Typhoon track unsupervised clustering induction method based on target perception depth autoencoder and double spatial effectiveness index

Abstract

The application discloses a typhoon path unsupervised clustering induction method based on a target perception depth autoencoder and a double-space effectiveness index, including obtaining historical typhoon path data and preset target sea area position information, performing trajectory resampling and constructing a multi-channel feature vector; inputting the multi-channel feature vector into a pre-trained target perception depth representation model, and outputting a corresponding path depth embedding vector; performing unsupervised clustering in a latent space based on the embedding vector, and fusing clustering evaluation results of the latent space and an original geographic coordinate space to construct a double-space comprehensive effectiveness index, thereby determining a classification cluster number, and outputting a typical path type library; finally, quantitatively evaluating the engineering adverse degree of each path type on the target sea area, and identifying and outputting an adverse path type. The application is beneficial to reducing the problem of clustering relying on manually specified cluster numbers, and is beneficial to improving the geographical form consistency of trajectory typing and the engineering pertinence.

Description

Technical Field

[0001] This invention belongs to the field of marine meteorological disaster risk analysis and flood control engineering technology, and in particular relates to an unsupervised clustering and summarization method for typhoon paths. Background Technology

[0002] In the planning and design of coastal flood and storm surge hazard assessments, the evolution of historical typhoon paths is a key input factor for determining the engineering design water level and unfavorable boundary conditions. Given the varying lifespans and highly nonlinear spatial paths of numerous historical typhoons, scientifically extracting features and classifying these trajectories helps to objectively construct representative meteorological driving fields in hydrodynamic numerical simulations. This process provides significant technical support for improving the accuracy and reliability of quantitative analysis of engineering disaster risks.

[0003] Current research on typhoon track classification largely relies on traditional statistical analysis and basic machine learning models. In practice, this typically involves performing isochronous or isochronous linear interpolation on discrete original track nodes, concatenating the coordinate sequences into a high-dimensional vector, performing principal component analysis for linear dimensionality reduction, and finally inputting the vector into a conventional clusterer for classification. However, this approach faces several bottlenecks: First, feature extraction heavily depends on linear coordinate mapping, easily losing key local geometric dynamics when dealing with track sets exhibiting large evolution scales and complex transformations. Second, these basic clustering models require pre-specifying the number of clusters, and evaluation metrics often rely solely on single-dimensional mathematical distance, resulting in poor spatial consistency of classification results on real geographic maps. Furthermore, the general classification results generated by the entire process are disconnected from the directional attributes of specific coastal protection projects, making them difficult to directly use as a quantitative basis for downstream storm surge risk assessment.

[0004] In summary, existing methods still fall short in terms of scale adaptability of feature representation and objectivity of classification decisions when processing massive amounts of highly heterogeneous trajectory data. Furthermore, generalized classification results are difficult to effectively match the specific engineering needs of particular coastal protection areas. Therefore, it is necessary to investigate an inductive method that can improve the quality of feature extraction and adaptive clustering capabilities of complex trajectory data, thereby enhancing the objectivity and engineering applicability of meteorological risk classification results. Summary of the Invention

[0005] The purpose of this invention is to provide an unsupervised clustering and induction method for typhoon paths based on a target-aware deep autoencoder and a dual spatial effectiveness index, in order to solve the aforementioned problems in the prior art.

[0006] Technical solution: An unsupervised clustering and induction method for typhoon paths based on a target-aware deep autoencoder and a dual-spatial effectiveness index, comprising:

[0007] Acquire historical typhoon path data and location information of preset target sea areas;

[0008] The historical typhoon path data is resampled, and combined with the location information of the preset target sea area, a multi-channel feature vector is constructed; the multi-channel feature vector includes target distance channel features of the distance from each node to the preset target sea area;

[0009] The multi-channel feature vector is input into a pre-trained target perception depth representation model, and the path depth embedding vector corresponding to the historical typhoon path data is output. During training, the target perception depth representation model weights the reconstruction error by the distance from each path node to the preset target sea area, so that the trajectory segments adjacent to the preset target sea area can obtain higher reconstruction fidelity.

[0010] Unsupervised clustering is performed in the latent space based on the path depth embedding vector. Clustering and evaluation are carried out separately for different numbers of candidate clusters. The clustering evaluation results of the latent space and the original geographic coordinate space are integrated to construct a dual-space comprehensive effectiveness index. The number of classification clusters is determined according to the dual-space comprehensive effectiveness index, and a typical typhoon path type library is output. The unsupervised clustering is a partitioning clustering method that requires a preset number of candidate clusters and outputs partitioning results. It is one of K-means clustering, Gaussian mixture model, or K-centroid clustering.

[0011] Based on the library of typical typhoon path types, the engineering adverseness of each typhoon path type to the preset target sea area is quantitatively assessed, and the adverse path types are identified. The typical typhoon paths corresponding to the identified adverse path types are then transformed into input meteorological boundary conditions for storm surge numerical simulation, which are used to estimate the design water level or classify the disaster risk level of coastal flood control and tide engineering.

[0012] In some embodiments, based on the typical typhoon path type library, the degree of unfavorability of each typhoon path type to the engineering path of the preset target sea area is quantitatively evaluated, and unfavorable path types are identified.

[0013] Furthermore, the trajectory resampling of historical typhoon path data has two parallel alternative implementation methods: one is linear curvature-weighted curvature adaptive arc-length resampling, which involves linearly weighting the arc length with normalized discrete curvature and accumulating it segment by segment to construct a curvature adaptive arc-length parameter; the other is curvature adaptive arc-length resampling with bounded saturation constraints, which introduces a basis density guarantee factor and a nonlinear saturation curvature transformation, allocating a fixed proportion of the resampling budget according to a uniform arc length and the remaining proportion according to the saturation curvature value, thus ensuring the basic spatial resolution of straight segments while suppressing sampling collapse at extreme sharp turns. These two implementation methods are mutually exclusive and can both preserve the local geometric features of the typhoon turning segment under the premise of a unified variable-length trajectory scale.

[0014] Beneficial effects: Addressing the technical challenge of losing local dynamic features in conventional linear dimensionality reduction for variable-length and highly heterogeneous trajectory data, this solution introduces curvature adaptive isoparametric resampling and multi-channel feature extraction techniques. This approach, through a composite mechanism of nonlinear saturated curvature transformation and basis density budget allocation, smoothly suppresses anomalous clustering of sampling points at extreme sharp turns while maintaining the basic spatial resolution of the entire trajectory. This processing, while unifying the scale of heterogeneous data, effectively preserves the second-order geometric structure information during trajectory evolution, providing multi-dimensional physical input for subsequent representation learning.

[0015] To address the issues of traditional clustering models relying on manually specifying the number of clusters and the monotonicity bias of single-space evaluation metrics leading to classification distortion, this solution constructs a dual-space cross-validation mechanism that integrates latent space compactness and physical coordinate space consistency. By superimposing baseline correction based on random permutations from an unstructured reference set, this mechanism helps filter out background noise caused by mechanical partitioning in the algorithm, captures the inflection point positions corresponding to the hierarchical structure of the data, thereby achieving adaptive recommendation of the number of classification clusters and improving the morphological overlap of trajectory classifications on real maps.

[0016] Furthermore, a target-aware deep representation model and a verification system with accompanying spatial weighting are proposed. By simultaneously injecting exponentially decaying node-level target proximity weights into the network reconstruction error and clustering evaluation layers, the feature mapping of the deep network is guided to focus more on the surrounding sea area of ​​the engineering target. This helps the final output path classification and quantified adverseness to better reflect and respond to the safety constraints of specific flood and tide control projects, providing a connection and support for the extrapolation of meteorological trajectory classification to water conservancy project risk. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall process of the typhoon path unsupervised clustering and induction method based on target-aware deep autoencoder and dual spatial effectiveness index of the present invention.

[0018] Figure 2 This is a schematic diagram of the typhoon path curvature adaptive resampling process of the present invention.

[0019] Figure 3 This is a schematic diagram illustrating the construction of multi-channel feature vectors based on isoparametric path data according to the present invention.

[0020] Figure 4 This is a schematic diagram of the network structure and training process of the target perception depth representation model of the present invention.

[0021] Figure 5 This is a schematic diagram illustrating the construction of a dual-space comprehensive effectiveness index based on latent space and original geographic coordinate space, and the determination of the number of classification clusters in this invention.

[0022] Figure 6 This is a schematic diagram illustrating two alternative calculation methods for the geographic coordinate space consistency term in this invention.

[0023] Figure 7 This is a schematic diagram illustrating the calculation of path unfavorability and identification of unfavorable path types based on a typical typhoon path type library, according to the present invention.

[0024] Figure 8 This is a schematic diagram illustrating how the unfavorable path type is transformed into a storm surge numerical simulation input and used to derive the engineering design water level in this invention. Detailed Implementation

[0025] Example 1: This example provides an unsupervised clustering and induction method for typhoon paths based on a target-aware deep autoencoder and a dual spatial effectiveness index, such as... Figure 1 As shown, the method includes the following steps:

[0026] Step 101: Obtain historical typhoon path data and corresponding typhoon intensity records, along with the location information of the preset target sea area;

[0027] In this step, the acquired historical typhoon path data is specifically represented as a raw data set containing point-by-point longitude and latitude records of each historical typhoon arranged in a time series. This data set constitutes the basis trajectory samples for subsequent cluster analysis. The location information of the preset target sea area is used to characterize the central geographic location of a specific coastal flood control project or water area, typically expressed as specific latitude and longitude coordinate parameters. By introducing the location information of the preset target sea area, an engineering prior is provided for subsequent trajectory evaluation, avoiding the problem of a disconnect between general typhoon path classifications and specific protection targets. The preset target sea area can be represented by one or more reference coordinate points, the geometric center point of the target protection area, or a discrete set of points on the boundary of the target protection area.

[0028] In some alternative implementations, historical typhoon path data can be obtained by reading authoritative best path dataset files released by meteorological departments. The location information of the preset target sea area can be set as the geographical center of a specific estuary tidal gate or dike project. In the case of multiple points / regions, the location information of the target sea area adopts the geometric center location.

[0029] Step 102: Resample the historical typhoon path data and combine it with the location information of the preset target sea area to construct a multi-channel feature vector; the multi-channel feature vector includes target distance channel features of the distance from each node to the preset target sea area;

[0030] Specifically, since the original historical typhoon path data varies in lifespan and number of recording points, direct mathematical measurement would lead to scale inconsistencies. Therefore, a trajectory resampling operation is used to map all variable-length paths into isoparametric path data with a fixed number of nodes, eliminating the heterogeneity of the input dimension.

[0031] Furthermore, displacement and curvature variables reflecting the path movement dynamics are extracted from the isoparametric path data, and the spherical spatial distances from each node of the isoparametric path data to the preset target sea area are hard-coded and concatenated into the feature matrix to generate the multi-channel feature vector. This feature vector explicitly integrates kinematic dynamics, geometric turning morphology, and distance priors relative to the engineering target, providing a multi-dimensional input foundation with physical structure guidance for subsequent deep networks.

[0032] Step 103: Input the multi-channel feature vector into the pre-trained target perception depth representation model and output the path depth embedding vector corresponding to the historical typhoon path data.

[0033] The target perception depth representation model weights the reconstruction error by the distance from each path node to the preset target sea area during training, so that the trajectory segments adjacent to the preset target sea area can obtain higher reconstruction fidelity.

[0034] The target perception depth representation model is a neural network framework whose internal loss function is customized based on the spherical distance to the target sea area with weighted attenuation. After feeding high-dimensional multi-channel feature vectors into the model, a layer-by-layer nonlinear compression mapping is performed, and a dimensionality-reduced path depth embedding vector is extracted at the bottleneck layer of the network. This path depth embedding vector is a low-dimensional representation. Due to the target distance weighting constraint applied during model training, this low-dimensional representation is more conducive to preserving the morphological details of trajectory segments close to the preset target sea area, thus making the extracted depth features more targeted towards specific engineering risks.

[0035] Step 104: Perform unsupervised clustering in the latent space based on the path depth embedding vector. Cluster and evaluate different numbers of candidate clusters separately. Construct a dual-space comprehensive effectiveness index by fusing the clustering evaluation results of the latent space and the original geographic coordinate space. Determine the number of classification clusters based on the dual-space comprehensive effectiveness index and output a typical typhoon path type library. The unsupervised clustering is a partitioning clustering method that requires a preset number of candidate clusters and outputs partitioning results, including K-means clustering, Gaussian mixture model, or K-centroid clustering.

[0036] In this step, unsupervised clustering is performed without relying on any manually assigned category labels. The latent space is an abstract feature space spanned by the path depth embedding vectors obtained from the aforementioned dimensionality reduction. Clustering in this space can measure the compactness and separation of the mathematical structure. Simultaneously, to prevent physical overlap and confusion of path directions in the clustering results of the abstract space on the real map, the morphological distribution consistency of each cluster in the original geographic coordinate space is extracted. The metrics of the two independent spaces are normalized by extreme values ​​and weighted equally to calculate the dual-space comprehensive effectiveness index. Because this index combines statistical superiority and physical... Figure 1 Consistency is assessed by adaptively determining the hierarchical boundaries of the data based on its extreme values, thereby outputting the recommended number of classification clusters with the best comprehensive evaluation under the dual-space index. Based on the results of this cluster division, representative mean paths and prototype typhoon samples within each cluster are summarized to form a typical typhoon path type library with physical guidance significance.

[0037] Step 105: Based on the typical typhoon path type library, quantitatively assess the engineering adverseness of each typhoon path type to the preset target sea area, identify the adverse path types, and convert the typical typhoon paths corresponding to the identified adverse path types into input meteorological boundary conditions for storm surge numerical simulation, which are used for the calculation of design water levels or the classification of disaster risk levels for coastal flood control and storm surge projects.

[0038] To transform the purely objective unsupervised clustering results into a selection criterion for storm surge simulation candidate scenarios that can be directly applied in engineering, it is necessary to reproject the typical typhoon path type library, which includes multiple typical paths, onto the surrounding area of ​​the preset target sea area. By statistically analyzing the probability tendency of various typhoons intruding into a specific influence radius around the preset target sea area, and combining the normalized disaster-causing intensity of typhoon samples within the subset, a comprehensive evaluation parameter that can measure the frequency and intensity of disasters, namely, the engineering adverseness, is calculated. Based on the calculated engineering adverseness value, all path types in the type library are sorted in descending order from high to low, and the top-ranked path categories are identified as the unfavorable path types facing the preset target sea area.

[0039] In some specific implementation scenarios, the output of the unfavorable path type can be transformed into one of the unfavorable meteorological boundary driving input conditions in the typhoon storm surge numerical simulation process, which can be used to support the estimation of the design water level of downstream water conservancy projects.

[0040] Example 2: Based on Example 1 above, this example further explains the quality control and cleaning process of the original optimal path dataset.

[0041] In one possible implementation, the following steps are included:

[0042] Step 201: Obtain the original optimal path dataset; remove unnamed typhoon records and duplicate records of sub-centers under the same typhoon number from the original optimal path dataset, and retain only the main center trajectory;

[0043] The original optimal path dataset can be obtained from publicly available meteorological data sources, such as the optimal path dataset for tropical cyclones in the Northwest Pacific released by the Shanghai Typhoon Institute of the China Meteorological Administration, or the optimal path dataset released by the Joint Typhoon Warning Center. Those skilled in the art can select the appropriate data source based on the ocean area where the target sea area is located.

[0044] After obtaining the original optimal path dataset, the name and number fields of each typhoon record are extracted. The name field of each record is checked for any identifier characters indicating an unnamed typhoon. If the result is yes, all time-series records corresponding to that typhoon are removed. For records with valid names, the typhoon number field is extracted and grouped for comparison. When multiple center coordinate records with the same timestamp under the same typhoon number are detected, deduplication logic is triggered. Pressure characteristics or built-in primary / secondary marker fields are extracted from each concurrent record to determine the primary center record, and redundant node data belonging to secondary centers are deleted. Through extraction, determination, and deduplication operations, each typhoon lifecycle event corresponds to a spatial evolution sequence.

[0045] In some alternative implementations, after deduplication of the secondary center is completed, the timestamp interval of the retained nodes can also be detected, and linear interpolation can be used to fill in the small number of time discontinuities caused by deduplication.

[0046] Step 202: Perform spatial validity filtering based on the set latitude and longitude activity domain boundaries, and filter out incomplete paths with a total number of continuous nodes below a set threshold, and use the retained complete typhoon paths as the historical typhoon path data.

[0047] The latitude and longitude activity domain boundary is set as a spatial filtering reference. Specifically, the latitude and longitude activity domain boundary can be defined as a rectangular area with latitude between 0 and 60 degrees and longitude between 90 and 180 degrees. Each main center trajectory retained from the previous steps is traversed, and it is determined whether the coordinate parameters of its trajectory nodes are located inside this rectangular area. If all coordinate points of a trajectory are located outside the boundary, a spatial filtering action is performed to remove it.

[0048] For trajectories filtered by spatial boundary conditions, the cumulative number of valid record nodes contained within them is counted. A threshold of 4 nodes can be selected. It is then determined whether the cumulative number is less than this threshold. If it is, the trajectory is considered a fragmented path lacking evolutionary analysis value and is removed. The remaining trajectory data, determined by spatial boundary conditions and node count, is encapsulated and used as historical typhoon path data for subsequent feature extraction steps.

[0049] Furthermore, when performing spatial validity filtering, supplementary filtering conditions can be added based on the completeness of meteorological observations, such as removing trajectory records that lack a central pressure or whose maximum wind speed attribute field ratio exceeds a preset limit coefficient.

[0050] Example 3: In one possible implementation, the historical typhoon path data is resampled, such as... Figure 2 As shown, it includes the following steps:

[0051] Step 301: Calculate the arc length of adjacent nodes in the historical typhoon path data based on the spherical distance, and extract the discrete curvature of the corresponding node based on the difference in azimuth angle of adjacent arc segments;

[0052] The latitude and longitude coordinate sequence is obtained from historical typhoon path data. The semi-sine cross algorithm is used to calculate the great circle physical distance between adjacent nodes on the Earth's surface, and the calculated great circle physical distance is used as the arc length. Based on the latitude and longitude coordinates of adjacent nodes, the corresponding heading and azimuth angles are calculated, and the absolute difference of the azimuth angle between two consecutive arc segments is extracted. The extracted absolute difference is divided by the arithmetic mean of the lengths of the two consecutive arc segments to extract the discrete curvature reflecting the local curvature of the trajectory.

[0053] Step 302: Normalize the discrete curvature;

[0054] Extract the discrete curvature set corresponding to all nodes along the entire path of a single typhoon, and search for the maximum curvature value in this set. Divide the discrete curvature of each node by this maximum curvature value, and add a small value (such as a numerical stability constant around 0.000000001) as a denominator compensation term to prevent division by zero errors. This yields a normalized discrete curvature with values ​​ranging from 0 to 1.

[0055] Step 303: The arc length is linearly weighted using the normalized discrete curvature, and the cumulative arc length parameter is obtained by accumulating it segment by segment, which is used as the curvature adaptive arc length parameter.

[0056] Extract the preset curvature emphasis coefficient, multiply the normalized discrete curvature by the curvature emphasis coefficient, and then add 1 to the product to form a weighting amplification factor. Multiply the weighting amplification factor by the original arc length to obtain the curvature-weighted updated arc length.

[0057] Specifically, the updated arc length after curvature weighting. ;

[0058] Where η is the preset curvature emphasis coefficient, and C norm Let be the normalized discrete curvature of the node, and L be the original arc length. In this embodiment, the curvature emphasis coefficient η is set to 5. Those skilled in the art can adaptively adjust this parameter according to the desired sampling density difference between turning segments and straight segments in actual trajectory data.

[0059] After calculating the updated arc length of all nodes on a single path, the updated arc lengths are accumulated segment by segment along the trajectory evolution direction to obtain a cumulative value. The cumulative value of each node is divided by the sum of the updated arc lengths of the entire segment to calculate a monotonically increasing cumulative arc length parameter with a value range between 0 and 1. This cumulative arc length parameter is then set as the curvature adaptive arc length parameter.

[0060] Step 304: Using the curvature adaptive arc length parameter as the interpolation reference, perform conformal segmented interpolation on the historical typhoon path data to obtain isoparametric path data with a fixed number of nodes.

[0061] A target parameter vector containing a fixed number of equally spaced values ​​is generated. This fixed number corresponds to a pre-defined total number of isoparametric nodes, which can be set to 32. Using the curvature adaptive arc length parameter as the interpolation independent variable and the original longitude and latitude sequences as dependent variables, a piecewise cubic Hermitian conformal interpolation algorithm is used to resample and evaluate the parameters at their corresponding positions in the target parameter vector. The resulting longitude and latitude numerical sequences with equally spaced parameters are then combined and assembled to obtain isoparametric path data with a fixed number of nodes.

[0062] The curvature-adaptive arc-length parameter constructed using this mechanism exhibits a larger parameter increment in high-curvature sections where the trajectory undergoes sharp turns, enabling the equally spaced target parameter vectors to project and match more physical interpolation nodes in this region. This mechanism causes resampling nodes to cluster and distribute in the typhoon's turning segment, preserving local turning geometry characteristics that indicate dynamic changes while unifying the dimensions of various heterogeneous path data.

[0063] In some alternative implementations, the process of performing shape-preserving segmented interpolation on the historical typhoon path data can be replaced by using a cubic spline interpolation algorithm to smooth the numerical sequence in order to obtain an isoparametric trajectory that meets other derivative continuity requirements.

[0064] Example 4: Unlike Example 3, this example provides an alternative trajectory resampling implementation with bounded saturation constraints. Addressing the technical problem that single linear curvature weighting may lead to excessive clustering of feature points or even sampling collapse when processing extreme sharp-turn typhoon paths, resulting in loss of spatial resolution for straight sea surface segments, this example replaces the specific calculation method for constructing the curvature adaptive arc length parameter using the discrete curvature and the arc segment length.

[0065] In one possible implementation, constructing a curvature-adaptive arc length parameter using the discrete curvature and the arc segment length includes the following steps:

[0066] Step 401: Extract the preset percentile from the discrete curvature sequence as a half-saturation threshold, and use the half-saturation threshold to perform a nonlinear saturation curvature transformation on the discrete curvature to obtain a saturation curvature value.

[0067] Obtain all discrete curvature value sequences corresponding to a single typhoon path and sort the discrete curvature value sequences in ascending order. The value at the 75th percentile of the discrete curvature value sequence is extracted as the half-saturation threshold. A saturation transformation function based on a nonlinear power-law proportional relationship is established. The discrete curvature values ​​of each node and the half-saturation threshold are input into this saturation transformation function to calculate the saturated curvature values ​​within the interval of 0 to 1. This saturation transformation function ensures that the input discrete curvature tends towards an asymptotic saturation value of 1 after exceeding the half-saturation threshold, thereby constraining the weighting of extreme curvature values.

[0068] Nonlinear saturation curvature numerical ;

[0069] in, The discrete curvature of the node, The extracted half-saturation threshold is p, which is a preset curvature sensitivity index. In this embodiment, the curvature sensitivity index p is set to 2. Those skilled in the art can adaptively adjust this parameter according to the actual requirements for curvature response sensitivity.

[0070] Step 402: Introduce a preset base density guarantee factor. Distribute a fixed proportion of the overall resampling budget according to a uniform arc length. Nonlinearly weight the remaining proportion according to the saturation curvature value on the arc length. Obtain the preset base density guarantee factor, whose value is constrained between 0 and 1, with a typical value of 0.3. Define the difference between the value 1 and the base density guarantee factor as a variable weight ratio. Calculate the composite attenuation coefficient by arithmetically adding the product of the base density guarantee factor, the variable weight ratio, and the saturation curvature value. Multiply the composite attenuation coefficient by the original arc length to obtain the weighted arc length after budget redistribution.

[0071] Weighted arc length ;

[0072] Where λ is the preset basis density guarantee factor, Ψ is the saturation curvature value after nonlinear transformation, and ΔL is the original arc length.

[0073] The above calculation logic divides the overall interpolation resampling budget into two basic parts. A proportion not less than the basis density guarantee factor is fixedly allocated to the uniform arc length to maintain the basic spatial resolution of the global straight line segment interval. The remaining proportion is dynamically tilted and allocated according to the saturation curvature value after nonlinear transformation.

[0074] Step 403: Accumulate the weighted arc segment lengths to obtain the curvature adaptive arc length parameter.

[0075] Following the evolution sequence of the trajectory nodes, the weighted arc segment lengths are summed segment by segment to obtain the cumulative path distance corresponding to each node. The cumulative path distance of all nodes is divided by the sum of the weighted arc segment lengths of the entire trajectory segment to output the normalized curvature adaptive arc length parameter, which is used in the subsequent shape-preserving segmented interpolation process.

[0076] Through the aforementioned bounded saturation mechanism, even if there are extreme sharp-turn nodes in the trajectory with discrete curvature far exceeding the half-saturation threshold, the output value after saturation transformation is constrained to an asymptotic upper limit of 1. Combined with the fixed proportion reserved for uniform arc length by the base density guarantee factor, the ratio of the sampling density multiple of extreme turning points relative to ordinary straight-line segment nodes is constrained to a bounded range not exceeding the reciprocal of the base density guarantee factor, effectively avoiding the sparse sampling point problem in other normal segments caused by the uncontrolled density multiple ratio in the unconstrained linear weighted scheme.

[0077] Example 5: Based on Example 3 or Example 4 above, this example further details the specific process of constructing the multi-channel feature vector. In one possible implementation, such as... Figure 3 As shown, it includes the following steps:

[0078] Step 501: Extract latitude and longitude channel features based on the coordinates of the isoparametric path data;

[0079] Obtain the isoparametric path data generated in the previous resampling step, and separate and extract the longitude and latitude coordinate sequences. Due to shape-preserving segmented interpolation and other techniques, all heterogeneously elongated single typhoon trajectories have been uniformly mapped to a pre-defined number of discrete interpolation nodes, which can be set to 32. Extract the absolute longitude values ​​corresponding to these 32 fixed nodes and combine them to form longitude channel features; simultaneously extract the absolute latitude values ​​corresponding to these 32 fixed nodes and combine them to form latitude channel features. The longitude and latitude channel features characterize the objective static spatial distribution of cyclone events in the Earth's spherical coordinate system, forming the underlying positional reference for subsequent feature spaces.

[0080] Step 502: Extract displacement increment channel features based on the latitude and longitude difference between adjacent nodes in the isoparametric path data;

[0081] Along the parameter evolution sequence, calculate the difference in longitudinal and latitudinal coordinates between forward and backward nodes in the isoparametric path data one by one.

[0082] By subtracting the longitude of the preceding node from the longitude of the subsequent node, the longitudinal displacement increment sequence of the path segment is obtained; similarly, by subtracting the latitude of the preceding node from the latitude of the subsequent node, the latitudinal displacement increment sequence of the path segment is obtained.

[0083] For an isoparametric trajectory consisting of 32 consecutive nodes, the length of the derived displacement increment sequence decreases accordingly to 31. The displacement increment channel characteristics are used to characterize the first derivative dynamic properties of the typhoon trajectory, which can quantitatively reflect the instantaneous evolution direction and absolute translational span of the cyclone system in various local time windows or path segments.

[0084] Step 503: Extract node curvature channel features based on the heading angle changes of adjacent arc segments in the isoparametric path data;

[0085] Based on the latitude and longitude coordinate parameters contained in the isoparametric path data, multiple spherical great circle arcs connecting each adjacent node are calculated.

[0086] For each great circle arc on the spherical surface, the corresponding absolute heading azimuth angle is analyzed based on the starting and ending directions. The heading azimuth angles of two consecutive spherical arcs connected in space are extracted, and the absolute difference between them is calculated. This absolute difference is used as a structural scalar to measure the degree of geometric transformation of the trajectory at that node.

[0087] To maintain the alignment of the length of the entire channel feature sequence with other feature channels, the second-order change values ​​of this term at the beginning and end nodes of the trajectory are uniformly set to 0. The values ​​of this term from all nodes are then concatenated to form a nodal curvature channel feature with the same length of 32. The nodal curvature channel feature explicitly characterizes the second-order dynamic geometric features of the trajectory on the Earth's surface, providing data support for detecting structural abrupt changes in typhoons, such as high-curvature reversals or sudden turns.

[0088] Step 504: Calculate the spherical distance from each node in the isoparametric path data to the preset target sea area, and extract the target distance channel features;

[0089] The central latitude and longitude coordinate system reference point of the preset target sea area is obtained. The semi-versus algorithm or other great circle distance calculation logic is invoked to calculate the great circle distances of all 32 nodes in the isoparametric path data relative to the reference point, without disrupting the path evolution timeline. The sequentially calculated great circle distances are then assembled and stitched together to form the target distance channel feature. As a dimensional sequence with independent physical meaning, the target distance channel feature directly injects the distance target constraint prior into the subsequent unsupervised learning module at the initial stage of feature construction, thereby establishing a numerical correlation between general meteorological spatial data and specific engineering protection objects.

[0090] Step 505: The latitude and longitude channel features, displacement increment channel features, node curvature channel features and target distance channel features are spliced ​​together and standardized dimension by dimension to obtain the multi-channel feature vector.

[0091] Tensor concatenation was used to hard-concatenate the six independent physical feature sequences extracted above along the matrix dimension. The concatenated sequence specifically includes longitude channel features of length 32, latitude channel features of length 32, meridional displacement increment channel features of length 31, latitudinal displacement increment channel features of length 31, nodal curvature channel features of length 32, and target distance channel features of length 32. Since the number of nodes in all individual trajectory data has been fixed by interpolation, the single multi-channel matrix formed by concatenating all the above one-dimensional feature sequences has a constant dimension of 190.

[0092] After tensor concatenation, an input sample set containing all historical path records is obtained. For each of the 190 independent dimensions, the arithmetic mean and overall standard deviation of the entire sample in that specific dimension are calculated. In the process of eliminating interference from different dimensional units, the observed values ​​of a single path in each dimension are subtracted from the corresponding arithmetic mean, and the residual is divided by the corresponding overall standard deviation, performing a dimension-wise standard score data standardization mapping. Through this process, the data distribution of all physical channels is adjusted to a state where the mean approaches 0 and the variance approaches 1, eliminating the risk of feature overload caused by the significant difference in magnitude between displacement differences and absolute distances. The data set output after the standardization operation is determined as the multi-channel feature vector.

[0093] In an alternative implementation, the extracted multi-channel feature tensor can be dynamically reduced or expanded based on the specific requirements of the hydrological scenario and the boundary conditions of the protective engineering analysis. For example, a typhoon movement speed channel corresponding to a node can be added to the tensor splicing sequence, or a near-center air pressure intensity channel corresponding to a node can be added, thereby enriching the mapping space of subsequent neural network representation learning with more meteorological element attributes.

[0094] Example 6: Based on Example 1 above, this example further details the specific execution process of training the target perception depth representation model through the following pre-construction steps. Since the depth representation model is an offline construction module, its training process constitutes a prerequisite foundational support step. In one possible implementation, such as... Figure 4 As shown, it includes the following steps:

[0095] Step 601: Construct a symmetric bottleneck autoencoder that includes an encoder, a bottleneck layer, and a decoder;

[0096] A neural network architecture for feature compression and representation extraction is established. The symmetric bottleneck autoencoder exhibits a funnel-shaped symmetric structure, with the number of neurons in its input and output layers consistent with the physical dimension of the input multi-channel feature vector, for example, set to 190 dimensions. The hidden layer at the center of the network is defined as the bottleneck layer, used to output the dimensionality-reduced features. An encoder module is constructed between the input and bottleneck layers, and a decoder module is constructed between the bottleneck and output layers. The number of neurons in the hidden layers within the encoder is compressed layer by layer according to a preset gradient, for example, using a topological sequence where the dimension is reduced from 190 to 48, 16, and finally shrinks to 6 at the bottleneck layer; the number of neurons in the hidden layers within the decoder adopts a reverse expansion structure. Each hidden layer applies a sigmoid activation function to the forward propagation tensor to perform nonlinear response processing, ensuring that the activation values ​​of each hidden layer fall within the range of 0 to 1, thus achieving mathematical compatibility with the subsequently used KL divergence sparsity regularization term.

[0097] Step 602: Obtain training samples, calculate the spherical distance from the training samples to the preset target sea area, and construct reconstruction weights with an exponential decay distribution based on the spherical distances;

[0098] An offline training sample set containing a large number of historical typhoon path features was obtained. This offline training sample set consists of a set of multi-channel feature vectors obtained after performing the aforementioned trajectory resampling and multi-channel feature construction steps on the historical typhoon path data. For each trajectory node of each isoparametric path in the training samples, the specific physical distance of that node relative to the center coordinates of the preset target sea area was obtained using the Earth's spherical distance calculation rules. To encourage the autoencoder to prioritize improving the reconstruction fidelity of the trajectory around the engineering target sea area during the reconstruction task, a distance attenuation control variable was introduced, and the weight values ​​distributed on each node were calculated.

[0099] The specific calculation formula is as follows: ;

[0100] Among them, w i,j The node-level reconstruction weights for the j-th node on the i-th trajectory are generated, where α is a preset weight magnitude control parameter, exp is the natural exponential operation, and D... i,j Let τ be the spherical distance from the j-th node on the i-th trajectory to the preset target sea area, and let τ be the preset spatial scale control parameter.

[0101] In this computational logic, the spatial scale control parameter and the weight amplitude control parameter determine the physical boundary of the target perception. In practical engineering settings, the weight amplitude control parameter can be assigned a value of 2.0, and the spatial scale control parameter can be assigned a value of 500. This allows pathpoints closer to the target sea area to receive non-linearly increasing high weight constraints, guiding computing power to tilt towards key monitoring areas.

[0102] Step 603: Apply the reconstruction weights to the reconstruction error between the network output features and the input features, and combine the network weight regularization term and the hidden layer activation sparse regularization term to construct the model training loss function;

[0103] Considering that the reconstruction weights are node-level scalars established based on the spatial location of path nodes, while the network's input and output features are high-dimensional tensors that concatenate multiple physical channels, a weight dimension mapping relationship needs to be established between the two. For the specific dimensions of longitude, latitude, node curvature, and target distance channels corresponding to a particular node in the feature tensor, the node-level reconstruction weights of that specific node are directly assigned to the corresponding feature dimensions. For the specific dimensions of the longitudinal and latitudinal displacement increment channels spanning adjacent nodes in the feature tensor, the node-level reconstruction weights of two adjacent nodes are extracted, and their arithmetic mean is calculated. This arithmetic mean is then assigned to the corresponding feature dimensions. After completing the above mapping, dimensional reconstruction weights consistent with the total number of dimensions in the feature vector are generated.

[0104] The dimensional reconstruction weights are multiplied by the squared reconstruction error of the corresponding dimension of the network and then summed. This summation is combined with the L2 norm regularization term for network weights to constrain overfitting of network parameters, and the KL divergence sparsity regularization term to constrain the activation frequency of hidden layer neurons. This results in a model training loss function used to guide the network's backpropagation. The dimensional reconstruction weights are bounded importance weighting coefficients (falling within the interval [1, 1+α]), which act on the reconstruction error after dimensional standardization. They only adjust the reconstruction priority of different nodes and samples, without changing the relative scale between feature channels. Dimensional standardization is a dimensional affine transformation of the data, and importance weighting is an element-wise scaling of the loss term. These two methods act on different objects and are independent of each other, without mathematical cancellation or weakening the network's ability to represent non-distance channels such as curvature. The specific calculation process is expressed by the following formula:

[0105] ;

[0106] in, Here, N is the total number of training samples, D is the total dimension of the multi-channel feature vector, l is the network layer index, j is the hidden neuron index, and Σ represents summation over the corresponding index range. To reconstruct the weights for the d-th dimension corresponding to the i-th trajectory, and These are the output and input feature values ​​of the network in the d-th dimension of the i-th trajectory, respectively, where β is the L2 weight regularization coefficient, and W... (l) Let be the connection weight matrix of the l-th layer network. Let be the square of the Frobenius norm of the matrix, γ be the sparsity regularization coefficient, KL be the relative entropy function, and ρ be the target sparsity parameter. represents the average actual activation probability of the j-th hidden layer neuron. Related hyperparameters can be preset, for example, the L2 weight regularization coefficient can be set to 0.001, the sparse regularization coefficient to 4, and the target sparsity to 0.15.

[0107] Step 604: Use the model training loss function to iteratively fine-tune the symmetric bottleneck autoencoder until the set convergence condition is met, and obtain the pre-trained target perception depth representation model.

[0108] A multi-channel feature tensor is input into the network. During the alternating forward and backward propagation iterations, an adaptive moment estimation optimization algorithm is used to update the network connection weights based on the gradient calculated from the model training loss function. Basic training control parameters are configured, such as configuring the training batch size to 64, setting the learning rate for the initial training phase to 0.001, and setting the learning rate for subsequent fine-tuning phases to 0.002.

[0109] During the iterative process, the relative rate of decline of the loss function between adjacent training epochs is continuously monitored. When it is determined that the relative rate of decline has been below a set tolerance threshold of 0.0001 for several consecutive epochs, a stopping condition is triggered to determine model convergence. The frozen network weights and topology at this point are saved as the output of the pre-trained target-aware deep representation model.

[0110] In an alternative implementation, the network topology of the symmetric bottleneck autoencoder can be replaced by a variational autoencoder, which uses latent variables in the latent space that follow a specific probability distribution to replace discrete deterministic feature mappings; or the topology can be replaced by a sequence autoencoder based on a long short-term memory structure, which uses latent state propagation to mine the time step dependencies of multi-channel features in the evolution direction.

[0111] Example 7: Based on Example 1 above, this example further details the construction of a dual-space comprehensive effectiveness index by fusing the clustering evaluation results of the latent space and the original geographic coordinate space. This example provides a basic dual-space evaluation mechanism that decouples the mathematical feature space and the physical map space. In one possible implementation, such as... Figure 5 and Figure 6 As shown, it includes the following steps:

[0112] Step 701: Evaluate the compactness of the cluster structure and the degree of separation between clusters in the latent space, and calculate the latent space separation term;

[0113] Obtain the path depth embedding vectors output from the previous steps. In the low-dimensional latent space formed by these path depth embedding vectors, a clustering algorithm based on the distance minimization criterion is used for classification. In this embodiment, the clustering algorithm is implemented using K-means clustering; however, those skilled in the art can also use other conventional segmentation clustering algorithms based on distance metrics. Obtain the partitioned cluster structure and call a preset internal evaluation index to numerically measure the cluster structure. Specifically, a conventional intra-cluster divergence to inter-cluster distance ratio algorithm is used to measure the average divergence from data points within each cluster to the cluster center, and the Euclidean distance between different cluster centers is calculated. The sum of the calculated average divergences is compared with the cluster center distance to extract the latent space separation term. This term is used to quantitatively evaluate the intra-cluster compactness and inter-cluster separation of the classified samples from the dimension of the mathematical feature space.

[0114] Step 702: Determine the cluster mean path for each cluster in the original geographic coordinate space;

[0115] After clustering, obtain the cluster assignment labels for each data sample and extract the isoparametric path data corresponding to each sample, retaining the actual latitude and longitude coordinates. In the original geographic coordinate space, for each independently divided cluster, extract all isoparametric path data contained within that cluster. Following a pre-defined fixed node index order, calculate the arithmetic mean of the longitude and latitude values ​​of all nodes at the same node index position. Assemble the sequence of the longitude and latitude averages obtained at all node positions within the same cluster to generate a virtual center evolution trajectory, which is defined as the cluster mean path.

[0116] Step 703: Calculate the spherical distance from each node on each path to the node corresponding to the cluster mean path of its respective cluster;

[0117] The algorithm iterates through all historical typhoon path samples, locates the target cluster to which a single path belongs, and retrieves the cluster mean path generated prior to that target cluster. Along the trajectory evolution sequence, it extracts the coordinates of nodes on each single path and the coordinates of nodes at the same parameter position on the cluster mean path. Using a pre-defined spherical great circle distance algorithm, it calculates the physical surface distance on the Earth's sphere for each pair of spatially corresponding nodes, outputting a sequence of spherical distances distributed along the single path.

[0118] Step 704: Perform a global equal-weighted average on the spherical distances of all paths to obtain the global average spherical distance, and use the global average spherical distance as the geographic coordinate space consistency term;

[0119] The spherical distances of all nodes along a single path are summed, and the sum is divided by the total number of nodes along that path to calculate the average offset distance of that path. The average offset distances of all sample paths under the current clustering scheme are then summed again and divided by the total number of samples involved in the calculation to obtain the global average spherical distance at the full sample level. This value reflects the degree of divergence in the intra-cluster distribution of the clustering scheme on the real physical map, and is directly defined as the geographic coordinate space consistency term. This term is presented as a distance scalar, with the unit set to kilometers; the smaller the value, the more consistent the physical trajectories of various typhoons are on the Earth's surface.

[0120] Step 705: Extract the second-order difference of the geographic coordinate spatial consistency term according to the candidate cluster number sequence to obtain the geographic consistency inflection point gain, perform extreme value normalization on the latent space separation term and the geographic consistency inflection point gain respectively, and add the two normalized indicators with equal weight to obtain the dual-space comprehensive effectiveness index.

[0121] The latent space separation term and the geographic coordinate space consistency term originate from different metric spaces, and their physical dimensions and numerical distribution ranges differ. Therefore, a dimensionless mapping operation is performed before index fusion. The geographic coordinate space consistency term monotonically decreases with the number of candidate clusters. Direct normalization fusion would result in a monotonic bias towards larger clusters. Therefore, before fusion, a second-order difference is extracted based on the candidate cluster number sequence to obtain a non-monotonic geographic consistency inflection point gain (see Example 9 for second-order difference calculation), which is used as the geographic space component participating in the fusion. The maximum and minimum values ​​in each index sequence are extracted, and a linear mapping rule is used to shift and compress the value ranges of the two indicators to the interval between 0 and 1, resulting in a normalized index set. The normalized values ​​corresponding to the latent space separation term and the normalized values ​​corresponding to the geographic consistency inflection point gain are assigned the same weight ratio coefficient, i.e., each is assigned a weight of 0.5. The weighted values ​​of the two indicators are arithmetically added to calculate and output the dual-space comprehensive effectiveness index.

[0122] Dual-space comprehensive effectiveness index ;

[0123] Where W1 is the latent space index weighting coefficient set to 0.5, and V latent The latent spatial separation term is the result of extreme value normalization, W2 is the geospatial indicator weighting coefficient set to 0.5, and V... geo This represents the geographic consistency inflection point gain after extreme value normalization.

[0124] When performing extreme value normalization, for the indicator term where a smaller value in the latent space separation term indicates a better clustering effect, a reverse extreme value normalization operation is used. This involves subtracting the current value from the maximum value of the term's sequence and then dividing by the range. For the indicator term where a larger value in the geographic consistency inflection point gain indicates more sufficient separation of geographic types, a forward extreme value normalization operation is used. After the above processing, both normalized indicators exhibit the same attribute: a larger value indicates a better clustering effect. This facilitates determining the optimal clustering scheme by finding the maximum value of the comprehensive index. This dual-space evaluation mechanism helps avoid the map shape distortion problem associated with relying solely on feature space indicators. Furthermore, by using the inflection point gain instead of the monotonic consistency term in the fusion process, it avoids the shift towards larger clusters caused by directly using the consistency term.

[0125] Example 8: Unlike Example 7, this example provides an alternative implementation for calculating the geographic coordinate spatial consistency term to ensure the clustering verification process has spatial assessment specificity for a particular target sea area. The specific difference lies in the introduction of a spatial weighted allocation mechanism based on the physical distance from the node to the preset target sea area. The remaining calculations of the potential spatial separation term and the final fusion process of the dual-space comprehensive effectiveness index are consistent with Example 7 and can be referred to the foregoing description; they will not be repeated here.

[0126] In one possible implementation, the following steps are included:

[0127] Step 801: Determine the cluster mean path for each cluster in the original geographic coordinate space;

[0128] Obtain the sample cluster allocation results after clustering and classification, and extract the isoparametric path data contained within each cluster. For each independent cluster, traverse according to the sequence index of the isoparametric nodes, and calculate the arithmetic mean of the longitude and latitude parameters of all nodes at the same index position. Combine and assemble the calculated longitude and latitude arithmetic mean sequences to generate a central trajectory sequence representing the overall geographical distribution of the cluster, and define this central trajectory sequence as the cluster mean path.

[0129] Step 802: Calculate the distance from the nodes of each path to the preset target sea area, and generate node-level target proximity weights with a decaying distribution based on the distances;

[0130] The isoparametric paths participating in the clustering evaluation are retrieved, and the absolute coordinates of each isoparametric node on the Earth's surface are obtained. Using the great circle distance algorithm, the physical spatial distance from each node to the center coordinates of the preset target sea area is calculated one by one. Based on the obtained physical spatial distance values, a nonlinear decay mapping relationship is constructed, where the value decreases as the distance increases. According to this mapping relationship, a corresponding weight value is assigned to each node, generating the node-level target proximity weights. By assigning higher weights to nodes around the target sea area, the clustering verification algorithm is forced to impose a more stringent numerical penalty on the trajectory divergence near the engineering protection area.

[0131] Step 803: Based on the pre-set weight amplitude control parameters and spatial scale control parameters, an exponential function is used to construct a spatial decay mechanism;

[0132] The pre-configured hyperparameters of the validation layer are obtained, specifically including weight magnitude control parameters and spatial scale control parameters. A nonlinear spatial decay mechanism is established using a negative exponential operation rule with the natural constant as the base. The weight magnitude control parameter determines the maximum gain upper limit that this spatial decay mechanism can output at the origin, avoiding an unbounded weight amplification effect on nodes at the center. The spatial scale control parameter, as the denominator, controls the decay gradient range of the exponential function. In the specific parameter configuration of the validation layer, the weight magnitude control parameter can be set to 1.5, and the spatial scale control parameter can be set to 800. Compared with the distance weight parameters of the representation learning layer, this parameter combination uses a larger spatial scale control parameter and a smaller weight magnitude control parameter to cover a wider near-shore area in the validation layer.

[0133] Step 804: Substitute the distance from each node of the path to the preset target sea area into the spatial attenuation mechanism, so that the nodes closer to the preset target sea area receive a non-linearly enhanced weight value, and the highest weight is subject to the bounded limitation of the weight amplitude control parameter, thus obtaining the node-level target proximity weight.

[0134] The distances from each node to the preset target sea area, calculated in the previous step, are used as independent variables and input into the formula of the spatial attenuation mechanism for numerical calculation. Nodes closer to the center of the target sea area receive larger exponential output values, which are then converted into non-linearly enhanced weight values. When a node is exactly located at the center of the target sea area (i.e., the distance is 0), the obtained weight value reaches the theoretical upper limit. This theoretical upper limit is equal to the arithmetic sum of the base value 1 and the weight amplitude control parameter, reflecting the bounded constraint characteristic of the weight allocation system. This calculation process is expressed by the following formula:

[0135] ;

[0136] Where, ω i,j For the node-level target proximity weights corresponding to the j-th node on the i-th trajectory, α g The pre-defined weighting magnitude control parameter, exp represents the natural exponent calculation process, and D... i,j τ is the distance from the j-th node on the i-th trajectory to the preset target sea area. g These are pre-set spatial scale control parameters.

[0137] Step 805: Using the node-level target proximity weight, perform a spatial weighted summation of the spherical distances from nodes on each path to the corresponding nodes of the cluster mean path of their respective clusters, and use the weighted calculation result as the geographic coordinate spatial consistency term.

[0138] Obtain the node coordinates on each sample path, retrieve the node coordinates of the cluster mean path of its respective cluster, and calculate the great circle distance between each pair of nodes at the same sequence position. Multiply the calculated great circle distance by the corresponding node-level target proximity weight to obtain a weighted distance product. Sum all weighted distance products on a single path, and simultaneously sum all node-level target proximity weights on that path, calculating the ratio of the sum of weighted distance products to the sum of weights to obtain the weighted deviation value at the single path level. Then, calculate the global arithmetic mean of the weighted deviation values ​​for all paths in the entire sample set, outputting a comprehensive spatial weighted evaluation value. Use this weighted calculation result as the geographic coordinate spatial consistency term reflecting the physical map overlap. This operation process is expressed by the following formula:

[0139] ;

[0140] Among them, G val The weighted calculation result obtained as the geographic coordinate spatial consistency term is Σ. i This represents the outermost summation operation performed on all paths within the entire sample range, Σ. j This represents the inner-level summation operation on all nodes along a specific path, ω. i,j H represents the node-level target proximity weight corresponding to the j-th node on the i-th trajectory. i,j The great circle distance is the spherical distance from the j-th node on the i-th trajectory to the node corresponding to the cluster mean path of the i-th cluster, and N is the total number of paths in the full sample participating in the evaluation.

[0141] This target-aware geographic compactness measurement method constitutes a feature verification mechanism for engineering applications. If the physical spatial orientations of clustered trajectories diverge and overlap in path segments close to the preset target sea area, the calculation result of the consistency term will exhibit a non-linear numerical increase due to the intervention of weighting values, thereby triggering numerical penalties and affecting the subsequent inflection point gain output. This guides the clustering search process to eliminate non-optimal solutions with overlapping far-sea orientations but chaotic near-sea orientations.

[0142] Example 9: Based on Example 7 or Example 8 above, this example further details the specific steps of determining the number of classification clusters according to the dual-space comprehensive effectiveness index. This example provides an adaptive cluster number determination mechanism based on discrete second-order difference and random permutation baseline correction. In one possible implementation, it includes the following steps:

[0143] Step 901: For different numbers of candidate clusters, calculate the corresponding dual-space comprehensive effectiveness index, extract the first-order descent difference of the geographic coordinate space consistency term, extract the discrete increment of the first-order descent difference to obtain the second-order difference, and use the second-order difference as the geographic consistency inflection point gain.

[0144] The geographic consistency inflection point gain after extreme value normalization is added to the latent space separation term after extreme value normalization to determine the candidate cluster number corresponding to the maximum sum, which is taken as the final number of classification clusters. A traversal evaluation range for the candidate cluster number is set. The traversal evaluation range can be set to a discrete integer sequence including 4, 5, 6, and 7. For each set candidate cluster number, the corresponding geographic coordinate space consistency term value is obtained. The first-order descent difference is calculated by subtracting the consistency term value of the current candidate cluster number from the consistency term value of the preceding cluster number corresponding to the current candidate cluster number. Subsequently, the discrete increment used to characterize the descent acceleration of the index is extracted by subtracting the subsequent first-order descent difference derived from the current candidate cluster number and the subsequent cluster number from the first-order descent difference. The calculated discrete increment is defined as the second-order difference and used as the geographic consistency inflection point gain. This process, by extracting the differential characteristics of the distance index, finds the natural boundary where the physical compactness decreases in a stepwise manner, avoiding the monotonic bias problem of directly using the distance index that tends towards small cluster numbers. The traversal range of the candidate cluster number can be preset according to the data scale, for example, set to an integer between 2 and 20, or set to an integer not exceeding the square root of the total number of samples; the effective calculation range of the geographic consistency inflection point gain is the interval after removing the first and last endpoints. When the range of a certain indicator sequence is lower than a preset threshold, unit mapping, zeroing, or setting the normalization result to a preset constant value is used to avoid division by zero in extreme value normalization. The specific calculation process is expressed by the following formula:

[0145] ;

[0146] Among them, E gain The calculated geographic consistency inflection point gain, where K is the number of candidate clusters currently evaluated, and C... geo (K-1) represents the geographic coordinate spatial consistency term under the preceding cluster number, C geo (K) represents the geographic coordinate spatial consistency term under the current cluster number, C geo (K+1) represents the spatial consistency term of geographic coordinates under subsequent cluster numbers.

[0147] Step 902: Add the geographic consistency inflection point gain to the latent space separation term after extreme value normalization, determine the number of candidate clusters corresponding to the maximum sum, and use it as the final number of classification clusters.

[0148] Obtain the geographic consistency inflection point gain sequence calculated for all candidate cluster numbers, and perform a positive extreme value normalization mapping operation on this sequence. Retrieve the previously constructed latent space separation term and perform an extreme value normalization mapping operation. Add the two normalized values ​​corresponding to the same candidate cluster number dimensionlessly, and output the candidate's comprehensive score value. By comparing the comprehensive score values ​​under different candidate cluster numbers, extract the set number corresponding to the sum reaching the global maximum value, and use it as the final number of classification clusters to stop the division. Based on the selected final number of classification clusters, extract the mean representative trajectory and the prototype sample closest to the cluster center for each cluster, and complete the construction of the type library.

[0149] In a specific verification scenario, the independent latent space index reaches a single-dimensional extreme value when the number of candidates is 5, while the independent geospatial inflection point gain reaches a single-dimensional extreme value when the number of candidates is 7. This step, through a compromise mechanism of adding the values ​​of the two spaces, ensures that the sum reaches its global maximum value when the number of candidates is 6, achieving a complementarity between separation and consistency, and avoiding the local convergence problem caused by a single evaluation space. In actual testing on historical typhoon samples in the Northwest Pacific, under the recommended number of clusters determined by the dual-space comprehensive effectiveness index of this invention, the consistency of intra-class path orientation of the clustering results on the real geographic map is improved compared to the scheme that relies solely on a single latent space index to determine the number of clusters, and the spatial crossover phenomenon between the mean paths of each type of cluster is significantly reduced. Those skilled in the art will understand that specific performance indicators may vary depending on the data set coverage period, the location of the target sea area, and the configuration of network hyperparameters.

[0150] Based on the above processing flow, in order to eliminate unstructured background interference caused by the monotonically decreasing intra-cluster distance due to the increase in the number of clusters, this embodiment also provides an alternative enhancement scheme based on a reference baseline. In an optional implementation, before calculating the first-order descent difference of the geographic coordinate spatial consistency term, a baseline correction step is further included, specifically comprising the following steps:

[0151] Step 903: Randomly permutate the isoparametric path samples corresponding to the historical typhoon path data at each node location dimension to generate an unstructured reference set with scrambled geographical structure.

[0152] The coordinate matrix data is constructed by resampling all historical typhoon path samples. For each independent isoparametric node position in this coordinate matrix, the longitude and latitude data columns are independently and randomly permuted using a random number generation function. This permutation operation preserves the boundary extreme value distribution characteristics of each dimension, but disrupts the inherent hierarchical correlation of the real samples in spatial evolution. The permuted data is packaged to generate an unstructured reference set with scrambled geographic structure. To reduce the random error introduced by a single random operation, a loop execution parameter can be set to repeat the above permutation operation 20 times, generating multiple parallel unstructured reference sets.

[0153] Step 904: Calculate the reference consistency term based on the unstructured reference set, and calculate the arithmetic mean of the reference consistency terms generated by multiple random permutations to obtain the mean of the reference consistency term.

[0154] The generated unstructured reference set is input into the same clustering evaluation calculation channel as the real data. For a given range of candidate clusters, the global cluster mean path and deviation distance corresponding to these random samples are calculated, outputting the distance evaluation index corresponding to the random unstructured state. The arithmetic mean of the evaluation indexes corresponding to the repeatedly generated unstructured reference set is calculated, and this arithmetic mean is defined as the reference consistency term.

[0155] To ensure that the reference baseline and the real samples are measured using the same criteria, the unstructured reference set undergoes the same processing procedure as the real data: First, the permuted unstructured reference set is input into the pre-trained target-aware depth representation model to obtain its path depth embedding vector. Then, it is clustered in the latent space with the same number of candidate clusters to obtain virtual cluster assignment labels for each reference sample. Next, based on these virtual cluster assignment labels, the deviation distance of each reference sample from the cluster mean path to its respective cluster is calculated in the original geographic coordinate space to obtain the reference consistency term. Thus, the reference consistency term and the geographic coordinate space consistency term of the real samples are under the same "embedding-clustering-geographic metric" framework, making them comparable.

[0156] Step 905: Calculate the difference between the mean of the geographic coordinate spatial consistency term and the mean of the reference consistency term for the real sample, and perform subsequent first-order descent difference and second-order difference calculations based on the difference to obtain the corrected geographic consistency inflection point gain.

[0157] The geographic coordinate spatial consistency term calculated based on real typhoon path data is extracted. This term's value is then subtracted from the mean of the reference consistency term obtained in the previous step to calculate the net consistency difference after removing the random spatial decreasing trend. This net consistency difference is substituted into the aforementioned first-order descent difference and second-order discrete increment operation logic to replace the original distance parameters in the calculation derivation. After difference stripping, the corrected geographic consistency inflection point gain is output. This correction process filters out the fictitious compactness enhancement variable caused by the algorithm's mechanical partitioning, ensuring that the feedback gain signal corresponds to the accelerated descent effect caused by the data geographic boundary.

[0158] Net consistency difference after removing random decreasing trend ;

[0159] in, To remove the net consistency difference after removing the random decreasing trend, The geographic coordinate spatial consistency term is calculated from real samples. This is the mean of the reference consistency term calculated based on the unstructured reference set.

[0160] Example 10: Based on Example 1 above, this example further details the application of quantitatively evaluating the screening index of the path unfavorability of each typhoon path type to the preset target sea area and identifying the unfavorable path type.

[0161] To facilitate the commercialization of algorithmic models for practical engineering flood and tide risk assessment, such as... Figure 7 and Figure 8 As shown, this embodiment includes the following steps:

[0162] Step 1001: Calculate the spherical distance from the path nodes of various typhoons in the typical typhoon path type library to the preset target sea area, and select typhoons that fall within the preset influence radius as a subset of influence paths;

[0163] Obtain the typical typhoon path type library output from the previous steps, and retrieve the pre-defined clusters within the library. For each cluster, obtain the physical coordinates of each node on the Earth's surface along its evolution timeline for all historical typhoon samples. Configure the preset target sea area center coordinates and preset influence radius. In specific flood control and tide engineering applications, the Yangtze River estuary can be set as the preset target sea area, with corresponding center coordinate parameters configured and the preset influence radius set to 500 kilometers. Using the spherical distance calculation rule, calculate the physical distance from all discrete nodes of each typhoon path to the preset center coordinates. Determine if there are any nodes in a single path whose distance value is less than or equal to the preset influence radius. If so, the typhoon is considered to have a substantial impact on the preset target sea area, and it is extracted and classified into the corresponding category of the subset of impact paths affecting specific sea areas.

[0164] Step 1002: Calculate the proportion of the subset of the impact paths in the total number of typhoons of the corresponding category to obtain the conditional impact probability, and extract the maximum wind speed information of the typhoons in the subset of the impact paths to calculate the normalized intensity mean.

[0165] For each cluster, the number of typhoon samples included in the impact path subset is statistically calculated. The conditional impact probability is calculated by dividing the number of samples in this subset by the total number of typhoon samples contained in the cluster. This conditional impact probability serves as a frequency indicator, quantifying the frequency tendency of a typhoon with a certain path pattern striking a predetermined target sea area. Further, the maximum near-center wind speed values ​​from the life history of each typhoon within the impact path subset are extracted from the original typhoon observation records. Upper and lower bound values ​​for the maximum wind speed are obtained across the entire sample range, and the wind speed values ​​within the subset are normalized using a maximum-minimum interval mapping rule. The arithmetic mean of the normalized wind speeds of all typhoons within the subset is calculated, and the normalized intensity mean is output. This mean serves as a meteorological destructive force indicator, quantifying the potential intensity level of extreme disasters caused by a certain path type.

[0166] Step 1003: Multiply the conditional influence probability by the normalized intensity mean to obtain the path unfavorability index (screening index). Based on the magnitude of the path unfavorability index, typhoon path types are ranked to identify unfavorable path types facing the preset target sea area. Specifically, typhoon path types ranked within a preset range or with a path unfavorability screening index not lower than a preset threshold are identified as unfavorable path types.

[0167] Obtain the conditional impact probability and normalized mean intensity for each typhoon path type. Multiply the conditional impact probability (representing frequency tendency) by the normalized mean intensity (representing disaster intensity) arithmetically to derive a comprehensive evaluation parameter, which is defined as the path unfavorability index I. TAI .

[0168] I TAI = P impact · V norm ;

[0169] Among them, P impact V represents the conditional influence probability of the corresponding category. norm This represents the normalized mean intensity of typhoons within the subset of impact paths for this category.

[0170] The path unfavorability index sequences corresponding to all clusters are compiled and sorted in descending order of numerical value. Based on the sorting results, the path types with the highest index values ​​in the sequence are extracted. In a specific verification scenario, the sorting results indicate that a certain near-shore northward turning trajectory has the highest path unfavorability index, while a certain northward turning trajectory in the Northwest Pacific has the second highest index. Based on this, these two trajectory types are identified and marked as the candidate unfavorable path types with the highest and second highest unfavorability indices for specific storm surge barrier projects. This assessment process transforms the trajectory classification from a purely mathematical perspective into a risk grading reference conclusion with engineering orientation. The path unfavorability index is a proxy screening indicator for engineering-oriented unfavorability, used to initially screen potentially unfavorable candidate path types from a large number of historical typhoons; the actual disaster severity of a typhoon to the target sea area is also related to factors such as path direction, maximum wind speed radius, movement speed, incident angle, astronomical tide level, and wind field structure. The final unfavorability level is based on the results of subsequent storm surge hydrodynamic numerical simulations. Furthermore, the engineering adverseness is a relative ranking index characterizing the potential impact of each path type on the target sea area, rather than a direct prediction of the actual storm surge disaster intensity.

[0171] Based on this, in order to realize the engineering application of the assessment results in the water conservancy project protection system, after identifying the unfavorable path type facing the preset target sea area, this embodiment also includes an operational application step, which includes the following steps:

[0172] Step 1004: Use the typical typhoon path corresponding to the identified unfavorable path type as the input meteorological boundary condition for storm surge numerical simulation, perform storm surge numerical simulation, and obtain simulation output results.

[0173] Extract the cluster mean path or the closest real-world typhoon sample data to the cluster center corresponding to the unfavorable path type in the cluster space. Perform spatiotemporal interpolation and format conversion on the extracted trajectory data, configuring it as a driving field file for a two-dimensional or three-dimensional storm surge ocean hydrodynamic numerical model. This trajectory data, along with the corresponding empirical formulas for wind pressure relationships, constitutes the input meteorological boundary conditions for driving the model calculations, used to reproduce the meteorological forcing force distribution under this unfavorable path in a virtual grid space.

[0174] The typical typhoon paths corresponding to the identified unfavorable path types are transformed into input meteorological boundary conditions for storm surge numerical simulation. Based on the output results of the storm surge numerical simulation, the design water level or disaster risk level is estimated for coastal flood control and storm surge projects located in the preset target sea area.

[0175] Step 1005: Based on the simulation output results of the storm surge numerical simulation, calculate the design water level and classify the disaster risk for the coastal flood control and storm surge engineering located in the preset target sea area.

[0176] Initiate iterative calculations using the numerical simulation program to monitor and extract the time series of hydrological variables within the geographic grid area where the target project is located. Obtain the extreme values ​​of storm surge rise and the highest water level parameters resulting from the superposition of the astronomical tide and storm surge. Using the above hydrodynamic calculation results, combined with the extreme value statistical distribution model, derive the design water level elevations that conform to engineering specifications for physical coastal flood control projects such as tide gates and seawalls within the fortification area. Simultaneously, use the inundation depth matrix to perform regional delineation of disaster risk levels.

[0177] In an alternative implementation, when performing the disaster intensity assessment and adverse metric calculation, for application scenarios with complete data from some hydrological stations, the maximum historically measured storm surge increase or the energy dissipation index per unit time can be used to replace the near-center maximum wind speed to calculate the normalized intensity mean, thereby providing a disaster intensity measurement scheme that is more in line with the actual ocean hydrodynamic response.

[0178] In some embodiments, the business application steps may also be:

[0179] The typical typhoon paths corresponding to the identified unfavorable path types are used to generate the input meteorological boundary driving field for the storm surge hydrodynamic numerical model after spatiotemporal interpolation and format conversion, and the storm surge hydrodynamic numerical simulation is then performed.

[0180] Based on the hydrodynamic calculation results of storm surge increase and superimposed water level output by the storm surge hydrodynamic numerical simulation, the design water level of the coastal flood control and storm surge physical project located in the preset target sea area is estimated, and disaster risk classification results are generated accordingly.

[0181] It should be noted that the various specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the present invention will not describe the various possible combinations separately.

Claims

1. An unsupervised clustering and induction method for typhoon paths based on a target-aware deep autoencoder and a dual spatial effectiveness index, characterized in that, include: Acquire historical typhoon path data and location information of preset target sea areas; The historical typhoon path data is resampled, and combined with the location information of the preset target sea area, a multi-channel feature vector is constructed. The multi-channel feature vector is input into a pre-trained target perception depth representation model, and the path depth embedding vector corresponding to the historical typhoon path data is output. Unsupervised clustering is performed in the latent space based on the path deep embedding vector. Clustering and evaluation are performed for different numbers of candidate clusters. The clustering evaluation results of the latent space and the original geographic coordinate space are merged to construct a dual-space comprehensive effectiveness index. The number of classification clusters is determined according to the dual-space comprehensive effectiveness index, and a typical typhoon path type library is output. Based on the typical typhoon path type library, calculate the path unfavorability screening index of each typhoon path type relative to the preset target sea area, and output the unfavorable path type and its corresponding representative typhoon path sample. The construction of a dual-space comprehensive effectiveness index by integrating the clustering evaluation results of the latent space and the original geographic coordinate space includes: evaluating the tightness of the cluster structure and the degree of separation between clusters in the latent space, and calculating the latent space separation term; evaluating the physical map form of the clustering results in the original geographic coordinate space, and calculating the geographic coordinate space consistency term; extracting the second-order difference of the geographic coordinate space consistency term as a function of the number of clusters for different candidate cluster numbers to obtain the geographic consistency inflection point gain; performing extreme value normalization on the latent space separation term and the geographic consistency inflection point gain respectively, and adding the two normalized indicators with equal weights to obtain the dual-space comprehensive effectiveness index; The pre-construction of the target perception depth representation model includes: constructing a symmetric bottleneck autoencoder containing an encoder, a bottleneck layer, and a decoder; acquiring training samples, calculating the spherical distance from the training samples to the preset target sea area, and constructing reconstruction weights with an exponentially decaying distribution based on the spherical distance; applying the reconstruction weights to the reconstruction error between the network output features and the input features, and constructing a model training loss function by combining the network weight regularization term and the hidden layer activation sparse regularization term; and iteratively fine-tuning the symmetric bottleneck autoencoder using the model training loss function until the set convergence condition is met, thereby obtaining the pre-trained target perception depth representation model. Constructing the dual-space comprehensive effectiveness index includes: extracting the second-order difference of the geographic coordinate spatial consistency term according to the candidate cluster number sequence to obtain the geographic consistency inflection point gain; performing extreme value normalization on the latent space separation degree term and the geographic consistency inflection point gain respectively; and adding the two normalized indicators with equal weights to obtain the dual-space comprehensive effectiveness index.

2. The method according to claim 1, characterized in that, The resampling of the historical typhoon path data includes: The arc length of adjacent nodes in the historical typhoon path data is calculated based on the spherical distance, and the discrete curvature of the corresponding node is extracted based on the difference in azimuth angle of adjacent arc segments. The discrete curvature and the arc segment length are used to construct the curvature adaptive arc length parameter; Using the curvature adaptive arc length parameter as the interpolation benchmark, the historical typhoon path data is subjected to shape-preserving segmented interpolation to obtain isoparametric path data with a fixed number of nodes.

3. The method according to claim 2, characterized in that, The construction of the curvature adaptive arc length parameter using the discrete curvature and the arc segment length specifically includes: The discrete curvature is normalized. The arc length is linearly weighted using the normalized discrete curvature, and the cumulative arc length parameter is obtained by accumulating it segment by segment, which is used as the curvature adaptive arc length parameter.

4. The method according to claim 2, characterized in that, The construction of the curvature adaptive arc length parameter using the discrete curvature and the arc segment length specifically includes: A preset percentile is extracted from the discrete curvature sequence as a half-saturation threshold, and the discrete curvature is subjected to a nonlinear saturation curvature transformation using the half-saturation threshold. A preset basis density guarantee factor is introduced, and a fixed proportion in the overall resampling budget is allocated according to a uniform arc length. The remaining proportion is nonlinearly weighted according to the value after the saturation curvature transformation. The weighted arc lengths are accumulated to obtain the curvature adaptive arc length parameter.

5. The method according to claim 2, characterized in that, The construction of the multi-channel feature vector specifically includes: Based on the coordinates of the isoparametric path data, latitude and longitude channel features are extracted; Displacement increment channel features are extracted based on the latitude and longitude differences between adjacent nodes in the isoparametric path data. Based on the heading angle variation of adjacent arc segments in the isoparametric path data, the curvature channel features of the nodes are extracted; Calculate the spherical distance from each node in the isoparametric path data to the preset target sea area, and extract the target distance channel features; The latitude and longitude channel features, displacement increment channel features, node curvature channel features, and target distance channel features are concatenated and standardized dimension by dimension to obtain the multi-channel feature vector.

6. The method according to claim 1, characterized in that, The calculation of the geographic coordinate spatial consistency term specifically includes: Determine the cluster mean path for each cluster in the original geographic coordinate space; Calculate the spherical distance from each node on each path to the node corresponding to the cluster mean path of its respective cluster; A global average spherical distance is obtained by performing a global equal-weighted average on the spherical distances of all paths, and this global average spherical distance is used as the geographic coordinate space consistency term.

7. The method according to claim 1, characterized in that, The calculation of the geographic coordinate spatial consistency term specifically includes: Determine the cluster mean path for each cluster in the original geographic coordinate space; Calculate the distance from each node of each path to the preset target sea area, and generate node-level target proximity weights with a decaying distribution based on the distance; Using the node-level target proximity weights, the spherical distances from nodes on each path to the corresponding nodes of the cluster mean path of their respective clusters are spatially weighted and summed, and the weighted calculation results are used as the geographic coordinate spatial consistency term.

8. The method according to claim 1, characterized in that, Calculate the path unfavorability screening index for each typhoon path type relative to the preset target sea area, and output the unfavorable path type, specifically including: Calculate the spherical distance from the path nodes of various typhoons in the typical typhoon path type library to the preset target sea area, and select typhoons that fall within the preset influence radius as a subset of influence paths; The proportion of the affected path subset in the total number of typhoons of the corresponding category is statistically analyzed to obtain the conditional impact probability, and the maximum wind speed information of typhoons within the affected path subset is extracted to calculate the normalized intensity mean. The conditional influence probability is multiplied by the normalized intensity mean to obtain the path unfavorability screening index. Based on the value of the path unfavorability index, each typhoon path type is sorted. Typhoon path types that are ranked within a preset range or whose path unfavorability screening index is not lower than a preset threshold are determined as unfavorable path types.

9. The method according to claim 7, characterized in that, The generation of node-level target proximity weights with a decaying distribution based on the distance specifically includes: Based on pre-set weight amplitude control parameters and spatial scale control parameters, an exponential function is used to construct a spatial decay mechanism; The distance from each node of each path to the preset target sea area is substituted into the spatial attenuation mechanism, so that the nodes closer to the preset target sea area receive a non-linearly enhanced weight value, and the highest weight is subject to the bounded limitation of the weight amplitude control parameter.

10. The method according to claim 8, characterized in that, After identifying the unfavorable path types for the preset target sea area, the process also includes an operational application step: For each type of candidate unfavorable path, the historical typhoon sample with the smallest distance from the corresponding cluster center and complete intensity record is selected from the included historical typhoon samples as the input meteorological boundary condition for storm surge numerical simulation. Based on the output of the storm surge numerical simulation, the design water level and disaster risk classification are carried out for the coastal flood control and storm surge projects located in the preset target sea area.

Images