A method for evaluating safety of a typical tropical marine ranching and predicting future safety situation

Abstract

The application provides a method for typical tropical marine ranching safety evaluation and future safety situation prediction, and relates to the technical field of ecological safety. The method comprises the following steps: determining an evaluation target; constructing a monitoring system, collecting and pre-processing to obtain a standardized time series index data set; setting a spatial control area and a time control benchmark to obtain an evaluation quantization benchmark system; constructing a BVPTL prediction model, outputting future safety situation prediction values and model feature importance parameters after training; establishing a three-level index system and a five-level quantization standard to form a safety evaluation model framework; using analytic hierarchy process and machine learning weight to conduct game combination weighting, and calculating comprehensive weights; then calculating current and future comprehensive safety indexes and dividing safety grades; and starting corresponding response and ecological restoration measures according to the current and predicted grades. The application realizes dynamic and accurate evaluation and early warning of the safety state of the tropical marine ranching.

Description

Technical Field

[0001] This application relates to the field of ecological security technology, and in particular to a method for safety assessment and future security prediction of a typical tropical marine ranch. Background Technology

[0002] Against the backdrop of a global strategy for the sustainable use of marine resources and the coordinated governance of ecological protection, marine ranches, as a complex system of resource enhancement and ecological restoration based on human intervention, have become an important solution to address the decline of global nearshore fishery resources and the degradation of marine ecosystems. It is worth noting that while tropical seas such as the South China Sea and the Beibu Gulf possess advantages such as rich biodiversity and high primary productivity, the spatial coupling between their ecological sensitivity and the intensity of human activities leads to a non-linear growth pattern in ecological security risks in these areas. Therefore, it is urgent to construct an adaptive assessment system to scientifically evaluate the ecological security of tropical marine ranches. Summary of the Invention

[0003] This application provides a method for safety assessment and future safety prediction of typical tropical marine ranches. From the perspective of the coupling of natural-economic-social systems, it clarifies the mechanism and regulation path of ecological security, so as to scientifically assess the ecological security of tropical marine ranches.

[0004] To achieve the above objectives, this application adopts the following technical solution: This application provides a method for safety assessment and future safety trend prediction of a typical tropical marine ranch, the method comprising: The assessment objectives are defined, and the evaluation and prediction cycles are divided into three time scales: short-term, medium-term, and long-term. The benchmark parameters for the evaluation dimensions and the benchmark parameters for the time cycles are output. The assessment objectives include the evaluation objectives for the three dimensions of ecological security, economic security, and social security in the comprehensive safety assessment of tropical marine ranches. Based on the aforementioned evaluation dimension benchmark parameters and time period benchmark parameters, an integrated three-dimensional monitoring system covering driving force, pressure, state, impact, and response is constructed, and corresponding indicator data are collected. All collected indicator data are preprocessed to obtain a standardized time series indicator dataset. Based on the aforementioned time period benchmark parameters and standardized time series index dataset, at least two natural control areas are established in adjacent sea areas that have not undergone artificial construction. At the same time, baseline environmental and biological resource data for at least 1 to 3 consecutive years before the construction of marine ranches are collected to form spatial control benchmarks and time series control benchmarks, thereby obtaining an evaluation quantitative benchmark system. A BVPTL hybrid prediction model integrating a BKA-VMD data preprocessing module, a Transformer-LSTM hybrid model, and a PLO hyperparameter optimization algorithm is constructed. The BVPTL hybrid prediction model is trained based on the standardized time series index data. Based on the trained BVPTL hybrid prediction model, the predicted values ​​of the security status of tropical marine ranches in the future time scale and the model feature importance parameters are obtained. Based on the aforementioned evaluation dimension benchmark parameters, standardized time series index dataset, and evaluation quantification benchmark system, a three-level DPSIR index system and a five-level quantification standard for tropical marine ranch safety evaluation are established, the calculation method for the comprehensive safety index is determined, and a safety evaluation model framework is obtained. A game-theoretic weighting model based on the Analytic Hierarchy Process (AHP) and BVPTL model machine learning weights is constructed. According to the security evaluation model framework, the standardized time series index dataset, and the model feature importance parameters, the comprehensive weight of each evaluation index is calculated, and the comprehensive weight dataset is output. Based on the comprehensive weighted dataset, the current comprehensive safety index of the tropical marine ranch is calculated, and the predicted comprehensive safety index for future times is calculated in combination with the predicted future safety situation value. Based on the current comprehensive safety index and the predicted comprehensive safety index, the current and future safety status of the marine ranch is divided into multiple safety levels, and the current safety level and the predicted future safety level are obtained. Based on the current security level and the predicted future security level, corresponding response and ecological restoration measures will be initiated.

[0005] Furthermore, all collected indicator data are preprocessed to obtain a standardized time-series indicator dataset, including: Min-max standardization is applied to all indicator data to eliminate dimensional differences. The formula is as follows: in, For the first The first sample Individual indicator values, and The first The minimum and maximum values ​​of each indicator. For the first The first sample A standardized time-series index value; For missing data, linear interpolation is used for time series data and kriging interpolation is used for spatial data. Indicators with a missing rate of more than 20% are re-collected.

[0006] Furthermore, the BKA-VMD data preprocessing module is used to adaptively decompose the input standardized time-series index dataset, and the specific process includes: Using each standardized time series index in the standardized time series index dataset as the input signal, variational mode decomposition is used to transform the input signal. Adaptive decomposition into eigenmode functions Solve the constrained variational optimization problem shown below: in, To perform the minimum value operation, For modal number index, The penalty coefficient is... Let k be the eigenmode function. For the first The central angular frequencies corresponding to each eigenmode function For time variables The first-order partial derivative operator, For Dirac delta function, is the base of the natural logarithm. The imaginary unit, It is a time variable; The constraints are: The number of modes in variational mode decomposition and penalty coefficient Optimize, The code is encoded as a two-dimensional vector. An optimal solution is searched within a preset boundary, with the fitness function being the sum of the sample entropies of the decomposed subsequences. The fitness function is expressed as: in, Let be the objective function used to evaluate the optimization effect of variational mode decomposition parameters, and let its value be the sum of the sample entropies of all intrinsic mode functions obtained by decomposition. This represents the k-th eigenmode function obtained from VMD decomposition. Calculated sample entropy; The optimization is achieved through iterative analysis of migration and attack behaviors until a preset maximum number of iterations is reached. or continuous Optimization stops when the fitness value no longer decreases, and the optimal parameter combination is output. The original signal is then decomposed using the parameter combination to obtain... One IMF component is used as the input to the Transformer-LSTM hybrid model; The Transformer-LSTM hybrid model takes the IMF component sequence obtained after decomposition by the BKA-VMD data preprocessing module as input and uses a cascaded fusion structure to extract temporal features; the cascaded fusion structure includes an LSTM branch and a Transformer branch; The LSTM branch uses a single-layer LSTM model to extract local temporal features from the IMF component sequence, and takes the last hidden state as a local feature. The Transformer branch first uses a linear mapping to upscale each component data in the IMF component sequence to a high-dimensional space, and adds position encoding to provide order information for each component in the IMF component sequence. The global dependency features are extracted by the Transformer encoder, and global average pooling is used to obtain global features. By fusing the output layer, the local features of the LSTM branch and the global features of the Transformer branch are concatenated, and then mapped to the prediction dimension through a fully connected layer to achieve the prediction of future security situation.

[0007] Furthermore, in the BVPTL hybrid prediction model, the PLO hyperparameter optimization algorithm is used to globally optimize the key hyperparameters of the Transformer-LSTM hybrid model, train the model, and perform prediction validation. Its specific process includes: Using the number of heads, initial learning rate, and L2 regularization coefficient in the self-attention mechanism of the Transformer-LSTM hybrid model as optimization target parameters, and minimizing the mean absolute percentage error (MAPE) as the optimization objective, global optimization is performed by simulating the particle motion mechanism in the aurora formation process. The core steps include rotational motion, aurora elliptical walking, adaptive weight fusion mechanism, and particle collision. The rotational motion includes: simulating the damped spiral motion of a charged particle in the Earth's magnetic field, the dynamics of which are characterized by a modified first-order linear differential equation, which is expressed as: in, The particle velocity vector The damping factor, The particle charge. For particle mass, Given the Earth's magnetic field strength, the general solution of this modified first-order linear differential equation is: ; Let be the particle velocity vector at time t. Let be the integration constant. is the base of the natural logarithm; The aurora elliptical walk includes: chaotic fluctuations of high-energy particles within the boundary of the aurora ellipse based on Levy flight simulation, wherein the particle displacement update formula is: in, This represents the particle displacement vector generated by the elliptical aurora walk. Let be the current position vector of the i-th particle in the d-dimensional solution space. The random numbers are uniformly distributed in the interval [0,1]. This is the current location of the population centroid. , To solve for the upper and lower bounds of the space; Let d be the random step size of the Levy distribution, calculated using the following formula: in, To conform to a mean of 0 and a variance of σ 2 A normally distributed random variable, These are the stability parameters for Levy flight. , It follows a normal distribution (Gaussian distribution). Let N(0,σ) be the standard deviation of the normal distribution. , It is a gamma function; The adaptive weight fusion mechanism includes determining time-varying weights using the following formula: in, These are the time-varying weighting coefficients for the rotational motion components. The hyperbolic tangent activation function is used. The number of times the current function is evaluated. The maximum number of function evaluations is preset. For the time-varying weighting coefficients of the aurora elliptical travel component, It is a natural exponential function; Based on the determined time-varying weights, the position is updated using the following formula: in, , is the rotational motion component, and the amplitude of the rotational motion component is according to Dynamic decay, This is the attenuation adjustment coefficient. Let i be the updated position vector of the i-th particle. Let be the current position vector of the i-th particle in the d-dimensional solution space. This is the particle displacement vector generated by rotational motion. Let be a d-dimensional random vector that follows a uniform distribution in the interval [0,1]. For Hadamah accumulation; The particle collisions include: simulated chaotic collisions between solar wind particles, triggered by the currently generated random number rand satisfying the following condition. and If the triggering condition is met, a collision operation is performed. The collision probability control factor is calculated using the following formula: When the trigger condition is met, the particle position is updated as follows: in, For the current particle In the The position of the dimension for A random arrangement of , where N is the population size. To specify with particles The particle numbers that collided; Iterative search continues until the preset maximum number of function evaluations MaxFEs is reached, and the optimal combination of hyperparameters that minimizes MAPE is output as the configuration of the Transformer-LSTM model.

[0008] Furthermore, in the BVPTL hybrid prediction model, the method of training the BVPTL hybrid prediction model based on the standardized time-series index data includes: The standardized time-series index dataset is divided into training set, validation set and test set according to a set ratio; Based on a regularized loss function, the training process sequentially performs forward propagation, loss calculation, backpropagation, and parameter update. The loss is accumulated during training, and training stops when the validation set loss no longer decreases for a set number of consecutive epochs. The regularized loss function is... Represented as: in, For the true value, For predicted values, The regularization coefficient is determined using the PLO hyperparameter optimization algorithm. For the model number One weight parameter, Where j is the number of samples in the batch, and j is the sample index; Based on the test set, the coefficient of determination is used to evaluate the model's predictive performance. Once the model's predictive performance reaches the set standard, the model obtained from the current training is used as the trained BVPTL hybrid prediction model.

[0009] Furthermore, the DPSIR three-level indicator system is divided into three levels: target layer, criterion layer, and indicator layer. The target layer is for the safety assessment of tropical marine ranches. The criterion layer includes five parallel sub-layers: driving force, pressure, state and resources, impact, and response. The indicator layer includes specific indicators corresponding to each parallel sub-layer, with each sub-layer corresponding to at least one specific indicator. The comprehensive security index is calculated as follows: in, This refers to the total number of specific indicators. The comprehensive weight of the j-th indicator is... Let be the standardized value of the j-th indicator.

[0010] Furthermore, the specific methods for calculating the comprehensive weight of each evaluation indicator include: Multiple judgment matrices are obtained, and their geometric or arithmetic mean is taken to form a comprehensive judgment matrix. The weight vector of the comprehensive judgment matrix is ​​calculated using the sum-product method or the square root method to obtain the subjective weights of each indicator using the analytic hierarchy process. The multiple judgment matrices are obtained by using the 1-9 scale method to perform pairwise comparisons and scores on the criterion layer and the indicator layer under each criterion. Feature importance analysis was performed on the BVPTL hybrid prediction model to extract objective weights for machine learning. The subjective weights of the analytic hierarchy process and the objective weights of machine learning are integrated to obtain the comprehensive weights of each evaluation index.

[0011] Furthermore, feature importance analysis is performed on the BVPTL hybrid prediction model to extract objective machine learning weights, including: For the test set samples, calculate the predicted values ​​of the BVPTL mixture prediction model. For the first Input metrics The partial derivatives of the sample and the mean of the absolute values ​​of all samples and time steps are normalized to obtain the first objective weight. Extract the attention weights of the Transformer multi-head attention layer, average the attention weights of each metric across all time steps to obtain the global attention weight of that metric, and use it as the second objective weight. The objective weights for machine learning are obtained based on the first objective weight and the second objective weight.

[0012] Furthermore, the subjective weights of the analytic hierarchy process and the objective weights of machine learning are integrated to obtain the comprehensive weights of each evaluation index, including: Calculate the Kendall coordination coefficient for the subjective weights of the analytic hierarchy process and the objective weights of machine learning. , the calculation formula is: where n is the total number of indicators, is the sum of ranks of indicator j in two weight rankings, is the average value of the sum of ranks of all indicators; When is greater than the set threshold, perform the integration operation; The integration operation includes constructing an optimization model of the weight combination, and solving the optimization model of the weight combination to obtain the comprehensive weight of each indicator; the optimization model of the weight combination is expressed as: where, is the comprehensive weight, is the k-th kind of weight, k = 1 is the subjective weight of the analytic hierarchy process, k = 2 is the objective weight of machine learning, and m is the number of weight sources.

[0013] Furthermore, the method for dividing the current and future safety states of the marine ranch into multiple safety levels according to the current comprehensive safety index and the predicted comprehensive safety index includes: If the comprehensive safety index MSEI ≤ 0.2, determine that the safety level is level 1; If 0.2 < MSEI ≤ 0.4, determine that the safety level is level 2; If 0.4 < MSEI ≤ 0.6, determine that the safety level is level 3; If 0.6 < MSEI ≤ 0.8, determine that the safety level is level 4; If MSEI ≥ 0.8, determine that the safety level is level 5; The corresponding responses and ecological restoration measures for each level are specifically as follows: If the safety level is level 1, immediately initiate the emergency response plan, carry out large-scale ecological restoration projects, and strengthen the monitoring frequency to once a month; If the safety level is level 2, initiate targeted restoration measures, update the monitoring data quarterly, and dynamically adjust the restoration plan; If the safety level is level 3, strengthen monitoring and control, conduct a comprehensive evaluation once every six months, and promptly identify and prevent potential risks; If the safety level is level 4, maintain the existing management measures, conduct a comprehensive evaluation once a year, and slightly adjust the management strategy according to the evaluation results; If the safety level is level 5, summarize the management experience and promote it as a demonstration ranch, conduct a comprehensive evaluation once every two years, and continuously optimize the management mode.

[0014] The method for safety assessment and future safety prediction of typical tropical marine ranches provided in this application has at least the following beneficial effects: (1) In the assessment objective determination step, this application clearly divides the comprehensive safety assessment objective of typical tropical marine ranches into three dimensions: ecological safety, economic safety, and social safety. It also sets evaluation and prediction cycles for three time scales: short-term, medium-term, and long-term. At the same time, it forms spatial and temporal comparison benchmarks by setting up at least two natural control areas and collecting baseline data for at least 1-3 years before construction. The DPSIR three-level indicator system and five-level quantitative standard constructed in this way, which covers the entire chain of driving force-pressure-state-impact-response and includes tropical characteristic indicators such as typhoon frequency and coral bleaching rate, can comprehensively and multi-scale reflect the comprehensive safety status of tropical marine ranches at different time evolution stages and under spatial comparison conditions, overcoming the shortcomings of existing technologies that only focus on short-term ecological benefits.

[0015] (2) In the data preprocessing and prediction stages, this application adopts a BVPTL hybrid prediction model that integrates the BKA-VMD data preprocessing module, the Transformer-LSTM hybrid model, and the PLO hyperparameter optimization algorithm. The VMD decomposition effectively reduces the non-stationarity of the original time series signal, and the PLO algorithm is used to globally optimize the model hyperparameters, thereby more effectively combining the global dependency capture capability of Transformer and the local time series feature extraction capability of LSTM.

[0016] (3) In the weight determination step, this application constructs a game-theoretic combination weighting model based on subjective weights from the Analytic Hierarchy Process (AHP) and objective weights from the BVPTL model's machine learning. The model feature importance parameters output by the BVPTL model (including gradient feature importance and attention weights) are used as objective weights and integrated with the AHP subjective weights through a game-theoretic combination weighting method. This approach combines expert subjective experience with objective information from time-series data, avoiding the subjective arbitrariness in weight determination, making the evaluation results more scientific and credible, and enabling dynamic adaptation to environmental changes.

[0017] (4) In the safety level classification and response steps, this application calculates the current comprehensive safety index based on the comprehensive weight dataset, and simultaneously calculates the predicted comprehensive safety index for future times by combining the future safety situation prediction value output by the BVPTL model, thus classifying the current and future safety status into multiple safety levels. Based on the current safety level and the predicted future safety level, corresponding response and ecological restoration measures are initiated, thereby establishing a dynamic adjustment and early warning mechanism based on real-time monitoring data and model prediction, realizing closed-loop management of assessment-prediction-early warning-feedback, which can promptly identify current and future risks and prompt countermeasures, significantly improving the initiative and emergency response efficiency of marine ranching management.

[0018] (5) This application adopts a weighted adaptive learning method based on the BVPTL model, combined with the expert experience of the analytic hierarchy process, and optimized through game theory. The weight determination process considers both the temporal characteristics and future trends of the indicators and incorporates expert experience, enabling it to dynamically adapt to environmental changes. At the same time, by combining multi-timescale assessment, control area benchmarks, the DPSIR system, and closed-loop response measures, the entire evaluation system has a high degree of scientific rigor, adaptability, and operability, providing strong technical support for the long-term safe operation and adaptive management of typical tropical marine ranches. Attached Figure Description

[0019] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0020] Figure 1 A flowchart illustrating a method for safety assessment and future safety prediction of a typical tropical marine ranch, provided as an embodiment of this application; Figure 2 A flowchart illustrating the process of obtaining a standardized time-series index dataset as provided in this application embodiment; Figure 3 A flowchart illustrating the comprehensive weight calculation of various evaluation indicators provided in the embodiments of this application.

[0021] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concepts of this application to those skilled in the art through reference to specific embodiments. Detailed Implementation

[0022] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0023] The collection, storage, use, processing, transmission, provision, and disclosure of relevant data and information in the technical solution of this application all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0024] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0025] This application provides a method for safety assessment and future safety prediction of a typical tropical marine ranch, such as... Figure 1 As shown, the specific implementation steps are as follows: S10 to S80.

[0026] S10: Determine the evaluation objectives and divide the evaluation and prediction cycles into three time scales: short-term, medium-term, and long-term. Output the benchmark parameters for the evaluation dimensions and the benchmark parameters for the time cycles.

[0027] In this embodiment, the benchmark parameters for the evaluation dimensions are used to define the core objective scope covered by the comprehensive safety assessment of tropical marine ranches, specifically including evaluation objectives in three dimensions: ecological security, economic security, and social security. The meaning of the benchmark parameters for each dimension is as follows: Ecological security objectives: These objectives focus on assessing biodiversity, ecosystem health, and physiological responses to typical environmental stresses to clarify the stability and degradation risk of marine ranch ecosystems. These baseline parameters provide an ecological basis for subsequent indicator selection and evaluation criteria development.

[0028] Economic security objective: To accurately quantify direct and indirect economic benefits and assess the economic sustainability of marine ranches in long-term operation. This benchmark parameter provides a quantitative basis for setting weights and thresholds for economic benefit indicators.

[0029] Social security objectives: A systematic assessment of job creation capacity, public environmental satisfaction, and community participation in governance is used to clarify the marine ranching's support for regional social development. This benchmark parameter provides a reference framework for the evaluation and dynamic adjustment of social impact indicators.

[0030] The time-cycle baseline parameters are used to define the time scales for evaluation and prediction, in order to match the response characteristics and management needs of marine ranch ecosystems at different successional stages. Specifically, they include three time scales: short-term, medium-term, and long-term. The meanings of the baseline parameters for each scale are as follows: The short-term timescale is set at 1-2 years. This phase focuses on monitoring the initial response patterns of the ecosystem after human interventions such as artificial reef deployment and restocking, to rapidly verify the effectiveness of management measures. Monitoring will be conducted quarterly, with data updates and simultaneous calibration of evaluation model parameters.

[0031] The medium-term timescale is set at 3-5 years. This stage comprehensively assesses the effects of resource enhancement and the initial stability of the ecosystem, conducts trend predictions based on initial data, identifies key nodes and potential risks in system succession, and focuses on analyzing the cumulative effects of tropical-specific stressors such as typhoons and high temperatures.

[0032] The long-term timescale is set at 5 years or more. This stage predicts the sustainability of ecosystem steady-state transitions and comprehensive benefits, providing a forward-looking scientific basis for strategic planning, industrial restructuring, and adaptive management. At the same time, the boundary conditions of the prediction model are revised in conjunction with the trend of climate warming.

[0033] S20: Based on the benchmark parameters of the evaluation dimensions and the benchmark parameters of the time period, construct an integrated three-dimensional monitoring system covering driving force, pressure, state, impact and response, and complete the collection of corresponding indicator data. Preprocess all collected indicator data to obtain a standardized time series indicator dataset.

[0034] In this embodiment, an integrated three-dimensional monitoring system covering the entire chain of driving force, pressure, state, impact, and response is constructed to ensure the timeliness, comparability, and completeness of the data. All monitoring data must be standardized before being stored in the database.

[0035] In some embodiments, such as Figure 2 The diagram shown is a flowchart for obtaining the standardized time-series index dataset provided in this application embodiment. Based on the constructed integrated air-space-ground-sea three-dimensional monitoring system covering the entire chain of driving force-pressure-state-impact-response, the standardized time-series index dataset can be obtained through the following steps S201-S206.

[0036] S201: Collect driving force data.

[0037] The driving force data collected in this embodiment includes the following indicators: The proportion of GDP of D1 sea area to the total GDP of the region: GDP data can be obtained from the region's marine economic authorities and the National Bureau of Statistics. D2 Leisure Tourism Development Intensity: The annual number of tourists received is obtained from the cultural and tourism departments, and the area occupied by tourism facilities is calculated by remote sensing. D3 Typhoon Annual Frequency: Historical typhoon data obtained from meteorological departments to quantify high-frequency natural disturbances; D4 Seasonal Peak Tourist Flow: Real-time monitoring of tourist density during holidays, reflecting the intensity of human disturbance; D5 Resident Population Density: The ratio of the annual population to the area, obtained from the local village committee or police station.

[0038] S202: Collect pressure data.

[0039] The pressure data collected in this embodiment includes the following indicators: Industrial wastewater discharge in area P1: obtained from the Environmental Statistics Yearbook of the ecological and environmental protection department; P2 Marine Aquaculture Area: Based on remote sensing image interpretation and administrative registration data; P3 The proportion of coastal waters with water quality meeting Class I and II standards: Water samples are collected monthly at fixed points to measure pH, DO, DIN, DIP, COD, and heavy metals, etc. Spatial uniformity control of sampling points is added (the distance between adjacent sampling points is ≥3km, and the density is increased to 1-2km in key ecological areas). Combined with satellite remote sensing image interpretation (covering the entire evaluation sea area), the proportion of sea areas meeting Class I and II standards is calculated by remote sensing inversion area method. P4 sediment pollution levels conform to the mass-area ratio of Class I and II marine sediments: sediment column samples are collected during the dry and wet seasons each year, and the TOC, TN, TP and heavy metal contents are determined by remote sensing inversion of area. P5 Typhoon Disturbance Intensity: Calculated based on typhoon data, including maximum wind speed, rainfall, and storm surge height. P6 Red Tide Outbreak Frequency and Area: Recorded using satellite remote sensing and on-site monitoring; P7 Diving Activity Intensity: Number of divers per unit coral reef area, quantifying human interference.

[0040] S203: Collect status data.

[0041] The status data collected in this embodiment includes the following indicator data: S1 marine primary productivity: verification by remote sensing inversion of chlorophyll a concentration combined with in-situ fluorescence method; S2 seawater seedling count: trawling sampling and microscopic counting of planktonic organisms; S3 Seafood Production: Enterprise Production Logs and Fisheries Department Statistics; S4 reef-building coral coverage and S5 key fish species abundance: underwater field survey (line-point interception method) and underwater video cross-section method; S6 Coral Bleaching Rate: Regular diving surveys are conducted to investigate the bleaching rate, reflecting the effects of high-temperature stress; S7 zooxanthellae density decline rate: laboratory analysis of coral samples to characterize the health of the symbiotic system.

[0042] S204: Collect impact data.

[0043] The impact data collected in this embodiment includes the following indicator data: Total marine output value after the disaster in I1: Inquire with the marine economic authorities of the region.

[0044] Number of talents introduced to I2 region: Check the website of the social security department of the region. Number of job openings in I3 region: Check the website of the government human resources and social security department.

[0045] Public environmental satisfaction in I4 area: Questionnaire survey and statistical bureau sampling data. The effective return rate of the questionnaire must be ≥80%, and the sampling covers fishermen at different income levels.

[0046] S205: Collect response data.

[0047] The response data collected in this embodiment includes the following metrics: Ecological restoration investment in R1 area includes special funds for artificial reefs, sewage treatment, coral transplantation, and restocking.

[0048] R2 Onshore observation station coverage density: the ratio of the number of observation stations to the sea area.

[0049] Annual number of fishermen receiving technical training in R3 region and number of marine fishery workers in R4 region: Statistics from the human resources and social security department and fishery association.

[0050] R5 Disaster Early Warning and Response Efficiency: Event Response Time and Effectiveness Assessment.

[0051] R6 Tourism Capacity Control Measures: Implementation status, such as tourist diversion and area closure.

[0052] R7 Marine Research Funding: Access financial statements from governments at all levels.

[0053] S206: Preprocess the collected data.

[0054] Min-max standardization is applied to all indicator data to eliminate dimensional differences. The formula is as follows: in, For the first The first sample Individual indicator values, and The first The minimum and maximum values ​​of each indicator. For the first The first sample A standardized time series index value.

[0055] For missing data, interpolation methods are used to supplement it (linear interpolation for time series data and Kriging interpolation for spatial data). For indicators with a missing rate of more than 20%, the data collection plan needs to be re-evaluated.

[0056] S30: Based on time-cycle benchmark parameters and standardized time-series index datasets, at least two natural control areas are established in adjacent sea areas that have not undergone artificial construction. At the same time, baseline environmental and biological resource data for at least 1 to 3 consecutive years before the construction of marine ranches are collected to form spatial control benchmarks and time-series control benchmarks, thereby obtaining an evaluation quantitative benchmark system.

[0057] Spatial baselines are used to eliminate the interference of spatial heterogeneity on evaluation results and provide a natural environmental background comparable to the marine ranching construction area. The specific establishment method is as follows: A natural control area is established in adjacent sea areas that have not undergone artificial construction. The environmental background of this area is highly similar to that of the core area of ​​the marine ranching, with differences in water quality, substrate type, and water depth not exceeding 10%. Simultaneously, an excellent marine ranching demonstration area that has passed acceptance is selected as a supplementary control. There are no fewer than two natural control areas to avoid the randomness brought about by a single control. The monitoring indicators and monitoring frequencies of the natural control areas and the excellent marine ranching demonstration areas are consistent with those of the marine ranching construction area. The resulting spatial baseline is used to quantify the net effect of artificial intervention measures relative to the natural state.

[0058] Time-series benchmarks are used to eliminate the interference of temporal evolution trends on evaluation results and provide historical baseline data before the construction of marine ranches. The specific establishment method is as follows: collect baseline environmental and biological resource data for at least 1-3 consecutive years prior to the construction of the marine ranch, including key indicators such as water quality, sediment, and biological communities. Data within this period constitutes a reliable time-series benchmark, used for longitudinal comparison with post-construction monitoring data to identify the actual changes in the ecosystem caused by human intervention.

[0059] By integrating the aforementioned spatial and time-series benchmarks, a unified quantitative evaluation benchmark system is formed. This system provides a benchmark reference for the subsequent formulation of quantitative standards for safety evaluation models, the classification of safety levels, and the initiation of dynamic response measures.

[0060] S40: Construct a BVPTL hybrid prediction model that integrates the BKA-VMD data preprocessing module, the Transformer-LSTM hybrid model, and the PLO hyperparameter optimization algorithm. Train the BVPTL hybrid prediction model based on standardized time series index data. Based on the trained BVPTL hybrid prediction model, obtain the predicted values ​​of the security status of tropical marine ranches in the future time scale and the model feature importance parameters.

[0061] To achieve accurate prediction of the safety status of marine ranches, this evaluation adopts the BVPTL (BKA-VMD-PLO-Transformer-LSTM) hybrid prediction model (hereinafter referred to as the BVPTL model). The BVPTL model integrates the BKA-VMD data preprocessing module, the PLO optimization algorithm, and the Transformer-LSTM hybrid model. It combines the global dependency capture capability of Transformer with the local temporal feature extraction capability of LSTM, and improves the global optimality of the model's hyperparameters through intelligent optimization algorithms, thereby improving the accuracy of long-term prediction.

[0062] The BVPTL model adopts a three-stage structure of BKA-VMD data preprocessing, PLO optimization, and Transformer-LSTM cascade fusion. The data processing flow of each structure is as follows.

[0063] The BKA-VMD data preprocessing module is used to adaptively decompose the original time series data (i.e., the standard time series index data processed in step S20), extract purer signal features, and reduce the impact of the non-stationarity of the original data on prediction accuracy.

[0064] The BKA-VMD data preprocessing module will input signals Adaptive decomposition into One intrinsic mode function (IMF) Essentially, it involves solving a constrained variational optimization problem: in, To perform the minimum value operation, For modal number index, The penalty coefficient is... Let k be the eigenmode function. For the first The central angular frequencies corresponding to each eigenmode function For time variables The first-order partial derivative operator, For Dirac delta function, is the base of the natural logarithm. The imaginary unit, It is a time variable; The constraints are: BKA Optimizes VMD Parameters: The Black Kite Algorithm (BKA) is used to optimize two key hyperparameters of VMD—the number of modes. and penalty coefficient Optimize. Encode the data as a two-dimensional vector, and search for the optimal solution within a preset boundary, using the sum of the sample entropies of the decomposed subsequences as the fitness function: in, Let be the objective function used to evaluate the performance of variational mode decomposition (VMD) parameter optimization, and let its value be the sum of the sample entropies of all intrinsic mode functions (IMFs) obtained from the decomposition. This represents the k-th eigenmode function obtained from VMD decomposition. Calculated sample entropy.

[0065] The optimization is achieved through iterative analysis of BKA's migration and attack behaviors until a preset maximum number of iterations is reached. or continuous Optimization stops when the fitness value no longer decreases, and the optimal parameter combination is finally output. The original signal is then decomposed using this parameter to obtain... The IMF components are used as inputs to the subsequent Transformer-LSTM hybrid model.

[0066] The Transformer-LSTM hybrid model takes the IMF components decomposed by the BKA-VMD data preprocessing module as input and uses a cascaded fusion structure to extract temporal features. The cascaded fusion structure includes LSTM and Transformer branches.

[0067] The LSTM branch uses a single-layer LSTM model to extract local temporal features from the input sequence, using the last hidden state as the representative of the local features. The input dimension of the LSTM is the number of indicators, the hidden layer dimension is set to 64, and the batch processing mode is set to batch_first=True.

[0068] The Transformer branch first uses a linear mapping to upscale the input data to a high-dimensional space, then adds positional encoding (adding order information to the input data to address the Transformer's inability to capture temporal order), then extracts global dependency features through the Transformer encoder, and finally uses global average pooling to obtain global feature representations. The Transformer encoder uses two encoder layers, with four multi-head attention heads and a model dimension of 64.

[0069] Finally, the local features of the LSTM branch and the global features of the Transformer branch are concatenated by the fusion output layer, and then mapped to the prediction dimension through a fully connected layer to achieve the prediction of future security situation.

[0070] To improve the prediction performance of the Transformer-LSTM hybrid model, the Aurora Optimization (PLO) algorithm is used to globally optimize the key hyperparameters of the model. The optimization process is as follows: Determine the target parameters for optimization. The PLO algorithm directly optimizes the three core hyperparameters of the Transformer-LSTM hybrid model: The number of heads in the self-attention mechanism directly affects the model's parallel computing capability and the granularity of dependency capture. Initial learning rate: controls the step size of parameter updates and is crucial to training stability and convergence speed; L2 regularization coefficient: Balances model fitting ability and generalization performance.

[0071] The optimization objective of PLO is to minimize the mean absolute percentage error (MAPE), which is expressed as: in, The total number of test samples, For the first The true value of each sample These are the model predictions. MAPE is sensitive to relative error and is particularly suitable for time-series tasks where the magnitude of the predicted values ​​varies greatly.

[0072] The PLO algorithm performs global optimization by simulating the particle motion mechanism during aurora formation. Its core sub-models include: Gyration Motion: Simulates the damped helical motion of charged particles in the Earth's magnetic field, whose dynamics are characterized by a modified first-order linear differential equation. in The particle velocity vector Damping factor (range of values) ), The particle charge. For particle mass, The strength of Earth's magnetic field; for The particle velocity vector at time t. Let be the integration constant. is the base of the natural logarithm.

[0073] The general solution of the first-order linear differential equation is: This demonstrates the physical law that velocity decays exponentially with time. - Aurora Oval Walk: Based on Levy flight simulations, this involves chaotic fluctuations of high-energy particles within the boundary of an aurora ellipse. Its displacement update formula is: in, This represents the particle displacement vector generated by the elliptical aurora walk. Let be the current position vector of the i-th particle in the d-dimensional solution space. The random numbers are uniformly distributed in the interval [0,1]. , To solve for the upper and lower bounds of space, for The random step size of the Levy distribution is calculated as follows: in, To conform to a mean of 0 and a variance of σ 2 A normally distributed random variable, These are the stability parameters for Levy flight. , It follows a normal distribution (Gaussian distribution). Let N(0,σ) be the standard deviation of the normal distribution. , This is a gamma function.

[0074] Adaptive weight fusion mechanism: PLO employs time-varying weight dynamic balancing exploration and development capabilities. in, These are the time-varying weighting coefficients for the rotational motion components. The hyperbolic tangent activation function is used. The number of times the current function is evaluated. The maximum number of function evaluations is preset. For the time-varying weighting coefficients of the aurora elliptical travel component, It is a natural exponential function.

[0075] The main formula for position update is: in For rotational motion components, their amplitudes are calculated as follows: Dynamic decay, , This is the attenuation adjustment coefficient. Let i be the updated position vector of the i-th particle. Let be the current position vector of the i-th particle in the d-dimensional solution space. This is the particle displacement vector generated by rotational motion. Let be a d-dimensional random vector that follows a uniform distribution in the interval [0,1]. For Hadama accumulation.

[0076] Particle Collision: Simulates chaotic collisions between solar wind particles, enhancing the ability to escape local optima. Its triggering conditions and position update rules are as follows: Triggering Condition: If the currently generated random number... satisfy and If so, a collision operation is performed. This is the collision probability control factor, which is dynamically adjusted as the iteration progresses. The calculation formula is: In the formula, This represents the number of times the function has been evaluated (i.e., the total number of fitness calculations performed in the current iteration). The maximum number of function evaluations is preset for the algorithm. As iterations progress, Gradually increase Value from Towards The incremental nature of the algorithm makes it easier to trigger particle collisions in later stages. Position perturbation formula: When the triggering condition is met, the particle position is updated using the following formula: in, For the current particle In the The position of the dimension for random arrangement ( (for population size) Specifying and Particles The particle numbers that collided. This represents the disturbance intensity factor.

[0077] The PLO algorithm iteratively searches until a preset maximum number of function evaluations is reached. The final output is the optimal combination of hyperparameters that minimizes MAPE, which serves as the final configuration of the Transformer-LSTM model.

[0078] This embodiment trains the BVPTL hybrid prediction model as follows: Time-series data is divided into training, validation, and test sets in a 7:2:1 ratio. The training set is used for model parameter learning, the validation set for hyperparameter tuning, and the test set for model performance evaluation. An L2 regularization term is added to the mean squared error (MSE) as the loss function. The L2 regularization term is added to the loss function to control model complexity. The regularization coefficient... Determined by the PLO algorithm, its expression is: in, For the true value, For predicted values, The regularization coefficient is determined using the PLO hyperparameter optimization algorithm. For the model number One weight parameter, is the number of samples in the batch, and i is the sample index.

[0079] The optimizer uses the Adam optimizer, the initial learning rate is determined by the PLO algorithm, and a learning rate decay strategy is adopted, which reduces the learning rate to 0.9 every 10 epochs.

[0080] During training, data is fed to the model in batches. First, forward propagation is performed, where batches of data are input into the model to calculate the predicted safety posture value. Then, a loss function is used to compare the predicted value with the true label. Next, backpropagation is performed, and the model automatically identifies the source of error. Finally, the optimizer adjusts the parameters based on the error. The loss is accumulated during training, and the validation set loss is calculated every 5 epochs. Training stops when the validation set loss no longer decreases for 10 consecutive epochs to avoid overfitting.

[0081] This embodiment uses the following determination coefficients ( Evaluate the model's predictive performance: in, The average of the true values. The model is required to... A value of ≥0.85 ensures that the prediction accuracy meets the evaluation requirements.

[0082] S50: Based on the evaluation dimension benchmark parameters, standardized time series index dataset, and evaluation quantification benchmark system, a three-level DPSIR index system and a five-level quantification standard for the safety evaluation of tropical marine ranches are established, the calculation method of the comprehensive safety index is determined, and the safety evaluation model framework is obtained.

[0083] In this embodiment, the DPSIR three-level index system for the safety assessment of tropical marine ranches is shown in Table 1.

[0084] Table 1. DPSIR Three-Level Index System for Safety Assessment of Tropical Marine Ranches

[0085] Based on the DPSIR three-level indicator system shown in Table 1, a quantitative standard for safety evaluation indicators of typical tropical marine ranches is established. Specifically, the following indicators are used to evaluate the marine ranch construction area over time: D1 (proportion of marine GDP to regional GDP), D2 (intensity of leisure tourism development), D4 (seasonal peak of tourist traffic), D5 (resident population density), P1 (regional industrial wastewater discharge), P2 (marine aquaculture area), P6 (frequency and area of ​​red tide outbreaks), P7 (intensity of diving activities), S1 (marine primary productivity), S2 (number of marine seedlings), S3 (seafood production), S4 (coverage of reef-building corals), S5 (number of key fish species), S6 (coral bleaching rate), and S7 (rate of decline in zooxanthellae density). These indicators are compared with those of natural control areas and excellent demonstration areas of similar marine ranches. The changes in the evaluation indicators of the natural control area and the marine ranch demonstration area were compared. The following indicators were used as standard values: D3 annual typhoon frequency, I1 total marine output value after disasters, I2 number of talents introduced to the area, I4 public environmental satisfaction, R1 ecological restoration investment, R2 onshore observation station coverage density, R4 number of marine fishery workers, R5 disaster early warning and response efficiency, R6 tourism carrying capacity control measures, R7 marine scientific research funding, P3 proportion of coastal waters meeting Class I and II seawater quality standards, P4 proportion of sediment pollution meeting Class I and II marine sediment quality standards, and P5 typhoon disturbance intensity. A five-level quantification was performed, as shown in Table 2.

[0086] Table 2 Safety Evaluation Index System and Quantitative Standards for Tropical Marine Ranches

[0087] The Marine Ranching Security Index (MSEI) is calculated using a weighted summation method. The formula is as follows: in, The total number of specific indicators (in this embodiment) ), For the first The combined weight of each indicator For the first The standardized value of each indicator.

[0088] S60: Construct a game-theoretic combined weighting model based on the weights of machine learning using the Analytic Hierarchy Process (AHP) and the BVPTL model. Calculate the comprehensive weights of each evaluation index according to the security evaluation model framework, the standardized time-series index dataset, and the importance parameters of the model features, and output the comprehensive weight dataset.

[0089] In some embodiments, such as Figure 3 The diagram shown is a flowchart for calculating the comprehensive weight of each evaluation indicator provided in an embodiment of this application. Step S60 can be performed by calculating the comprehensive weight of each evaluation indicator through the following steps S601-603.

[0090] S601: Calculate the subjective weights of the analytic hierarchy process, including the following steps S6011-S6015.

[0091] S6011: Construct a hierarchical model, decomposing the evaluation objective into three levels, as follows: Target layer (highest layer): Comprehensive evaluation of the construction effects of artificial reef-type marine ranches; Criterion layer (middle layer): driving force data, pressure data, status and resource data, impact data, response data; Indicator layer (bottom layer): 30 specific indicators under each criterion.

[0092] S6012: Construct a comprehensive judgment matrix.

[0093] In practical implementation, 7–15 experts in marine ecology, environmental science, and fisheries economics can be invited. Expert selection criteria should include holding a senior professional title or above, having over 5 years of experience in tropical marine research, and having industry backgrounds covering academic, regulatory, and practical fields. The recommended number of experts is 10–12. Multiple rounds of feedback should be conducted using the Delphi method to converge opinions and avoid individual bias. Pairwise comparisons should be made on the five criteria: driving force, pressure, state and resources, impact, and response, using a 1-9 scale for scoring to form a judgment matrix. The expert scores should be aggregated, and the geometric or arithmetic mean of the judgment matrices from multiple experts should be taken to form a comprehensive judgment matrix.

[0094] S6013: Calculate the weight vector.

[0095] This embodiment presents two methods for calculating the weight vector as follows: Method 1, Sum-product method: Normalize the judgment matrix by column; sum by row and normalize again to obtain the weight vector; Method 2, Root Mean Method (suitable for programming implementation): Calculate the geometric mean of each row of elements; normalize to obtain the weights.

[0096] The weight vectors calculated by the two methods above are the subjective weights of the corresponding indicators in the analytic hierarchy process.

[0097] S6014: Consistency check.

[0098] Calculate the consistency index (CI): in Calculate element by element and then sum. To determine the order of a matrix, the criterion layer =5, Indicator Layer This refers to the number of indicators under the corresponding criteria.

[0099] The random consistency index (RI, RI = 0.58 when n = 3) is obtained from the table; the consistency ratio is calculated. If CR < 0.1, the test passes; otherwise, the judgment matrix needs to be adjusted.

[0100] S6015: Calculate the weights of the indicator layer.

[0101] Repeat steps S6012-S6014 to calculate the weights of specific indicators under each criterion, ultimately forming a comprehensive weight table. The indicator layer weights are as follows: .

[0102] S602: Extract machine learning weights from the BVPTL model.

[0103] Objective weights are extracted through feature importance analysis using the BVPTL model. The specific steps are as follows: S6021: Calculate the first objective weight based on gradient-based feature importance.

[0104] Calculate the mean gradient of the BVPTL model output with respect to each input metric. The larger the absolute value of the gradient, the greater the influence of that metric on the model output, i.e., the higher its weight. Specifically, for the test set samples, calculate the model's predicted values. For the first Input metrics The partial derivatives of the first objective weights are obtained by taking the mean of their absolute values ​​over all samples and time steps and normalizing them. : in, For the sample size, For time step, For the total number of indicators, For the first The first sample The model output at the time step is for the . The partial derivatives of each indicator.

[0105] S6022: Calculate the second objective weight based on the feature importance of attention weight.

[0106] The attention weights of the Transformer multi-head attention layer are extracted, and the attention weights of each metric are averaged across all time steps to obtain the global attention weight of that metric, which serves as the second objective weight and complements the first objective weight.

[0107] S6023: Obtain the machine learning objective weights based on the first objective weight and the second objective weight.

[0108] This embodiment fuses the first objective weight and the second objective weight to obtain the final machine learning objective weight. The fusion method can be either an arithmetic average or a weighted average. In the case of a weighted average, the specific weight coefficients can be adjusted based on the prediction accuracy of the validation set. The resulting machine learning objective weight comprehensively reflects the contribution of each indicator to the model's prediction and is entirely data-driven, avoiding interference from subjective factors.

[0109] S603: Weight integration.

[0110] Integrating subjective weights using game theory combinatorial weighting method ( ) and machine learning objective weights ( The specific steps are as follows: S6031: Construct an optimization model for the weight combination, expressed as: in, For comprehensive weighting, For the first Type weight ( For AHP weights, (For machine learning weights).

[0111] S6032: Solve the optimization model to obtain the comprehensive weights. This ensures that the overall weighting takes into account both expert experience and the objective characteristics of the data.

[0112] S6033: Weight consistency test.

[0113] Calculate the Kendall coordination coefficient for the two weights (AHP weights and machine learning weights). To assess its ordering consistency, the formula is: in, The number of weight sources (i.e., AHP weights and machine learning weights). For the total number of indicators, As an indicator The sum of ranks in two weighted sorts (i.e., the sum of ranks under the two weights). This is the average of the rank sums of all indicators. When If the two sources of weights are considered to have good consistency, then the reasons for the inconsistency need to be analyzed and the weight calculation process adjusted.

[0114] In some embodiments, a dynamic update mechanism is also designed when constructing the comprehensive weight dataset. This dynamic update mechanism specifically involves setting a weight update cycle and updating the weights based on new monitoring data. Short-term updates (quarterly): Only deep learning weights are updated. Keeping the AHP weights unchanged, this approach is suitable for scenarios where indicator characteristics fluctuate in the short term. Mid-term update (annually): Update deep learning weights And conduct a new expert survey, adjust the judgment matrix for AHP weights, and update the AHP weights; Long-term updates (every 3 years): Reconstruct the evaluation index system, recalculate AHP weights and extract machine learning weights to ensure that the evaluation system adapts to the long-term evolution of marine ranches.

[0115] S70: Based on the comprehensive weighted dataset, calculate the current comprehensive safety index of the tropical marine ranch, and combine it with the future safety situation prediction value to calculate the predicted comprehensive safety index for future times. Based on the current comprehensive safety index and the predicted comprehensive safety index, divide the current and future safety status of the marine ranch into multiple safety levels to obtain the current safety level and the future predicted safety level.

[0116] This embodiment establishes an adaptive management mechanism based on monitoring data and model predictions through step S70. Based on the comprehensive ecological security assessment results (divided into levels 1-5), different levels of response and remediation measures are initiated to achieve precise management.

[0117] In some embodiments, the safety status of marine ranches is divided into 5 levels based on a comprehensive safety index, as shown in Table 3.

[0118] Table 3 Safety Level Classification Standards

[0119] In practice, after obtaining the grading results, the scores of each criterion layer and indicator layer can be displayed using visualization methods such as heat maps and radar charts, which can intuitively present the safety status and shortcomings of marine ranches and provide a clear reference for management decisions.

[0120] S80: Based on the current security level and the predicted future security level, initiate corresponding level of response and ecological restoration measures.

[0121] This embodiment establishes an adaptive management mechanism based on monitoring data and model predictions through step S80. The security situation prediction value for future time points from the BVPTL model in step 4 is input into the comprehensive security index calculation process to obtain the predicted security level for future times. Based on the dynamic changes between the current evaluation level and the predicted level, corresponding response and remediation measures are initiated. a) Level 1 (extremely unsafe): Immediately activate the emergency response plan, suspend human activities that may cause further impact (such as tourism and fishing), carry out large-scale ecological restoration projects (such as coral transplantation and artificial reef deployment), and increase the monitoring frequency to once a month.

[0122] b) Level 2 (Unsafe): Initiate targeted remediation measures, such as reducing aquaculture density, strengthening wastewater treatment, and carrying out stock enhancement and release. Update monitoring data quarterly and adjust the remediation plan accordingly.

[0123] c) Level 3 (Critical Safety): Strengthen monitoring and control, such as restricting tourist flow, optimizing aquaculture models, conducting a comprehensive evaluation every six months, and promptly identifying potential risks.

[0124] d) Level 4 (Safety): Maintain existing management measures, conduct a comprehensive evaluation annually, and fine-tune management strategies based on the evaluation results.

[0125] e) Level 5 (Extremely Safe): Summarize management experience, promote it as a model ranch, conduct a comprehensive evaluation every 2 years, and continuously optimize the management model.

[0126] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A method for safety assessment and future safety trend prediction of a typical tropical marine ranch, characterized in that, The method includes: The assessment objectives are defined, and the evaluation and prediction cycles are divided into three time scales: short-term, medium-term, and long-term. The benchmark parameters for the evaluation dimensions and the benchmark parameters for the time cycles are output. The assessment objectives include the evaluation objectives for the three dimensions of ecological security, economic security, and social security in the comprehensive safety assessment of tropical marine ranches. Based on the aforementioned evaluation dimension benchmark parameters and time period benchmark parameters, an integrated three-dimensional monitoring system covering driving force, pressure, state, impact, and response is constructed, and corresponding indicator data are collected. All collected indicator data are preprocessed to obtain a standardized time series indicator dataset. Based on the aforementioned time period benchmark parameters and standardized time series index dataset, at least two natural control areas are established in adjacent sea areas that have not undergone artificial construction. At the same time, baseline environmental and biological resource data for at least 1 to 3 consecutive years before the construction of marine ranches are collected to form spatial control benchmarks and time series control benchmarks, thereby obtaining an evaluation quantitative benchmark system. A BVPTL hybrid prediction model integrating a BKA-VMD data preprocessing module, a Transformer-LSTM hybrid model, and a PLO hyperparameter optimization algorithm is constructed. The BVPTL hybrid prediction model is trained based on the standardized time series index data. Based on the trained BVPTL hybrid prediction model, the predicted values ​​of the security status of tropical marine ranches in the future time scale and the model feature importance parameters are obtained. Based on the aforementioned evaluation dimension benchmark parameters, standardized time series index dataset, and evaluation quantification benchmark system, a three-level DPSIR index system and a five-level quantification standard for tropical marine ranch safety evaluation are established, the calculation method for the comprehensive safety index is determined, and a safety evaluation model framework is obtained. A game-theoretic weighting model based on the Analytic Hierarchy Process (AHP) and BVPTL model machine learning weights is constructed. According to the security evaluation model framework, the standardized time series index dataset, and the model feature importance parameters, the comprehensive weight of each evaluation index is calculated, and the comprehensive weight dataset is output. Based on the comprehensive weighted dataset, the current comprehensive safety index of the tropical marine ranch is calculated, and the predicted comprehensive safety index for future times is calculated in combination with the predicted future safety situation value. Based on the current comprehensive safety index and the predicted comprehensive safety index, the current and future safety status of the marine ranch is divided into multiple safety levels, and the current safety level and the predicted future safety level are obtained. Based on the current security level and the predicted future security level, corresponding response and ecological restoration measures will be initiated.

2. The method according to claim 1, characterized in that, All collected indicator data are preprocessed to obtain a standardized time-series indicator dataset, including: Min-max standardization is applied to all indicator data to eliminate dimensional differences. The formula is as follows: in, For the first The first sample Individual indicator values, and The first The minimum and maximum values ​​of each indicator. For the first The first sample A standardized time-series index value; For missing data, linear interpolation is used for time series data and kriging interpolation is used for spatial data. Indicators with a missing rate of more than 20% are re-collected.

3. The method according to claim 1, characterized in that, The BKA-VMD data preprocessing module is used to adaptively decompose the input standardized time-series index dataset. The specific process includes: Using each standardized time series index in the standardized time series index dataset as the input signal, variational mode decomposition is used to transform the input signal. Adaptive decomposition into eigenmode functions Solve the constrained variational optimization problem shown below: in, To perform the minimum value operation, For modal number index, The penalty coefficient is... Let k be the eigenmode function. For the first The central angular frequencies corresponding to each eigenmode function For time variables The first-order partial derivative operator, For Dirac delta function, is the base of the natural logarithm. The imaginary unit, It is a time variable; The constraints are: The number of modes in variational mode decomposition and penalty coefficient Optimize, The code is encoded as a two-dimensional vector. An optimal solution is searched within a preset boundary, with the fitness function being the sum of the sample entropies of the decomposed subsequences. The fitness function is expressed as: in, Let be the objective function used to evaluate the optimization effect of variational mode decomposition parameters, and let its value be the sum of the sample entropies of all intrinsic mode functions obtained by decomposition. This represents the k-th eigenmode function obtained from VMD decomposition. Calculated sample entropy; The optimization is achieved through iterative analysis of migration and attack behaviors until a preset maximum number of iterations is reached. or continuous Optimization stops when the fitness value no longer decreases, and the optimal parameter combination is output. The original signal is then decomposed using the parameter combination to obtain... One IMF component is used as the input to the Transformer-LSTM hybrid model; The Transformer-LSTM hybrid model takes the IMF component sequence obtained after decomposition by the BKA-VMD data preprocessing module as input and uses a cascaded fusion structure to extract temporal features; the cascaded fusion structure includes an LSTM branch and a Transformer branch; The LSTM branch uses a single-layer LSTM model to extract local temporal features from the IMF component sequence, and takes the last hidden state as a local feature. The Transformer branch first uses a linear mapping to upscale each component data in the IMF component sequence to a high-dimensional space, and adds position encoding to provide order information for each component in the IMF component sequence. The global dependency features are extracted by the Transformer encoder, and global average pooling is used to obtain global features. By fusing the output layer, the local features of the LSTM branch and the global features of the Transformer branch are concatenated, and then mapped to the prediction dimension through a fully connected layer to achieve the prediction of future security situation.

4. The method according to claim 1, characterized in that, In the BVPTL hybrid prediction model, the PLO hyperparameter optimization algorithm is used to globally optimize the key hyperparameters of the Transformer-LSTM hybrid model, train the model, and perform prediction validation. Its specific process includes: Using the number of heads, initial learning rate, and L2 regularization coefficient in the self-attention mechanism of the Transformer-LSTM hybrid model as optimization target parameters, and minimizing the mean absolute percentage error (MAPE) as the optimization objective, global optimization is performed by simulating the particle motion mechanism in the aurora formation process. The core steps include rotational motion, aurora elliptical walking, adaptive weight fusion mechanism, and particle collision. The rotational motion includes: simulating the damped spiral motion of a charged particle in the Earth's magnetic field, the dynamics of which are characterized by a modified first-order linear differential equation, which is expressed as: in, The particle velocity vector The damping factor, The particle charge. For particle mass, Given the Earth's magnetic field strength, the general solution of this modified first-order linear differential equation is: ; Let be the particle velocity vector at time t. Let be the integration constant. is the base of the natural logarithm; The aurora elliptical walk includes: chaotic fluctuations of high-energy particles within the boundary of the aurora ellipse based on Levy flight simulation, wherein the particle displacement update formula is: in, This represents the particle displacement vector generated by the elliptical aurora walk. Let be the current position vector of the i-th particle in the d-dimensional solution space. The random numbers are uniformly distributed in the interval [0,1]. This is the current location of the population centroid. , To solve for the upper and lower bounds of the space; Let d be the random step size of the Levy distribution, calculated using the following formula: in, To conform to a mean of 0 and a variance of σ 2 A normally distributed random variable, These are the stability parameters for Levy flight. , It follows a normal distribution (Gaussian distribution). Let N(0,σ) be the standard deviation of the normal distribution. , It is a gamma function; The adaptive weight fusion mechanism includes determining time-varying weights using the following formula: in, These are the time-varying weighting coefficients for the rotational motion components. The hyperbolic tangent activation function is used. The number of times the current function is evaluated. The maximum number of function evaluations is preset. For the time-varying weighting coefficients of the aurora elliptical travel component, It is a natural exponential function; Based on the determined time-varying weights, the position is updated using the following formula: in, , is the rotational motion component, and the amplitude of the rotational motion component is according to Dynamic decay, This is the attenuation adjustment coefficient. Let i be the updated position vector of the i-th particle. Let be the current position vector of the i-th particle in the d-dimensional solution space. This represents the particle displacement vector generated by rotational motion. Let be a d-dimensional random vector that follows a uniform distribution in the interval [0,1]. For Hadamah accumulation; The particle collisions include: simulated chaotic collisions between solar wind particles, triggered by the currently generated random number rand satisfying the following condition. and If the triggering condition is met, a collision operation is performed. The collision probability control factor is calculated using the following formula: When the trigger condition is met, the particle position is updated as follows: in, For the current particle In the The position of the dimension for A random arrangement, where N is the population size. To specify with particles The particle numbers that collided; Iterative search continues until the preset maximum number of function evaluations MaxFEs is reached, and the optimal combination of hyperparameters that minimizes MAPE is output as the configuration of the Transformer-LSTM model.

5. The method according to claim 1, characterized in that, In the BVPTL hybrid prediction model, the methods for training the BVPTL hybrid prediction model based on the standardized time-series index data include: The standardized time-series index dataset is divided into training set, validation set and test set according to a set ratio; Based on a regularized loss function, the training process sequentially performs forward propagation, loss calculation, backpropagation, and parameter update. The loss is accumulated during training, and training stops when the validation set loss no longer decreases for a set number of consecutive epochs. The regularized loss function is... Represented as: in, For the true value, For predicted values, The regularization coefficient is determined using the PLO hyperparameter optimization algorithm. For the model number One weight parameter, Where j is the number of samples in the batch, and j is the sample index; Based on the test set, the coefficient of determination is used to evaluate the model's predictive performance. Once the model's predictive performance reaches the set standard, the model obtained from the current training is used as the trained BVPTL hybrid prediction model.

6. The method according to claim 1, characterized in that, The DPSIR three - level index system is divided into three levels: the target layer, the criterion layer, and the index layer. Among them, the target layer is the safety assessment of tropical marine ranches; the criterion layer includes 5 parallel sub - layers: driving force, pressure, state and resources, impact, and response; the index layer includes specific indicators corresponding to each parallel sub - layer, and each sub - layer corresponds to at least one specific indicator. The calculation method of the comprehensive safety index is as follows: in, This refers to the total number of specific indicators. The comprehensive weight of the j-th indicator is... Let be the standardized value of the j-th indicator.

7. The method according to claim 6, characterized in that, The specific methods for calculating the comprehensive weights of each evaluation index include: Obtain multiple judgment matrices, take the geometric mean or arithmetic mean of the multiple judgment matrices to form a comprehensive judgment matrix; use the sum - product method or the root - extraction method to calculate the weight vector of the comprehensive judgment matrix to obtain the subjective weights of each index by the analytic hierarchy process. Among them, the multiple judgment matrices are obtained by pairwise comparison and scoring of the criterion layer and the index layer under each criterion using the 1 - 9 scale method. Extract the objective weights of machine learning by performing feature importance analysis on the BVPTL hybrid prediction model. Integrate the subjective weights of the analytic hierarchy process and the objective weights of machine learning to obtain the comprehensive weights of each evaluation index.

8. The method according to claim 7, characterized in that, Extracting the objective weights of machine learning by performing feature importance analysis on the BVPTL hybrid prediction model includes: For the test set samples, calculate the predicted values ​​of the BVPTL mixture prediction model. For the first Input metrics The partial derivatives of the sample and the mean of the absolute values ​​of all samples and time steps are normalized to obtain the first objective weight. Extract the attention weights of the Transformer multi - head attention layer, average the attention weights of each index at all time steps to obtain the global attention weight of the index, and use it as the second objective weight. Based on the first objective weight and the second objective weight, obtain the objective weights of machine learning.

9. The method according to claim 7, characterized in that, Integrating the subjective weights of the analytic hierarchy process and the objective weights of machine learning to obtain the comprehensive weights of each evaluation index includes: Calculate the Kendall coordination coefficient for the subjective weights of the analytic hierarchy process and the objective weights of machine learning. The calculation formula is: Where n is the total number of indicators. Let j be the rank sum of the index j in the two weighted rankings. This is the average of the rank sums of all indicators; when If the value exceeds the set threshold, an integration operation will be performed. The integration operation includes constructing an optimization model for the weight combination and solving the optimization model of the weight combination to obtain the comprehensive weights of each index. The optimization model of the weight combination is expressed as: in, The comprehensive weight of the j-th indicator is... Let be the k-th weight, k=1 be the subjective weight of the analytic hierarchy process, k=2 be the objective weight of machine learning, and m be the number of weight sources.

10. The method according to claim 1, characterized in that, The method for dividing the safety status of the marine ranch at the current and future moments into multiple safety levels according to the current comprehensive safety index and the predicted comprehensive safety index includes: If the comprehensive safety index MSEI ≤ 0.2, then determine the safety level as level 1. If 0.2 < MSEI ≤ 0.4, then determine the safety level as level 2. If 0.4 < MSEI ≤ 0.6, then determine the safety level as level 3. If 0.6 < MSEI ≤ 0.8, then determine the safety level as level 4. If MSEI ≥ 0.8, then determine the safety level as level 5. The corresponding responses and ecological restoration measures for each level are as follows: If the safety level is level 1, immediately activate the emergency response plan, carry out large - scale ecological restoration projects, and strengthen the monitoring frequency to once a month. If the safety level is level 2, initiate targeted restoration measures, update the monitoring data quarterly, and dynamically adjust the restoration plan. If the safety level is level 3, strengthen monitoring and control, conduct a comprehensive evaluation every six months, and promptly identify and prevent potential risks. If the safety level is level 4, maintain the existing management measures, conduct a comprehensive evaluation once a year, and fine - tune the management strategy according to the evaluation results. If the safety level is 5, the management experience will be summarized and promoted as a model ranch. A comprehensive evaluation will be carried out every 2 years to continuously optimize the management model.

Images