Wave height prediction system and method for deep-sea intelligent aquaculture platform

Abstract

The present application discloses a wave height prediction system and method for deep-sea intelligent aquaculture platforms, which includes the following steps: obtaining multi-source marine environmental data, processing and generating high-resolution gridded data; performing spatiotemporal feature extraction, building an association network, applying nonlinear manifold learning algorithm to reduce dimensionality, and generating a low-dimensional high-information feature data set; extracting dynamic characteristics of the wave field, reconstructing the high-resolution wave field, performing multi-scale fusion and data assimilation, and generating a complete wave field data set; building a multi-scale computing grid, performing adaptive model selection, realizing model coupling, performing parallel computing and multi-scale coupling simulation; performing probability prediction and uncertainty quantification based on simulation results, and performing result visualization and impact analysis. The present application improves the accuracy and efficiency of wave height prediction and provides reliable decision support for deep-sea aquaculture platforms.

Description

Technical Field

[0001] The present invention relates to wave height simulation prediction, and in particular to a wave height prediction system and method for a deep-sea intelligent aquaculture platform. Background Art

[0002] In the operation of deep-sea intelligent aquaculture platforms, accurate wave height prediction is of vital importance to ensure facility safety, optimize production management and improve economic benefits. With the rapid development of the global marine economy and the increasing scarcity of offshore aquaculture resources, deep-sea aquaculture is becoming an important direction for the sustainable development of the aquaculture industry. However, the deep-sea environment is complex and changeable, and extreme weather events occur frequently, posing severe challenges to the safety and production of aquaculture platforms. Accurate wave height prediction can not only provide decision-making support for the daily operation of the platform, such as optimizing feed delivery time and adjusting equipment configuration, but also provide timely warnings before extreme weather arrives, providing key information for emergency response and disaster prevention and mitigation. In addition, high-quality wave height prediction can also guide the long-term planning and design optimization of the platform, improve the overall wind and wave resistance, thereby reducing operational risks and improving the return on investment.

[0003] At present, the research on deep sea wave height prediction mainly focuses on methods such as numerical simulation, statistical analysis and machine learning. Traditional numerical simulation methods, such as the SWAN (Simulating Waves Nearshore) model and the WAVEWATCH III model, predict wave height by solving the dynamic equations of the wave spectrum. These models perform well in large-scale ocean environment simulations, but have limitations in dealing with complex terrain and local strong nonlinear processes. Statistical analysis methods, such as time series analysis and extreme value statistics, can extract statistical laws of wave height changes from historical data, but it is difficult to accurately capture sudden events and non-stationary processes. In recent years, machine learning methods have been widely used in the field of wave height prediction, such as support vector machines (SVM), artificial neural networks (ANN) and long short-term memory networks (LSTM). These methods can effectively handle nonlinear relationships and show good performance in short-term predictions. However, they usually require a large amount of high-quality training data and still face challenges in long-term predictions and extreme event predictions.

[0004] Although existing research has made some progress, there are still many technical difficulties in the application of wave height prediction in deep-sea intelligent aquaculture platforms. In deep-sea environments, the temporal and spatial resolution, accuracy and coverage of data sources such as satellite remote sensing, buoy observations and numerical simulations vary. How to effectively integrate these data to provide continuous and consistent input information still needs in-depth research. In addition, in deep-sea environments, factors such as large-scale weather systems, mesoscale vortices and local terrain effects jointly affect wave evolution. Existing models are difficult to simultaneously take into account the accurate description of these different scale processes, and accurate simulation of cross-scale wave interactions remains difficult. Furthermore, traditional methods often perform poorly in dealing with sudden strong wind and wave events, especially when the prediction time is long, and the prediction accuracy of non-stationary extreme events needs to be improved. In addition, the quantification and propagation of prediction uncertainty also face challenges. Existing methods focus on point prediction, and the reliability assessment and risk quantification of prediction results are insufficient, making it difficult to provide comprehensive risk assessment information for aquaculture platforms. Solving these technical problems is of great significance to improving the wave height prediction capabilities of deep-sea intelligent aquaculture platforms and ensuring the safety of aquaculture production. Therefore, research and innovation are needed. Summary of the invention

[0005] The purpose of the invention is to provide a wave height prediction system and method for deep-sea intelligent aquaculture platforms to solve the above-mentioned problems existing in the prior art.

[0006] Technical solution, according to one aspect of the present application, a wave height prediction method for a deep-sea intelligent aquaculture platform comprises the following steps:

[0007] Step S1, acquiring multi-source marine environmental data, performing spatiotemporal alignment, quality control, data filling, outlier processing, spatiotemporal interpolation and scale conversion, and generating high-resolution gridded marine environmental data;

[0008] Step S2, obtaining high-resolution gridded marine environment data, performing spatiotemporal feature extraction, and obtaining a main spatiotemporal modal data set; constructing a marine environment element association network based on the main spatiotemporal modal data set to generate a network feature data set; applying a nonlinear manifold learning algorithm to the network feature data set and the main spatiotemporal modal data set to perform dimensionality reduction processing to obtain a feature data set after dimensionality reduction; integrating the feature data set after dimensionality reduction, the network feature data set, and the main spatiotemporal modal data set to generate a final low-dimensional and high-information feature data set;

[0009] Step S3, obtaining a low-dimensional high-information feature data set, applying a dynamic mode decomposition method to extract dynamic features of the wave field, and generating a dynamic feature data set; reconstructing a high-resolution wave field based on the dynamic feature data set to obtain a reconstructed high-resolution wave field data set and performing multi-scale fusion to obtain a multi-scale fused wave field data set; performing data assimilation using the multi-scale fused wave field data set and real-time observation data to obtain an enhanced wave field data set; integrating the enhanced wave field data set and the dynamic feature data set to perform spatiotemporal interpolation and extrapolation of the wave field to generate a complete wave field data set;

[0010] Step S4, constructing a multi-scale computational grid based on the complete wave field data set to obtain a grid data set; performing adaptive model selection according to the grid data set and the complete wave field data set to generate a model configuration data set; using the model configuration data set to achieve bidirectional coupling between models to create a coupling configuration data set; performing parallel computing based on the coupling configuration data set, the model configuration data set and the grid data set to form a parallel configuration data set; integrating the parallel configuration data set, the coupling configuration data set, the model configuration data set and the grid data set to perform multi-scale coupling simulation to generate a coupling simulation result data set;

[0011] Step S5: Based on the coupled simulation result data set, perform probability prediction and uncertainty quantification, obtain and perform result visualization and impact analysis based on the probability prediction data set and the prediction evaluation data set.

[0012] According to one aspect of the present application, a wave height prediction system for a deep-sea intelligent aquaculture platform is provided, comprising:

[0013] at least one processor; and

[0014] a memory communicatively connected to at least one of the processors; wherein,

[0015] The memory stores instructions executable by the processor, and the instructions are used to be executed by the processor to implement the wave height prediction method for the deep-sea intelligent aquaculture platform described in any of the above technical solutions.

[0016] Beneficial effect: This wave height prediction method provides a high-precision, comprehensive and dynamic decision-making support tool for deep-sea intelligent aquaculture platforms through the comprehensive application of innovative technologies. Multi-source data processing technology solves the problem of sparse deep-sea data and provides high-quality input. The nonlinear dimensionality reduction method effectively captures the characteristics of complex marine environments and improves prediction efficiency and accuracy. Dynamic pattern decomposition and multi-scale fusion achieve high-resolution wave field reconstruction and provide comprehensive sea area information. Multi-model coupling takes into account large-scale environmental changes and local wave characteristics to ensure prediction consistency. Probabilistic prediction and uncertainty quantification provide a complete risk assessment and enhance the platform's risk management capabilities.

[0017] The system can simultaneously provide long-term large-scale trends and short-term high-precision local forecasts, including extreme event assessments. This enables aquaculture platform managers to develop refined operational strategies, such as optimizing work arrangements, adjusting equipment configurations, and responding to extreme weather in a timely manner. This method improves platform safety and economic benefits, and can predict wave trends for the next week and local fine changes within 24 hours, providing comprehensive support for management decisions. By providing reliable and timely early warning information, this method provides key technical support for the development of deep-sea aquaculture. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a flow chart of the present invention.

[0019] Figure 2 It is a flow chart of step S1 of the present invention.

[0020] Figure 3 It is a flow chart of step S2 of the present invention.

[0021] Figure 4 It is a flow chart of step S3 of the present invention.

[0022] Figure 5 It is a flow chart of step S4 of the present invention. DETAILED DESCRIPTION

[0023] like Figure 1 As shown, according to one aspect of the present application, a wave height prediction method for a deep-sea intelligent aquaculture platform comprises the following steps:

[0024] Step S1, acquiring multi-source marine environmental data, performing spatiotemporal alignment, quality control, data filling, outlier processing, spatiotemporal interpolation and scale conversion, and generating high-resolution gridded marine environmental data;

[0025] Step S2, obtaining high-resolution gridded marine environment data, performing spatiotemporal feature extraction, and obtaining a main spatiotemporal modal data set; constructing a marine environment element association network based on the main spatiotemporal modal data set to generate a network feature data set; applying a nonlinear manifold learning algorithm to the network feature data set and the main spatiotemporal modal data set to perform dimensionality reduction processing to obtain a feature data set after dimensionality reduction; integrating the feature data set after dimensionality reduction, the network feature data set, and the main spatiotemporal modal data set to generate a final low-dimensional and high-information feature data set;

[0026] Step S3, obtaining a low-dimensional high-information feature data set, applying a dynamic mode decomposition method to extract dynamic features of the wave field, and generating a dynamic feature data set; reconstructing a high-resolution wave field based on the dynamic feature data set to obtain a reconstructed high-resolution wave field data set and performing multi-scale fusion to obtain a multi-scale fused wave field data set; performing data assimilation using the multi-scale fused wave field data set and real-time observation data to obtain an enhanced wave field data set; integrating the enhanced wave field data set and the dynamic feature data set to perform spatiotemporal interpolation and extrapolation of the wave field to generate a complete wave field data set;

[0027] Step S4, constructing a multi-scale computational grid based on the complete wave field data set to obtain a grid data set; performing adaptive model selection according to the grid data set and the complete wave field data set to generate a model configuration data set; using the model configuration data set to achieve bidirectional coupling between models to create a coupling configuration data set; performing parallel computing based on the coupling configuration data set, the model configuration data set and the grid data set to form a parallel configuration data set; integrating the parallel configuration data set, the coupling configuration data set, the model configuration data set and the grid data set to perform multi-scale coupling simulation to generate a coupling simulation result data set;

[0028] Step S5: Based on the coupled simulation result data set, perform probability prediction and uncertainty quantification, obtain and perform result visualization and impact analysis based on the probability prediction data set and the prediction evaluation data set.

[0029] Through the efficient processing and integration technology of multi-source marine environmental data, the problem of sparse and inconsistent data in deep and deep sea areas is solved, providing high-quality input data for prediction. Through dimensionality reduction and feature extraction methods, such as nonlinear manifold learning algorithms, the key features of the complex marine environment are effectively captured, and the computational efficiency and accuracy of subsequent prediction models are improved. Through the application of dynamic mode decomposition and multi-scale fusion technology, accurate reconstruction from limited observations to high-resolution wave fields is achieved, providing comprehensive wave information in the surrounding sea areas for the aquaculture platform. The realization of multi-model coupling and parallel computing enables the system to simultaneously consider large-scale marine environmental changes and local fine wave characteristics, providing consistent prediction results from global to local. The introduction of probabilistic prediction and uncertainty quantification methods not only provides point prediction values, but also gives complete wave height probability distribution and risk assessment information, improving the risk management capabilities of aquaculture platforms.

[0030] For deep-sea intelligent aquaculture platforms, more accurate and timely wave height warnings can be obtained, potential risks can be fully understood, and various complex sea conditions can be calmly dealt with. For example, the system can simultaneously provide large-scale wave trends for the next week and high-precision local forecasts within 24 hours around the platform, including the probability of occurrence and impact assessment of extreme events. This enables aquaculture platform managers to formulate more sophisticated and flexible operating strategies, such as optimizing daily operation arrangements, adjusting equipment configurations, and even making timely emergency response decisions before extreme weather arrives. By providing comprehensive and reliable decision-making support information, the safety and economic benefits of deep-sea intelligent aquaculture platforms are improved, providing key technical support for promoting the expansion of marine aquaculture to the deep sea.

[0031] like Figure 2 As shown, according to one aspect of the present application, step S1 is specifically:

[0032] Step S11, receiving raw ocean environment data from multiple data sources and storing them in a distributed database to form an original multi-source ocean environment data set; wherein the raw ocean environment data includes sea surface temperature, sea surface height and sea surface wind field data transmitted by a satellite remote sensing system, wave height, wave direction and period data transmitted in real time by a buoy observation network, and ocean environment forecast data output by a numerical simulation system;

[0033] Step S12, read the original multi-source marine environment data set, use the dynamic time warping algorithm to time align the data with different time resolutions, and generate unified time series data; use the geo-referencing technology to spatially align the data with different spatial resolutions, and create data with a unified spatial grid; apply the outlier detection method based on the interquartile range to identify and mark the outliers in the data, and obtain a data set with anomaly marks; use the data quality index evaluation system to assign a quality mark to each data point, and generate a data set with a quality mark; integrate and store the processed data into a marine environment data set in a unified format;

[0034] Step S13, obtaining a marine environment data set in a unified format, using a tensor completion algorithm to repair missing data, and obtaining a preliminary repaired data set; correcting the marked outliers to generate an outlier-corrected data set; denoising the corrected data to obtain a denoised data set, and integrating it into a complete and consistent marine environment data set;

[0035] Step S14, read a complete and consistent marine environment data set, apply the Kriging interpolation method to perform spatial interpolation on the data of discrete observation points to generate a continuous spatial distribution field; use the wavelet transform method to convert the time series data into a time-uniform data set; use multi-scale analysis technology to convert data of different spatial resolutions to a unified high-resolution grid to obtain a high-resolution gridded marine environment data set.

[0036] Through the acquisition, processing and integration of multi-source marine environmental data, the generation of high-quality, high-resolution gridded marine environmental data is achieved, laying a solid data foundation for subsequent wave height prediction. Specifically, the dynamic time warping algorithm and georeferencing technology are used to solve the inconsistency of different data sources in terms of temporal and spatial resolution, ensuring the temporal and spatial alignment of the data. Through the outlier detection method based on the interquartile range and the data quality index evaluation system, outliers are effectively identified and processed, improving the reliability of the data. The application of the tensor completion algorithm solves the common missing problem of marine observation data, while the wavelet transform denoising method effectively reduces the noise interference in the data. The use of the Kriging interpolation method realizes the conversion from discrete observation points to continuous spatial distribution fields, improving the spatial resolution and coverage of the data. The comprehensive application of these technologies makes the generated marine environmental data have the characteristics of high temporal and spatial resolution, high quality and high reliability, providing comprehensive and accurate environmental parameter input for the wave height prediction of deep-sea intelligent aquaculture platforms. In practical applications, this means that environmental changes in local sea areas, such as wave anomalies caused by sudden weather events, can be captured more accurately, thereby providing more reliable data support for the safety management and operation decisions of aquaculture platforms.

[0037] like Figure 3 As shown, according to one aspect of the present application, step S2 is specifically:

[0038] Step S21, read the high-resolution gridded marine environment data set, standardize the data, eliminate the dimensional differences between different variables, and obtain a standardized data set; calculate the covariance matrix of the standardized data set, and perform eigenvalue decomposition on the matrix; select the first N eigenvectors with cumulative variance contribution exceeding 95% as the main space-time modes; project the original data onto these main modes to obtain the corresponding time coefficient sequence; store the extracted main space-time modes and time coefficient sequence as a space-time modal data set;

[0039] Step S22, obtaining a spatiotemporal modal data set, calculating the mutual information between the time coefficient sequences of different environmental elements, quantifying the nonlinear correlation between the elements, and generating a mutual information matrix; based on the mutual information matrix, using an adaptive threshold method to determine network connections and form a preliminary network structure; applying a community detection algorithm to identify tightly connected subgroups in the network and obtain a network community structure; calculating the centrality index of each node in the network community structure, including degree centrality, betweenness centrality, and eigenvector centrality, completing the construction of an association network between marine environmental elements, and storing the constructed association network and its topological characteristics as a network feature data set;

[0040] Step S23, read the network feature data set and the spatiotemporal modal data set, use the t-SNE algorithm to map the high-dimensional features to the low-dimensional space, and obtain the preliminary dimensionality reduction results; use the UMAP algorithm to optimize the low-dimensional representation and generate the optimized low-dimensional representation; calculate the quality index of the dimensionality reduction result; select the best low-dimensional representation method and generate the feature data set after dimensionality reduction;

[0041] Step S24, obtaining the feature data set after dimensionality reduction, the network feature data set and the spatiotemporal modal data set, using the feature importance ranking method based on random forest to screen the most representative features, and obtain the feature importance ranking results; using the multi-view learning algorithm to fuse feature information from different sources to generate a comprehensive feature representation; using sparse coding technology to compress the feature representation to obtain a compressed feature representation; combining the processed features into a unified feature vector, and storing it as the final low-dimensional and high-information feature data set.

[0042] Through a series of data processing and feature extraction techniques, the conversion from high-dimensional marine environmental data to low-dimensional high-information feature data sets is realized, which improves the computational efficiency and prediction accuracy of the subsequent wave height prediction model. Specifically, by performing eigenvalue decomposition on the covariance matrix and selecting the main spatiotemporal modes, the main change patterns of the marine environmental data are effectively captured and redundant information is reduced. The method of constructing a marine environmental element association network based on mutual information reveals the complex nonlinear correlation between different environmental factors and provides more comprehensive environmental element interaction information for wave height prediction. The combined application of t-SNE and UMAP algorithms realizes the nonlinear mapping of high-dimensional data to low-dimensional space, retains the local structure and global relationship of the data, and helps to discover potential wave height influencing factors. The use of random forest feature importance ranking and multi-view learning algorithms further screens and integrates the most representative features, improving the information content and expression ability of the features. The combined effect of these technologies is to generate a low-dimensional but high-information feature data set, which not only retains the key information of the original marine environmental data, but also reduces the data dimension. For the wave height prediction of deep-sea intelligent aquaculture platforms, this means that large-scale marine environmental data can be processed and analyzed more efficiently, and the environmental factors that have the greatest impact on wave height changes can be quickly identified, thereby improving the calculation speed and prediction accuracy of the prediction model. At the same time, the reduced-dimensional data is easier to visualize and interpret, which helps aquaculture platform managers to more intuitively understand and use the prediction results and make more reasonable operational decisions.

[0043] like Figure 4 As shown, according to one aspect of the present application, step S3 is specifically:

[0044] Step S31, read a low-dimensional high-information feature data set, construct a time delay matrix, organize the feature data into a series of snapshots to form a snapshot matrix; perform singular value decomposition on the snapshot matrix, extract the main dynamic mode, and obtain a dynamic mode matrix; calculate the growth rate and frequency corresponding to each dynamic mode, and identify the key dynamic characteristics; obtain a simplified model that describes the main dynamic behavior of the wave field through screening and reconstruction; store the extracted dynamic mode and its parameters as a dynamic feature data set;

[0045] Step S32, obtaining a dynamic feature data set, analyzing the wave field characteristics and constructing a sparse basis suitable for the wave field characteristics, and constructing a sparse representation dictionary; constructing a measurement matrix and an observation vector based on limited observation points; using an orthogonal matching pursuit algorithm to solve the sparse reconstruction problem, and obtaining a representation of the wave field on the sparse basis; converting the sparse representation back to the physical space through an inverse transformation, and obtaining reconstructed high-resolution wave field data; and storing the reconstruction result as a high-resolution wave field data set;

[0046] Step S33, reading the high-resolution wave field data set and the original multi-scale observation data, using wavelet packet transform to decompose the data of different resolutions into multiple scales and directions, and obtaining multi-scale decomposition coefficients; applying an adaptive weight fusion algorithm on each sub-band to integrate information from different sources, and generating fused sub-band coefficients; using an inverse wavelet packet transform, reconstructing the fused sub-band coefficients into a wave field of uniform resolution; using an edge-preserving filter to remove artifacts introduced in the fusion process; and saving the processed data as a multi-scale fused wave field data set;

[0047] Step S34, obtaining a multi-scale fusion wave field data set and real-time observation data, constructing an initial set of the ensemble Kalman filter, and generating an observation set using a perturbation observation method; performing a forecasting step, using the extracted dynamic model to perform a short-term forecast on each ensemble member to obtain a forecast set; performing an analysis step, updating the forecast set according to the observation data, and generating an analysis set; calculating the ensemble average as the best estimate, and using the ensemble discreteness to estimate uncertainty; storing the assimilated wave field data and its uncertainty estimate as an enhanced wave field data set;

[0048] Step S35, integrating the generated enhanced wave field data set and dynamic feature data set, using the result of dynamic mode decomposition, constructing the spatiotemporal evolution model of the wave field; applying the Kriging interpolation method in the spatial domain to fill the observation blank areas and obtain a spatially complete data set; using the Kalman smoother in the time domain to smooth the historical data and perform short-term prediction to generate a time-continuous data set; combining the results of spatial interpolation and time extrapolation to generate continuous wave field data covering all predetermined areas and all time periods; saving the processed data as a complete wave field data set.

[0049] Through technologies such as dynamic mode decomposition, multi-scale fusion and data assimilation, the generation of a complete wave field data set with high precision and high spatiotemporal resolution is achieved, providing comprehensive and accurate wave information for wave height prediction of deep-sea intelligent aquaculture platforms. Specifically, the application of the dynamic mode decomposition method effectively extracts the main dynamic characteristics of the wave field, captures the key patterns of wave evolution, and lays the foundation for subsequent wave field reconstruction. The use of the sparse reconstruction algorithm, combined with an adaptive multi-scale dictionary and an optimized measurement matrix, achieves accurate reconstruction of the wave field from limited observation points to high-resolution, effectively solving the problem of sparse observation points in deep-sea areas. The application of wavelet packet transform and adaptive weight fusion algorithm achieves effective integration of multi-scale and multi-source data, and improves the spatiotemporal resolution and accuracy of the reconstructed wave field. The introduction of the ensemble Kalman filter assimilation technology, by combining model prediction and real-time observation data, continuously corrects and optimizes wave field estimation, and improves the accuracy and reliability of prediction. The combined effect of these technologies is to generate a high-quality, comprehensive and complete wave field data set, which not only contains large-scale wave evolution trends, but also retains the detailed characteristics of local areas. For deep-sea smart aquaculture platforms, this means that more accurate and continuous wave height information can be obtained, which can not only accurately predict the wave height changes at the platform's location, but also fully grasp the wave conditions in the surrounding waters. This is of great significance for the safety management, equipment adjustment and daily operations of the aquaculture platform, and can help platform managers identify potential dangerous wave conditions in advance, take preventive measures in a timely manner, and maximize the safety of aquaculture facilities and personnel.

[0050] like Figure 5 As shown, according to one aspect of the present application, step S4 is specifically:

[0051] Step S41, read the generated complete wave field data set, apply the quadtree algorithm to adaptively divide the calculation domain according to the spatial variation characteristics of the wave field, and obtain preliminary grid division; use unstructured grid generation algorithms, such as Delaunay triangulation, to create high-resolution grids in nearshore and key areas to form local fine grids; use regular grids in open sea areas to maintain computational efficiency and generate global coarse grids; use grid smoothing technology to optimize grid quality and ensure smooth transition between grids; store the generated multi-scale grid data as a grid data set;

[0052] Step S42, obtaining the created grid data set and the complete wave field data set, calculating the characteristic parameters of each grid unit, and forming a grid characteristic data set; using a decision tree algorithm, selecting the most suitable wave model for each grid unit according to the grid characteristic data set, and obtaining a preliminary model allocation result; constructing a transition area at the junction of the models to ensure smooth switching between different models, and generating an optimized model allocation; generating a model configuration file containing the model type and parameter settings of each grid unit and saving it as a model configuration data set; wherein the characteristic parameters include water depth, wave steepness, and nonlinear parameters;

[0053] Step S43, read the generated model configuration data set, build a model coupling framework, define the data exchange interface between models, and obtain the coupling framework structure; build a spatial interpolation module for data transmission between grids of different resolutions, and generate data mapping relationships between grids; build a time integration scheme, coordinate the time steps of different models, and create a time synchronization mechanism; for overlapping areas, use the Schwarz alternation method to iteratively solve, ensure the continuity of boundary conditions, and obtain a boundary processing scheme; use a feedback mechanism to allow the results of fine-scale models to affect the calculation of large-scale models; store the coupling settings and data exchange rules as a coupling configuration data set;

[0054] Step S44, obtaining a coupling configuration data set, a model configuration data set and a grid data set, using a METIS graph partitioning algorithm to decompose the computational domain into load-balanced sub-regions, and generating a domain decomposition scheme; allocating computing resources to each sub-region, establishing an MPI communication topology, and forming a parallel computing architecture; and storing the parallel computing settings as a parallel configuration data set;

[0055] Step S45, integrating the parallel configuration data set, the coupling configuration data set, the model configuration data set and the grid data set, initializing each sub-model, interpolating the complete wave field data onto the multi-scale grid as the initial condition, and obtaining the initialized data set; iteratively executing the calculation of each sub-model according to the preset time step, performing data exchange and boundary condition update between models at each time step, and generating an intermediate result data set; using an adaptive time step control algorithm, dynamically adjusting the step size according to the calculation stability and accuracy requirements, and forming a time step sequence; regularly executing checkpoint saving to facilitate breakpoint continuation and create a checkpoint data set; collecting the calculation results of each sub-area and merging them into a unified output data; storing the intermediate results in the simulation process and the final high-precision wave field prediction results as a coupled simulation result data set.

[0056] By constructing multi-scale computational grids, adaptive model selection, and multi-model coupling, efficient and accurate wave height prediction simulation is achieved, providing reliable wave forecast information for deep-sea intelligent aquaculture platforms. Specifically, the combined application of quadtree algorithm and unstructured grid generation algorithm creates a multi-scale computational grid that takes into account both computational efficiency and local accuracy, which can perform high-resolution simulation of the sea area around the aquaculture platform while ensuring global simulation efficiency. The adaptive model selection method based on the decision tree algorithm selects the most suitable wave model for each grid unit, effectively improving the overall accuracy and computational efficiency of the simulation. The construction of the model coupling framework, especially the application of the Schwarz alternation method, realizes seamless connection and two-way feedback between wave models of different scales and types, ensuring the continuity and consistency of the simulation results. The use of the METIS graph partitioning algorithm realizes the load balancing distribution of computing tasks and improves the efficiency of parallel computing. The combined effect of these technologies is to realize a highly optimized and efficient multi-scale coupled wave simulation system. For deep-sea intelligent aquaculture platforms, this means that high-precision wave height prediction results can be obtained within a reasonable computing time, which can not only accurately simulate the impact of large-scale marine environment on waves, but also finely depict the wave characteristics of the local sea area around the platform. This multi-scale, high-precision simulation capability enables aquaculture platform managers to more comprehensively understand and predict wave changes under complex sea conditions, and calmly respond to various marine environmental challenges. For example, the system can simultaneously predict the propagation of large-scale storm surges in the open sea and their specific impact on the wave height in the local area of ​​the aquaculture platform, providing accurate and timely decision-making support for the platform's emergency response and risk management.

[0057] According to one aspect of the present application, in step S5, the process of performing probability prediction and uncertainty quantification based on the coupled simulation result data set is specifically as follows:

[0058] Step S51, read the coupled simulation result data set, use the perturbation observation method to generate multiple sets of perturbation fields of initial conditions and boundary conditions, and form a perturbation data set; use the random physical parameterization method to assign probability distribution to key physical parameters in the model, and create a parameter perturbation set; run multiple coupled simulations, each time using different initial conditions, boundary conditions and physical parameter combinations, to generate multiple sets of simulation results; collect all simulation results, form an ensemble forecast member set and store it as an ensemble forecast data set;

[0059] Step S52, obtaining an ensemble forecast data set, calculating the likelihood function of each ensemble member relative to the historical observation data, and obtaining a likelihood value sequence; using the Markov chain Monte Carlo method to estimate the posterior probability distribution of each model, and forming a posterior probability data set; performing weighted averaging on the prediction results of all ensemble members according to the posterior probability, and generating a comprehensive prediction result; calculating the credible interval and probability density function of the comprehensive prediction result; and storing the weighted average result and related statistics as a Bayesian average prediction data set;

[0060] Step S53, read the Bayesian average prediction data set, use the peak overthreshold method to extract extreme value samples from the prediction data to form an extreme value sample set; fit the generalized Pareto distribution and estimate the probability distribution parameters of the extreme value to obtain a distribution parameter set; calculate the design wave heights of different regression periods to generate a design wave height data set; use the bootstrap method to estimate the confidence interval of the extreme value prediction; store the extreme value analysis results as an extreme value prediction data set;

[0061] Step S54, obtaining an extreme value prediction data set and a Bayesian average prediction data set, using the Sobol method to calculate the global sensitivity index of each input parameter to the prediction result, and obtaining the parameter sensitivity ranking; using the FAST algorithm to analyze the interaction effect between parameters and generate a parameter interaction matrix; constructing a response surface model that describes the relationship between input parameters and output results, and forming a response surface function; based on the response surface model, using the Monte Carlo method to perform large sample simulation to obtain the probability distribution of the prediction result; storing the sensitivity analysis results and the probability distribution information as a sensitivity analysis data set;

[0062] Step S55, integrating the sensitivity analysis data set, the extreme value prediction data set and the Bayesian average prediction data set, using the ensemble Kalman filter method to assimilate the latest observation data, updating the prediction results and their uncertainties, and obtaining an updated prediction set; using the covariance inflation technique to generate a corrected prediction set; calculating the spatiotemporal correlation structure of the prediction results, using the EOF decomposition method to extract the main uncertainty modes, and forming an uncertainty mode data set; generating a probability prediction product and storing it as a probability prediction data set, the probability prediction product includes a prediction mean field, a standard deviation field, a probability distribution map and a credible interval;

[0063] Step S56, read the probability prediction data set, use at least two scoring methods to evaluate the accuracy of the probability prediction, and obtain a skill scoring data set; analyze the reliability and resolution of the prediction through the attribute graph and the reliability graph, and generate a reliability analysis result; calculate the spatial distribution and time evolution characteristics of the prediction deviation to form a deviation distribution data set; use Taylor graph and Murphy-Winkler decomposition to evaluate the performance of the prediction system; store the evaluation results and error statistics as a prediction evaluation data set; wherein the scoring methods include Brier scoring, sorted probability scoring, and continuous probability scoring.

[0064] According to one aspect of the present application, in step S5, result visualization and impact analysis are performed based on the probability prediction data set and the prediction evaluation data set, specifically:

[0065] Step S57, reading the probability prediction data set and the prediction evaluation data set, using the contour map technology to draw the spatial distribution map of the predicted wave height field, and generating spatial distribution visualization data;

[0066] Step S58, generating a heat map of wave height changes by a color mapping method to form a heat map data set; constructing a time series animation to display the time evolution process of the wave field and creating animation sequence data;

[0067] Step S59: for the probability information, use box plots and violin plots to display the predicted uncertainty distribution, and obtain probability distribution visualization data;

[0068] Use parallel coordinates to display the relationship between multiple predictor variables and form multivariate relationship graph data; save these visualizations as visualization datasets.

[0069] Through the techniques of probability prediction, uncertainty quantification and result visualization, a comprehensive and reliable analysis and display of wave height prediction results are achieved, providing a powerful tool for decision support of deep-sea intelligent aquaculture platforms. Specifically, the application of perturbation observation method and random physical parameterization method generates a large number of simulation results, effectively quantifying the impact of initial conditions, boundary conditions and physical parameter uncertainties on the prediction results. The use of Markov chain Monte Carlo method realizes the accurate estimation of the posterior probability of the model and improves the reliability of the prediction results. The application of peak overthreshold method and generalized Pareto distribution enables the system to accurately assess the probability and risk level of extreme wave height events, which is crucial for the safety management of aquaculture platforms. The combination of Sobol method and FAST algorithm realizes a comprehensive sensitivity analysis and helps identify the most influential factors for wave height prediction. The introduction of ensemble Kalman filter method improves the timeliness and accuracy of prediction by assimilating the latest observation data, continuously updating and optimizing the prediction results. The use of multiple scoring methods, such as Brier scoring and continuous probability scoring, provides an objective and comprehensive evaluation of the prediction results. Various advanced visualization techniques, such as contour maps, heat maps, and probability distribution maps, enable complex forecast results to be presented intuitively and clearly. In short, a comprehensive, reliable, and easy-to-understand wave height prediction and risk assessment system has been built. For deep-sea intelligent aquaculture platforms, this means that managers can not only obtain accurate wave height prediction values, but also have a clear understanding of the uncertainty of the prediction and the possible range of risks. For example, the system can provide a probability distribution map of wave height exceeding a certain threshold in the next 24 hours, while marking high-risk areas and time periods. This comprehensive information support enables the aquaculture platform to formulate more sophisticated and flexible emergency plans, such as taking different levels of protective measures according to different probability levels, thereby maximizing the platform's operational efficiency while ensuring safety.

[0070] According to one aspect of the present application, step S13 is specifically:

[0071] Step S131, read the marine environment data set in a unified format, organize the marine environment data into a three-order tensor structure, where time, space and variables correspond to three dimensions respectively; use wavelet transform to decompose the tensor into sub-tensors of multiple scales; apply the nuclear norm minimization algorithm to complete each sub-tensor; reconstruct the complete tensor through inverse wavelet transform to obtain a preliminary repair data set;

[0072] Step S132, obtaining a preliminary repaired data set; using a sliding window method to divide the data subsets in time and space dimensions; applying robust principal component analysis to each subset to decompose the data into a low-rank part and a sparse abnormal part; identifying and marking outliers based on Mahalanobis distance and local heteroscedasticity estimation; using local Gaussian process regression to correct the marked outliers to generate an outlier-corrected data set;

[0073] Step S133, read the data set after outlier correction, build a correlation network between variables, and use mutual information and conditional mutual information to quantify the variable relationship; based on the network structure, build a multivariate empirical mode decomposition algorithm with adaptive weights, and apply soft threshold wavelet denoising to each decomposition mode; reconstruct the denoised signal to obtain the final complete and consistent marine environment data set.

[0074] Through the techniques of tensor completion, outlier correction and denoising, high-quality restoration and optimization of marine environmental data are achieved, providing a reliable and consistent data basis for the wave height prediction of deep-sea intelligent aquaculture platforms. Specifically, the combined application of wavelet transform and nuclear norm minimization algorithm realizes the multi-scale decomposition and effective completion of the third-order tensor structure of marine environmental data, overcomes the limitations of traditional methods in processing high-dimensional data, and effectively fills the data gaps. The use of robust principal component analysis realizes the accurate identification and correction of outliers by decomposing the data into low-rank parts and sparse abnormal parts, thereby improving the reliability of the data. The application of local Gaussian process regression considers the local correlation of the data, making the outlier correction more accurate and reasonable. The combination of adaptive weighted multivariate empirical mode decomposition algorithm and soft threshold wavelet denoising technology realizes the effective removal of noise of various scales in the data, while retaining the key features and change trends of the data. The combined effect of these technologies is to generate a high-quality and consistent marine environmental data set, which not only repairs the missing data, but also removes outliers and noise interference. For the wave height prediction of deep-sea intelligent aquaculture platforms, this means that more complete and reliable input data can be obtained. For example, when processing satellite remote sensing data, it can effectively fill in data gaps caused by cloud cover, correct outliers caused by sensor failures, and remove noise caused by atmospheric interference. This high-quality data foundation improves the accuracy and reliability of subsequent wave height predictions. Especially in deep and deep sea areas, due to the sparse observation sites, data quality issues are more prominent. The application of this embodiment can improve the quantity and quality of available data and provide more comprehensive and accurate marine environmental information support for aquaculture platforms.

[0075] According to one aspect of the present application, step S23 is specifically:

[0076] Step S231, read the network feature data set and the spatiotemporal modal data set, construct a K nearest neighbor graph, use an adaptive bandwidth algorithm to determine the local neighborhood, calculate the geodesic distance matrix, use the fast marching method to improve the computational efficiency, apply local linear embedding to perform preliminary dimensionality reduction, and obtain an initial low-dimensional representation; based on the initial low-dimensional representation, adaptively adjust the neighbor relationship to generate an optimized neighbor graph structure;

[0077] Step S232, obtaining the optimized neighbor graph structure and the initial low-dimensional representation, using an adaptive learning rate scheduling strategy to dynamically adjust the gradient descent process; accelerating the gradient through momentum terms and Nesterov; using the Barnes-Hut approximation algorithm to accelerate high-dimensional similarity calculations; using a multi-scale strategy to gradually optimize embedding from coarse to fine, and obtaining a t-SNE dimensionality reduction result;

[0078] Step S233, read the t-SNE dimension reduction result, construct a fuzzy topological representation, and use adaptive fuzziness parameters to capture local structures; construct an optimization target based on Riemannian geometry, and apply random negative sampling technology and parallel computing strategy to accelerate the optimization process, and generate UMAP dimension reduction results;

[0079] Step S234, obtain the t-SNE dimensionality reduction results and UMAP results, use the local preservation projection method to map the two results to a common low-dimensional space; apply the weighted voting mechanism to integrate the advantages of different methods; use the manifold alignment technology to ensure the consistency of the results of different methods; calculate the dimensionality reduction quality index and select the best representation method to generate the final feature data set after dimensionality reduction.

[0080] By combining multiple dimensionality reduction techniques, efficient dimensionality reduction and feature extraction of marine environmental data are achieved, providing low-dimensional but information-rich feature representation for wave height prediction of deep-sea intelligent aquaculture platforms. Specifically, the application of adaptive bandwidth algorithm and geodesic distance calculation improves the accuracy of local linear embedding and better preserves the local structural information of the data. The adaptive learning rate scheduling strategy and Barnes-Hut approximation algorithm adopted in the t-SNE algorithm improve the efficiency of the dimensionality reduction process and the large-scale data processing capability. The fuzzy topological representation and Riemannian geometric optimization objective introduced in the UMAP algorithm enable the dimensionality reduction results to better preserve the global structure of the data. The use of the local preservation projection method realizes the effective fusion of different dimensionality reduction results and comprehensively utilizes the advantages of various methods. The combined effect of these technologies is to generate a feature data set that retains key information and reduces dimensionality. For the wave height prediction of deep-sea intelligent aquaculture platforms, this means that the complexity of data processing and analysis can be reduced while retaining the key information of marine environmental data. For example, high-dimensional marine environmental data containing dozens of variables can be reduced to a few key dimensions, which may represent the main factors affecting wave height, such as wind field intensity, current direction, and water temperature gradient. This dimensionality reduction not only improves the computational efficiency of subsequent model training and prediction, but also helps to identify and understand the key environmental factors that affect wave height. Especially when dealing with long time series and large-scale spatial data, the advantages of this method are more obvious. It can quickly extract the most critical features for wave height prediction from massive data, providing more timely and accurate early warning information for aquaculture platforms. In addition, the data after dimensionality reduction is easier to visualize, allowing aquaculture platform managers to intuitively understand the complex changes in the marine environment and make more informed operational decisions.

[0081] According to one aspect of the present application, step S32 is specifically:

[0082] Step S321, read the dynamic feature data set obtained in S31, and design an adaptive multi-scale dictionary. Use wavelet packet transform to decompose the wave field into multi-scale representation. Apply K-SVD algorithm to each scale to learn local sparse dictionary. Introduce dynamic features as constraints to optimize dictionary atoms. Combine multi-scale dictionary to form a complete adaptive sparse representation system.

[0083] Step S322, obtain the adaptive multi-scale dictionary of S321, and construct an optimized measurement matrix. Design a structured random matrix based on the mutual interference minimization criterion. Use the annealing algorithm to optimize the matrix elements and improve the reconstruction performance. Introduce dynamic feature information and adjust the measurement matrix structure. Generate the final optimized measurement matrix and the corresponding observation vector.

[0084] Step S323, read the optimized measurement matrix and observation vector of S322, and execute the improved sparse reconstruction algorithm. Design an iterative soft threshold algorithm based on block coordinate descent (BCD-IST). Introduce an adaptive regularization parameter to balance sparsity and reconstruction error. Use a warm start strategy to accelerate algorithm convergence. Apply parallel computing technology to improve the efficiency of solving large-scale problems and obtain a sparse representation of the wave field.

[0085] Step S324, obtain the sparse representation of the wave field in S323, and perform inverse transformation reconstruction. Use the adaptive multi-scale dictionary generated in S321 to map the sparse representation back to the original space. Apply the inverse wavelet packet transform to synthesize the multi-scale components. Use the local regression technique to eliminate the block effect and artifacts. Through iterative post-processing, improve the physical consistency of the reconstruction result and generate the final high-resolution wave field data set.

[0086] Through the design of adaptive multi-scale dictionary, the construction of optimized measurement matrix and the improved sparse reconstruction algorithm, high-precision and high-efficiency wave field reconstruction is achieved, providing comprehensive and detailed wave information for deep-sea intelligent aquaculture platforms. Specifically, the design of adaptive multi-scale dictionary, combined with wavelet packet transform and K-SVD algorithm, creates a dictionary that can effectively represent the characteristics of waves at different scales, improving the accuracy and adaptability of reconstruction. The construction of optimized measurement matrix improves the performance of compressed sensing through the mutual interference minimization criterion and annealing algorithm, making it possible to reconstruct the complete wave field from limited observation points. The improved sparse reconstruction algorithm, especially the application of iterative soft threshold algorithm based on block coordinate descent (BCD-IST), improves the solution efficiency and reconstruction quality. The introduction of parallel computing technology enables the system to efficiently process large-scale wave field data. The combined effect of these technologies is to achieve accurate reconstruction from sparse observations to high-resolution wave fields. For deep-sea intelligent aquaculture platforms, this means that detailed wave information covering the entire sea area of ​​concern can be obtained from limited observation data. For example, even in the far sea area where observation sites are sparse, this method can reconstruct a high-resolution wave height distribution map, providing aquaculture platforms with comprehensive surrounding sea wave conditions. This comprehensive wave field information not only helps to accurately predict the wave height changes at the platform location, but also helps managers understand the wave propagation trends in the adjacent waters, providing an important reference for the platform's position adjustment and deployment of protective measures. Especially in complex sea conditions, such as during typhoons or storm surges, this high-resolution wave field reconstruction capability can help aquaculture platforms better assess and respond to potential risks and improve safety management.

[0087] According to one aspect of the present application, step S34 is further:

[0088] Step S341, read the multi-scale fusion wave field data set and real-time observation data obtained in S33, and construct the initial set of the ensemble Kalman filter (EnKF). Use the orthogonal sampling method to generate the initial disturbance field to ensure the representativeness of the ensemble members. Apply the covariance localization technology to reduce the influence of sampling errors. Implement the adaptive inflation method to dynamically adjust the ensemble discreteness. Generate the initialized filter set and observation set.

[0089] Step S342, obtain the initialized filter set of S341, and perform the ensemble forecasting step. Use the dynamic model extracted in S31 to make short-term forecasts for each ensemble member. Apply random perturbation techniques to simulate model errors. Implement adaptive time step control to ensure numerical stability. Generate a forecast ensemble, including the state vector and error covariance estimate of each member.

[0090] Step S343, read the forecast set of S342 and the observation set of S341, and perform the analysis step. Calculate the collective Jacobian matrix of the observation operator and linearize the observation process. Use the local transformation set Kalman filter (LETKF) algorithm to analyze in the observation space. Apply the iterative set smoothing technique to process nonlinear observations. Generate an analysis set, including the updated state vector and error covariance.

[0091] Step S344: Get the analysis set of S343 and perform post-processing and diagnosis. Calculate the ensemble mean as the best estimate and use the ensemble dispersion to estimate uncertainty. Apply bias correction techniques to eliminate systematic errors. Perform consistency checks to ensure that physical constraints are met. Generate diagnostic statistics such as innovation sequences and analysis increments to form the final assimilated wave field dataset and its uncertainty estimate.

[0092] Through the techniques of ensemble Kalman filter assimilation, adaptive inflation method and iterative ensemble smoothing, high-precision and real-time updated wave field data assimilation is achieved, providing dynamic and accurate wave height prediction support for deep-sea intelligent aquaculture platforms. Specifically, the application of orthogonal sampling method ensures the representativeness of the initial set and improves the performance of the filter. The use of covariance localization technology effectively reduces the impact of sampling error and improves the stability of assimilation. The introduction of adaptive inflation method dynamically adjusts the ensemble discreteness and avoids the problem of filter divergence. The application of local transform ensemble Kalman filter (LETKF) algorithm improves the assimilation performance of nonlinear system and better handles the complex wave dynamics process. The use of iterative ensemble smoothing technology effectively handles nonlinear observations and further improves the accuracy of assimilation results. The combined effect of these technologies is to build a highly accurate and real-time updated wave field data assimilation system. For deep-sea intelligent aquaculture platforms, this means that the latest and most accurate wave height prediction information can be continuously obtained. For example, the system can integrate multi-source data such as satellite remote sensing, buoy observations and numerical model forecasts in real time to continuously correct and optimize wave height prediction results. This dynamically updated prediction capability is particularly important for responding to the rapidly changing marine environment, especially when extreme weather events such as typhoons occur. The system can quickly absorb the latest observation information, adjust the prediction results in time, and provide more accurate and timely warnings for the aquaculture platform. In addition, this embodiment can also provide uncertainty estimates for the prediction, such as the credible interval of the wave height prediction, which helps platform managers better assess risks and formulate more flexible and reasonable response strategies.

[0093] According to one aspect of the present application, step S43 is specifically:

[0094] Step S431, read the model configuration data set generated in S42, and build an adaptive coupling framework. Design a graph-based representation of model dependencies. Implement a dynamic load balancing algorithm to optimize computing resource allocation. Develop an adaptive time step coordination mechanism to ensure synchronization between different models. Generate a coupling framework structure and resource allocation plan.

[0095] Step S432, obtain the coupling framework structure of S431, and implement an efficient spatial interpolation module. Design a conservative interpolation method based on radial basis function (RBF). Use the fast multipole method (FMM) to accelerate large-scale interpolation operations. Implement an adaptive grid refinement strategy to improve the interpolation accuracy of key areas. Generate data mapping relationships and interpolation operators between grids.

[0096] Step S433, read the data mapping relationship between the grids in S432, and develop an intelligent boundary processing solution. Design a boundary condition analysis method based on eigendecomposition. Implement an adaptive Schwarz alternation algorithm to improve boundary matching efficiency. Introduce a machine learning model to predict the optimal relaxation parameters. Generate a boundary processing strategy and iterative solution.

[0097] Step S434: Obtain the boundary processing strategy of S433 to implement feedback mechanism and scale interaction. Design a multi-scale information transmission protocol to ensure effective communication between large and small scale models. Implement a key variable identification algorithm based on sensitivity analysis. Develop an adaptive upper and lower scale parameterization scheme to dynamically adjust the model interaction strength. Generate scale interaction strategy and feedback mechanism configuration.

[0098] Step S435: Integrate all outputs from S431 to S434 to build a complete bidirectional coupling system. Implement a distributed communication layer to support efficient data exchange. Develop error detection and recovery mechanisms to improve system stability. Design a visual monitoring interface to display the coupling process in real time. Generate the final coupling configuration data set, including all coupling settings and data exchange rules.

[0099] By building an adaptive coupling framework, achieving efficient spatial interpolation, developing intelligent boundary processing schemes and feedback mechanisms, an efficient and accurate multi-model coupling system is realized, providing comprehensive and consistent wave height prediction capabilities for deep-sea intelligent aquaculture platforms. Specifically, the application of graph-based model dependency representation and dynamic load balancing algorithm optimizes the allocation of computing resources and improves the overall efficiency of the coupling system. The combination of the conservative interpolation method based on radial basis function (RBF) and the fast multipole method (FMM) achieves efficient and accurate spatial interpolation and ensures the accuracy of data transmission between grids of different resolutions. The use of the adaptive Schwarz alternation algorithm improves the efficiency of model boundary matching and ensures the continuity of results. The development of multi-scale information transfer protocols and adaptive upper and lower scale parameterization schemes enables effective communication and dynamic adjustment between large and small scale models and captures cross-scale wave interactions. The combined effect of these technologies is to build a highly optimized and adaptive multi-model coupling system. For deep-sea intelligent aquaculture platforms, this means that consistent results of large-scale ocean environment and local fine wave prediction can be obtained at the same time. For example, the system can combine global-scale ocean environment models, regional-scale wave models, and high-resolution local models around the platform to provide comprehensive predictions from large-scale weather systems to micro-scale waves around the platform. This multi-scale coupled prediction capability enables the aquaculture platform to simultaneously grasp the large-scale environmental change trends in the open sea and the fine wave conditions around the platform, providing comprehensive support for the platform's long-term planning and short-term operational decisions. Especially in complex sea conditions, such as when a typhoon passes, the system can accurately simulate the impact of large-scale meteorological drives on local waves, providing the platform with more accurate risk assessment and early warning information.

[0100] According to one aspect of the present application, step S52 is specifically:

[0101] Step S521, read the ensemble forecast data set generated in S51, and calculate the model likelihood function. Design an adaptive kernel density estimation method to construct the probability distribution of the observed data. Use particle filtering technology to estimate the state space representation of each model. Apply variational Bayesian inference to calculate the marginal likelihood. Generate a likelihood value sequence for each model.

[0102] Step S522, obtain the likelihood value sequence of S521, and perform improved MCMC sampling. Implement the adaptive Metropolis-Hastings algorithm to improve sampling efficiency. Use the parallel tempering strategy to avoid local optimality. Apply the Hamiltonian Monte Carlo method to improve the sampling performance of high-dimensional space. Generate posterior probability samples of model weights.

[0103] Step S523: read the posterior probability sample of S522 and perform Bayesian model combination. Construct a nonparametric mixture model based on the Dirichlet process. Implement the variational inference algorithm to estimate the optimal mixture weight. Apply cross-validation technology to evaluate the generalization performance of the combination model. Generate comprehensive prediction results and model weight distribution.

[0104] Step S524, obtain the comprehensive prediction results of S523 and calculate the prediction uncertainty. Use multivariate Gaussian process regression to estimate the probability distribution of the prediction. Apply the Monte Carlo integration method to calculate the credible interval of the prediction results. Implement conditional risk assessment and quantify the probability of extreme events. Generate a Bayesian average prediction data set containing weighted average results, probability distributions, and statistics.

[0105] Through the use of adaptive kernel density estimation, improved MCMC sampling, Bayesian model combination and multivariate Gaussian process regression, high-precision and reliable probabilistic wave height prediction is achieved, providing comprehensive risk assessment and decision support for deep-sea intelligent aquaculture platforms. Specifically, the application of adaptive kernel density estimation method improves the estimation accuracy of the probability distribution of observation data and better captures the statistical characteristics of complex sea conditions. The improved MCMC sampling technology, especially the use of adaptive Metropolis-Hastings algorithm and parallel tempering strategy, improves the efficiency and quality of posterior probability sampling. The introduction of non-parametric mixed model based on Dirichlet process realizes flexible and adaptive model combination, effectively utilizing the advantages of multiple prediction models. The application of multivariate Gaussian process regression enables the system to generate continuous probability prediction distribution and fully quantify the uncertainty of prediction. The combined effect of these technologies is to build a highly accurate and reliable probabilistic wave height prediction system. For deep-sea intelligent aquaculture platforms, this means that not only point prediction values ​​can be obtained, but also complete wave height probability distribution information can be obtained. For example, the system can provide the probability of wave height exceeding a specific threshold for each hour in the next 24 hours, as well as wave height interval predictions at different confidence levels. This probabilistic prediction information improves the risk management capabilities of aquaculture platforms. Managers can formulate graded response strategies based on different probability levels, such as starting a level one alert when the probability of wave height exceeding the safety threshold reaches 30%, and starting a level two alert when it reaches 50%. In addition, by analyzing the contribution weights of different models, it is also possible to identify the prediction model that performs best under different sea conditions and continuously optimize the prediction system. This comprehensive and detailed probabilistic prediction capability enables aquaculture platforms to maximize production efficiency while ensuring safety, and achieve refined management in a complex and changing marine environment.

[0106] According to one aspect of the present application, step S55 is specifically:

[0107] Step S551, read the sensitivity analysis data set of S54, the extreme value prediction data set of S53 and the Bayesian average prediction data set of S52, and perform ensemble Kalman filter assimilation. Construct an augmented state vector, including model parameters and state variables. Implement the local ensemble transform Kalman filter (LETKF) algorithm to improve the assimilation performance of nonlinear systems. Apply covariance inflation technology to avoid filter divergence. Generate an updated prediction set and parameter distribution.

[0108] Step S552, obtain the updated prediction set of S551, and perform high-order moment correction. Use Gaussian anamorphosis technology to convert non-Gaussian distribution into approximate Gaussian distribution. Apply multivariate rank histogram method to diagnose and correct high-order statistical moments. Implement multivariate dependency structure modeling based on copula. Generate a corrected prediction probability distribution.

[0109] Step S553: ​​read the corrected predicted probability distribution of S552 and extract the main uncertainty modes. Apply functional principal component analysis (FPCA) to capture the main change patterns of the spatiotemporal field. Use sparse PCA technology to improve the interpretability of the modes. Implement dynamic mode decomposition (DMD) to identify the key features of time evolution. Generate the main modes of uncertainty and their spatiotemporal distribution.

[0110] Step S554: Obtain the main mode of uncertainty in S553 and construct a multi-scale uncertainty quantification framework. Use wavelet decomposition to decompose uncertainty into components of different scales. Apply a conditional autoregressive model to each scale component to capture temporal correlation. Implement spatial random field simulation to describe the spatial structure of uncertainty. Reconstruct the complete multi-scale uncertainty field through inverse wavelet transform.

[0111] Step S555, read the multi-scale uncertainty field of S554 and generate a probability prediction product. Use the Monte Carlo method to generate a large number of possible prediction scenarios through repeated sampling. Calculate the statistical moments of each grid point, including mean, variance, skewness and kurtosis. Apply kernel density estimation to construct the probability density function of each location. Generate the predicted mean field, standard deviation field, probability distribution map and credible interval.

[0112] Step S556: Obtain the probability prediction product of S555, and perform extreme value analysis and risk assessment. Use the peak over threshold (POT) method to extract extreme value samples. Apply the generalized Pareto distribution (GPD) to fit the extreme value distribution. Calculate the risk level and confidence interval of different return periods. Generate an extreme value risk map and risk assessment report.

[0113] Step S557: Integrate all outputs from S551 to S556 to construct a comprehensive uncertainty quantification report. Design an automated report generation system to integrate various statistical charts and analysis results. Implement an interactive visualization interface to allow users to explore uncertainty information at different levels. Develop an explanation generation module based on natural language processing to provide text descriptions for key findings. Generate a final probability prediction dataset containing all uncertainty quantification results.

[0114] Through the techniques of ensemble Kalman filter assimilation, high-order moment correction, multi-scale uncertainty quantification and extreme value analysis, comprehensive and accurate probability prediction and risk assessment are achieved, providing a highly reliable decision support tool for deep-sea intelligent aquaculture platforms. Specifically, the application of the local ensemble transform Kalman filter (LETKF) algorithm improves the assimilation performance of nonlinear systems and better handles complex wave dynamics. The use of Gaussian transform and multivariate rank histogram methods effectively corrects high-order statistical moments and improves the processing capacity of non-Gaussian distribution. The combination of functional principal component analysis (FPCA) and dynamic mode decomposition (DMD) realizes the extraction of the main change modes of the spatiotemporal field and captures the key dynamic characteristics of the wave field. The application of wavelet decomposition and conditional autoregressive model constructs a multi-scale uncertainty quantification framework to comprehensively characterize the prediction uncertainty at different spatiotemporal scales. The use of peak overthreshold (POT) method and generalized Pareto distribution (GPD) realizes high-precision extreme wave height analysis and risk assessment. The combined effect of these technologies is to build a comprehensive, accurate and dynamic probability prediction and risk assessment system. For deep-sea intelligent aquaculture platforms, this means that more comprehensive and reliable wave height forecasts and risk information can be obtained. For example, the system can provide a forecast of the wave height probability distribution every 6 hours in the next week, and also give the probability of occurrence and potential impact of extreme events. This detailed probability forecast and risk assessment information enables aquaculture platforms to formulate more sophisticated and flexible operation strategies. Platform managers can optimize daily operation arrangements, adjust equipment configurations, and even make temporary evacuation decisions based on forecasts of different time scales and risk levels. Especially before extreme weather events occur, the system can provide accurate risk assessments to help the platform find the best balance between ensuring safety and maintaining production. In addition, the system's automated report generation and interactive visualization interface enable complex forecast results to be presented intuitively and clearly, making it easier for decision makers to quickly understand and use this information.

[0115] According to one aspect of the present application, a wave height prediction system for a deep-sea intelligent aquaculture platform is provided, comprising:

[0116] at least one processor; and

[0117] a memory communicatively connected to at least one of the processors; wherein,

[0118] The memory stores instructions executable by the processor, and the instructions are used to be executed by the processor to implement the wave height prediction method for the deep-sea intelligent aquaculture platform described in any one of the above embodiments.

[0119] The preferred embodiments of the present invention are described in detail above; however, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all belong to the protection scope of the present invention.

Claims

1. A wave height prediction method for a deep-sea intelligent aquaculture platform, characterized in that: The steps include: Step S1, acquiring multi-source marine environmental data, performing spatiotemporal alignment, quality control, data filling, outlier processing, spatiotemporal interpolation and scale conversion, and generating high-resolution gridded marine environmental data; Step S2, obtaining high-resolution gridded marine environment data, performing spatiotemporal feature extraction, and obtaining a main spatiotemporal modal data set; constructing a marine environment element association network based on the main spatiotemporal modal data set to generate a network feature data set; applying a nonlinear manifold learning algorithm to the network feature data set and the main spatiotemporal modal data set to perform dimensionality reduction processing to obtain a feature data set after dimensionality reduction; integrating the feature data set after dimensionality reduction, the network feature data set, and the main spatiotemporal modal data set to generate a final low-dimensional and high-information feature data set; Step S3, obtaining a low-dimensional and high-information feature data set, applying a dynamic mode decomposition method to extract the dynamic features of the wave field, and generating a dynamic feature data set; Reconstruct a high-resolution wave field based on the dynamic feature data set to obtain a reconstructed high-resolution wave field data set and perform multi-scale fusion to obtain a multi-scale fused wave field data set; use the multi-scale fused wave field data set and real-time observation data to perform data assimilation to obtain an enhanced wave field data set; integrate the enhanced wave field data set and the dynamic feature data set to perform spatiotemporal interpolation and extrapolation of the wave field to generate a complete wave field data set; Step S4, constructing a multi-scale computational grid based on the complete wave field data set to obtain a grid data set; performing adaptive model selection according to the grid data set and the complete wave field data set to generate a model configuration data set; using the model configuration data set to achieve bidirectional coupling between models and create a coupling configuration data set; Perform parallel calculation based on the coupling configuration data set, the model configuration data set and the grid data set to form a parallel configuration data set; integrate the parallel configuration data set, the coupling configuration data set, the model configuration data set and the grid data set to perform multi-scale coupling simulation and generate a coupling simulation result data set; Step S5: Based on the coupled simulation result data set, perform probability prediction and uncertainty quantification, obtain and perform result visualization and impact analysis based on the probability prediction data set and the prediction evaluation data set.

2. The wave height prediction method for deep sea intelligent aquaculture platform according to claim 1, characterized in that: Step S1 is specifically as follows: Step S11, receiving raw ocean environment data from multiple data sources and storing them in a distributed database to form an original multi-source ocean environment data set; wherein the raw ocean environment data includes sea surface temperature, sea surface height and sea surface wind field data transmitted by a satellite remote sensing system, wave height, wave direction and period data transmitted in real time by a buoy observation network, and ocean environment forecast data output by a numerical simulation system; Step S12, read the original multi-source marine environment data set, use the dynamic time warping algorithm to time align the data with different time resolutions, and generate unified time series data; use the geo-referencing technology to spatially align the data with different spatial resolutions, and create data with a unified spatial grid; apply the outlier detection method based on the interquartile range to identify and mark the outliers in the data, and obtain a data set with anomaly marks; use the data quality index evaluation system to assign a quality mark to each data point, and generate a data set with a quality mark; integrate and store the processed data into a marine environment data set in a unified format; Step S13, obtaining a marine environment data set in a unified format, using a tensor completion algorithm to repair missing data, and obtaining a preliminary repaired data set; correcting the marked outliers to generate an outlier-corrected data set; denoising the corrected data to obtain a denoised data set, and integrating it into a complete and consistent marine environment data set; Step S14, read a complete and consistent marine environment data set, apply the Kriging interpolation method to perform spatial interpolation on the data of discrete observation points to generate a continuous spatial distribution field; use the wavelet transform method to convert the time series data into a time-uniform data set; use multi-scale analysis technology to convert data of different spatial resolutions to a unified high-resolution grid to obtain a high-resolution gridded marine environment data set.

3. The wave height prediction method for deep sea intelligent aquaculture platform according to claim 2, characterized in that: Step S2 is specifically as follows: Step S21, read the high-resolution gridded marine environment data set, standardize the data, eliminate the dimensional differences between different variables, and obtain a standardized data set; calculate the covariance matrix of the standardized data set, and perform eigenvalue decomposition on the matrix; select the first N eigenvectors with cumulative variance contribution exceeding 95% as the main space-time modes; project the original data onto these main modes to obtain the corresponding time coefficient sequence; store the extracted main space-time modes and time coefficient sequence as a space-time modal data set; Step S22, obtaining a spatiotemporal modal data set, calculating the mutual information between the time coefficient sequences of different environmental elements, quantifying the nonlinear correlation between the elements, and generating a mutual information matrix; Based on the mutual information matrix, the adaptive threshold method is used to determine the network connections and form a preliminary network structure; Apply community detection algorithms to identify densely connected subgroups in the network and obtain the network community structure; Calculate the centrality index of each node in the network community structure, including degree centrality, betweenness centrality and eigenvector centrality, complete the construction of the association network between marine environmental elements, and store the constructed association network and its topological characteristics as a network feature dataset; Step S23, read the network feature data set and the spatiotemporal modal data set, use the t-SNE algorithm to map the high-dimensional features to the low-dimensional space, and obtain the preliminary dimensionality reduction results; use the UMAP algorithm to optimize the low-dimensional representation and generate the optimized low-dimensional representation; calculate the quality index of the dimensionality reduction result; select the best low-dimensional representation method and generate the feature data set after dimensionality reduction; Step S24, obtaining a feature data set after dimensionality reduction, a network feature data set, and a spatiotemporal modal data set, using a feature importance ranking method based on a random forest to screen the most representative features and obtain a feature importance ranking result; using a multi-view learning algorithm to fuse feature information from different sources and generate a comprehensive feature representation; The feature representation is compressed using sparse coding technology to obtain a compressed feature representation; the processed features are combined into a unified feature vector and stored as the final low-dimensional and high-information feature data set.

4. The wave height prediction method for deep sea intelligent aquaculture platform according to claim 3, characterized in that: Step S3 is specifically as follows: Step S31, read a low-dimensional high-information feature data set, construct a time delay matrix, organize the feature data into a series of snapshots, and form a snapshot matrix; Perform singular value decomposition on the snapshot matrix, extract the main dynamic modes, and obtain the dynamic mode matrix; calculate the growth rate and frequency corresponding to each dynamic mode, and identify the key dynamic features; Through screening and reconstruction, a simplified model describing the main dynamic behaviors of the wave field is obtained; the extracted dynamic modes and their parameters are stored as a dynamic feature data set; Step S32, obtaining a dynamic feature data set, analyzing the wave field characteristics and constructing a sparse basis suitable for the wave field characteristics, and constructing a sparse representation dictionary; Construct measurement matrices and observation vectors based on limited observation points; use orthogonal matching pursuit algorithm to solve sparse reconstruction problem and obtain the representation of wave field on sparse basis; convert sparse representation back to physical space through inverse transformation to obtain reconstructed high-resolution wave field data; store the reconstruction results as high-resolution wave field data set; Step S33, reading the high-resolution wave field data set and the original multi-scale observation data, using wavelet packet transform to decompose the data of different resolutions into multiple scales and directions, and obtaining multi-scale decomposition coefficients; applying an adaptive weight fusion algorithm on each sub-band to integrate information from different sources, and generating fused sub-band coefficients; using an inverse wavelet packet transform, reconstructing the fused sub-band coefficients into a wave field of uniform resolution; using an edge-preserving filter to remove artifacts introduced in the fusion process; and saving the processed data as a multi-scale fused wave field data set; Step S34, obtaining a multi-scale fusion wave field data set and real-time observation data, constructing an initial set of the ensemble Kalman filter, and generating an observation set using a perturbation observation method; performing a forecasting step, using the extracted dynamic model to perform a short-term forecast on each ensemble member to obtain a forecast set; performing an analysis step, updating the forecast set according to the observation data, and generating an analysis set; calculating the ensemble average as the best estimate, and using the ensemble discreteness to estimate uncertainty; storing the assimilated wave field data and its uncertainty estimate as an enhanced wave field data set; Step S35, integrating the generated enhanced wave field data set and dynamic feature data set, using the result of dynamic mode decomposition, constructing the spatiotemporal evolution model of the wave field; applying the Kriging interpolation method in the spatial domain to fill the observation blank areas and obtain a spatially complete data set; using the Kalman smoother in the time domain to smooth the historical data and perform short-term prediction to generate a time-continuous data set; combining the results of spatial interpolation and time extrapolation to generate continuous wave field data covering all predetermined areas and all time periods; saving the processed data as a complete wave field data set.

5. The wave height prediction method for deep sea intelligent aquaculture platform according to claim 4, characterized in that: Step S4 is specifically as follows: Step S41, read the generated complete wave field data set, apply the quadtree algorithm to adaptively divide the calculation domain according to the spatial variation characteristics of the wave field, and obtain preliminary grid division; use the unstructured grid generation algorithm in the nearshore and key areas to create a high-resolution grid and form a local fine grid; use a regular grid in the open sea to maintain calculation efficiency and generate a global coarse grid; Use mesh smoothing technology to optimize mesh quality and ensure smooth transition between meshes; store the generated multi-scale mesh data as mesh datasets; Step S42, obtaining the created grid data set and the complete wave field data set, calculating the characteristic parameters of each grid unit, and forming a grid characteristic data set; Using the decision tree algorithm, the most appropriate wave model is selected for each grid cell according to the grid feature data set to obtain the preliminary model allocation result; a transition area is constructed at the model junction to ensure smooth switching between different models and generate optimized model allocation; a model configuration file containing the model type and parameter settings for each grid cell is generated and saved as a model configuration data set; the characteristic parameters include water depth, wave steepness and nonlinear parameters; Step S43, read the generated model configuration data set, build a model coupling framework, define the data exchange interface between models, and obtain the coupling framework structure; build a spatial interpolation module for data transmission between grids of different resolutions, and generate data mapping relationships between grids; build a time integration scheme, coordinate the time steps of different models, and create a time synchronization mechanism; for overlapping areas, use the Schwarz alternation method to iteratively solve to ensure the continuity of boundary conditions and obtain a boundary processing scheme; use a feedback mechanism to allow the results of fine-scale models to affect the calculation of large-scale models; store the coupling settings and data exchange rules as a coupling configuration data set; Step S44, obtaining a coupling configuration data set, a model configuration data set and a grid data set, using the METIS graph partitioning algorithm to decompose the computational domain into load-balanced sub-regions, and generating a domain decomposition scheme; allocating computing resources to each sub-region, establishing an MPI communication topology, and forming a parallel computing architecture; and storing the parallel computing settings as a parallel configuration data set; Step S45, integrating the parallel configuration data set, the coupling configuration data set, the model configuration data set and the grid data set, initializing each sub-model, interpolating the complete wave field data onto the multi-scale grid as the initial condition, and obtaining the initialized data set; iteratively executing the calculation of each sub-model according to the preset time step, performing data exchange and boundary condition update between models at each time step, and generating an intermediate result data set; using an adaptive time step control algorithm, dynamically adjusting the step size according to the calculation stability and accuracy requirements, and forming a time step sequence; regularly executing checkpoint saving to facilitate breakpoint continuation and create a checkpoint data set; collecting the calculation results of each sub-area and merging them into a unified output data; storing the intermediate results in the simulation process and the final high-precision wave field prediction results as a coupled simulation result data set.

6. The wave height prediction method for deep sea intelligent aquaculture platform according to claim 4, characterized in that: In step S5, the process of probability prediction and uncertainty quantification is performed based on the coupled simulation result data set, specifically: Step S51, reading the coupled simulation result data set, using the perturbation observation method to generate multiple sets of perturbation fields of initial conditions and boundary conditions to form a perturbation data set; using the random physical parameterization method to assign probability distribution to key physical parameters in the model, and create a parameter perturbation set; running multiple coupled simulations, each time using different initial conditions, boundary conditions and physical parameter combinations, to generate multiple sets of simulation results; Collect all simulation results to form an ensemble forecast member set and store it as an ensemble forecast dataset; Step S52, obtaining an ensemble forecast data set, calculating the likelihood function of each ensemble member relative to the historical observation data, and obtaining a likelihood value sequence; using the Markov chain Monte Carlo method to estimate the posterior probability distribution of each model, and forming a posterior probability data set; performing weighted averaging on the prediction results of all ensemble members according to the posterior probability, and generating a comprehensive prediction result; calculating the credible interval and probability density function of the comprehensive prediction result; and storing the weighted average result and related statistics as a Bayesian average prediction data set; Step S53, read the Bayesian average prediction data set, use the peak overthreshold method to extract extreme value samples from the prediction data to form an extreme value sample set; fit the generalized Pareto distribution and estimate the probability distribution parameters of the extreme value to obtain a distribution parameter set; calculate the design wave heights of different regression periods to generate a design wave height data set; use the bootstrap method to estimate the confidence interval of the extreme value prediction; The extreme value analysis results are stored as an extreme value prediction data set; Step S54, obtaining an extreme value prediction data set and a Bayesian average prediction data set, using the Sobol method to calculate the global sensitivity index of each input parameter to the prediction result, and obtaining a parameter sensitivity ranking; The FAST algorithm is used to analyze the interaction effects between parameters and generate a parameter interaction matrix; a response surface model describing the relationship between input parameters and output results is constructed to form a response surface function; Based on the response surface model, the Monte Carlo method is used to perform large sample simulation to obtain the probability distribution of the prediction results; the sensitivity analysis results and probability distribution information are stored as a sensitivity analysis data set; Step S55, integrating the sensitivity analysis data set, the extreme value prediction data set and the Bayesian average prediction data set, using the ensemble Kalman filter method to assimilate the latest observation data, updating the prediction results and their uncertainties, and obtaining an updated prediction set; using the covariance inflation technique to generate a corrected prediction set; calculating the spatiotemporal correlation structure of the prediction results, using the EOF decomposition method to extract the main uncertainty modes, and forming an uncertainty mode data set; generating a probability prediction product and storing it as a probability prediction data set, the probability prediction product includes a prediction mean field, a standard deviation field, a probability distribution map and a credible interval; Step S56, read the probability prediction data set, use at least two scoring methods to evaluate the accuracy of the probability prediction, and obtain a skill scoring data set; analyze the reliability and resolution of the prediction through the attribute graph and the reliability graph, and generate a reliability analysis result; calculate the spatial distribution and time evolution characteristics of the prediction deviation to form a deviation distribution data set; use Taylor graph and Murphy-Winkler decomposition to evaluate the performance of the prediction system; store the evaluation results and error statistics as a prediction evaluation data set; wherein the scoring methods include Brier scoring, sorted probability scoring, and continuous probability scoring.

7. The wave height prediction method for deep sea intelligent aquaculture platform according to claim 5, characterized in that: In step S5, result visualization and impact analysis are performed based on the probability prediction data set and the prediction evaluation data set, specifically: Step S57, reading the probability prediction data set and the prediction evaluation data set, using the contour map technology to draw the spatial distribution map of the predicted wave height field, and generating spatial distribution visualization data; Step S58, generating a heat map of wave height changes by a color mapping method to form a heat map data set; constructing a time series animation to display the time evolution process of the wave field and creating animation sequence data; Step S59: for the probability information, use box plots and violin plots to display the predicted uncertainty distribution, and obtain probability distribution visualization data; Use parallel coordinates to show the relationship between multiple predictor variables and form multivariate relationship graph data; Save these visualizations as a visualization dataset.

8. The wave height prediction method for deep sea intelligent aquaculture platform according to claim 7, characterized in that: The step S13 is specifically as follows: Step S131, read the marine environment data set in a unified format, organize the marine environment data into a three-order tensor structure, where time, space and variables correspond to three dimensions respectively; use wavelet transform to decompose the tensor into sub-tensors of multiple scales; apply the nuclear norm minimization algorithm to complete each sub-tensor; reconstruct the complete tensor through inverse wavelet transform to obtain a preliminary repair data set; Step S132, obtaining a preliminary repaired data set; using a sliding window method to divide the data subsets in time and space dimensions; applying robust principal component analysis to each subset to decompose the data into a low-rank part and a sparse abnormal part; identifying and marking outliers based on Mahalanobis distance and local heteroscedasticity estimation; using local Gaussian process regression to correct the marked outliers to generate an outlier-corrected data set; Step S133, read the data set after outlier correction, build a correlation network between variables, and use mutual information and conditional mutual information to quantify the variable relationship; based on the network structure, build a multivariate empirical mode decomposition algorithm with adaptive weights, and apply soft threshold wavelet denoising to each decomposition mode; reconstruct the denoised signal to obtain the final complete and consistent marine environment data set.

9. The wave height prediction method for deep sea intelligent aquaculture platform according to claim 7, characterized in that: The step S23 is specifically as follows: Step S231, read the network feature data set and the spatiotemporal modal data set, construct a K nearest neighbor graph, use an adaptive bandwidth algorithm to determine the local neighborhood, calculate the geodesic distance matrix, use the fast marching method to improve the computational efficiency, apply local linear embedding to perform preliminary dimensionality reduction, and obtain an initial low-dimensional representation; based on the initial low-dimensional representation, adaptively adjust the neighbor relationship to generate an optimized neighbor graph structure; Step S232, obtain the optimized neighbor graph structure and the initial low-dimensional representation, use the adaptive learning rate scheduling strategy to dynamically adjust the gradient descent process; accelerate the gradient through the momentum term and Nesterov; use the Barnes-Hut approximation algorithm to accelerate the high-dimensional similarity calculation; use the multi-scale strategy to gradually optimize the embedding from coarse to fine, and obtain the t-SNE dimensionality reduction result; Step S233, read the t-SNE dimension reduction result, construct a fuzzy topological representation, and use adaptive fuzziness parameters to capture local structure; construct an optimization target based on Riemannian geometry, and apply random negative sampling technology and parallel computing strategy to accelerate the optimization process, and generate UMAP dimension reduction results; Step S234, obtain the t-SNE dimensionality reduction results and UMAP results, use the local preservation projection method to map the two results to a common low-dimensional space; apply the weighted voting mechanism to integrate the advantages of different methods; use the manifold alignment technology to ensure the consistency of the results of different methods; calculate the dimensionality reduction quality index and select the best representation method to generate the final feature data set after dimensionality reduction.

10. A wave height prediction system for deep sea intelligent aquaculture platform, characterized in that: include: at least one processor; as well as a memory communicatively connected to at least one of the processors; wherein, The memory stores instructions executable by the processor, and the instructions are used to be executed by the processor to implement the wave height prediction method for a deep-sea intelligent aquaculture platform as described in any one of claims 1 to 9.

Images