prediction at the i-th time step; x i ,y i Indicates the coordinates of the typhoon center position observed at the i-th time step; Indicates the predicted typhoon center position coordinates at the i-th time step; represents the water increase value at the shore station d at the i-th time step, as observed and predicted by the model; represents the maximum water increase value observed and predicted by the model at the i-th time step; L p represents the root mean square error loss function of the typhoon minimum central pressure prediction result; L v represents the root mean square error loss function of the typhoon maximum sustained wind speed prediction; L r represents the root mean square error loss function of typhoon radius prediction; L d represents the root mean square error loss function of the typhoon center position prediction; L s represents the root mean square error loss function of storm surge water increase prediction; α1, α2, α3, α4, α5 are balance coefficients;

[0089] According to the principles of fluid mechanics, the constraints of the continuity equation and momentum equation are introduced in the training process of the spatiotemporal series prediction model. The form of the physical constraint term is:

[0090]

[0091] in, It is the physical quantity value calculated according to the physical equation. is the physical quantity value predicted by the model, α is the weight coefficient of the physical constraint term; n represents the number of samples;

[0092] After each iteration, the predicted typhoon state parameters Input into the classic storm surge numerical model to simulate the spatial distribution of water increase during this period:

[0093]

[0094] Among them, H s is the water height field simulated at time i, ADCIRC is an unstructured grid fluid dynamics model coupled with wave and tide modules, C d is the seabed friction coefficient, n1 is the Manning coefficient, etc., and DEM is high-precision geographic information data;

[0095] Calculation simulation results H s The mean square error with the observed value is:

[0096]

[0097] Among them, H o (x, y, i) represents the actual water height field observed at the position (x, y) at the i-th moment;

[0098] L assim Introduced as weights into the next iterative training as follows:

[0099] L total′ =L total+β·L assim

[0100] Where β is the assimilation intensity parameter; L total′ Indicates the loss value of the next iteration of training.

[0101] An electronic device includes a processor, a memory, a user interface, and a network interface. The memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory so that the electronic device performs a storm surge intelligent prediction method based on deep learning.

[0102] A computer-readable storage medium stores instructions, which, when executed, execute a storm surge intelligent prediction method based on deep learning.

[0103] The beneficial effects of the technical solution provided by this application are:

[0104] 1. Construct a typhoon and storm surge forecasting network based on a physical mechanism enhancement module. This approach, incorporating "physical mechanism enhancement" into a neural network, breaks the limitations of traditional forecasting methods and introduces physical information into deep learning models, effectively improving the model's interpretability and accuracy. This approach uses physical mechanism enhancement to explore the synergistic changes between different spatial points, variables, and systems in meteorological and oceanographic observation data, introducing these as physical information constraints into the model and optimizing the model input.

[0105] 2. Combining graph networks with attention mechanisms, we accurately model the dynamic interactions between typhoons and the ocean, and between the ocean and land. By integrating machine learning with numerical model prediction results, we significantly improve the model's interpretability and ability to predict extreme cases, breaking through the bottlenecks of existing methods in simulating complex ocean-air interactions and extrapolating future scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0106] The present application will be further described below with reference to the accompanying drawings and embodiments, in which:

[0107] Figure 1 It is a step diagram in the embodiment of the present application;

[0108] Figure 2 It is a schematic diagram of the structure of an electronic device in an embodiment of the present application. DETAILED DESCRIPTION

[0109] In order to have a clearer understanding of the technical features, purposes and effects of this application, the specific implementation methods of this application are now described in detail with reference to the accompanying drawings.

[0110] The embodiments of the present application provide a storm surge intelligent prediction method based on deep learning.

[0111] Please refer to Figure 1 , Figure 1 This is a step diagram of a storm surge intelligent prediction method based on deep learning in an embodiment of the present application, including:

[0112] S1: Acquire and preprocess multi-source heterogeneous data on typhoon storm surge disasters; multi-source heterogeneous data include: meteorological and oceanographic observation data and high-precision geographic information data;

[0113] S2: Build a typhoon and storm surge prediction network; the typhoon and storm surge prediction network includes: a multi-scale spatiotemporal feature module, a physical mechanism enhancement module, and a spatiotemporal sequence prediction model based on a fusion attention mechanism and graph convolutional network;

[0114] S3: Training typhoon and storm surge prediction networks using multi-source heterogeneous data;

[0115] S4: Obtaining real-time meteorological and tidal data; inputting the real-time meteorological and tidal data into the trained typhoon and storm surge prediction network to obtain typhoon and storm surge status prediction results.

[0116] As an example, an AI-based model is used to predict typhoons and storm surges. First, multi-scale spatiotemporal features are extracted to capture key information about the typhoon and ocean environment. Then, a spatiotemporal sequence prediction model is constructed that integrates an attention mechanism and a graph convolutional network to predict typhoon paths, intensities, and storm surges.

[0117] As an example, we propose a MEOIF physical mechanism enhancement method for neural networks, which introduces physical mechanisms and significantly improves model interpretability. Finally, by assimilating the prediction results with traditional numerical models, we enhance the physical consistency and interpretability of the predictions, improving the model's performance in extreme situations.

[0118] As an example, a comprehensive storm surge disaster risk assessment system was established. First, a multi-level assessment indicator system was constructed, covering multiple dimensions, including the hazard-bearing body, the disaster-prone environment, and the hazard-causing factors. Monte Carlo simulation was used to analyze the uncertainty of the assessment results and provide a probabilistic distribution of risk. Indicator weights were determined through a combination of subjective and objective methods, balancing expert experience and data characteristics. Finally, combined with the forecast results, dynamic risk assessment was implemented under multiple scenarios, supporting pre-disaster warning and emergency decision-making during a disaster.

[0119] This application provides an embodiment as follows, developing a comprehensive intelligent analysis platform. First, the aggregation and integration of multi-source heterogeneous data is achieved to build a unified storm surge thematic data warehouse. Secondly, based on ontology engineering and knowledge graph technology, a knowledge base in the field of storm surge disasters is constructed to support intelligent risk analysis and reasoning. Finally, an integrated analysis and visualization system is developed to achieve full-process intelligence from data management, risk analysis to decision support, and provide mobile services for the public to enhance social participation and risk prevention awareness.

[0120] Step S1 includes:

[0121] Preprocessing includes: outlier detection, missing value interpolation, and detrending.

[0122] As an example, the multi-source, heterogeneous data required for storm surge hazard analysis is collected and processed. Specifically, this involves collecting long-term meteorological and oceanographic observation data, such as typhoon track, intensity, and tide level observations; integrating high-precision geographic information data, including topography, landforms, and land use; and acquiring high-temporal and spatial resolution socioeconomic data, such as population distribution and GDP. All collected data undergoes systematic preprocessing, including outlier detection, missing value interpolation, and detrending analysis, to ensure data quality and consistency.

[0123] Meteorological and oceanographic data: This paper selects the northwest Pacific and South China Sea (0°N-40°N, 100°E-150°E) as the research area. By cooperating with the Tropical Cyclone Data Center (TCDC) of the China Meteorological Administration, the best path dataset of all typhoons that have reached or exceeded the tropical storm (TS) level in the region since 1949 was obtained. This dataset uses the WGS84 geographic coordinate system with a time resolution of 6 hours. The elements provided include: typhoon center position (longitude and latitude), near-center maximum sustained wind speed (MSW), minimum central pressure (MCP), level 7 wind circle radius (R15), level 10 wind circle radius (R25), etc., as shown in Table 1.

[0124] Table 1 Attribute description of the TCDC typhoon best track dataset

[0125] Observation time Year, month, day, and time (UTC) Once every 6 hours Central longitude °E 0.1 -180°E-180°E Central latitude °N 0.1 -90°N-90°N Maximum sustained wind speed m / s 1m / s Two-minute average Minimum central air pressure hPa 1hPa Radius of level 7 wind circle km 10km Northeast, Southeast, Northwest, Radius of level 10 wind circle km 10km Northeast, Southeast, Northwest,

[0126] At the same time, hourly tide level observations from 76 coastal tide gauges nationwide since 1970 were collected, provided by the Marine Forecasting Center of the Ministry of Natural Resources. All data are in UTC time and calibrated to the local mean sea level datum. Compared to global tide databases such as GLOSS and PSMSL, domestic tide level data is more comprehensive and continuous, and has been rigorously verified by the marine authorities.

[0127] Geographic information data: To support the refined assessment of storm surge disasters, the present invention brings together a variety of high-resolution geographic information data. By integrating the national 1:50,000 and 1:100,000 basic geographic data, detailed coastlines, administrative boundaries, river systems and other key land features were extracted. Based on high-resolution remote sensing images, deep learning methods such as U-Net were used to extract the outlines and attribute information of houses in the typhoon's key impact areas. The SRTM v4 90m resolution global digital elevation model (DEM) data released by NASA was selected and corrected using ICESat / GLA14 laser altimetry data, with a relative elevation accuracy of better than 10m. The underlying surface roughness classification map of key areas was also obtained, and the current land use status was divided into paddy fields, irrigated land, dry land, gardens, woodlands, grasslands, urban land, roads, bare land and other types. The introduction of the above-mentioned geographic information data has greatly improved the level of refinement of risk analysis.

[0128] Socioeconomic Data: Storm surge disasters have profound impacts on all aspects of the socioeconomic landscape. Accurately assessing disaster risk requires understanding the spatial distribution of the population and economy in the affected area. This requires fully utilizing multi-source data fusion methods, improving the spatial and temporal resolution of data beyond traditional population and economic censuses. By sharing data, dynamic demographic information, such as the size of the floating population and the distribution of population heat, was obtained for coastal counties and cities by time period. Nighttime light remote sensing data, combined with land use data, was used to infer the spatial distribution of GDP in coastal areas. Local statistical yearbooks and national economic accounting data at the county, city, and township levels in coastal areas were compiled to obtain macroeconomic indicators such as per capita GDP, per capita disposable income, and the proportion of the primary, secondary, and tertiary industries. Furthermore, socioeconomic attribute data, such as population structure and educational attainment, were collected for key coastal cities through field surveys and questionnaires. These data were then weighted by population and mapped to each grid cell. Through these methods, a high-resolution socioeconomic dataset, as shown in Table 2, was generated, significantly enhancing the targeted and operational nature of disaster assessments.

[0129] Table 2 Spatiotemporal scales of the coastal socioeconomic dataset

[0130]

[0131] All data was standardized to 2020 to reflect the latest socioeconomic conditions. To eliminate heterogeneity in data from different sources, the CGCS2000 geodetic coordinate system and the 1985 National Elevation Datum were uniformly adopted, along with the WGS_1984_Albers equal product conic projection and the GeoTIFF raster and Shapefile vector data formats for storage and management. By establishing a rigorous data quality assessment system, data quality was evaluated across five dimensions: completeness, logical consistency, spatial position accuracy, attribute accuracy, and temporal accuracy. Through systematic quality inspection and cleaning, data integrity and consistency have been significantly improved.

[0132] As an example, in order to eliminate the impact of different data sources and improve data quality, it is also necessary to carry out systematic preprocessing of the data. All collected data must be systematically preprocessed, including outlier detection, missing value interpolation, detrending analysis, etc., to ensure data quality and consistency.

[0133] (1) Outlier detection: The box plot and Z-score method are used to identify outliers in typhoon path parameters and tide data, and manual review is performed in combination with expert knowledge. l , Q u The definition is as follows:

[0134] Q l =Q1-k(Q3-Q1)

[0135] Q u =Q3+k(Q3-Q1)

[0136] In the formula, Q1 and Q3 are the first and third quartiles of the data, respectively, and k is 1.5. The Z-score method is based on the Gaussian distribution assumption and determines that values that deviate from the mean by more than 3 standard deviations are abnormal:

[0137] (2) Missing value interpolation: For the period of missing data, cubic spline interpolation is used for path parameters and piecewise linear interpolation is used for tide levels to ensure data continuity.

[0138] (3) Detrending: To eliminate the long-term trend impact of global warming and sea level rise, the non-parametric LOWESS fitting method is used to detrend the MCP and tide series. The weighted regression formula of LOWESS is:

[0139]

[0140] in, represents the fitted value at point x0, β0 and β1 are the fitted constant term and slope, W i (x0) is the weight function, y iis the observed value. In this method, W is the cubic weight function.

[0141] Step S2 includes:

[0142] As an example, predicting typhoons and storm surges requires learning the spatiotemporal patterns of their evolution from historical data. Targeted extraction of key predictive factors and forecast targets can simplify model complexity, improve computational efficiency, and enhance forecast accuracy. To this end, a multi-scale spatiotemporal feature extraction method is proposed.

[0143] The multi-scale spatiotemporal feature module is used to extract the spatiotemporal features of typhoons and ocean environments from multi-source heterogeneous data. The spatiotemporal features include: typhoon statistical features, environmental field spatial features, and terrain field spatial features. The specific steps are as follows:

[0144] As an embodiment, for the time dimension, the present invention adopts a pyramid pooling method on the optimal typhoon path, and extracts statistical features within the sliding time window with a radius of 6h, 12h, 24h, and 48h with each time step as the center, including: moving distance, direction change, intensity change, etc., to characterize the multi-scale dynamic change trend of typhoon path and intensity.

[0145] Through the multi-scale spatiotemporal feature module, typhoon statistical characteristics are extracted with each time step as the center. Typhoon statistical characteristics include: movement distance, direction change and intensity change;

[0146] Assume that the typhoon center position at the i-th time step is (x i ,y i ), the distance to the next step is d i , the direction change is θ i , the maximum sustained wind speed change of the typhoon is Δv i , the change in the lowest central pressure is Δp i , the statistical characteristics within the time scale T can be expressed as:

[0147]

[0148]

[0149] Where t is the current time step; T represents the size of the sliding window, T = 6, 12, 24 or 48; D T represents the average distance the typhoon center moves within the time scale T; d i Indicates the distance from the typhoon center in the i-th time step to the next step; Θ T represents the average value of the typhoon direction change within the time scale T; θ i Indicates the change in typhoon direction at the i-th time step; ΔV TΔV represents the average value of the change in the maximum sustained wind speed (MSW) of the typhoon within the time scale T; T It represents the average value of the change of the typhoon's minimum central pressure (MCP) within the time scale T;

[0150] For the tide level observation sequence r=(r1,r2,…,r t ), extract the maximum water surge S within the sliding window T at each tide gauge station max and cumulative water increase S sum :

[0151] S max =max r i

[0152]

[0153] Where r i is the tide level at the i-th moment, r (MSL,i) is the astronomical mean tide level at the i-th moment;

[0154] Through the multi-scale spatiotemporal feature module, the environmental field spatial features and topographic field spatial features of meteorological and ocean observation data are extracted;

[0155] For the spatial dimension, the typhoon center position (x t ,y t ) is the origin, and the preset range from the origin center is divided into 8 sectors, each sector is 45°; let s represent the eight sectors, s = 1, 2, ..., 8;

[0156] set up They represent the average longitudinal and latitudinal wind speeds, average air pressure, average humidity, average tide level, and average sea surface temperature in the sth region at time t, respectively. The spatial characteristics of the environmental field at time t are represented by six 8-dimensional vectors as follows:

[0157]

[0158] In the typhoon prediction area, N typhoon landing paths are divided, each path L n Use buffer (L (n,1) , L (n,r) ) indicates that l and r represent the left and right boundaries of the buffer respectively;

[0159] Extract the weighted average coastline tortuosity C within each buffer zone n and weighted mean elevation H n :

[0160]

[0161] Where n = 1, 2, ..., N; k is Ln Number of internal DEM grid cells, c i and h i are the coastline tortuosity and elevation values of the i-th grid, w i is the grid area weight;

[0162] (C n ,H n ) as the spatial features of the terrain field; the spatiotemporal features are used as the input of the spatiotemporal sequence prediction model.

[0163] As an example, the above-mentioned spatiotemporal characteristics not only take into account the multi-scale evolution trends of typhoons and storm surges on a temporal scale, but also depict the heterogeneous distribution of environmental and topographic fields in space, providing rich input information for the construction of prediction models.

[0164] Step S2 further includes:

[0165] The spatial characteristics of the environmental field at time t are expressed by matrix X as follows:

[0166]

[0167] Normalize the matrix and scale the data to the [0,1] interval to obtain the normalized matrix X norm ;

[0168] As an embodiment, normalization is performed to eliminate the influence of dimensions between different variables.

[0169] To X norm Process and construct the covariance matrix C, whose element c j,k , defined as:

[0170]

[0171] Where C is a 6×6 matrix, representing the correlation between six meteorological and oceanographic physical quantities; Represents the normalized matrix X norm The mean of the elements in the jth column;

[0172] Perform eigendecomposition on the covariance matrix C to obtain the eigenvalue λ l and the corresponding eigenvector e l , meeting Ce l =λ l e l , and arrange the eigenvalues in descending order;

[0173] Select the eigenvectors corresponding to the first q eigenvalues so that the cumulative contribution rate reaches the preset threshold. The cumulative contribution rate formula is:

[0174]

[0175] The first q vectors constitute the principal component matrix E q =[e1,e 2, ...,e q ];

[0176] For multivariate time series, considering a d-dimensional continuous-time dynamic system, the information flow is calculated as follows:

[0177] dx=F(x,t)dt+B(x,t)dω

[0178] Where F=(F1,F2,...,F d ) T is any nonlinear function about x and t, ω is the standard Wiener process vector, B={b ij} is the perturbation amplitude matrix;

[0179] For a d-dimensional continuous-time dynamic system, from the variable x j to x i Information flow T j→i The calculation formula is:

[0180]

[0181] Among them, dx ji Indicates dx1...dx i-1 dx i+1 ...dx j-1 dx j+1 ...dx n , which means that the difference between x i and x j The integral differential element when integrating all variables except ; E is the mathematical expectation; ρ i =ρ i (x i ) is x i The marginal probability density function, ρ ji is x j In x i Probability density function under the conditions; represents the d-2 dimensional real space;

[0182] If T j→i ≠0, then the variable x j For variable x i There is a significant causal effect, and these key variables are grouped into a key variable set V key ;

[0183] E q As a new feature with Vkey The variables in are used as inputs to the spatiotemporal series prediction model.

[0184] As an example, E q As a new feature with V key The variables in the dataset are used as inputs to the spatiotemporal series prediction model. During the training process, the model can not only learn the main change patterns of meteorological and ocean physical quantities, but also focus on key influencing factors, thereby effectively introducing physical mechanisms and improving the interpretability of the model for storm surge prediction.

[0185] Step S2 further includes:

[0186] As an embodiment, an encoder-decoder architecture is used to construct a spatiotemporal sequence prediction model that integrates an attention mechanism and a graph convolutional network to jointly predict typhoon path, intensity, and storm surge.

[0187] The spatiotemporal sequence prediction model adopts an encoder-decoder architecture;

[0188] The encoder uses two layers of stacked bidirectional gated recurrent units to encode the spatiotemporal sequences of typhoons and storm surges respectively.

[0189] Assume that the temporal and spatial characteristics of the typhoon at time i are input as The temporal and spatial characteristic input of the storm surge at time i is: The spatiotemporal characteristics of the typhoon at time i are The hidden state Storm surge characteristics The hidden state

[0190] Introduce the attention layer and calculate the attention weights of each spatial region of the typhoon and the storm surge:

[0191]

[0192] in, Indicates the size of attention; tanh is a nonlinear function; represents the normalized attention weight; It is the storm surge hidden state that aggregates spatial attention; represents the attention weight of the storm surge on the r-th typhoon spatial region at time i; is the typhoon hidden state of the r-th spatial region, r=1,2,…,8; W s ,W t , b a is the attention layer parameter;

[0193] Will and Concatenate at all time steps and input into a graph convolutional network to model the evolution of the typhoon spatial area in the time dimension:

[0194]

[0195] Among them, H (l) is the regional feature matrix of the lth layer, A is the normalized adjacency matrix, for The degree matrix of (l) is a trainable parameter; ReLU represents the activation function;

[0196] Stack L layers of graph convolutional networks to obtain the regional representation H that integrates spatiotemporal associations r ;

[0197] The decoder uses two LSTM layers, with the i-th moment As input, predict the typhoon state parameters and storm surge state at time i+1;

[0198] Typhoon status parameters include: typhoon intensity and typhoon center location;

[0199] For the prediction of typhoon intensity, the decoding process is as follows:

[0200]

[0201] in, is the hidden state of LSTM of typhoon intensity, MLP p Contains two fully connected layers and ReLU activation; p (i+1) ,v (i+1) , They represent the predicted minimum central pressure of the typhoon at time i+1, the predicted maximum sustained wind speed of the typhoon, the radius of the typhoon level 7 wind circle, and the radius of the typhoon level 10 wind circle respectively;

[0202] For the prediction of the typhoon center position, H r Introducing decoding:

[0203]

[0204] where x (i+1) ,y (i+1) Respectively represent the horizontal and vertical coordinates of the typhoon;

[0205] The water increase at site j on the shore of the storm surge and the largest water increase in the region For prediction, use a dual LSTM decoder:

[0206]

[0207] in, represents the hidden state of the LSTM prediction model for storm surge shore segment site j at time i; Represents the hidden state of the LSTM model used to predict the maximum water increase in the entire storm surge area at time i.

[0208] As an embodiment, the autoregressive chain decoding is used to achieve the joint forecast of typhoon and storm surge.

[0209] Step S3 includes:

[0210] The sliding window method is used to sample historical multi-source heterogeneous data and construct training samples;

[0211] Assume that the window length is T, then a training sample formed by sampling at time t is: (X t ,Y t )

[0212] Among them, X t =[x (t-T+1) ,…,x t ] is the input sequence, To predict the target sequence, the typhoon state parameters and storm surge state at the future time T0 are included;

[0213] Select the root mean square error as the loss function:

[0214]

[0215] L total =α1L p +α2L v +α3L r +α4L d +α5L s

[0216] Among them, p i , represents the lowest central pressure of the typhoon observed and predicted by the model at the i-th time step; v i , represents the maximum sustained wind speed of the typhoon observed and predicted by the model at the i-th time step; r i , represents the typhoon radius of observation and model prediction at the i-th time step; x i ,y i Indicates the coordinates of the typhoon center position observed at the i-th time step; Indicates the predicted typhoon center position coordinates at the i-th time step; represents the water increase value at the shore segment station j at the i-th time step, as observed and predicted by the model; represents the maximum water increase value observed and predicted by the model at the i-th time step; Lp represents the root mean square error loss function of the typhoon minimum central pressure prediction result; L v represents the root mean square error loss function of the typhoon maximum sustained wind speed prediction; L r represents the root mean square error loss function of typhoon radius prediction; L d represents the root mean square error loss function of the typhoon center position prediction; L s represents the root mean square error loss function of storm surge water increase prediction; α1, α2, α3, α4, α5 are balance coefficients;

[0217] According to the principles of fluid mechanics, the constraints of the continuity equation and momentum equation are introduced in the training process of the spatiotemporal series prediction model. The form of the physical constraint term is:

[0218]

[0219] in, It is the physical quantity value calculated according to the physical equation. is the physical quantity value predicted by the model, α is the weight coefficient of the physical constraint term; n represents the number of samples;

[0220] After each iteration, the predicted typhoon state parameters Input into the classic storm surge numerical model to simulate the spatial distribution of water increase during this period:

[0221]

[0222] Among them, H s is the water height field simulated at time i, ADCIRC is an unstructured grid fluid dynamics model coupled with wave and tide modules, C d is the seabed friction coefficient, n1 is the Manning coefficient, etc., and DEM is high-precision geographic information data;

[0223] Calculation simulation results H s The mean square error with the observed value is:

[0224]

[0225] Among them, H o (x, y, i) represents the actual water height field observed at the position (x, y) at the i-th moment;

[0226] L assim Introduced as weights into the next iterative training as follows:

[0227] L total′ =L total +β·L assim

[0228] Where β is the assimilation intensity parameter; L total′ Indicates the loss value of the next iteration of training.

[0229] As an example, through this "observation-prediction-simulation" assimilation cycle training, the model can better conform to the physical mechanism of storm surge evolution while maintaining the data-driven advantages of deep learning, thereby having stronger generalization capabilities.

[0230] This application also discloses an electronic device. Figure 2 , Figure 2 Schematic diagram of the structure of an electronic device disclosed in an embodiment of the present application. The electronic device 500 may include: at least one processor 501, at least one network interface 504, a user interface 503, a memory 505, and at least one communication bus 502.

[0231] The communication bus 502 is used to implement the connection and communication between these components.

[0232] The user interface 503 may include a display screen, and the optional user interface 503 may also include a standard wired interface or a wireless interface.

[0233] The network interface 504 may optionally include a standard wired interface or a wireless interface (such as a WI-FI interface).

[0234] The present application also discloses a computer-readable storage medium storing a plurality of instructions suitable for loading by a processor to execute the above-mentioned storm surge intelligent prediction method based on deep learning.

[0235] The above are merely exemplary embodiments of the present disclosure and are not intended to limit the scope of the present disclosure. In other words, any equivalent changes and modifications made according to the teachings of the present disclosure are still within the scope of the present disclosure.

[0236] This application is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art not described herein. The description and examples are to be considered as exemplary only, and the scope and spirit of the present disclosure are to be defined by the claims.

Claims

1. A storm surge intelligent prediction method based on deep learning, characterized in that: The method comprises the following steps: S1: Acquire and preprocess multi-source heterogeneous data on typhoon storm surge disasters; multi-source heterogeneous data include: meteorological and oceanographic observation data and high-precision geographic information data; S2: Build a typhoon and storm surge prediction network; the typhoon and storm surge prediction network includes: a multi-scale spatiotemporal feature module, a physical mechanism enhancement module, and a spatiotemporal sequence prediction model based on a fusion attention mechanism and graph convolutional network; S3: Training typhoon and storm surge prediction networks using multi-source heterogeneous data; S4: Obtaining real-time meteorological and tidal data; inputting the real-time meteorological and tidal data into the trained typhoon and storm surge prediction network to obtain typhoon and storm surge status prediction results.

2. A storm surge intelligent prediction method based on deep learning according to claim 1, characterized in that: Step S1 includes: Preprocessing includes: outlier detection, missing value interpolation, and detrending.

3. The storm surge intelligent prediction method based on deep learning according to claim 1, characterized in that: Step S2 includes: The multi-scale spatiotemporal feature module is used to extract the spatiotemporal features of typhoons and ocean environments from multi-source heterogeneous data. The spatiotemporal features include: typhoon statistical features, environmental field spatial features, and terrain field spatial features. The specific steps are as follows: Through the multi-scale spatiotemporal feature module, typhoon statistical characteristics are extracted with each time step as the center. Typhoon statistical characteristics include: movement distance, direction change and intensity change; Assume that the typhoon center position at the i-th time step is (x i ,y i ), the distance to the next step is d i , the direction change is θ i , the maximum sustained wind speed change of the typhoon is Δv i , the change in the lowest central pressure is Δp i , the statistical characteristics within the time scale T can be expressed as: Where t is the current time step; T represents the size of the sliding window, T = 6, 12, 24 or 48; D T represents the average distance the typhoon center moves within the time scale T; d i Indicates the distance from the typhoon center in the i-th time step to the next step; Θ T represents the average value of the typhoon direction change within the time scale T; θ i Indicates the change in typhoon direction at the i-th time step; ΔV T ΔV represents the average value of the change in the maximum sustained wind speed (MSW) of the typhoon within the time scale T; T It represents the average value of the change of the typhoon's minimum central pressure (MCP) within the time scale T; For the tide level observation sequence r=(r1,r2,…,r t ), extract the maximum water surge S within the sliding window T at each tide station max and cumulative water increase S sum : S max =max r i Where r i is the tide level at the i-th moment, r (MSL,i) is the astronomical mean tide level at the i-th moment; Through the multi-scale spatiotemporal feature module, the environmental field spatial features and topographic field spatial features of meteorological and ocean observation data are extracted; For the spatial dimension, the typhoon center position (x t ,y t ) is the origin, and the preset range from the origin center is divided into 8 sectors, each sector is 45°; let s represent the eight sectors, s = 1, 2, ..., 8; set up They represent the average longitudinal and latitudinal wind speeds, average air pressure, average humidity, average tide level, and average sea surface temperature in the sth region at time t, respectively. The spatial characteristics of the environmental field at time t are represented by six 8-dimensional vectors as follows: In the typhoon prediction area, N typhoon landing paths are divided, each path L n Use buffer (L (n,1) , L (n,r) ) indicates that l and r represent the left and right boundaries of the buffer respectively; Extract the weighted average coastline tortuosity C within each buffer zone n and weighted mean elevation H n : Where n = 1, 2, ..., N; k is L n Number of internal DEM grid cells, c i and h i are the coastline tortuosity and elevation values of the i-th grid, w i is the grid area weight; (C n ,H n ) as the spatial features of the terrain field; the spatiotemporal features are used as the input of the spatiotemporal sequence prediction model.

4. The storm surge intelligent prediction method based on deep learning according to claim 1, characterized in that: Step S2 further includes: The spatial characteristics of the environmental field at time t are expressed by matrix X as follows: Normalize the matrix and scale the data to the [0,1] interval to obtain the normalized matrix X norm ; To X norm Process and construct the covariance matrix C, whose element c j,k , defined as: Where C is a 6×6 matrix, representing the correlation between six meteorological and oceanographic physical quantities; Represents the normalized matrix X norm The mean of the elements in the jth column; Perform eigendecomposition on the covariance matrix C to obtain the eigenvalue λ l and the corresponding eigenvector e l , meeting Ce l =λ l e l , and arrange the eigenvalues in descending order; Select the eigenvectors corresponding to the first q eigenvalues so that the cumulative contribution rate reaches the preset threshold. The cumulative contribution rate formula is: The first q vectors constitute the principal component matrix E q =[e1,e 2, ...,e q ]; For multivariate time series, considering a d-dimensional continuous-time dynamic system, the information flow is calculated as follows: dx=F(x,t)dt+B(x,t)dω Where F=(F1,F2,...,F d ) T is any nonlinear function about x and t, ω is the standard Wiener process vector, B={b ij } is the perturbation amplitude matrix; For a d-dimensional continuous-time dynamic system, from the variable x j to x i Information flow T j→i The calculation formula is: Among them, dx ji Indicates dx1...dx i-1 dx i+1 ...dx j-1 dx j+1 ...dx n , which means that the difference between x i and x j The integral differential element when integrating all variables except ; E is the mathematical expectation; ρ i =ρ i (x i ) is x i The marginal probability density function, ρ ji is x j In x i Probability density function under the conditions; represents the d-2 dimensional real space; If T j→i ≠0, then the variable x j For variable x i There is a significant causal effect, and these key variables are grouped into a key variable set V key ; E q As a new feature with V key The variables in are used as inputs to the spatiotemporal series prediction model.

5. The storm surge intelligent prediction method based on deep learning according to claim 1, characterized in that: Step S2 further includes: The spatiotemporal sequence prediction model adopts an encoder-decoder architecture; The encoder uses two layers of stacked bidirectional gated recurrent units to encode the spatiotemporal sequences of typhoons and storm surges respectively. Assume that the temporal and spatial characteristics of the typhoon at time i are input as The temporal and spatial characteristic input of the storm surge at time i is: The spatiotemporal characteristics of the typhoon at time i are The hidden state Storm surge characteristics The hidden state Introduce the attention layer and calculate the attention weights of each spatial region of the typhoon and the storm surge: in, Indicates the size of attention; tanh is a nonlinear function; represents the normalized attention weight; It is the storm surge hidden state that aggregates spatial attention; represents the attention weight of the storm surge on the r-th typhoon spatial region at time i; is the typhoon hidden state of the r-th spatial region, r=1,2,…,8; W s ,W t , b a is the attention layer parameter; Will and Concatenate at all time steps and input into a graph convolutional network to model the evolution of the typhoon spatial area in the time dimension: Among them, H (l) is the regional feature matrix of the lth layer, A is the normalized adjacency matrix, for The degree matrix of (l) is a trainable parameter; ReLU represents the activation function; Stack L layers of graph convolutional networks to obtain the regional representation H that integrates spatiotemporal associations r ; The decoder uses two LSTM layers, with the i-th moment As input, predict the typhoon state parameters and storm surge state at time i+1; Typhoon status parameters include: typhoon intensity and typhoon center location; For the prediction of typhoon intensity, the decoding process is as follows: in, is the hidden state of LSTM of typhoon intensity, MLP p Contains two fully connected layers and ReLU activation; They represent the predicted minimum central pressure of the typhoon at time i+1, the predicted maximum sustained wind speed of the typhoon, the radius of the typhoon level 7 wind circle, and the radius of the typhoon level 10 wind circle respectively; For the prediction of the typhoon center position, H r Introducing decoding: where x (i+1) ,y (i+1) Respectively represent the horizontal and vertical coordinates of the typhoon; The water increase at site j on the shore of the storm surge and the largest water increase in the region For prediction, use a dual LSTM decoder: in, represents the hidden state of the LSTM prediction model for storm surge shore segment site j at time i; Represents the hidden state of the LSTM model used to predict the maximum water increase in the entire storm surge area at time i.

6. A storm surge intelligent prediction method based on deep learning according to claim 1, characterized in that: Step S3 includes: The sliding window method is used to sample historical multi-source heterogeneous data and construct training samples; Assume that the window length is T, then a training sample formed by sampling at time t is: (X t ,Y t ) Among them, X t =[x (t-T+1) ,…,x t ] is the input sequence, To predict the target sequence, the typhoon state parameters and storm surge state at the future time T0 are included; Select the root mean square error as the loss function: L total =α1L p +α2L v +α3L r +α4L d +α5L s in, represents the lowest central pressure of the typhoon at the i-th time step, as observed and predicted by the model; represents the maximum sustained wind speed of the typhoon at the i-th time step, as observed and predicted by the model; represents the typhoon radius of observation and model prediction at the i-th time step; x i ,y i Indicates the coordinates of the typhoon center position observed at the i-th time step; Indicates the predicted typhoon center position coordinates at the i-th time step; represents the water increase value at the shore segment station j at the i-th time step, as observed and predicted by the model; represents the maximum water increase value observed and predicted by the model at the i-th time step; L p represents the root mean square error loss function of the typhoon minimum central pressure prediction result; L v represents the root mean square error loss function of the typhoon maximum sustained wind speed prediction; L r represents the root mean square error loss function of typhoon radius prediction; L d represents the root mean square error loss function of the typhoon center position prediction; L s represents the root mean square error loss function of storm surge water increase prediction; α1, α2, α3, α4, α5 are balance coefficients; According to the principles of fluid mechanics, the constraints of the continuity equation and momentum equation are introduced in the training process of the spatiotemporal series prediction model. The form of the physical constraint term is: in, It is the physical quantity value calculated according to the physical equation. is the physical quantity value predicted by the model, α is the weight coefficient of the physical constraint term; n represents the number of samples; After each iteration, the predicted typhoon state parameters Input into the classic storm surge numerical model to simulate the spatial distribution of water increase during this period: Among them, H s is the water height field simulated at time i, ADCIRC is an unstructured grid fluid dynamics model coupled with wave and tide modules, C d is the seabed friction coefficient, n1 is the Manning coefficient, etc., and DEM is high-precision geographic information data; Calculation simulation results H s The mean square error with the observed value is: Among them, H o (x, y, i) represents the actual water height field observed at the position (x, y) at the i-th moment; L assim Introduced as weights into the next iterative training as follows: L total′ =L total +β·L assim Where β is the assimilation intensity parameter; L total′ Indicates the loss value of the next iteration of training.

7. An electronic device, characterized in that: It includes a processor, a memory, a user interface and a network interface, the memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory so that the electronic device executes the method according to any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium stores instructions, and when the instructions are executed by a computer, the method according to any one of claims 1 to 6 is executed.

Images