Weight optimization-based multi-algorithm fused cyclone similarity identification method and storm surge forecasting method

Abstract

The invention discloses a multi-algorithm fusion cyclone similarity identification method and a storm surge forecasting method based on weight optimization, relates to the field of cyclone similarity identification, and establishes a similarity calculation function by constructing a multi-dimensional time sequence containing a plurality of key factors and combining a weighted dynamic time warping algorithm. In this way, multi-dimensional comprehensive characterization of the cyclonic space-time evolution process is achieved, and the identification deviation caused by the fact that a traditional method only depends on a single feature is overcome. On this basis, a genetic algorithm, a particle swarm optimization algorithm and a differential evolution algorithm are respectively adopted to carry out iterative optimization on the weight vectors, minimization of a deviation coefficient of a candidate similar cyclone set is taken as a target, and a global optimal weight vector corresponding to each algorithm is searched; and based on a preset fusion strategy, determining a target optimal weight vector from the plurality of globally optimal solutions. According to the closed-loop recognition process, multi-dimensional feature modeling and scientific weight optimization are effectively fused, and objectivity and accuracy of cyclone similarity recognition are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of cyclone similarity recognition, and more particularly to a multi-algorithm fusion method for cyclone similarity recognition and storm surge forecasting based on weight optimization. Background Technology

[0002] Located on the western coast of the Northwest Pacific Ocean, my country is one of the regions with the most frequent tropical cyclone activity globally. With intensifying climate change, the frequency and intensity of tropical and extratropical cyclones are increasing, leading to storm surge disasters that pose a serious threat to coastal areas. Analogical analysis methods based on the evolutionary characteristics of historically similar cyclones are valuable for operational early warning systems because they can reflect the overall behavioral patterns of cyclone systems. However, existing cyclone similarity identification technologies still have significant shortcomings in feature representation and weight determination, limiting their application in high-precision disaster simulation.

[0003] First, existing methods suffer from a lack of simplistic feature dimensions in their similarity criteria. Traditional methods often rely on single indicators such as cyclone path trajectory or center intensity for matching, failing to fully utilize the multidimensional dynamic information contained in the path points, such as center location, movement speed, maximum wind speed at the center, and minimum air pressure at the center. Such single-dimensional or low-dimensional matching strategies cannot comprehensively characterize the spatiotemporal evolution of cyclones, resulting in insufficient ability to identify complex path transitions or abrupt changes in intensity, and making the similarity cyclone screening results prone to bias.

[0004] Second, the determination of the weights of key factors relies on subjective experience and lacks a scientific optimization mechanism. Current practices commonly use expert weighting or equal-weighted averaging to set the weights of each feature parameter, resulting in fixed weight values ​​and a lack of data-driven support. Since the contribution of each factor to similarity judgment varies across different cyclone events, manually set weights are unlikely to approximate the optimal discriminant function. Furthermore, existing methods often use Euclidean distance or simple dynamic time warping (DTW) for similarity calculation, failing to consider the relative importance differences of each factor in time series matching, thus affecting the accuracy and robustness of similarity ranking.

[0005] Against this backdrop, developing a cyclone similarity identification method that can integrate multi-dimensional path features and automatically optimize weights based on historical matching results is of great significance for improving the objectivity and accuracy of similar cyclone screening. Furthermore, by using the complete path information of the identified optimal similar cyclone (including wind field, pressure field, and measured open-boundary tide level at subsequent path points) as driving and boundary conditions, and inputting it into a refined storm surge numerical model, rapid analog forecasting of potential storm surge processes triggered by the current cyclone can be achieved. This compensates for the real-time limitations of traditional physical models and provides efficient and reliable disaster early warning support for coastal areas. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention proposes a multi-algorithm fusion cyclone similarity recognition method based on weight optimization, comprising: Historical cyclone data is acquired, which includes path records of multiple historical cyclones. Each path record consists of multiple path points distributed at equal time intervals, and each path point contains multiple key factors. The values ​​of each key factor in all path records are globally standardized. For each historical cyclone path record, a multidimensional time series characterizing the spatiotemporal evolution of the cyclone is constructed based on the standardized key factors corresponding to the continuous path points it contains. A weight vector containing the weight coefficients of each key factor is constructed as the parameter structure for weight optimization. Based on this weight vector and the multidimensional time series of cyclones, a similarity calculation function using a weighted dynamic time warping algorithm is constructed to calculate the similarity value between two cyclones. Genetic algorithm, particle swarm optimization algorithm, and differential evolution algorithm are used respectively to initialize the population based on the parameter structure, and the weight vector in the population is iteratively optimized. During the iteration process, each algorithm constructs a candidate similar cyclone set based on the current weight vector and the similarity calculation function, and calculates the deviation coefficient corresponding to the set. The deviation coefficient is used as the optimization objective, and the optimal solution is searched by minimizing the deviation coefficient. Finally, the global optimal weight vector corresponding to each algorithm is obtained. Based on the preset fusion strategy, the target optimal weight vector is determined from it. Substitute the target optimal weight vector into the similarity calculation function, recalculate the similarity value between the current measured cyclone and each historical cyclone, generate a similarity ranking list, and select the historical cyclone with the highest similarity value as the optimal similar cyclone.

[0007] Furthermore, the key factors include: the longitude of the cyclone's center, the latitude of the center, the maximum wind speed at the center, the minimum air pressure at the center, and the horizontal movement speed.

[0008] Furthermore, a genetic algorithm is used to initialize the population based on the aforementioned parameter structure, and the weight vectors in the population are iteratively optimized, specifically including: Population initialization: Multiple individuals with weight vectors are randomly generated based on the parameter structure to form the initial population; Calculate fitness value: For each individual with a weight vector in the population, substitute it into the similarity calculation function, and calculate the similarity value between it and all historical cyclones based on the current measured cyclone, forming a similarity ranking list; select the first preset number of historical cyclones according to the ranking list to form a candidate similar cyclone set, calculate the deviation coefficient corresponding to the set, and take the reciprocal of the deviation coefficient as the fitness value of the individual. Parent selection: Sort individuals in the population in descending order according to their fitness values. The higher the fitness value, the better the consistency of the candidate similar cyclone set corresponding to the weight vector. Use roulette wheel selection to select several individuals as parents. Perform crossover operation: randomly pair up parent individuals, perform arithmetic crossover on the weight vectors of each pair of parent individuals to generate two new offspring individuals; wherein, each offspring individual is obtained by using a random number between 0 and 1 as the combination coefficient to perform a linear weighted combination of the two parent weight vectors. Perform mutation operation: Apply random perturbation to the weight components of a preset proportion of offspring individuals, with the perturbation amplitude following a Gaussian distribution with a mean of zero; after perturbation, perform non-negativity correction and normalization on the weight vector to ensure that each component is non-negative and the sum is 1. Iterative update: The newly generated offspring individuals are used as the next generation population, and the calculation of fitness values, selection of parent individuals, crossover and mutation operations are repeatedly performed; in each generation iteration, the global optimal weight vector is updated and recorded. The global optimal weight vector is the weight vector with the highest fitness among all individuals in the current generation; when the deviation coefficient corresponding to the optimal weight vector changes less than a preset threshold within several consecutive generations, or when the maximum number of iterations is reached, the optimization is terminated, and the global optimal weight vector is output as the optimal weight vector obtained by the genetic algorithm.

[0009] Furthermore, a particle swarm optimization algorithm is used to initialize the population based on the aforementioned parameter structure, and the weight vectors in the population are iteratively optimized, specifically including: Initialize the particle swarm: Based on the parameter structure, randomly generate multiple particles to form an initial swarm. Each particle corresponds to a weight vector, and a velocity vector is initialized for each particle. Calculate the fitness value: For each particle's corresponding weight vector, substitute it into the similarity calculation function, and calculate its similarity value with all historical cyclones based on the current measured cyclone, forming a similarity ranking list; select the first preset number of historical cyclones according to the ranking list to form a candidate similar cyclone set, calculate the deviation coefficient corresponding to the set, and take the reciprocal of the deviation coefficient as the fitness value of the particle. Update individual and global optimal positions: For each particle, record the weight vector with the highest fitness value obtained in all iterations since initialization as the individual's historical optimal position; at the same time, maintain a global optimal position in the entire population, which corresponds to the weight vector with the highest fitness achieved by all particles since initialization. Update velocity and position: For each particle, calculate the updated velocity vector based on its current velocity, the deviation of its current weight vector from its individual historical best position, and the deviation from its global best position; update its corresponding weight vector based on this velocity vector. Constraint processing: The updated weight vector is corrected for nonnegativity and normalized to ensure that each component is nonnegative and the sum is 1; Iterative update: Repeatedly perform fitness value calculation, optimal position update, velocity and position update and constraint processing operations; in each iteration, update and record the global optimal weight vector; when the deviation coefficient corresponding to the optimal weight vector changes less than a preset threshold within several consecutive generations, or when the maximum number of iterations is reached, the optimization is terminated, and the global optimal weight vector is output as the optimal weight vector obtained by particle swarm optimization.

[0010] Furthermore, the population is initialized based on the parameter structure using the differential evolution algorithm, and the weight vector in the population is iteratively optimized, specifically including: Population initialization: Based on the parameter structure, a preset number of weight vectors, i.e. individuals, are randomly generated to form the initial population; Calculate fitness value: For each weight vector in the population, substitute it into the similarity calculation function, take the current measured cyclone as the benchmark, calculate its similarity value with all historical cyclones, and form a similarity ranking list; select the first preset number of historical cyclones according to the ranking list to form a candidate similar cyclone set, calculate the deviation coefficient corresponding to the set, and take the reciprocal of the deviation coefficient as the fitness value of the individual. Generate a test vector: For each weight vector in the population, randomly select three other weight vectors that are different from each other as the parent, and generate a mutation vector accordingly; then perform a crossover operation between the mutation vector and the current weight vector to generate a test vector. Boundary constraints and normalization: The generated test vectors are corrected for non-negativity and normalized to ensure that each component is non-negative and the sum is 1; Selection operation: Compare the fitness value of the test vector with the current weight vector used to generate the test vector, and select the vector with higher fitness to enter the next generation of the population; Iterative update: Repeatedly execute the calculation of fitness values, generation of trial vectors, boundary constraints and normalization processing, and selection operations; in each generation, update and record the weight vector with the highest fitness among all individuals up to the current generation as the global optimal weight vector; when the deviation coefficient corresponding to the global optimal weight vector changes less than a preset threshold within several consecutive generations, or when the maximum number of iterations is reached, the optimization is terminated, and the global optimal weight vector is output as the optimal weight vector obtained by the differential evolution algorithm.

[0011] Furthermore, the calculation of the deviation coefficient corresponding to this set is specifically as follows: Calculate the average similarity value between each historical cyclone and the current measured cyclone in the candidate similar cyclone set; calculate the mean of the squares of the differences between each similarity value and the average value to obtain the variance; take the square root of the variance to obtain the standard deviation; divide the standard deviation by the average similarity value to obtain the coefficient of variation corresponding to the set.

[0012] Furthermore, the step of determining the target optimal weight vector based on the preset fusion strategy involves substituting the globally optimal weight vectors obtained by the genetic algorithm, particle swarm optimization algorithm, and differential evolution algorithm into the similarity calculation function to calculate the deviation coefficients of the corresponding candidate similar cyclone sets; and selecting the globally optimal weight vector with the smallest deviation coefficient as the target optimal weight vector.

[0013] This invention also proposes a storm surge forecasting method, including: Based on the multi-algorithm fusion cyclone similarity recognition method described above, the optimal similar cyclone that is most similar to the currently measured cyclone is selected; Obtain the observed path record of the currently measured cyclone, as well as the complete path record of the best similar cyclone; Align the latest observation time of the current measured cyclone with its corresponding relative development time point in the optimal similar cyclone path; Each path point of the optimal similar cyclone after the alignment time is taken as a subsequent path point, and the key factors corresponding to each subsequent path point are extracted. The key factors include: cyclone center position, maximum wind speed at the center, maximum wind speed radius, and horizontal movement speed. The Jelesnianski wind field model and the Holland pressure model were used to calculate the wind field distribution and pressure field distribution at the corresponding subsequent path points using key factors. The wind field distribution, pressure field distribution, and the measured open boundary tide level of the optimal similar cyclone at the corresponding path point are used as driving and boundary conditions and input into the ADCIRC refined storm surge forecast model for numerical simulation, and the storm surge water level field corresponding to each subsequent path point is output. The storm surge water level fields corresponding to each subsequent path point of all outputs are used as the storm surge forecast results corresponding to the current measured cyclone.

[0014] Furthermore, the Jelesnianski wind field model is used to calculate the wind field distribution at corresponding subsequent path points using key factors. Specifically, based on the key factors of subsequent path points, the following steps are performed: Calculate the spatial distance between the center of each grid point in the simulation area and the center of the cyclone. Using the radius of maximum wind speed as the core structure, and combining the horizontal movement speed and the wind field inflow angle, the magnitude and direction of the wind speed at each grid point are calculated. Based on the spatial distance between the center of each grid point and the center of the cyclone, and the magnitude and direction of the wind speed at each grid point, an asymmetric wind field distribution that evolves as the cyclone center moves is generated.

[0015] Furthermore, the Holland pressure model was used to calculate the pressure field distribution at the corresponding subsequent path points using key factors, specifically: Obtain key factors corresponding to subsequent path points, including: cyclone center location, minimum central pressure, ambient background pressure, and radius of maximum wind speed; The spatial distance between the center of each grid point and the center of the cyclone within the simulation area is calculated using key factors. Using the radius of maximum wind speed as the core structure, combined with the pressure difference between the lowest central pressure and the ambient background pressure, as well as the spatial distance between the center of each grid point and the center of the cyclone, the air pressure value at each grid point is calculated based on the nonlinear law that air pressure decreases with distance. A dynamic pressure field is constructed based on the pressure values ​​at each grid point, with the cyclone center as the low-pressure core and distributed in an axisymmetric manner.

[0016] Compared with the prior art, the present invention has at least the following beneficial effects:

[0017] (1) This invention constructs a multidimensional time series containing multiple key factors and combines it with a weighted dynamic time warping algorithm to establish a similarity calculation function. This achieves a multidimensional comprehensive representation of the spatiotemporal evolution of cyclones, overcoming the identification bias caused by traditional methods relying on only a single feature, and significantly improving the ability to characterize complex path turning points and intensity abrupt events. On this basis, the weight vector is iteratively optimized using genetic algorithm, particle swarm optimization algorithm, and differential evolution algorithm, respectively, with the goal of minimizing the deviation coefficient of the candidate similar cyclone set, and searching for the globally optimal weight vector corresponding to each algorithm; then, based on a preset fusion strategy, the target optimal weight vector is determined from multiple globally optimal solutions. This multi-algorithm collaborative optimization mechanism avoids the problem of single optimization methods easily getting trapped in local optima or unstable convergence. Finally, based on the target optimal weight vector, the historical cyclone with the highest similarity to the currently measured cyclone is selected as the optimal similar cyclone. This closed-loop identification process effectively integrates multidimensional feature modeling and scientific weight optimization, significantly improving the objectivity, accuracy, and business applicability of cyclone similarity identification.

[0018] (2) In the weight optimization process, this invention calculates the average and variance of the similarity values ​​between each candidate similar cyclone set and the currently measured cyclone, and further obtains the ratio of the standard deviation to the mean to obtain the deviation coefficient as the optimization target. This achieves a quantitative assessment of the consistency within the similar cyclone set, overcomes the problem of neglecting discreteness caused by relying solely on the mean similarity value in traditional methods, and significantly improves the statistical rationality and robustness of the optimization target. On this basis, the globally optimal weight vectors obtained by the genetic algorithm, particle swarm optimization algorithm, and differential evolution algorithm are substituted into the similarity function, and the deviation coefficients of their corresponding candidate sets are compared. The smallest one is selected as the target optimal weight vector. This fusion strategy effectively avoids the convergence bias or local optimum risk that may exist in a single optimization algorithm, and significantly enhances the stability and global optimum of the weight determination process.

[0019] (3) After identifying the optimal similar cyclone, this invention obtains its complete path record and aligns the latest observation time of the currently measured cyclone with the corresponding relative development time point in the path of the similar cyclone. This achieves dynamic matching of the cyclone life cycle, overcomes the path extrapolation deviation caused by inconsistent time starting points in traditional analogy methods, and significantly improves the temporal consistency of subsequent path predictions. On this basis, the key factors corresponding to each path point of the optimal similar cyclone after the alignment time are extracted as the basis for extrapolating the future path of the current cyclone, and the wind field and pressure field distribution of each subsequent path point are calculated by combining the Jelesnianski wind field model and the Holland pressure model. This analogy extrapolation mechanism does not rely on numerical weather prediction models to provide future wind field and pressure field, avoids dependence on the assimilation of the initial atmospheric field, and significantly improves the rapid response capability and operational operability of storm surge forecasts.

[0020] (4) This invention employs the Jelesnianski wind field model. Based on the key factors of subsequent path points, it calculates the spatial distance between the center of each grid point in the simulation area and the cyclone center. Combined with the maximum wind speed radius, horizontal movement speed, and wind field inflow angle, it constructs an asymmetric wind field distribution. This achieves a high-fidelity reconstruction of the typical wind field structure of tropical cyclones, significantly improving the physical realism of wind field forcing. On this basis, the generated wind field distribution is used as a driving condition input into the ADCIRC refined model, effectively enhancing the accuracy of storm surge simulation, especially showing stronger adaptability and predictive ability in near-shore complex terrain areas.

[0021] (5) This invention employs the Holland pressure model, calculating the pressure values ​​at each grid point based on the cyclone center location, minimum central pressure, ambient background pressure, and maximum wind speed radius at subsequent path points, thus constructing an axisymmetric pressure field with the cyclone center as the low-pressure core. This achieves a reasonable modeling of the cyclone pressure structure, overcoming the pressure gradient force distortion problem caused by the assumption of a uniform or linear pressure field, and significantly improving the accuracy of pressure field driving. Furthermore, the wind field distribution, pressure field distribution, and the measured open boundary tide level of the optimal similar cyclone at the corresponding path point are used together as driving and boundary conditions input into the ADCIRC model for numerical simulation to output the storm surge water level field. This multi-source data fusion driving mechanism fully combines the advantages of analogical deduction and physical models, achieving high-precision forecasts with minute-level rapid response, and significantly improving the intelligence and operational level of the marine disaster early warning system. Attached Figure Description

[0022] Figure 1 This is a flowchart of a multi-algorithm fusion cyclone similarity recognition method based on weight optimization.

[0023] Figure 2 This is a flowchart of a storm surge forecasting method. Detailed Implementation

[0024] The following are specific embodiments of the present invention, which are described in conjunction with the accompanying drawings. However, the present invention is not limited to these embodiments.

[0025] This invention aims to address the problem of low matching accuracy in existing cyclone similarity recognition methods due to their single feature dimension and subjective weight setting. Figure 1 As shown in the figure, this invention proposes a multi-algorithm fusion cyclone similarity recognition method based on weight optimization, including: Historical cyclone data is acquired, which includes path records of multiple historical cyclones. Each path record consists of multiple path points distributed at equal time intervals, and each path point contains multiple key factors. The values ​​of each key factor in all path records are globally standardized. For each historical cyclone path record, a multidimensional time series characterizing the spatiotemporal evolution of the cyclone is constructed based on the standardized key factors corresponding to the continuous path points it contains. The key factors include: the longitude of the cyclone's center, the latitude of the center, the maximum wind speed at the center, the minimum air pressure at the center, and the speed of horizontal movement.

[0026] In this embodiment, based on the track records of tropical and extratropical cyclones collected by the China Meteorological Administration from 1945 to 2024, track records of multiple historical cyclones were obtained. Each track record was sampled at a fixed time interval of 6 hours. Each track point included measured data of five key factors: center longitude, center latitude, maximum wind speed at center, minimum central pressure, and horizontal movement speed, forming a multidimensional cyclone feature database with unified spatiotemporal resolution. The "maximum wind speed at center" refers to the maximum sustained wind speed (usually a 1-minute or 10-minute average wind speed, depending on the data source standard) occurring within the eyewall region of the cyclone, located in a ring-shaped area at a certain distance from the cyclone center, used to characterize cyclone intensity. The "maximum wind speed radius" is the horizontal distance from the cyclone center to the location of this maximum wind speed, reflecting the compactness of the wind circle structure. The above parameters are all derived from the China Meteorological Administration Typhoon Yearbook and internationally used datasets (such as IBTACS and CMA-STI), and are provided in the form of discrete path points. They can be directly used to construct multidimensional time series that characterize the spatiotemporal evolution of cyclones, and support subsequent similarity identification and storm surge forecast calculations.

[0027] In this embodiment, the values ​​of key factors (including center longitude, center latitude, maximum center wind speed, minimum center pressure, and horizontal movement speed) in all path records are globally standardized. Specifically, for each key factor, its overall mean and overall standard deviation across all historical path points are used to normalize all observed values ​​of that factor: the original observed value is subtracted from the overall mean of the factor, and then divided by the overall standard deviation of the factor, thereby eliminating the influence of differences in dimensions and numerical ranges among different factors. After standardization, the values ​​of each factor are converted into dimensionless relative variables, constructing a standardized multidimensional feature vector, which is then arranged in chronological order to ultimately generate a high-resolution multidimensional time series characterizing the spatiotemporal evolution of the cyclone for subsequent similarity calculations. Wherein: The formula for multidimensional feature vectors is: ; In the formula, , , , , These represent the longitude, latitude, maximum wind speed, minimum air pressure, and horizontal movement speed of the cyclone center corresponding to the path point, respectively.

[0028] The formula for multidimensional time series is: ; In the formula, This represents the multidimensional time series corresponding to the path record of any cyclone. This indicates the number of path points in the path record. Let represent the multidimensional feature vector corresponding to the i-th path point.

[0029] A weight vector containing the weight coefficients of each key factor is constructed as the parameter structure for weight optimization. Based on this weight vector and the multidimensional time series of cyclones, a similarity calculation function using a weighted dynamic time warping algorithm is constructed to calculate the similarity value between two cyclones. Specifically: Let the multidimensional time series of the currently measured cyclone to be compared be: ; Suppose a multidimensional time series of a historical cyclone is as follows: ; in: , These represent the number of waypoints for the corresponding cyclone; and These represent the number of cyclones currently being measured. The multidimensional feature vector of the path point and the first historical cyclone Multidimensional feature vectors of path points.

[0030] The local distance between two pathpoints is calculated using weighted Euclidean distance: ; in: to The weights are the weight coefficients for each key factor (longitude of the cyclone center, latitude of the center, maximum wind speed at the center, minimum air pressure at the center, and horizontal movement speed), forming a weight vector: And satisfy ; Based on the weighted dynamic time warping algorithm, the cumulative distance matrix M is constructed, and its recursive relationship is as follows: ; and ; The boundary conditions are: This indicates that the cumulative distance when the starting path points of the two sequences are aligned is zero; In the formula, , This indicates the current measured cyclone's front The cumulative distance between each waypoint and the first waypoint of a historical cyclone; In the formula, , This represents the cumulative distance by aligning the first j path points of a historical cyclone sequentially with the first path point of the currently measured cyclone.

[0031] Finally, the last element of the cumulative distance matrix is ​​taken as the minimum cumulative distance between the two sequences: ; The matching path corresponding to the minimum cumulative distance is the optimal nonlinear alignment relationship between the two cyclone path points, reflecting the structural similarity between them during the evolution process.

[0032] Based on the aforementioned weighted dynamic time warping algorithm, a similarity calculation function is constructed, which accumulates the distance. Converting to normalized similarity values: The formula for the similarity calculation function is as follows: ; In the formula, This represents the empirical scale parameter. This function is the similarity calculation function using the weighted dynamic time warping algorithm; its output value represents the degree of similarity between two cyclones, with a larger value indicating higher similarity. This represents the similarity value.

[0033] Genetic algorithm, particle swarm optimization algorithm, and differential evolution algorithm are used respectively to initialize the population based on the parameter structure, and the weight vector in the population is iteratively optimized. During the iteration process, each algorithm constructs a candidate similar cyclone set based on the current weight vector and the similarity calculation function, and calculates the deviation coefficient corresponding to the set. The deviation coefficient is used as the optimization objective, and the optimal solution is searched by minimizing the deviation coefficient. Finally, the global optimal weight vector corresponding to each algorithm is obtained. Based on the preset fusion strategy, the target optimal weight vector is determined from it. A genetic algorithm is used to initialize the population based on the aforementioned parameter structure, and the weight vectors in the population are iteratively optimized, specifically including: Population initialization: Multiple individuals with weight vectors are randomly generated based on the parameter structure to form the initial population; Calculate fitness value: For each individual with a weight vector in the population, substitute it into the similarity calculation function, and calculate the similarity value between it and all historical cyclones based on the current measured cyclone, forming a similarity ranking list; select the first preset number of historical cyclones according to the ranking list to form a candidate similar cyclone set, calculate the deviation coefficient corresponding to the set, and take the reciprocal of the deviation coefficient as the fitness value of the individual. Parent selection: Sort individuals in the population in descending order according to their fitness values. The higher the fitness value, the better the consistency of the candidate similar cyclone set corresponding to the weight vector. Use roulette wheel selection to select several individuals as parents. Perform crossover operation: randomly pair up parent individuals, perform arithmetic crossover on the weight vectors of each pair of parent individuals to generate two new offspring individuals; wherein, each offspring individual is obtained by using a random number between 0 and 1 as the combination coefficient to perform a linear weighted combination of the two parent weight vectors. For example, the following is an illustration: Assume the two parent individuals (weight vectors) are as follows: Parent 1: ; Father 2: ; Randomly generate a combination coefficient ,for example .

[0034] Perform a linear combination to generate two offspring: Offspring 1: ; Offspring 2: ; This is equivalent to taking two symmetrical points on the line connecting the parent vectors and generating two new weight vectors.

[0035] Perform mutation operation: Apply random perturbation to the weight components of a preset proportion of offspring individuals, with the perturbation amplitude following a Gaussian distribution with a mean of zero; after perturbation, perform non-negativity correction and normalization on the weight vector to ensure that each component is non-negative and the sum is 1. Iterative update: The newly generated offspring individuals are used as the next generation population, and the calculation of fitness values, selection of parent individuals, crossover and mutation operations are repeatedly performed; in each generation iteration, the global optimal weight vector is updated and recorded. The global optimal weight vector is the weight vector with the highest fitness among all individuals in the current generation; when the deviation coefficient corresponding to the optimal weight vector changes less than a preset threshold within several consecutive generations, or when the maximum number of iterations is reached, the optimization is terminated, and the global optimal weight vector is output as the optimal weight vector obtained by the genetic algorithm.

[0036] The particle swarm optimization algorithm is used to initialize the population based on the aforementioned parameter structure, and the weight vector in the population is iteratively optimized, specifically including: Initialize the particle swarm: Based on the parameter structure, randomly generate multiple particles to form an initial swarm. Each particle corresponds to a weight vector, and a velocity vector is initialized for each particle. Calculate the fitness value: For each particle's corresponding weight vector, substitute it into the similarity calculation function, and calculate its similarity value with all historical cyclones based on the current measured cyclone, forming a similarity ranking list; select the first preset number of historical cyclones according to the ranking list to form a candidate similar cyclone set, calculate the deviation coefficient corresponding to the set, and take the reciprocal of the deviation coefficient as the fitness value of the particle. Update individual and global optimal positions: For each particle, record the weight vector with the highest fitness value obtained in all iterations since initialization as the individual's historical optimal position; at the same time, maintain a global optimal position in the entire population, which corresponds to the weight vector with the highest fitness achieved by all particles since initialization. Update velocity and position: For each particle, calculate the updated velocity vector based on its current velocity, the deviation of its current weight vector from its individual historical best position, and the deviation from its global best position; update its corresponding weight vector based on this velocity vector. Constraint processing: The updated weight vector is corrected for nonnegativity and normalized to ensure that each component is nonnegative and the sum is 1; Iterative update: Repeatedly perform fitness value calculation, optimal position update, velocity and position update and constraint processing operations; in each iteration, update and record the global optimal weight vector; when the deviation coefficient corresponding to the optimal weight vector changes less than a preset threshold within several consecutive generations, or when the maximum number of iterations is reached, the optimization is terminated, and the global optimal weight vector is output as the optimal weight vector obtained by particle swarm optimization.

[0037] The population is initialized based on the parameter structure using the differential evolution algorithm, and the weight vector in the population is iteratively optimized, specifically including: Population initialization: Based on the parameter structure, a preset number of weight vectors, i.e. individuals, are randomly generated to form the initial population; Calculate fitness value: For each weight vector in the population, substitute it into the similarity calculation function, take the current measured cyclone as the benchmark, calculate its similarity value with all historical cyclones, and form a similarity ranking list; select the first preset number of historical cyclones according to the ranking list to form a candidate similar cyclone set, calculate the deviation coefficient corresponding to the set, and take the reciprocal of the deviation coefficient as the fitness value of the individual. Generate a test vector: For each weight vector in the population, randomly select three other weight vectors that are different from each other as the parent, and generate a mutation vector accordingly; then perform a crossover operation between the mutation vector and the current weight vector to generate a test vector. Specifically, the process of generating test vectors is as follows: For each weight vector in the population (called the "target vector"), three distinct other weight vectors are randomly selected as parent individuals, denoted as... And generate a mutation vector according to the differential mutation strategy. The formula for generating the formula is: ; In the formula, This represents the scaling factor, and its value range is typically 100%. It is used to control the perturbation amplitude of the difference vector.

[0038] Subsequently, the mutation vector With the current target vector (let's assume it's...) A crossover operation is performed to generate an experimental vector. The crossover operation uses a preset crossover rate CR (e.g., 0.8) as the probability, and determines whether the components of the experimental vector come from the mutated vector or the target vector dimension by dimension. An example is given below (for simplicity, it is assumed that...). (for three-dimensional vectors) Assumption: Target vector ; Mutation vector After truncating negative values ​​to 0, we get: ; Assuming a crossover rate CR = 0.8, and randomly generating crossover decisions for each dimension (e.g., "Yes" for dimension 1, "No" for dimension 2, and "Yes" for dimension 3), the experimental vector is constructed as follows: First dimension: Inherited from the mutation vector → 0.7; The second dimension: inherited from the target vector → 0.5; The third dimension: inherited from the mutation vector → 0.44; The experimental vector is obtained: .

[0039] Boundary constraints and normalization: The generated test vectors are corrected for non-negativity and normalized to ensure that each component is non-negative and the sum is 1; Selection operation: Compare the fitness value of the test vector with the current weight vector used to generate the test vector, and select the vector with higher fitness to enter the next generation of the population; Iterative update: Repeatedly execute the calculation of fitness values, generation of trial vectors, boundary constraints and normalization processing, and selection operations; in each generation, update and record the weight vector with the highest fitness among all individuals up to the current generation as the global optimal weight vector; when the deviation coefficient corresponding to the global optimal weight vector changes less than a preset threshold within several consecutive generations, or when the maximum number of iterations is reached, the optimization is terminated, and the global optimal weight vector is output as the optimal weight vector obtained by the differential evolution algorithm.

[0040] The calculation of the deviation coefficient corresponding to this set is specifically as follows: Calculate the average similarity value between each historical cyclone in the candidate similar cyclone set and the current measured cyclone (specifically: calculate the similarity value between each historical cyclone in the set and the current measured cyclone, and then calculate the average of these similarity values); calculate the mean of the squares of the differences between each similarity value and the average value to obtain the variance; take the square root of the variance to obtain the standard deviation; divide the standard deviation by the average of the similarity values ​​to obtain the coefficient of variation corresponding to the set.

[0041] To overcome the limitations of traditional subjective weighting methods in weight allocation, this invention introduces the coefficient of variation (COP). The objective function is used as the optimization objective function for genetic algorithms, particle swarm optimization algorithms, and differential evolution algorithms. This objective function measures the relative dispersion of the similarity values ​​between historical cyclones and currently measured cyclones in the candidate similar cyclone set. A smaller value indicates higher consistency in similarity within the set and stronger system coordination. During optimization, the algorithm iteratively adjusts the weight distribution of each key factor to minimize this deviation coefficient, thereby selecting a set of candidate similar cyclones that are highly consistent with the currently measured cyclones in terms of key factors and have a stable internal similarity distribution, achieving precise weight optimization.

[0042] The method for determining the target optimal weight vector based on a preset fusion strategy involves substituting the globally optimal weight vectors obtained by the genetic algorithm, particle swarm optimization algorithm, and differential evolution algorithm into the similarity calculation function to calculate the deviation coefficients of their corresponding candidate similar cyclone sets; and selecting the globally optimal weight vector with the smallest deviation coefficient as the target optimal weight vector.

[0043] In the weight optimization process, this invention calculates the average and variance of the similarity values ​​between each candidate set of similar cyclones and the currently measured cyclones, and further obtains the ratio of the standard deviation to the mean, using the deviation coefficient as the optimization target. This achieves a quantitative assessment of the consistency within the set of similar cyclones, overcoming the problem of neglecting dispersion caused by traditional methods relying solely on the mean of similarity, and significantly improving the statistical rationality and robustness of the optimization target. Based on this, the globally optimal weight vectors obtained by the genetic algorithm, particle swarm optimization algorithm, and differential evolution algorithm are substituted into the similarity function, and the deviation coefficients of their corresponding candidate sets are compared, selecting the smallest as the target optimal weight vector. This fusion strategy effectively avoids the convergence bias or local optimum risks that may exist in a single optimization algorithm, significantly enhancing the stability and global optimality of the weight determination process.

[0044] Substitute the target optimal weight vector into the similarity calculation function, recalculate the similarity value between the current measured cyclone and each historical cyclone, generate a similarity ranking list, and select the historical cyclone with the highest similarity value as the optimal similar cyclone.

[0045] This invention constructs a multidimensional time series containing multiple key factors and combines it with a weighted dynamic time warping algorithm to establish a similarity calculation function. This achieves a multidimensional comprehensive representation of the spatiotemporal evolution of cyclones, overcoming the identification bias caused by traditional methods relying on only a single feature, and significantly improving the ability to characterize complex path transitions and abrupt intensity changes. Based on this, genetic algorithms, particle swarm optimization algorithms, and differential evolution algorithms are used to iteratively optimize the weight vector, aiming to minimize the deviation coefficient of the candidate similar cyclone set, searching for the globally optimal weight vector corresponding to each algorithm; then, based on a preset fusion strategy, the target optimal weight vector is determined from multiple globally optimal solutions. This multi-algorithm collaborative optimization mechanism avoids the problem of single optimization methods easily getting trapped in local optima or having unstable convergence. Finally, based on the target optimal weight vector, the historical cyclone with the highest similarity to the currently measured cyclone is selected as the optimal similar cyclone. This closed-loop identification process effectively integrates multidimensional feature modeling and scientific weight optimization, significantly improving the objectivity, accuracy, and business applicability of cyclone similarity identification.

[0046] Example 2 To overcome the shortcomings of traditional storm surge forecasting models, which rely on numerical weather prediction and lack real-time performance, such as... Figure 2 As shown in the figure, this invention also proposes a storm surge forecasting method, including: Based on the multi-algorithm fusion cyclone similarity recognition method described above, the optimal similar cyclone that is most similar to the currently measured cyclone is selected; Obtain the observed path record of the currently measured cyclone, as well as the complete path record of the best similar cyclone; Align the latest observation time of the current measured cyclone with its corresponding relative development time point in the optimal similar cyclone path; In this embodiment, aligning the latest observation time of the current measured cyclone with its corresponding relative development time point in the optimal similar cyclone path means: based on the similarity matching result generated by the process of "substituting the target optimal weight vector into the similarity calculation function, recalculating the similarity value between the current measured cyclone and each historical cyclone, generating a similarity ranking list, and selecting the historical cyclone with the highest similarity as the optimal similar cyclone" in the multi-algorithm fusion cyclone similarity recognition method, the nonlinear spatiotemporal correspondence between the current measured cyclone and the optimal similar cyclone is obtained (wherein, the similarity value is determined by the minimum cumulative distance of the weighted dynamic time warping algorithm, and the corresponding optimal matching path is the nonlinear spatiotemporal correspondence between the current measured cyclone and the optimal similar cyclone); according to this correspondence, the path point position matched by the latest observation time of the current measured cyclone (such as time T) in the optimal similar cyclone path is determined, that is, the "relative development time point", thereby realizing the alignment of the two in the life cycle stage (such as generation, enhancement, maturity or decay). This alignment provides an accurate time reference for subsequent calls to measured storm surge data of the best similar cyclones, supporting analog forecasting.

[0047] Each path point of the optimal similar cyclone after the alignment time is taken as a subsequent path point, and the key factors corresponding to each subsequent path point are extracted. The key factors include: cyclone center position, maximum wind speed at the center, maximum wind speed radius, and horizontal movement speed. The Jelesnianski wind field model and the Holland pressure model were used to calculate the wind field distribution and pressure field distribution at the corresponding subsequent path points using key factors. The Jelesnianski wind field model is used to calculate the wind field distribution at subsequent path points using key factors. Specifically, based on the key factors of subsequent path points, the following steps are performed: Calculate the spatial distance between the center of each grid point in the simulation area and the center of the cyclone. Using the radius of maximum wind speed as the core structure, and combining the horizontal movement speed and the wind field inflow angle, the magnitude and direction of the wind speed at each grid point are calculated. Based on the spatial distance between the center of each grid point and the center of the cyclone, and the magnitude and direction of the wind speed at each grid point, an asymmetric wind field distribution that evolves as the cyclone center moves is generated.

[0048] This invention employs the Jelesnianski wind field model, calculating the spatial distance between the center of each grid point within the simulation area and the cyclone center based on key factors of subsequent path points. It then combines the maximum wind speed radius, horizontal movement speed, and wind inflow angle to construct an asymmetric wind field distribution. This achieves a high-fidelity reconstruction of the typical wind field structure of tropical cyclones, significantly improving the physical realism of wind forcing. Furthermore, the generated wind field distribution is used as a driving condition input into the ADCIRC refined model, effectively enhancing the accuracy of storm surge simulation, especially demonstrating stronger adaptability and predictive ability in near-shore complex terrain areas.

[0049] The Holland pressure model was used to calculate the pressure field distribution at the corresponding subsequent path points using key factors, specifically: Obtain key factors corresponding to subsequent path points, including: cyclone center location, minimum central pressure, ambient background pressure, and radius of maximum wind speed; The spatial distance between the center of each grid point and the center of the cyclone within the simulation area is calculated using key factors. Using the radius of maximum wind speed as the core structure, combined with the pressure difference between the lowest central pressure and the ambient background pressure, as well as the spatial distance between the center of each grid point and the center of the cyclone, the air pressure value at each grid point is calculated based on the nonlinear law that air pressure decreases with distance. A dynamic pressure field is constructed based on the pressure values ​​at each grid point, with the cyclone center as the low-pressure core and distributed in an axisymmetric manner.

[0050] This invention employs the Holland pressure model, calculating the pressure values ​​at each grid point based on the cyclone center location, minimum central pressure, ambient background pressure, and maximum wind speed radius at subsequent path points, thus constructing an axisymmetric pressure field with the cyclone center as the low-pressure core. This achieves a reasonable modeling of the cyclone's pressure structure, overcoming the pressure gradient force distortion problem caused by the assumption of a uniform or linear pressure field, and significantly improving the accuracy of pressure field driving. Furthermore, the wind field distribution, pressure field distribution, and the measured open-boundary tide level of the optimal similar cyclone at the corresponding path points are used as driving and boundary conditions input into the ADCIRC model for numerical simulation to output the storm surge water level field. This multi-source data fusion driving mechanism fully combines the advantages of analogical deduction and physical models, achieving high-precision forecasts with minute-level rapid response, significantly improving the intelligence and operational level of the marine disaster early warning system.

[0051] The wind field distribution, pressure field distribution, and the measured open boundary tide level of the optimal similar cyclone at the corresponding path point are used as driving and boundary conditions and input into the ADCIRC refined storm surge forecast model for numerical simulation, and the storm surge water level field corresponding to each subsequent path point is output. In this embodiment, the measured open boundary tide level of the optimal similar cyclone at the corresponding path point refers to the actual tide level data at the open boundary section provided by the ocean observation station or reanalysis data, corresponding to the time of each path point in the historical observation path of the optimal similar cyclone. This data has deducted the astronomical tide component and retained the storm surge increase caused by the cyclone, serving as the open boundary water level driving condition for the ADCIRC refined storm surge forecast model.

[0052] The storm surge water level fields corresponding to each subsequent path point of all outputs are used as the storm surge forecast results corresponding to the current measured cyclone.

[0053] The ADCIRC model employs an unstructured triangular mesh and the Galerkin finite element method to solve the shallow water dynamic equations, enabling high-precision simulation of storm surge water level fields. This invention utilizes its mature physical mechanisms to achieve analogous forecasting of storm surges that may be triggered by currently observed cyclones.

[0054] This invention, after identifying the optimal similar cyclone, acquires its complete path record and aligns the latest observation time of the currently measured cyclone with the corresponding relative development time point in the similar cyclone's path. This achieves dynamic matching of the cyclone's life cycle, overcoming path extrapolation biases caused by inconsistent time starting points in traditional analogy methods, and significantly improving the temporal consistency of subsequent path predictions. Based on this, key factors corresponding to each path point after the alignment time of the optimal similar cyclone are extracted as the basis for extrapolating the current cyclone's future path. Combined with the Jelesnianski wind field model and the Holland pressure model, the wind and pressure field distributions of each subsequent path point are calculated. This analogy extrapolation mechanism does not rely on numerical weather prediction models to provide future wind and pressure fields, avoiding dependence on the assimilation of the initial atmospheric field, and significantly improving the rapid response capability and operational operability of storm surge forecasts.

[0055] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0056] Furthermore, in this invention, descriptions involving terms such as "first," "second," and "a" are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0057] In this invention, unless otherwise explicitly specified and limited, the terms "connection," "fixed," etc., should be interpreted broadly. For example, "fixed" can mean a fixed connection, a detachable connection, or an integral part; it can mean a mechanical connection or an electrical connection; it can mean a direct connection or an indirect connection through an intermediate medium; it can mean the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0058] Furthermore, the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are feasible for those skilled in the art. If the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

Claims

1. A multi-algorithm fusion cyclone similarity recognition method based on weight optimization, characterized in that, include: Historical cyclone data is acquired, which includes path records of multiple historical cyclones. Each path record consists of multiple path points distributed at equal time intervals, and each path point contains multiple key factors. The values ​​of each key factor in all path records are globally standardized. For each historical cyclone's path record, a multidimensional time series characterizing the spatiotemporal evolution of the cyclone is constructed based on the standardized key factors corresponding to the continuous path points it contains. A weight vector containing the weight coefficients of each key factor is constructed as the parameter structure for weight optimization. Based on this weight vector and the multidimensional time series of cyclones, a similarity calculation function using a weighted dynamic time warping algorithm is constructed to calculate the similarity value between two cyclones. Genetic algorithm, particle swarm optimization algorithm, and differential evolution algorithm are used respectively to initialize the population based on the parameter structure, and the weight vector in the population is iteratively optimized. During the iteration process, each algorithm constructs a candidate similar cyclone set based on the current weight vector and the similarity calculation function, and calculates the deviation coefficient corresponding to the set. The deviation coefficient is used as the optimization objective, and the optimal solution is searched by minimizing the deviation coefficient. Finally, the global optimal weight vector corresponding to each algorithm is obtained. Based on the preset fusion strategy, the target optimal weight vector is determined from it. Substitute the target optimal weight vector into the similarity calculation function, recalculate the similarity value between the current measured cyclone and each historical cyclone, generate a similarity ranking list, and select the historical cyclone with the highest similarity value as the optimal similar cyclone.

2. The multi-algorithm fusion cyclone similarity recognition method based on weight optimization according to claim 1, characterized in that, The key factors include: the longitude of the cyclone's center, the latitude of the center, the maximum wind speed at the center, the minimum air pressure at the center, and the speed of horizontal movement.

3. The multi-algorithm fusion cyclone similarity recognition method based on weight optimization according to claim 2, characterized in that, A genetic algorithm is used to initialize the population based on the aforementioned parameter structure, and the weight vectors in the population are iteratively optimized, specifically including: Population initialization: Multiple individuals with weight vectors are randomly generated based on the parameter structure to form the initial population; Calculate fitness value: For each individual with a weight vector in the population, substitute it into the similarity calculation function, and calculate the similarity value between it and all historical cyclones based on the current measured cyclone, forming a similarity ranking list; select the first preset number of historical cyclones according to the ranking list to form a candidate similar cyclone set, calculate the deviation coefficient corresponding to the set, and take the reciprocal of the deviation coefficient as the fitness value of the individual. Parent selection: Sort individuals in the population in descending order according to their fitness values. The higher the fitness value, the better the consistency of the candidate similar cyclone set corresponding to the weight vector. Use roulette wheel selection to select several individuals as parents. Perform crossover operation: randomly pair up parent individuals, perform arithmetic crossover on the weight vectors of each pair of parent individuals to generate two new offspring individuals; wherein, each offspring individual is obtained by using a random number between 0 and 1 as the combination coefficient to perform a linear weighted combination of the two parent weight vectors. Perform mutation operation: Apply random perturbation to the weight components of a preset proportion of offspring individuals, with the perturbation amplitude following a Gaussian distribution with a mean of zero; after perturbation, perform non-negativity correction and normalization on the weight vector to ensure that each component is non-negative and the sum is 1. Iterative update: The newly generated offspring individuals are used as the next generation population, and the calculation of fitness values, selection of parent individuals, crossover and mutation operations are repeatedly performed; in each generation iteration, the global optimal weight vector is updated and recorded. The global optimal weight vector is the weight vector with the highest fitness among all individuals in the current generation; when the deviation coefficient corresponding to the optimal weight vector changes less than a preset threshold within several consecutive generations, or when the maximum number of iterations is reached, the optimization is terminated, and the global optimal weight vector is output as the optimal weight vector obtained by the genetic algorithm.

4. The multi-algorithm fusion cyclone similarity recognition method based on weight optimization according to claim 2, characterized in that, The particle swarm optimization algorithm is used to initialize the population based on the aforementioned parameter structure, and the weight vector in the population is iteratively optimized, specifically including: Initialize the particle swarm: Based on the parameter structure, randomly generate multiple particles to form an initial swarm. Each particle corresponds to a weight vector, and a velocity vector is initialized for each particle. Calculate the fitness value: For each particle's corresponding weight vector, substitute it into the similarity calculation function, and calculate its similarity value with all historical cyclones based on the current measured cyclone, forming a similarity ranking list; select the first preset number of historical cyclones according to the ranking list to form a candidate similar cyclone set, calculate the deviation coefficient corresponding to the set, and take the reciprocal of the deviation coefficient as the fitness value of the particle. Update individual and global optimal positions: For each particle, record the weight vector with the highest fitness value obtained in all iterations since initialization as the individual's historical optimal position; at the same time, maintain a global optimal position in the entire population, which corresponds to the weight vector with the highest fitness achieved by all particles since initialization. Update velocity and position: For each particle, calculate the updated velocity vector based on its current velocity, the deviation of its current weight vector from its individual historical best position, and the deviation from its global best position; update its corresponding weight vector based on this velocity vector. Constraint processing: The updated weight vector is corrected for nonnegativity and normalized to ensure that each component is nonnegative and the sum is 1; Iterative update: Repeatedly perform fitness value calculation, optimal position update, velocity and position update and constraint processing operations; in each iteration, update and record the global optimal weight vector; when the deviation coefficient corresponding to the optimal weight vector changes less than a preset threshold within several consecutive generations, or when the maximum number of iterations is reached, the optimization is terminated, and the global optimal weight vector is output as the optimal weight vector obtained by particle swarm optimization.

5. The multi-algorithm fusion cyclone similarity recognition method based on weight optimization according to claim 2, characterized in that, The population is initialized based on the parameter structure using the differential evolution algorithm, and the weight vector in the population is iteratively optimized, specifically including: Population initialization: Based on the parameter structure, a preset number of weight vectors, i.e. individuals, are randomly generated to form the initial population; Calculate fitness value: For each weight vector in the population, substitute it into the similarity calculation function, take the current measured cyclone as the benchmark, calculate its similarity value with all historical cyclones, and form a similarity ranking list; select the first preset number of historical cyclones according to the ranking list to form a candidate similar cyclone set, calculate the deviation coefficient corresponding to the set, and take the reciprocal of the deviation coefficient as the fitness value of the individual. Generate a test vector: For each weight vector in the population, randomly select three other weight vectors that are different from each other as the parent, and generate a mutation vector accordingly; then perform a crossover operation between the mutation vector and the current weight vector to generate a test vector. Boundary constraints and normalization: The generated test vectors are corrected for non-negativity and normalized to ensure that each component is non-negative and the sum is 1; Selection operation: Compare the fitness value of the test vector with the current weight vector used to generate the test vector, and select the vector with higher fitness to enter the next generation of the population; Iterative update: Repeatedly execute the calculation of fitness values, generation of trial vectors, boundary constraints and normalization processing, and selection operations; in each generation, update and record the weight vector with the highest fitness among all individuals up to the current generation as the global optimal weight vector; when the deviation coefficient corresponding to the global optimal weight vector changes less than a preset threshold within several consecutive generations, or when the maximum number of iterations is reached, the optimization is terminated, and the global optimal weight vector is output as the optimal weight vector obtained by the differential evolution algorithm.

6. A multi-algorithm fusion cyclone similarity recognition method based on weight optimization according to any one of claims 3 to 5, characterized in that, The calculation of the deviation coefficient corresponding to this set is specifically as follows: Calculate the average similarity value between each historical cyclone and the current measured cyclone in the candidate similar cyclone set; calculate the mean of the squares of the differences between each similarity value and the average value to obtain the variance; take the square root of the variance to obtain the standard deviation; divide the standard deviation by the average similarity value to obtain the coefficient of variation corresponding to the set.

7. The multi-algorithm fusion cyclone similarity recognition method based on weight optimization according to claim 1, characterized in that, The method for determining the target optimal weight vector based on a preset fusion strategy involves substituting the globally optimal weight vectors obtained by the genetic algorithm, particle swarm optimization algorithm, and differential evolution algorithm into the similarity calculation function to calculate the deviation coefficients of their corresponding candidate similar cyclone sets; and selecting the globally optimal weight vector with the smallest deviation coefficient as the target optimal weight vector.

8. A storm surge forecasting method, characterized in that, include: The multi-algorithm fusion cyclone similarity recognition method according to claim 1 selects the optimal similar cyclone that is most similar to the currently measured cyclone; Obtain the observed path record of the currently measured cyclone, as well as the complete path record of the best similar cyclone; Align the latest observation time of the current measured cyclone with its corresponding relative development time point in the optimal similar cyclone path; Each path point of the optimal similar cyclone after the alignment time is taken as the subsequent path point, and the key factors corresponding to each subsequent path point are extracted respectively. The key factors include: cyclone center location, maximum wind speed at the center, maximum wind speed radius, and horizontal movement speed; The Jelesnianski wind field model and the Holland pressure model were used to calculate the wind field distribution and pressure field distribution at the corresponding subsequent path points using key factors. The wind field distribution, pressure field distribution, and the measured open boundary tide level of the optimal similar cyclone at the corresponding path point are used as driving and boundary conditions and input into the ADCIRC refined storm surge forecast model for numerical simulation, and the storm surge water level field corresponding to each subsequent path point is output. The storm surge water level fields corresponding to each subsequent path point of all outputs are used as the storm surge forecast results corresponding to the current measured cyclone.

9. A storm surge forecasting method according to claim 8, characterized in that, The Jelesnianski wind field model is used to calculate the wind field distribution at subsequent path points using key factors. Specifically, based on the key factors of subsequent path points, the following steps are performed: Calculate the spatial distance between the center of each grid point in the simulation area and the center of the cyclone. Using the radius of maximum wind speed as the core structure, and combining the horizontal movement speed and the wind field inflow angle, the magnitude and direction of the wind speed at each grid point are calculated. Based on the spatial distance between the center of each grid point and the center of the cyclone, and the magnitude and direction of the wind speed at each grid point, an asymmetric wind field distribution that evolves as the cyclone center moves is generated.

10. A storm surge forecasting method according to claim 9, characterized in that, The Holland pressure model was used to calculate the pressure field distribution at the corresponding subsequent path points using key factors, specifically: Obtain key factors corresponding to subsequent path points, including: cyclone center location, minimum central pressure, ambient background pressure, and radius of maximum wind speed; The spatial distance between the center of each grid point and the center of the cyclone within the simulation area is calculated using key factors. Using the radius of maximum wind speed as the core structure, combined with the pressure difference between the lowest central pressure and the ambient background pressure, as well as the spatial distance between the center of each grid point and the center of the cyclone, the air pressure value at each grid point is calculated based on the nonlinear law that air pressure decreases with distance. A dynamic pressure field is constructed based on the pressure values ​​at each grid point, with the cyclone center as the low-pressure core and distributed in an axisymmetric manner.

Images