A coastal water level prediction method based on multi-modal observation data

Abstract

The present application provides a kind of coastal water level prediction method based on multi-modal observation data, belong to coastal water level prediction technical field, the present application is by collecting the multi-modal marine observation data of different sensors, the spatial and temporal alignment of heterogeneous data is realized by establishing multi-scale space-time registration matrix, constructs marine dynamic process feature extractor to identify astronomical tide storm surge and wave characteristics and calculate nonlinear coupling parameters, using adaptive space-time fusion algorithm according to the weight of data quality dynamic adjustment generation space-time consistency dataset, establishes marine dynamic coupling strength discrimination model to determine modeling strategy, finally generates the coastal water level prediction product containing prediction value confidence interval and risk warning, solves the technical problem that multi-modal marine observation data is difficult to effectively fuse under the condition of space-time scale mismatch, resulting in insufficient coastal water level prediction accuracy.

Description

Technical Field

[0001] This invention belongs to the field of coastal water level prediction technology, and more specifically, relates to a coastal water level prediction method based on multimodal observation data. Background Technology

[0002] Coastal water level prediction is a crucial technical support for marine disaster prevention and mitigation and marine engineering construction. Traditional methods mainly rely on single-sensor data for prediction, such as using only tide gauge observation data to build harmonic analysis models or using only satellite remote sensing data for statistical regression analysis. These methods play a fundamental role in applications such as port and shipping management, coastal engineering design, and marine disaster early warning. However, these traditional technologies face significant drawbacks. The observation data acquired by different sensors have different sampling frequencies: satellite remote sensing data is sampled on an hourly basis, tide gauge data on a minute basis, and high-frequency radar data on a second basis. At the same time, the spatial resolution spans a huge scale from meters to kilometers, making it impossible to directly fuse heterogeneous data. A single data source is also insufficient to comprehensively characterize the complex coupling effects of various marine dynamic processes such as astronomical tides, storm surges, and waves. In existing technologies, due to the lack of effective multi-scale spatiotemporal registration mechanisms and adaptive fusion algorithms, it is impossible to achieve precise alignment of data from different sensors on the time axis and spatial grid. The nonlinear coupling relationships between various dynamic factors are difficult to quantify accurately, especially under extreme weather conditions, which significantly increases prediction errors, resulting in severely insufficient reliability and accuracy of prediction results. In other words, existing technologies suffer from the technical problem of insufficient accuracy in coastal water level prediction due to the difficulty in effectively fusing multimodal ocean observation data under conditions of mismatched spatiotemporal scales. Summary of the Invention

[0003] In view of this, the present invention provides a coastal water level prediction method based on multimodal observation data, which can solve the technical problem in the prior art that the multimodal marine observation data is difficult to effectively fuse under the condition of mismatch in spatiotemporal scale, resulting in insufficient accuracy of coastal water level prediction.

[0004] This invention is implemented as follows: It provides a coastal water level prediction method based on multimodal observation data, comprising: collecting multimodal ocean observation data from different sensors, including satellite remote sensing data, tide gauge data, and high-frequency radar data; establishing a multi-scale spatiotemporal registration matrix to align observation data with different sampling frequencies along the time axis, and unifying data with different spatial resolutions into the same spatial grid system using bilinear interpolation; constructing an ocean dynamic process feature extractor to extract astronomical tide features, storm surge features, and wave features, and calculating nonlinear coupling strength parameters and nonstationarity indices; employing an adaptive spatiotemporal fusion algorithm to fuse the multimodal ocean observation data, determining dynamic weight allocation coefficients based on the data quality assessment results of each sensor, and generating a fused spatiotemporally consistent dataset; establishing an ocean dynamic coupling strength discrimination model, initiating a nonlinear modeling process based on the nonlinear coupling strength parameters, and calculating the nonlinear coupling strength parameters and nonstationarity indices based on the nonlinear coupling strength parameters and nonstationarity indices; and using an adaptive spatiotemporal fusion algorithm to fuse the multimodal ocean observation data, determining dynamic weight allocation coefficients based on the data quality assessment results of each sensor, and generating a fused spatiotemporally consistent dataset; establishing an ocean dynamic coupling strength discrimination model, initiating a nonlinear modeling process based on the nonlinear coupling strength parameters, and calculating nonlinear coupling strength parameters based on the nonlinear coupling strength parameters and nonstationarity indices. The stability index adds an extreme weather condition processing module; a sparse matrix compressed sensing reconstruction mechanism is constructed, which reconstructs a high-dimensional original matrix from a low-dimensional observation matrix using the sparsity assumption of the spatiotemporal consistency dataset; an eigenvalue sensitivity analysis mechanism is designed to calculate the sensitivity of the eigenvalues ​​of the multi-scale spatiotemporal registration matrix to changes in nonlinear coupling strength parameters and adjust the dynamic weight allocation coefficient control strategy accordingly; a multi-level prediction framework is constructed, which automatically selects the appropriate prediction algorithm and parameter configuration according to the prediction duration, and generates a preliminary prediction result set using the high-dimensional original matrix and the dynamic weight allocation coefficient control strategy; the preliminary prediction result set is iteratively optimized using a marine dynamic response optimization model, and the prediction parameters are adjusted through residual analysis and error feedback mechanisms to output the final prediction result; based on the final prediction result, a multi-timescale coastal water level prediction product is generated, including predicted values, confidence intervals, and risk warning levels, forming a complete coastal water level prediction output.

[0005] The multi-scale spatiotemporal registration matrix refers to a mathematical tool used to solve the problem of spatiotemporal scale mismatch between data from different sensors. It achieves a unified representation of heterogeneous data by establishing a time interpolation function and a spatial mapping relationship. The bilinear interpolation method refers to a spatial data resampling technique that estimates the value of the target location by weighted averaging of adjacent grid points.

[0006] The ocean dynamic process feature extractor refers to an algorithm module used to identify and quantify various physical processes in the ocean, which can automatically identify the periodic characteristics of astronomical tides, the sudden characteristics of storm surges, and the random characteristics of waves.

[0007] The astronomical tide characteristics refer to the periodic water level changes caused by the gravitational influence of celestial bodies, including the amplitude and phase information of semi-diurnal and diurnal tides. The storm surge characteristics refer to the statistical characteristics of abnormal sea level rise and fall caused by meteorological factors, including the degree of influence of wind speed, air pressure and wind direction on water level. The wave characteristics refer to the statistical descriptive parameters of sea surface fluctuations, including physical quantities such as significant wave height, period and wave direction.

[0008] The nonlinear coupling strength parameter refers to a dimensionless index that measures the intensity of interaction between different ocean dynamic factors. It is calculated by dividing the covariance of each factor by the square root of the product of their respective variances. The nonstationarity index refers to an index that describes the degree of change of the statistical characteristics of a time series over time. It is quantified by the coefficient of variation of the mean and variance within a sliding window.

[0009] The adaptive spatiotemporal fusion algorithm refers to a data processing method that dynamically adjusts the fusion weights based on data quality and reliability. It optimizes the fusion effect by evaluating the signal-to-noise ratio and consistency of each sensor's data in real time. The dynamic weight allocation coefficient refers to the fusion weight value calculated in real time based on the data quality of each sensor. The calculation basis includes data accuracy, timeliness, and spatial coverage factors.

[0010] The spatiotemporal consistency dataset refers to a multimodal data set with a unified spatiotemporal benchmark formed after processing with a multi-scale spatiotemporal registration matrix. The ocean dynamic coupling strength discrimination model refers to a classifier built based on statistical learning theory, used to determine whether a nonlinear modeling process needs to be initiated under the current marine environment.

[0011] The extreme weather conditions processing module refers to the water level prediction algorithm enhancement module designed for extreme weather conditions such as typhoons and cold waves. It includes an extreme weather identification unit, a nonlinear dynamic amplification unit, and a prediction error compensation unit. The extreme weather identification unit judges the current marine environment state based on the non-stationarity index, the nonlinear dynamic amplification unit performs nonlinear coupling calculations on astronomical tide characteristics, storm surge characteristics, and wave characteristics, and the prediction error compensation unit corrects the prediction results based on the statistical law of prediction deviations of historical extreme weather events.

[0012] The data flow of the extreme weather condition processing module is as follows: it receives astronomical tide features, storm surge features, wave features and non-stationarity index from the ocean dynamic process feature extractor, processes them internally and then passes the enhanced feature data to the adaptive spatiotemporal fusion algorithm. At the same time, it passes the extreme weather identification results to the ocean dynamic coupling strength discrimination model to determine whether to start the nonlinear modeling process.

[0013] The sparse matrix compressed sensing reconstruction mechanism refers to a method of reconstructing original high-dimensional data from a small amount of observation data by utilizing the principle of signal sparsity. The low-dimensional observation matrix refers to a low-dimensional data matrix obtained through random sampling or compressed measurement. The high-dimensional original matrix refers to the original data matrix containing complete information.

[0014] The eigenvalue sensitivity analysis mechanism refers to a mathematical analysis tool for evaluating the degree of response of matrix eigenvalues ​​to parameter disturbances. It quantifies sensitivity by calculating the partial derivatives of eigenvalues ​​with respect to parameters. The dynamic weight allocation coefficient control strategy refers to a control method that adjusts weight allocation based on the results of the eigenvalue sensitivity analysis mechanism.

[0015] The multi-level prediction framework refers to a hierarchical prediction architecture, including a short-term prediction layer, a medium-term prediction layer, and a long-term prediction layer. Different algorithm strategies and model parameters are selected according to the prediction duration. The prediction algorithm refers to the mathematical algorithm used to perform water level prediction calculations, including time series analysis algorithms, machine learning algorithms, and physical model algorithms. The parameter configuration refers to the set of parameters that control the behavior of the prediction algorithm.

[0016] The marine dynamic response optimization model refers to a hybrid model that combines physical mechanisms and data-driven methods. It uses a meta-learning-based optimization algorithm to learn strategies, learns update rules through a meta-optimizer, adopts an LSTM recurrent network parameterized optimizer, reduces the number of parameters by using coordinate sharing, and improves the optimization effect through preprocessing techniques.

[0017] The specific structure of the marine dynamic response optimization model is a multi-layer encoder-decoder architecture. The encoder part uses a multi-head attention mechanism to process the preliminary prediction result set, and the decoder part uses a sparse attention mechanism to generate the final prediction result. The number of attention heads in the model is determined according to the nonlinear coupling strength parameter, the sparsity parameter is adaptively adjusted according to the non-stationarity index and the data dimension, and the hidden layer dimension is dynamically configured according to the prediction duration.

[0018] The steps for establishing the training dataset of the marine dynamic response optimization model specifically include collecting historical multimodal marine observation data and corresponding measured water levels, dividing the data into training and validation sets according to time series, performing standardized preprocessing and outlier detection on the data, constructing input feature vectors and target prediction sequences, expanding the number of training samples through data augmentation techniques, and establishing standard data subsets under different sea conditions.

[0019] The specific steps of training the marine dynamic response optimization model include initializing model parameters and meta-optimizer state, using a gradient-based meta-learning algorithm for multi-task learning, automatically adjusting the learning rate and update direction through an LSTM parameterized optimizer, using a coordinate sharing mechanism to share parameters among different tasks to reduce overfitting, employing early stopping strategies and regularization techniques to improve model generalization ability, evaluating model performance through cross-validation, and optimizing hyperparameters.

[0020] This invention achieves a unified representation of data with different sampling frequencies and spatial resolutions by establishing a multi-scale spatiotemporal registration matrix. It employs an adaptive spatiotemporal fusion algorithm to dynamically adjust fusion weights based on data quality, constructs a marine dynamic coupling strength discrimination model to identify nonlinear modeling needs, designs a sparse matrix compressed sensing reconstruction mechanism to recover high-dimensional original data from low-dimensional observations, establishes a multi-level prediction framework to automatically select prediction algorithms, and uses a marine dynamic response optimization model for iterative optimization, thus solving the problem of effective fusion of multimodal marine observation data. This scheme unifies spatial resolution to the same grid system through bilinear interpolation, aligns the time axis of heterogeneous data using a temporal interpolation function to eliminate the obstacle of spatiotemporal scale mismatch, adjusts adaptive weight allocation coefficients in real time based on signal-to-noise ratio and consistency to ensure the quality and reliability of fused data, calculates nonlinear coupling strength parameters and non-stationarity exponents to enable the model to accurately identify complex marine dynamic processes, uses an eigenvalue sensitivity analysis mechanism for fine-tuning parameters, employs a multi-layer encoder-decoder architecture combined with a multi-head attention mechanism to effectively process multi-source data features, and uses a meta-learning-based optimization algorithm to improve the model's adaptability by learning historical trajectories. In summary, this invention achieves effective integration of heterogeneous observation data through multi-scale spatiotemporal registration and adaptive fusion processing, enhances feature extraction capabilities through dynamic coupling strength discrimination and compressed sensing reconstruction, and improves prediction accuracy through a multi-level prediction framework and iterative optimization mechanism. Thus, it solves the technical problem mentioned in the background art of insufficient accuracy in coastal water level prediction due to the difficulty in effectively fusing multimodal ocean observation data under conditions of spatiotemporal scale mismatch. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention.

[0022] Figure 2 This is a diagram illustrating the iterative convergence process of the marine dynamic response optimization model in the embodiment.

[0023] Figure 3 This is a comparison chart of data quality and accuracy in the examples.

[0024] Figure 4 This is a comparison chart showing the improvement in training efficiency in the embodiments.

[0025] Figure 5This is a schematic diagram illustrating the impact of the nonlinear coupling strength parameter on prediction accuracy in the embodiment.

[0026] Figure 6 This is a schematic diagram comparing the performance of the present invention with that of the prior art in the embodiments. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0028] like Figure 1 The diagram shown is a flowchart of a coastal water level prediction method based on multimodal observation data provided by the present invention. This method includes the following steps:

[0029] S01. Collect multimodal ocean observation data from different sensors, including satellite remote sensing data, tide gauge data, and high-frequency radar data. The sampling frequency of satellite remote sensing data is on the hourly level, the sampling frequency of tide gauge data is on the minute level, and the sampling frequency of high-frequency radar data is on the second level.

[0030] S02. Establish a multi-scale spatiotemporal registration matrix, align the observation data with different sampling frequencies along the time axis, and unify the spatial resolution data from meter level to kilometer level into the same spatial grid system through bilinear interpolation.

[0031] S03. Construct a feature extractor for ocean dynamic processes, extract astronomical tide features, storm surge features and wave features, and calculate the nonlinear coupling strength parameters and nonstationarity index of each dynamic factor;

[0032] S04. An adaptive spatiotemporal fusion algorithm is used to fuse multimodal ocean observation data. Dynamic weight allocation coefficients are determined based on the data quality assessment results of each sensor, and a fused spatiotemporally consistent dataset is generated.

[0033] S05. Establish a marine dynamic coupling strength discrimination model. When the nonlinear coupling strength parameter ∈ [0.75, 1.0], start the nonlinear modeling process. When the nonstationarity index > 0.6, add an extreme weather condition processing module.

[0034] S06. Construct a sparse matrix compressed sensing reconstruction mechanism. Utilize the sparsity assumption of the spatiotemporal consistent dataset to reconstruct a high-dimensional original matrix from a low-dimensional observation matrix. The compressed sensing algorithm is triggered when the ratio of the dimension of the observation matrix to the dimension of the original matrix is ​​∈ [0.1, 0.3).

[0035] S07. Design an eigenvalue sensitivity analysis mechanism to calculate the sensitivity of the eigenvalues ​​of the multi-scale spatiotemporal registration matrix to changes in the nonlinear coupling strength parameter and adjust the dynamic weight allocation coefficient control strategy accordingly.

[0036] S08. Construct a multi-level prediction framework, automatically select the corresponding prediction algorithm and parameter configuration according to the prediction duration, and generate a preliminary prediction result set using a high-dimensional original matrix and dynamic weight allocation coefficient control strategy.

[0037] S09. Iteratively optimize the preliminary prediction result set using the marine dynamic response optimization model, adjust the prediction parameters through residual analysis and error feedback mechanism, and trigger the model retraining process when the prediction error is >15%, and output the final prediction result.

[0038] S10. Generate coastal water level forecast products at multiple time scales based on the final forecast results, including forecast values, confidence intervals, and risk warning levels, forming a complete coastal water level forecast output.

[0039] Among them, the multi-scale spatiotemporal registration matrix is ​​a mathematical tool used to solve the problem of spatiotemporal scale mismatch between data from different sensors. It achieves a unified representation of heterogeneous data by establishing a temporal interpolation function and a spatial mapping relationship. The bilinear interpolation method is a spatial data resampling technique that estimates the numerical value of a target location by weighted averaging of adjacent grid points. The ocean dynamic process feature extractor is an algorithm module used to identify and quantify various physical processes in the ocean, capable of automatically identifying the periodic characteristics of astronomical tides, the suddenness of storm surges, and the randomness of wave characteristics.

[0040] Astronomical tidal characteristics refer to the periodic water level changes caused by the gravitational influence of celestial bodies, including the amplitude and phase information of semi-diurnal and diurnal tides. Storm surge characteristics refer to the statistical properties of abnormal sea level rises and falls caused by meteorological factors, including the degree of influence of wind speed, air pressure, and wind direction on water levels. Wave characteristics refer to the statistical descriptive parameters of sea surface fluctuations, including physical quantities such as significant wave height, period, and wave direction. Nonlinear coupling strength parameters are dimensionless indices that measure the intensity of interaction between different ocean dynamic factors; they are calculated by dividing the covariance of each factor by the square root of the product of their respective variances.

[0041] Nonstationarity index is an indicator describing the degree of change of statistical characteristics of a time series over time, quantified by the coefficient of variation of the mean and variance within a sliding window. Adaptive spatiotemporal fusion algorithm is a data processing method that dynamically adjusts fusion weights based on data quality and reliability, optimizing the fusion effect by real-time evaluation of the signal-to-noise ratio and consistency of data from each sensor. The dynamic weight allocation coefficient is a fusion weight value calculated in real-time based on the data quality of each sensor, taking into account factors such as data accuracy, timeliness, and spatial coverage.

[0042] The spatiotemporal consistency dataset is a multimodal data set with a unified spatiotemporal benchmark formed after processing with a multi-scale spatiotemporal registration matrix. The ocean dynamic coupling strength discrimination model is a classifier built based on statistical learning theory, used to determine whether a nonlinear modeling process needs to be initiated under the current ocean environment. The extreme weather condition processing module is an enhancement module for water level prediction algorithms designed for extreme meteorological conditions such as typhoons and cold waves. This module includes an extreme weather identification unit, a nonlinear dynamic amplification unit, and a prediction error compensation unit. The extreme weather identification unit judges the current ocean environment state based on the nonstationarity index. When the nonstationarity index > 0.6, the nonlinear dynamic amplification unit is activated. The nonlinear dynamic amplification unit performs nonlinear coupling calculations on astronomical tide characteristics, storm surge characteristics, and wave characteristics. The prediction error compensation unit corrects the prediction results based on the statistical regularity of prediction deviations from historical extreme weather events. The data flow of the extreme weather condition processing module is as follows: it receives astronomical tide features, storm surge features, wave features and non-stationarity index from the ocean dynamic process feature extractor. After internal processing, the enhanced feature data is passed to the adaptive spatiotemporal fusion algorithm. At the same time, the extreme weather identification results are passed to the ocean dynamic coupling strength discrimination model to determine whether to start the nonlinear modeling process.

[0043] Sparse matrix compressed sensing reconstruction mechanism is a mathematical method that utilizes the principle of signal sparsity to reconstruct original high-dimensional data from a small amount of observation data. This is achieved by optimizing the solution to minimize... Norm problems enable data reconstruction. The low-dimensional observation matrix is ​​a data matrix with lower dimensionality obtained through random sampling or compressed measurements, while the high-dimensional original matrix is ​​the original data matrix containing complete information. Compressed sensing algorithms are signal reconstruction algorithms based on sparse representation theory, utilizing greedy tracking or convex optimization methods to recover the original signal from compressed observations.

[0044] Eigenvalue sensitivity analysis is a mathematical analysis tool used to evaluate the degree of response of matrix eigenvalues ​​to parameter perturbations. It quantifies sensitivity by calculating the partial derivatives of the eigenvalues ​​with respect to the parameters. The dynamic weight allocation coefficient control strategy is a control method that adjusts weight allocation based on the results of the eigenvalue sensitivity analysis mechanism. When the sensitivity value > 0.8, the corresponding weight is increased; when the sensitivity value < 0.2, the corresponding weight is decreased.

[0045] A multi-level prediction framework is a hierarchical prediction architecture that includes short-term, medium-term, and long-term prediction layers, selecting different algorithm strategies and model parameters based on the prediction duration. Prediction algorithms are mathematical algorithms used to perform water level prediction calculations, including time series analysis algorithms, machine learning algorithms, and physical model algorithms. Parameter configuration is a set of parameters that control the behavior of the prediction algorithm, including window length, learning rate, and regularization parameters.

[0046] The preliminary prediction result set is a dataset containing multiple prediction candidate results generated by a multi-level prediction framework. The marine dynamic response optimization model is a hybrid model combining physical mechanisms and data-driven methods. It utilizes a meta-learning-based optimization algorithm to learn strategies, learns update rules through a meta-optimizer, employs an LSTM recurrent network parameterized optimizer, reduces the number of parameters through coordinate sharing, and improves optimization performance through preprocessing techniques.

[0047] Residual analysis is a statistical method for evaluating model performance by calculating the difference between predicted and measured values. Error feedback mechanisms are adaptive control methods that automatically adjust model parameters based on prediction errors. Prediction parameters are a set of key parameters controlling the predictive behavior of ocean dynamic response optimization models, including time window length, feature weights, and activation function parameters. The model retraining process is the complete procedure for recollecting training data and updating model parameters when prediction performance deteriorates.

[0048] The final prediction result is the optimal prediction output obtained after iterative optimization using the marine dynamic response optimization model. The coastal water level prediction product is a comprehensive prediction result that includes predicted values, uncertainty quantification, and risk assessment information. The coastal water level prediction output is the final system output result formed based on the coastal water level prediction product.

[0049] In addition, the present invention also provides a method for forming a coastal water level prediction system by means of a computer, wherein the computer is provided with a storage medium, the storage medium storing program instructions, and the program instructions are used to execute the above-mentioned method when the computer is run.

[0050] The specific implementation methods of the above steps are described in detail below.

[0051] The specific implementation of step S01 involves synchronous acquisition and preprocessing of heterogeneous marine data through a multi-type sensor network. The purpose of this step is to acquire raw observation data covering different spatiotemporal scales to support subsequent fusion analysis. First, a sensor network topology is established, logically grouping satellite remote sensing systems, shore-based tide gauge stations, and high-frequency radar arrays according to their spatial distribution characteristics. This ensures sufficient spatial overlap in the observation areas of each sensor for subsequent spatiotemporal registration. Then, the data acquisition parameters for each sensor are configured. The sampling period for the satellite remote sensing system is set to 1 hour to capture large-scale marine environmental changes; the sampling period for the tide gauge system is set to 1 to 10 minutes to monitor mesoscale water level fluctuations; and the sampling period for the high-frequency radar system is set to 1 to 30 seconds to acquire small-scale wave motion characteristics. Next, a preliminary data quality inspection process is implemented to determine the validity of the acquired raw data and remove outliers exceeding the physically possible range. This step employs an outlier detection algorithm based on statistical distribution, identifying abnormal observations by calculating the deviation of data points from the population mean. Data points with a difference exceeding three standard deviations from the mean are marked as suspicious. Subsequently, timestamps and spatial coordinate labels were added to the data from each sensor to establish a unified spatiotemporal indexing system. This indexing system uses Coordinated Universal Time (UTC) as the time reference and the WGS84 coordinate system as the spatial reference to ensure consistency of data from different sources within the spatiotemporal reference frame. Finally, the preprocessed multimodal observation data was stored in a distributed database system, employing a time-series database architecture to support efficient time-series data query and retrieval operations.

[0052] The specific implementation of step S02 is to achieve spatiotemporal alignment of heterogeneous data by constructing a multi-dimensional interpolation mapping relationship. This step aims to eliminate the differences in temporal and spatial resolution between observation data from different sensors. First, the temporal sampling characteristics of various observation data are analyzed to identify the temporal scale differences between hourly sampling of satellite remote sensing data, minute-level sampling of tide gauge data, and second-level sampling of high-frequency radar data. A time scale transformation matrix is ​​established, which projects data with different sampling frequencies onto a unified time reference axis through a time interpolation function. The time interpolation function uses cubic spline interpolation, which can maintain the local variation characteristics of the data while ensuring the continuity of the interpolation curve. The interval of the interpolation nodes is determined according to the target temporal resolution, typically set to 5 to 15 minutes to balance computational efficiency and data accuracy. Then, the spatial resolution mismatch problem is addressed. The spatial resolution of satellite remote sensing data is typically on the order of kilometers, while that of high-frequency radar data is on the order of hundreds of meters. Tide gauge data is a point measurement, requiring the unification of these data at different spatial scales into a regular grid system. The spatial grid system employs an equally spaced latitude and longitude grid, with the grid spacing ranging from 500 to 2000 meters, the specific values ​​determined based on the spatial extent of the study area and computational resource constraints. Spatial data interpolation utilizes a bilinear interpolation algorithm, which weights four known data points surrounding the target grid point, with the weighting coefficients inversely proportional to the distance, calculated using Euclidean distance. Next, a spatiotemporal registration accuracy evaluation mechanism is established, quantifying registration errors by comparing measurements from different sensors in overlapping observation areas. A parameter adjustment process is triggered when the registration error exceeds 10%. Finally, a multi-scale spatiotemporal registration matrix is ​​generated, containing temporal interpolation coefficients and spatial mapping relationships, providing a unified spatiotemporal reference framework for subsequent data fusion.

[0053] The specific implementation of step S03 involves extracting key physical features of ocean dynamic processes through multi-level signal decomposition and feature recognition algorithms. The purpose of this step is to separate the contributions of different physical processes from the raw observation data and quantify their statistical characteristics. First, frequency domain analysis is performed on the spatiotemporally registered water level observation sequence. The fast Fourier transform method is used to convert the time-domain signal to the frequency domain, identifying the astronomical tidal components with periodic characteristics. The identification of astronomical tidal components is based on tidal harmonic analysis theory. The amplitude and phase information of semi-diurnal and diurnal tides are extracted by identifying the main periodic components of 12.42-hour and 24.84-hour periods, with amplitude accuracy required to be at the centimeter level and phase accuracy at the degree level. Then, storm surge characteristics are extracted. These characteristics manifest as abrupt changes and trends in the water level sequence. The empirical mode decomposition method is used to decompose the water level signal into multiple intrinsic mode functions, extracting the components with a variation period ranging from several hours to several days as storm surge signals. The quantitative indicators of storm surge characteristics include maximum rise amplitude, duration, and rate of rise. These indicators are obtained through statistical analysis of the separated storm surge components. Next, wave features are extracted, and individual waves are identified using the zero-upward crossing method. Parameters such as significant wave height, mean period, and spectral peak period are statistically analyzed. Significant wave height is defined as the average of the maximum third of the wave height, typically varying between 0.5 and 5 meters. Then, the nonlinear coupling strength parameter is calculated. This parameter is obtained by analyzing the covariance matrix among astronomical tides, storm surges, and waves. Specifically, the time series of any two components are extracted, their normalized cross-correlation coefficients are calculated, and the average of all paired cross-correlation coefficients is taken as the overall coupling strength index. This index ranges from 0 to 1; a value close to 1 indicates a strong coupling state. Simultaneously, the nonstationarity index is calculated using a sliding window method to analyze the statistical characteristics of the time series. The window length is set to 24 to 72 hours. Within each window, the mean and variance are calculated, and the coefficient of variation of these local statistics is used as a measure of nonstationarity. A coefficient of variation exceeding 0.6 indicates that the series has significant nonstationar characteristics, at which point an enhanced processing module for extreme weather conditions needs to be activated.

[0054] The specific implementation of step S04 involves achieving optimal fusion of multi-source data through a dynamic weight optimization algorithm. This step aims to comprehensively utilize the observation information from various sensors to generate high-quality, consistent data products. First, a data quality assessment system is established, which comprehensively scores the data from each sensor across three dimensions: measurement accuracy, time delay, and spatial coverage. Measurement accuracy is assessed by analyzing observation noise levels and systematic error characteristics, using the standard deviation of repeated observations as the accuracy indicator; a smaller standard deviation indicates higher accuracy. Time delay assessment considers the time interval between observation and usability, with real-time observation data having a higher weight than delayed data. Spatial coverage assessment is obtained by calculating the overlap ratio between the effective observation range of the sensors and the target area; data with high coverage has a greater weight in the fusion process. Then, dynamic weight allocation coefficients are calculated based on the quality assessment results. A normalized exponential weighting method is used to convert the comprehensive quality score of each sensor into fusion weights, with the sum of all weights constrained to 1. The weight calculation formula considers the nonlinear response characteristics of data quality; the weight of data with higher quality scores increases exponentially, with the exponential parameter set between 1.5 and 2.5 to achieve a reasonable weight allocation gradient. Next, a weighted fusion calculation is performed. For each grid point in the spatial grid system, the observations from different sensors are weighted and averaged according to their corresponding dynamic weights to generate a fused water level estimate. The fusion process also includes uncertainty propagation calculation, estimating the uncertainty range of the fusion result using the error covariance matrix propagation theory, providing reliability information for subsequent predictive analysis. Finally, the consistency of the fusion results is checked. Cross-validation is used to evaluate the consistency between the fused data and the original observation data. When the consistency index is below 0.85, a fusion parameter adjustment mechanism is triggered to re-optimize the weight allocation strategy. The spatiotemporally consistent dataset generated after fusion processing has a uniform sampling interval in time and a regular grid distribution in space, providing standardized input data for subsequent modeling.

[0055] The specific implementation of step S05 involves determining the configuration strategy of the prediction model through threshold discrimination and conditional branching logic. The purpose of this step is to adaptively select a suitable modeling method based on the dynamic characteristics of the marine environment. First, the nonlinear coupling strength parameter calculated in step S03 is compared with a preset threshold. When the parameter value is within the range of 0.75 to 1.0, it indicates that there are strong nonlinear interactions between marine dynamic processes. At this time, the linear model cannot accurately describe the system behavior, and the nonlinear modeling process needs to be initiated. The nonlinear modeling process adopts a nonlinear mapping method based on neural networks, which captures complex dynamic coupling relationships through multi-layer nonlinear transformations. Then, the nonstationarity index is evaluated. When the index exceeds 0.6, it indicates that the current marine environment is in a rapidly changing state or is affected by extreme weather systems. At this time, it is necessary to add an extreme weather condition processing module to enhance the model's response capability to abnormal situations. The extreme weather condition processing module's extreme weather identification unit employs a threshold-based multi-index discrimination method, comprehensively analyzing meteorological elements such as wind speed, pressure change rate, and precipitation intensity. This module is activated when any one of these elements exceeds the extreme event standard. The wind speed threshold is typically set at 25 m / s, the pressure change rate threshold at 5 hPa, and the precipitation intensity threshold at 50 mm / s. The nonlinear dynamic amplification unit receives astronomical tides, storm surges, and waves as inputs. It uses multiplicative interaction terms and higher-order polynomial terms to express nonlinear coupling effects and introduces cross-correlation functions to capture the time-delay response relationships between different dynamic factors. The prediction error compensation unit establishes an error correction model based on the statistical regularity of prediction deviations from historical extreme weather events. This model learns patterns of systematic deviations from historical data using machine learning methods and performs real-time corrections to the model output during the prediction phase, typically within the range of 5% to 20%. The data flow within the module follows a sequential processing principle: first, the extreme weather identification unit determines the current environmental state; then, the identification results and original features are passed to the nonlinear dynamic amplification unit for feature enhancement; finally, the prediction error compensation unit corrects the errors in the enhanced features. The processed feature data is passed to the adaptive spatiotemporal fusion algorithm for further integration, while the extreme weather identification results are passed as flags to subsequent processes to guide the selection of modeling strategies.

[0056] The specific implementation of step S06 involves recovering the complete high-dimensional data structure from dimensionality-reduced observations using sparse optimization theory. This step aims to reconstruct the complete information field of the marine environment under limited observation conditions. First, the sparsity characteristics of the spatiotemporally consistent dataset are analyzed. The ocean level field typically exhibits a sparse representation in the wavelet transform or discrete cosine transform domains, meaning most transform coefficients are close to zero, with only a few coefficients having significant values. Based on this characteristic, the compressed sensing theoretical framework is used for data reconstruction. Then, the compression ratio of the observation matrix is ​​determined by calculating the ratio of the observation matrix dimension to the original matrix dimension. When this ratio is within the range of 0.1 to 0.3, it indicates that the dimension of the observed data is significantly lower than the dimension of the original data. At this point, the compressed sensing algorithm is triggered to achieve data reconstruction. The core of the compressed sensing algorithm is to solve an optimization problem, aiming to find the sparsest solution that is compatible with the observed data. Sparsity is achieved through… The norm is used as a metric, defined as the sum of the absolute values ​​of the elements of a vector. Optimization is achieved using either the Orthogonal Matching Pursuit (ORP) algorithm or the Basis Pursuit (BPS) algorithm. The ORP algorithm iteratively constructs a sparse representation by selecting the dictionary atoms most relevant to the residuals. Each iteration selects one atom and updates the residuals; the number of iterations is typically set to 1.5 to 2 times the sparsity. The BPS algorithm... The norm minimization problem is transformed into a linear programming problem, and the optimal solution is obtained through the simplex method or interior point method. Next, the reconstruction results are evaluated for quality using normalized mean square error (MSE). The reconstruction accuracy is quantified by comparing the differences between the reconstructed data and the complete observation data in the overlapping regions. A MSE less than 0.05 is considered to meet the reconstruction quality requirements. Finally, the high-dimensional original matrix obtained from the reconstruction is passed as complete environmental field information to the subsequent prediction process. This matrix contains complete evolutionary information of all grid points in space over time, providing a sufficient data foundation for establishing an accurate prediction model.

[0057] The specific implementation of step S07 involves evaluating the impact of system parameters on model stability using matrix perturbation analysis. The purpose of this step is to identify key influencing factors and optimize the weight allocation strategy. First, the multi-scale spatiotemporal registration matrix is ​​decomposed into eigenvalues ​​and eigenvectors. Eigenvalues ​​reflect the scaling characteristics of the matrix in different directions, with larger eigenvalues ​​corresponding to major change patterns. Then, the sensitivity of the eigenvalues ​​to the nonlinear coupling strength parameter is calculated using the finite difference method. This involves slightly perturbing the nonlinear coupling strength parameter and observing the magnitude of the eigenvalue changes to quantify the sensitivity. Specifically, a 1% perturbation is added or removed from the original parameter value, and the relative rate of change of the eigenvalues ​​before and after the perturbation is calculated as the sensitivity index. The magnitude of the sensitivity value reflects the degree of influence of the parameter on the system's dynamic characteristics; high sensitivity means that a small change in the parameter leads to a significant change in system behavior. Next, the dynamic weight allocation coefficient control strategy is adjusted based on the sensitivity analysis results. When the sensitivity value corresponding to a certain sensor exceeds 0.8, it indicates that the sensor data has a strong ability to characterize the system state, and its weight in the fusion process should be increased. The weight increase is typically 10% to 30% of the original value. Conversely, when the sensitivity value is below 0.2, it indicates that the sensor data contributes little information, and its weight should be appropriately reduced to avoid introducing unnecessary noise. The weight reduction is typically 10% to 20% of the original value. The adjusted weight coefficients need to be renormalized to ensure that the sum of all weights is 1. Finally, the updated weight control strategy is applied to subsequent prediction calculations, dynamically adjusting the influence of each data source to improve the overall performance and stability of the prediction model.

[0058] The specific implementation of step S08 is to achieve water level prediction tasks at different time scales through a hierarchical prediction architecture. This step adaptively configures model parameters and algorithm strategies based on the differences in prediction duration. First, the prediction task is divided into three levels according to time scale: a short-term prediction layer for prediction needs from 1 hour to 24 hours, a medium-term prediction layer for prediction needs from 1 day to 7 days, and a long-term prediction layer for prediction needs from 7 days to 30 days. The prediction algorithms at different levels have different characteristics and applicable ranges. The short-term prediction layer uses a time series method based on an autoregressive moving average model. This method uses the autocorrelation structure of historical observation sequences for extrapolation prediction. The model order is determined based on the partial autocorrelation function, typically set to 12 to 48 to capture the changing characteristics of daily and semi-daily cycles. The medium-term prediction layer uses a deep learning method based on a long short-term memory network. This network structure can learn long-term dependencies and selectively retains and forgets historical information through a gating mechanism. The number of network layers is set to 2 to 4, and the number of hidden units in each layer is set to 64 to 256, with the specific values ​​determined based on the scale of the training data. The long-term prediction layer employs a physics-based numerical simulation method. This method predicts future states by solving the ocean dynamics governing equations, which include continuity, momentum, and state equations. The equations are numerically discretized and solved using finite difference or finite element methods. Then, based on the user-inputted prediction duration, the appropriate prediction level is automatically selected, and the corresponding model parameter configuration is loaded. The parameter configuration file includes the model structure definition, trained weight parameters, and hyperparameter settings. Next, the high-dimensional original matrix and dynamic weight allocation coefficients are fed as input features into the selected prediction algorithm. The algorithm outputs a sequence of predicted water levels for future times. For short-term predictions, the output time step is typically 10 minutes to 1 hour; for medium-term predictions, the output time step is 1 hour to 6 hours; and for long-term predictions, the output time step is 6 hours to 24 hours. Finally, the output results from each prediction level are aggregated to form a preliminary prediction result set. This set contains multiple candidate prediction sequences, providing a foundation for subsequent optimization.

[0059] Step S09 is specifically implemented by refining the preliminary prediction results using a hybrid optimization model. The purpose of this step is to integrate data-driven methods and physical mechanism constraints to improve prediction accuracy and reliability. First, the set of preliminary prediction results is input into the marine dynamic response optimization model. This model employs an encoder-decoder architecture. The encoder uses a multi-head attention mechanism to process multiple candidate prediction sequences. This attention mechanism automatically learns the relative importance of different prediction results and performs weighted fusion. The number of attention heads is dynamically determined based on the nonlinear coupling strength parameter; when the coupling strength is high, more attention heads are used to capture complex interaction relationships. The number of heads typically ranges from 4 to 16. The decoder uses a sparse attention mechanism to generate the final prediction results. Sparse attention reduces computational complexity by limiting each query to focusing only on a subset of key-value pairs. The sparsity parameter is adaptively adjusted based on the non-stationarity index and data dimension; higher non-stationarity results in lower sparsity to retain more historical information. The hidden layer dimension of the model is dynamically configured according to the prediction duration. Short-term predictions use a smaller hidden layer dimension, such as 128, while long-term predictions use a larger hidden layer dimension, such as 512, to enhance the model's expressive power. Then, residual analysis is performed to calculate the difference between the model's predicted values ​​and historical measured values. By analyzing the statistical distribution characteristics of the residuals, systematic biases and random errors are identified. The residual analysis results are used in an error feedback mechanism, which automatically adjusts the model's prediction parameters based on the magnitude and pattern of the residuals. Adjustment strategies include learning rate decay, regularization strength modification, and feature weight updates. Next, the prediction error is evaluated by calculating the mean absolute percentage error of the predicted values. When this error exceeds 15%, it indicates that the current model performance cannot meet the accuracy requirements, triggering a model retraining process. The retraining process includes re-collecting the latest observation data, expanding the training dataset, re-initializing the model parameters, and training the model using a gradient descent-based optimization algorithm. Early stopping is employed during training to prevent overfitting. After iterative optimization, the final prediction result is output. This result integrates the advantages of multiple prediction methods and has been tested against physical constraints, demonstrating high accuracy and stability.

[0060] The specific implementation of step S10 involves transforming the optimized prediction results into standardized products for users. This step aims to provide actionable decision support information. First, the final prediction results undergo post-processing, including time series smoothing, outlier correction, and physical plausibility checks, to ensure the continuity and consistency of the prediction data. Then, multi-timescale prediction products are generated, providing prediction data with different time resolutions for different application scenarios. Real-time monitoring applications require predictions with a resolution of 10 minutes to 1 hour, waterway management applications require predictions with a resolution of 1 hour to 6 hours, and planning and design applications require daily or weekly statistical predictions. Next, uncertainty quantification is performed, using Monte Carlo methods or ensemble forecasting methods to estimate the confidence intervals of the prediction results. The calculation of the confidence intervals considers the combined effects of model parameter uncertainty, input data uncertainty, and physical process uncertainty, typically providing interval estimates at 90% and 95% confidence levels. Subsequently, a risk warning level classification standard was established. Based on the relationship between the predicted water level and the warning water level, the risk was divided into four levels: a predicted water level 0.5 meters below the warning water level is considered normal; a predicted water level within 0.5 meters above or below the warning water level is considered a state of concern; a predicted water level 0.5 to 1 meter above the warning water level is considered a warning; and a predicted water level more than 1 meter above the warning water level is considered a severe warning. Risk level information is displayed graphically, using a combination of color coding and numerical labeling for intuitive presentation. Finally, the predicted values, confidence intervals, and risk warning levels are summarized to form a complete coastal water level prediction product, which is released externally through a visualization platform and data interface, providing a scientific basis for marine disaster prevention and mitigation, port operations, and coastal zone management.

[0061] The detailed structure of the marine dynamic response optimization model adopts a multi-layer encoder-decoder architecture, which consists of four main components: an input embedding layer, a multi-head attention encoding layer, a sparse attention decoding layer, and an output projection layer. The input embedding layer is responsible for converting the initial prediction result set into a high-dimensional feature representation, achieved through a combination of linear transformation and positional encoding. The linear transformation maps the original prediction values ​​to a 512-dimensional feature space, while the positional encoding uses sine and cosine functions to preserve the sequential information of the time series. The multi-head attention encoding layer contains a stack of six identical encoders, each containing a multi-head self-attention sublayer and a feedforward neural network sublayer. The number of attention heads is dynamically adjusted according to the nonlinear coupling strength parameter: eight attention heads are used when the coupling strength parameter is in the range of 0.75 to 0.85, and sixteen attention heads are used when the coupling strength parameter is in the range of 0.85 to 1.0. Each attention head has a dimension of 64. The self-attention mechanism achieves feature fusion by calculating the similarity between the query vector, key vector, and value vector. The similarity calculation uses a scaled dot product method, with the scaling factor being the reciprocal of the square root of the attention head dimension. The feedforward neural network sublayer contains two linear transformations and one non-linear activation function. The first linear transformation expands the input from 512 dimensions to 2048 dimensions, and the activation function uses a Gaussian error linear unit function. The second linear transformation compresses the feature dimension back to 512 dimensions. Each sublayer is followed by residual connections and layer normalization operations to stabilize the training process. The sparse attention decoding layer also contains a stack of six decoders, each containing a masked multi-head self-attention sublayer, an encoder-decoder attention sublayer, and a feedforward neural network sublayer. The masking mechanism ensures that the decoding process can only access historical information and cannot peek into future information. Sparse attention restricts each query to focus only on key-value pairs within a fixed range using a local window. The window size is determined by the non-stationarity index: 32 for a non-stationarity index less than 0.4, 64 for an index between 0.4 and 0.6, and 128 for an index greater than 0.6. The encoder-decoder attention sublayer enables the decoder to pay cross-modal attention to the encoder output, allowing the decoder to fully utilize the global feature information extracted by the encoder. The output projection layer maps the decoder output to the target prediction dimension through a linear transformation. For water level prediction tasks, the output dimension is typically the prediction time step multiplied by the number of spatial grid points. The model employs a meta-learning-based training strategy. The meta-optimizer uses a Long Short-Term Memory (LSTM) network for parameterization, containing 256 hidden units. This network learns long-term patterns of the optimized trajectory and predicts the optimal parameter update direction. The coordinate sharing mechanism improves parameter efficiency by sharing some model parameters across different tasks. The sharing layer includes the input embedding layer and the first two encoder layers, while the task-specific layers include the subsequent encoder layers and all decoder layers.Preprocessing techniques include gradient clipping and learning rate warm-up. The gradient clipping threshold is set to 1.0 to prevent gradient explosion. The learning rate is linearly increased from 0 to the target learning rate in the early stages of training to stabilize the training process. The number of warm-up steps is set to 4000.

[0062] The detailed steps for establishing the training dataset for the marine dynamic response optimization model are as follows: First, historical multimodal marine observation data are collected, including satellite remote sensing data, tide gauge observation data, and high-frequency radar data from the past 5 to 10 years. The collected data needs to cover observation samples from different seasons, weather conditions, and sea states to ensure the representativeness of the dataset. Simultaneously, measured water levels for the corresponding time periods are collected as training targets. These measured values ​​are derived from high-precision observation records from coastal tide gauge stations, with a data sampling interval of 10 minutes and an observation accuracy requirement of centimeter-level. Next, the collected raw data undergoes quality control, removing data segments with obvious observation errors or instrument malfunctions. Suspicious data points are identified through cross-validation and physical constraint checks. Missing data are filled using linear interpolation or extrapolation from data from neighboring stations. Then, the data is divided into training and validation sets according to time series, with the training set accounting for 80% of the total data and the validation set accounting for 20%. The division maintains temporal continuity and ensures that the training and validation sets cover similar marine environmental conditions. The data then undergoes standardized preprocessing using a zero-mean, unit-variance standardization method. This transforms the data to a standard normal distribution by subtracting the mean and dividing by the standard deviation. Standardization parameters are obtained statistically from the training set and saved for data transformation during the testing phase. When constructing the input feature vector, multimodal observation data that has undergone spatiotemporal registration and fusion processing, extracted ocean dynamic process features, nonlinear coupling strength parameters, and nonstationarity indices are combined into a high-dimensional feature vector. The dimension of the feature vector is typically between 200 and 500, determined based on the specific application scenario. When constructing the target prediction sequence, a future water level sequence of appropriate length is selected as the learning target based on the required prediction duration. The target sequence length for short-term prediction is 24 to 72 time steps, for medium-term prediction it is 72 to 168 time steps, and for long-term prediction it is 168 to 720 time steps. To expand the number of training samples, data augmentation techniques were employed to generate more training examples. Augmentation methods included time-sliding window sampling, adding Gaussian noise perturbations, and time-series transformations. Sliding window sampling generated multiple overlapping samples by changing the start time, with the noise perturbation amplitude controlled within 10% of the observation accuracy. Time-series transformations included scaling, translation, and rotation to enhance the model's adaptability to different change patterns. Finally, standard data subsets under different sea state conditions were established. Based on marine environmental parameters such as wind speed, wave height, and astronomical tide type, the dataset was divided into three subsets: calm sea state, moderate sea state, and severe sea state. During model training, data from different subsets were sampled at a certain proportion to ensure the model's balanced learning ability across various sea states.

[0063] Specifically, the principle of this invention is as follows: This invention can solve the technical problem of insufficient prediction accuracy in multimodal data fusion. Its principle lies in establishing a complete data registration, fusion, optimization, and prediction chain. First, a time interpolation function and spatial mapping relationship are constructed through a multi-scale spatiotemporal registration matrix, unifying hourly sampling from satellite remote sensing, minute-level sampling from tide gauges, and second-level sampling from high-frequency radar to the same time reference. Simultaneously, bilinear interpolation is used to map spatial resolution data from meters to kilometers onto a unified grid. This registration process mathematically eliminates the spatiotemporal inconsistencies of heterogeneous data, laying the foundation for subsequent fusion. Second, the adaptive spatiotemporal fusion algorithm calculates dynamic weight allocation coefficients by real-time evaluation of the signal-to-noise ratio and consistency of each sensor's data, avoiding the adverse effects of low-quality data on the fusion results. Furthermore, the sparse matrix compression sensing reconstruction mechanism leverages the sparsity of ocean data to reconstruct a high-dimensional original matrix from a low-dimensional observation matrix by optimizing the minimum norm problem. Even with incomplete observation data, it can recover crucial information. The eigenvalue sensitivity analysis mechanism identifies the parameters with the greatest impact on prediction results by calculating the partial derivatives of matrix eigenvalues ​​with respect to parameter perturbations and adjusts the weight control strategy accordingly, achieving refined optimization of the parameter space. Finally, the multi-level prediction framework automatically selects appropriate algorithms and parameter configurations based on different prediction durations (short-term, medium-term, and long-term). The meta-optimizer predicts the optimal parameter update direction, the coordinate sharing mechanism reuses parameters across different tasks to reduce overfitting, and the residual analysis and error feedback mechanisms automatically adjust model parameters based on prediction deviations. When the prediction error exceeds a threshold, a model retraining process is triggered. The entire technical solution forms a complete closed loop from data registration to feature extraction, fusion prediction, and iterative optimization. Each step addresses specific challenges in multimodal data fusion, thus effectively improving the accuracy of coastal water level prediction.

[0064] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0065] In this embodiment, the specific implementation methods of steps S01-S02, S08, and S10 are the same as those described above, and will not be repeated in detail here.

[0066] The specific implementation of step S03 involves quantitatively calculating the coupling strength and non-stationary characteristics of the ocean dynamic process. This step involves the calculation of two key parameters: the nonlinear coupling strength parameter and the nonlinear coupling strength parameter. The calculation formula is expressed as follows:

[0067] ;

[0068] In the formula, This is a dimensionless nonlinear coupling strength parameter, with a value ranging from 0 to 1; This is a sequence of astronomical tide characteristics, in meters; This is a storm surge characteristic sequence, in meters; This is a wave characteristic sequence, with units in meters; represents the standard deviation of the astronomical tide characteristic sequence, in meters; This represents the standard deviation of the storm surge characteristic sequence, in meters. represents the standard deviation of the wave characteristic sequence, in meters; For characteristic sequences and The covariance function, in meters squared; and For feature sequence index, where The value can be between 1 and 2. Values Up to 3; the denominator 1.0 is the normalization constant, which is dimensionless.

[0069] Nonstationarity index The calculation formula is expressed as follows:

[0070] ;

[0071] In the formula, It is a dimensionless nonstationary index; represents the standard deviation of the mean sequence within the sliding window, in meters; This is a local mean sequence within a sliding window, in meters; For reference, the mean of the entire sequence is usually taken, and the unit is meters; This represents the maximum value of the sliding window mean, in meters. is the standard deviation of the variance sequence within the sliding window, in meters squared; This represents the local variance sequence within the sliding window, expressed in squared meters. For reference variance, the variance of the entire series is usually taken, and the unit is the square of meters; This represents the maximum variance of the sliding window, expressed in meters squared. This represents the standard deviation function.

[0072] The parameter acquisition method is as follows: The experimental method was used to obtain the spectrum, including step 1, which was to perform a fast Fourier transform on the water level sequence to obtain the spectrum; step 2, which was to identify the frequency components with periods of 12.42 hours and 24.84 hours; and step 3, which was to reconstruct the time series of astronomical tide components through inverse transform. The storm surge feature sequence was obtained through an experimental method, including step 1, performing empirical mode decomposition on the residual sequence after removing astronomical tides; step 2, extracting intrinsic mode functions with characteristic periods ranging from 6 hours to 72 hours; and step 3, superimposing these functions to obtain the storm surge feature sequence. The method used is experimental, including step 1, identifying the peaks and troughs of continuous rise and fall in the water level sequence; step 2, calculating the time interval and amplitude difference between adjacent peaks; and step 3, statistically analyzing the effective wave height sequence as wave characteristics. , , These are the standard deviations of the corresponding feature sequences, calculated from the entire sequence; The expected value is obtained by calculating the product of the deviations of two feature sequences from their respective means; the default sliding window length is 48 hours, and the default step size is 6 hours. The sequence is obtained by calculating the average water level within each sliding window; The sequence is obtained by calculating the variance of the water level within each sliding window; and The mean and variance are calculated for the entire sequence.

[0073] The specific implementation of step S04 is to achieve multi-source data fusion through dynamic weight optimization. First, the comprehensive quality score of each sensor is calculated. The calculation formula is expressed as follows:

[0074] ;

[0075] In the formula, For the first Dimensionless overall quality score of each sensor; For the first The measurement accuracy score of each sensor, in points, with a value range of 0 to 100; For the first The time delay score for each sensor is expressed in points, with a value ranging from 0 to 100. For the first Spatial coverage score for each sensor, in points, ranging from 0 to 100; 0.5 is the sensor index; 0.5, 0.3, and 0.2 are weighting coefficients, dimensionless; 100 is the normalization constant, in minutes.

[0076] Dynamic weight allocation coefficient The calculation formula is expressed as follows:

[0077] ;

[0078] In the formula, For the first Dimensionless dynamic weight allocation coefficients for each sensor; For the first The overall quality score of each sensor is dimensionless. The maximum value of the quality score for all sensors, dimensionless; The weighted response exponent is dimensionless and has an empirical value of 2.0. This represents the total number of sensors; For sensor indexing; It is a natural exponential function.

[0079] Water level after fusion The calculation formula is expressed as follows:

[0080] ;

[0081] In the formula, The dimensionless water level value after fusion; For the first The sensor is located in space. and time The observed water level is in meters; This is a water level scaling parameter, typically taken as 10 meters; The coordinates are eastward, and the unit is meters. The coordinates are north-facing, and the unit is meters. This refers to time, expressed in hours.

[0082] The parameter acquisition method is as follows: By analyzing the observation noise level and systematic error characteristics, a score of 0 to 100 is obtained by mapping the inverse of the standard deviation of repeated observations. The score is calculated based on the time interval from observation to availability of data; the shorter the time interval, the higher the score. The score is obtained by calculating the overlap ratio between the effective observation range of the sensor and the target area; the higher the coverage, the higher the score. The value is determined based on the degree of difference in data quality. When the quality difference between sensors is large, a larger value is taken, usually between 1.5 and 2.5. This refers to the actual number of sensors involved in the fusion process. Determined based on the tidal range of the study area.

[0083] The specific implementation method of step S05 is the same as described above, and will not be repeated in detail here.

[0084] The specific implementation of step S06 is to reconstruct high-dimensional data from low-dimensional observations through sparse optimization, and the optimization objective function of compressed sensing reconstruction. The statement is as follows:

[0085] ;

[0086] In the formula, The objective function value is dimensionless. The observation matrix has dimensions of . The unit is meters; The measurement matrix has dimensions of . Dimensionless; The original high-dimensional matrix to be reconstructed has a dimension of . The unit is meters; This is a water level scaling parameter, typically taken as the maximum value of the observed data, in meters; This is the regularization parameter, dimensionless, with an empirical value of 0.01; for The square of the norm; for Norm; This represents the number of observation points; This represents the total number of points in the complete spatial grid. This represents the number of time steps.

[0087] The parameter acquisition method is as follows: The actual sensor observation data matrix is ​​extracted from the observation location using the fusion results of step S04; The random Gaussian measurement matrix is ​​obtained by generating random numbers that follow a standard normal distribution and normalizing them by row. Determined based on noise level; the greater the noise... The larger the value, the more likely it is to be in the range of 0.001 to 0.1; the reconstruction algorithm uses the orthogonal matching pursuit algorithm, and the number of iterations is set to... to ; Take the observation matrix The maximum absolute value of all elements in the set.

[0088] The specific implementation of step S07 is to evaluate the parameter sensitivity through eigenvalue perturbation analysis, i.e., eigenvalue sensitivity. The calculation formula is expressed as follows:

[0089] ;

[0090] In the formula, The sensitivity is a dimensionless eigenvalue. The first multi-scale spatiotemporal registration matrix The matrix has 10 eigenvalues, which are dimensionless, and has been normalized. It is the largest eigenvalue, dimensionless; This is a nonlinear coupling strength parameter, dimensionless. The reference coupling strength is typically taken as 0.5, which is dimensionless. This is the perturbation amount for the coupling strength parameter, which defaults to 0.01 and is dimensionless. Represents the partial derivative operator; The coupling strength parameter is... The first time One eigenvalue; For feature value index.

[0091] Weighting adjustment coefficient The calculation formula is expressed as follows:

[0092] ;

[0093] In the formula, For the first Dimensionless weighting adjustment coefficients for each sensor; For the first The characteristic sensitivity of each sensor is dimensionless. This is the minimum sensitivity of all sensors, and it is dimensionless. This represents the maximum value of the sensitivity of all sensors, and is dimensionless. To adjust the intensity parameter, which is dimensionless, it is usually taken as 0.3; 0.5 and 1.0 are normalization constants, which are dimensionless.

[0094] The parameter acquisition method is as follows: The matrix is ​​obtained by eigenvalue decomposition of the multi-scale spatiotemporal registration matrix. A symmetric matrix, The matrix represents the total number of spatiotemporal nodes, and its eigenvalues ​​are normalized to be dimensionless; the perturbation quantity... Take 1% of the original parameter value; The value is determined based on the system stability requirements; a smaller value is taken when higher stability is required. By calculating with the first The response of the eigenvalues ​​associated with the sensor data to parameter perturbations is obtained.

[0095] The specific implementation of step S09 is to optimize the prediction results through residual feedback, thereby reducing the prediction error. The calculation formula is expressed as follows:

[0096] ;

[0097] In the formula, The mean absolute percentage error is expressed as a percentage. For a moment The predicted water level value, in meters; For a moment The measured water level value, in meters; This is a water level scaling parameter, with a value of 10 meters. To predict the total number of time steps; For time step index; For the first Each predicted time point is in hours; It is an absolute value function.

[0098] The parameter acquisition method is as follows: The predicted values ​​output by the marine dynamic response optimization model; For actual measured data from tide gauge stations; when When the percentage exceeds 15%, the model retraining process is triggered.

[0099] It should be explained that the principle behind the formula for calculating the nonlinear coupling strength parameter is based on statistical correlation theory. It assesses the interaction strength by quantifying the covariance relationship between different ocean dynamic factors. The core of this formula is the Pearson correlation coefficient, which, for any pair of feature sequences, is expressed as follows: ,in For characteristic sequences and The correlation coefficient, dimensionless, reflects the degree of linear correlation between the two sequences. The formula uses the average of the normalized cross-correlation coefficients as the overall coupling index, which eliminates the influence of differences in the magnitude of each factor, making the coupling strength assessment comparable. Covariance function. It captures the synchronous changing trends of two sequences, and its calculation is based on the expected value of the product of the sequences' deviations from their respective means and the product of their standard deviations. The normalization process makes the results dimensionless. The double summation structure enables ergonomic statistics for all paired factors, with a summation range of... up to 2 and Dividing by 3 ensures that each pair of factors is calculated only once; the division by 3 is because there are a total of 3 combinations for the three factors. , and The effectiveness of this formula lies in its ability to accurately identify the nonlinear coupling state between astronomical tides, storm surges, and waves in the marine environment. When the coupling strength parameter is close to 1, it indicates the existence of strong interaction. At this point, a nonlinear modeling process needs to be initiated to improve prediction accuracy. Compared with the traditional linear independent processing method, this formula enables the model to adaptively judge the complexity of the marine dynamic process and select an appropriate modeling strategy, significantly improving the prediction accuracy under complex sea conditions.

[0100] The principle behind the nonstationarity index calculation formula is based on a quantitative method of changing statistical characteristics of time series. It assesses the nonstationarity of the series by analyzing the fluctuations of local statistics within a sliding window. This formula comprehensively considers two dimensions: mean change and variance change. Mean change reflects the trend of water level drift, while variance change reflects the time-varying characteristics of fluctuation amplitude. The standard deviation function in the formula... Calculate the dispersion of the sliding window mean sequence to characterize the magnitude of the mean change over time. The dispersion of the sliding window variance series is calculated to characterize the magnitude of volatility changes over time. The numerator is the standard deviation of the sliding window statistic, representing the magnitude of change, while the denominator is the maximum value of the statistic, providing a normalization benchmark. The weighted average of the two values ​​gives a comprehensive measure of nonstationarity, with each weighted at 0.5 to indicate that changes in the mean and variance are equally important. This formula is effective in sensitively capturing rapid changes in the marine environment and the impact of extreme weather events. When the nonstationarity index exceeds 0.6, it indicates a significant change in the statistical characteristics of the time series. In this case, an extreme weather condition processing module needs to be added to enhance the model's responsiveness to abnormal situations. Compared to traditional prediction methods that assume stationarity, this formula allows the model to assess the environmental state in real time and dynamically adjust the prediction strategy. Especially under extreme weather conditions such as typhoons and cold waves, it can promptly initiate enhanced processing procedures, significantly improving the reliability of extreme event predictions and the timeliness of early warnings.

[0101] The principle behind the comprehensive quality score calculation formula is based on multi-dimensional weighted evaluation theory. It comprehensively evaluates sensor data quality by considering three key indicators: measurement accuracy, time delay, and spatial coverage. The formula uses a linear weighting method, with weight coefficients of 0.5, 0.3, and 0.2 reflecting the relative importance of each indicator. Measurement accuracy has the highest weight because it directly determines data reliability, followed by time delay because it affects data timeliness, and spatial coverage has the lowest weight because it mainly affects data integrity. Each scoring item is normalized by dividing by 100, ensuring the comprehensive score is within the range of 0 to 1, facilitating subsequent exponential weighting calculations. The effect of this formula is that it can quantify the overall performance level of each sensor, providing a scientific basis for dynamic weight allocation. Compared to single-dimensional evaluation methods, this multi-dimensional evaluation system can more comprehensively reflect the actual value of sensor data, making weight allocation more reasonable and accurate.

[0102] The principle behind the fusion water level calculation formula is based on the weighted average theory, which optimizes the combination of multi-source data by dynamically weighting observations from different sensors. The formula adopts a linear weighted sum form, defining the normalized observed water level. ,in For the first The dimensionless water levels observed by each sensor are normalized using a normalization operation, eliminating dimensions and resulting in a dimensionless value for the fused water level. This facilitates application under different regions and tidal range conditions. The fused water level is obtained through weighted summation. Weighting coefficients The weighting is dynamically determined based on data quality, with high-quality data contributing a larger weight and low-quality data contributing a smaller weight. The sum of all weights equals 1 to ensure the fused value remains within a reasonable range. This formula effectively utilizes complementary information from multiple sensors, automatically adapting to spatiotemporal changes in data quality through dynamic weight adjustments. Compared to single data sources or simple averaging methods, this weighted fusion strategy significantly improves the accuracy and reliability of the fused data, reduces the impact of individual sensor failures or errors on the overall results, and provides more stable and accurate input data for the predictive model.

[0103] The principle of the compressed sensing reconstruction objective function is based on sparse optimization theory. It recovers the high-dimensional original signal from low-dimensional observations by jointly minimizing the reconstruction error and sparse constraints. This formula includes a data fidelity term. and sparse regularization terms Two parts. The data fidelity item adopts... Norm measures the degree of fit between the reconstructed results and the observed data, molecule Calculate the squared Euclidean distance between the normalized observations and the reconstructed values, and the denominator is... Normalization ensures comparability across different data sizes, guaranteeing that the reconstructed signal is compatible with actual observations. The sparsity regularization term employs... The norm promotes the sparsity of solutions; the norm for a matrix... Defined as the sum of the absolute values ​​of all elements divided by the total number of elements. ,in For matrix No. Line number Column elements, For spatial grid point indexing, The norm serves as the time-step index, and its convex property makes the optimization problem solvable, allowing the selection of the simplest representation from an infinite number of possible solutions. Regularization parameter. The weights of the two terms are balanced; the larger this parameter is, the higher the requirement for sparsity. (Measurement matrix) A mapping relationship is established between the observation space and the original space. The effect of this formula is that it can reconstruct complete marine environmental field information under the condition of limited number of sensors or limited observation locations. It utilizes the sparsity of the water level field in the transform domain to achieve accurate recovery of high-dimensional data. Compared with traditional interpolation methods, this formula can maintain the inherent structural features and detailed information of the data. It can still achieve high-quality reconstruction when the observation dimension is only 10% to 30% of the original dimension, which significantly improves the spatial coverage and data utilization efficiency under limited observation conditions, and provides complete spatiotemporal data support for the establishment of accurate prediction models.

[0104] The principle behind the eigenvalue sensitivity calculation formula is based on matrix perturbation theory, using the finite difference method to approximate the eigenvalue's response to parameter changes. This formula employs numerical differentiation to evaluate the eigenvalue change caused by parameter perturbation; the general numerical differentiation formula is as follows: ,in Let be the function to be differentiated. As the independent variable, For a small increment, in this problem, it means calculating the partial derivatives of the eigenvalues ​​with respect to the coupling strength parameter. (Molecular) The relative change in eigenvalues ​​reflects the change in the system's dynamic characteristics, and the denominator... The relative perturbation of the parameter characterizes the magnitude of the parameter change. Eigenvalues This reflects the main variation pattern of the spatiotemporal registration matrix. Larger eigenvalues ​​correspond to the dominant spatiotemporal evolution pattern, and their changes are directly related to the system's state representation capability. Normalization processing... and Eliminating the influence of dimensions makes different systems comparable, and the disturbance quantity Setting the value to 1% of the original value ensures the effectiveness of the linear approximation. The formula effectively identifies the sensor data source that has the most significant impact on system behavior. A higher sensitivity value indicates that the data source has a significant impact on the spatiotemporal registration results and its fusion weight should be increased. Conversely, a lower sensitivity value indicates that the data source contributes less information and its weight can be appropriately reduced. Compared to empirical weight setting methods, this formula provides a quantitative basis for weight optimization based on mathematical analysis, making the weight allocation strategy more scientific and reasonable, and significantly improving the stability of multi-source data fusion and the overall performance of the prediction model.

[0105] The principle behind the prediction error calculation formula is based on statistical error analysis theory. It quantifies prediction accuracy by calculating the average absolute percentage deviation between the predicted and measured values. This formula uses an absolute value function. To avoid the cancellation of positive and negative errors, and to accurately reflect the degree of prediction deviation, the scaling parameter is used. Dimensionless conversion makes the errors of water levels of different magnitudes comparable, and multiplying by 100% to convert to a percentage format facilitates understanding and application. (Time series summation) Calculate the cumulative error across all prediction times and divide by the total number of time steps. The average error level is given, and this averaging operation eliminates the influence of individual outliers, providing a stable performance evaluation. The formula structure is intuitive and easy to understand, and the threshold set to 15% provides a quantitative standard for model performance evaluation and retraining triggering. The time frame indicates that the model accuracy has decreased and an update is needed. The effectiveness of this formula lies in its ability to objectively evaluate the predictive performance of the marine dynamic response optimization model, providing a basis for model parameter adjustment and retraining decisions. When the error exceeds a threshold, it automatically triggers the model update process to ensure the continuous and reliable operation of the prediction system. Compared to traditional fixed-model prediction methods, the error feedback mechanism supported by this formula enables the prediction system to have self-diagnosis and adaptive optimization capabilities. It can dynamically adjust the model configuration based on the actual prediction results, significantly improving the long-term stability and adaptability of the prediction system to environmental changes, ensuring that the prediction products always maintain high accuracy and usability.

[0106] To better understand and implement this invention, a specific application scenario is provided below as Example 2: A technical team needs to make high-precision predictions of water level changes in a coastal area over the next 72 hours. This area is frequently affected by typhoons and experiences complex tidal and wave interactions. The technical team decided to use a coastal water level prediction method based on multimodal observation data to solve this problem.

[0107] The technical team first collected multimodal oceanographic observation data for the area. Satellite remote sensing data, sampled hourly, provided information on sea level anomalies. Tide gauge data, sampled minutely at three fixed coastal stations, provided real-time water level measurements. High-frequency radar data, sampled ten times per second, covered 50 kilometers of the nearshore area. The data covers a wide sea area and provides information on sea surface currents and waves. These data differ by thousands of times in time scale and vary in spatial resolution from 10m to 5000m, requiring standardized processing.

[0108] The technical team established a multi-scale spatiotemporal registration matrix to address the data heterogeneity issue. Time axis alignment employed a cubic spline interpolation function to standardize all data to 10-minute time intervals. Spatial registration used bilinear interpolation to unify the spatial resolution to a 500m × 500m grid system. During the registration process, the team discovered missing data in certain cloud-covered areas of the satellite remote sensing data, which was filled using spatiotemporal interpolation algorithms. Upon completion of registration, a unified dataset containing 168 time steps and 120 × 80 spatial grid points was generated.

[0109] In the ocean dynamic process feature extraction phase, the technical team constructed a feature extractor. Astronomical tidal characteristic analysis showed that the region is mainly controlled by semi-diurnal tides. The tidal amplitude is 1.85m, and the phase lag angle is 127°. (Diurnal tide) The tidal amplitude was 0.42 m, and the phase lag angle was 89°. Storm surge characteristic analysis indicates that we are currently in the outer influence period of a typhoon, with wind speeds reaching 18 m / s, air pressure decreasing by 12 hPa, and wind direction northeast. Based on historical statistical models, the storm surge is estimated to be approximately 0.65 m. Wave characteristics show a significant wave height of 2.3 m, an average period of 7.8 s, and a dominant wave direction eastward. The technical team calculated the nonlinear coupling strength parameters: the coupling strength between the astronomical tide and storm surge is 0.82, and the coupling strength between storm surge and waves is 0.76, indicating significant nonlinear interactions in the current ocean dynamic processes.

[0110] The technical team further calculated the non-stationarity index. Using a 60-minute sliding window, they calculated the mean and variance of the water level data within each window, and then calculated the coefficient of variation of these statistics. The results showed that the non-stationarity index was 0.68, exceeding the threshold of 0.6, indicating that the current sea state is in a strongly non-stationary state, requiring the activation of the extreme weather condition handling module.

[0111] During the data fusion phase, the technical team employed an adaptive spatiotemporal fusion algorithm. First, the quality of the data from each sensor was assessed, yielding a signal-to-noise ratio (SNR) of 23 dB for satellite remote sensing data, 31 dB for tide gauge data, and 18 dB for high-frequency radar data. Based on these assessment results, the team calculated dynamic weight allocation coefficients. As shown in Table 1, the weight coefficients for different sensors vary in the fusion of different marine elements.

[0112] Table 1 Weighting coefficients for data fusion from various sensors

[0113]

[0114] The fused spatiotemporal consistent dataset contains multimodal information under a unified spatiotemporal benchmark, with a data integrity rate of 96.5%.

[0115] The technical team initiated a model for discriminant analysis of the coupling strength of marine dynamics. Because the nonlinear coupling strength parameter 0.82 lies within the range... Within the system, the nonlinear modeling process was automatically initiated. Simultaneously, due to the nonstationarity index of 0.68 being greater than 0.6, an extreme weather condition processing module was added. This module first confirms the current state of being under the influence of the outer periphery of a typhoon through the extreme weather identification unit, and then activates the nonlinear dynamic amplification unit to perform nonlinear coupling calculations on astronomical tides, storm surges, and wave characteristics. The calculation results show that the nonlinear interaction increases the expected maximum water level by 0.37m compared to the linear superposition result. The prediction error compensation unit corrects the prediction results based on the statistical patterns of eight similar typhoon events in history, with a correction margin of -0.12m.

[0116] During the compressed sensing reconstruction phase, the technical team discovered that the original data matrix had a dimension of 9600, while the observation matrix after preliminary processing had a dimension of 1920, a dimension ratio of 0.2, which met the triggering condition. The system initiated the sparse matrix compressed sensing reconstruction mechanism, using the orthogonal matching pursuit algorithm to reconstruct the high-dimensional original matrix from the low-dimensional observation matrix. The reconstruction process converged after 87 iterations, with a reconstruction error of 3.8%, successfully recovering the main features of the original data.

[0117] Eigenvalue sensitivity analysis showed that the first three principal eigenvalues ​​of the multi-scale spatiotemporal registration matrix were 342.5, 187.3, and 95.8, respectively, with sensitivities to changes in the nonlinear coupling strength parameter of 0.86, 0.72, and 0.31. Based on a dynamic weight allocation coefficient control strategy, the technical team increased the weights of the modes corresponding to the first two eigenvalues, adjusting them from 0.35 and 0.28 to 0.42 and 0.33, respectively, while decreasing the weight of the mode corresponding to the third eigenvalue from 0.22 to 0.16.

[0118] The multi-level prediction framework automatically configures a three-layer prediction structure based on a 72-hour prediction duration. The short-term prediction layer handles predictions for the next 0 to 24 hours, employing an LSTM neural network algorithm with a time window length of 120 time steps and a learning rate of 0.001. The medium-term prediction layer handles predictions for the next 24 to 48 hours, using a time series decomposition algorithm combined with support vector regression, a window length of 240 time steps, and a regularization parameter of 0.05. The long-term prediction layer handles predictions for the next 48 to 72 hours, employing a numerical ocean model coupled with a data-driven approach, with an integration time step of 5 minutes. Each of the three layers generates preliminary prediction results, forming a preliminary prediction result set containing 432 prediction time points.

[0119] The technical team constructed an ocean dynamic response optimization model to iteratively optimize the preliminary prediction results. This model employs a multi-layer encoder-decoder architecture. The encoder uses eight attention heads to process the preliminary prediction result set; this number is determined by a mapping function based on the nonlinear coupling strength parameter of 0.82. The sparse attention mechanism in the decoder has a sparsity parameter set to 0.15, adaptively calculated based on a non-stationarity index of 0.68 and a data dimension of 9600. The hidden layer dimension is configured to 512 dimensions based on a 72-hour prediction duration. Model training utilizes a meta-learning-based optimization algorithm. The meta-optimizer is parameterized through an LSTM network with 256 hidden units. A coordinate sharing mechanism reuses 68% of the parameters across different sea state tasks, significantly improving training efficiency.

[0120] like Figure 2As shown, the optimized model converged after 15 iterations. During the optimization process, the technical team found that the prediction error was relatively large from the 3rd to the 8th hour, reaching 18.3%, exceeding the 15% threshold. The system automatically triggered the model retraining process, increasing the weight of the training samples for this period and adjusting the weight coefficient of the storm surge feature in the feature extractor from 0.28 to 0.35. After retraining, the prediction error for this period decreased to 11.7%. Ultimately, the mean absolute error for the entire 72-hour prediction period was 0.13m, and the root mean square error was 0.18m.

[0121] The technical team generated a multi-timescale coastal water level forecast product. The forecast results show that two high tides will occur within the next 72 hours, with the first high tide occurring in the 14th hour, with a predicted water level of 3.87m and a confidence interval of [insert confidence interval here]. The risk warning level is Level II. The second high tide is expected in the 38th hour, with a predicted water level of 4.23m and a confidence interval of [missing information]. The risk warning level is level three. Low tide forecast indicates the lowest water level will occur in the 26th hour, with a predicted value of 0.68m and a confidence interval of [missing information]. m.

[0122] To address the technical problem of insufficient accuracy in coastal water level prediction caused by the difficulty in effectively fusing multimodal ocean observation data under conditions of mismatched spatiotemporal scales in existing technologies, this invention achieves deep fusion of heterogeneous data and accurate modeling of nonlinear ocean dynamic processes through innovative technical means such as multi-scale spatiotemporal registration matrix, adaptive spatiotemporal fusion algorithm, and meta-learning-based optimization model. Technical performance comparison experiments show that, under normal sea conditions, the mean absolute prediction error of this invention is reduced by 16.8% compared to the traditional single-modal method and by 12.3% compared to the simple multi-source data overlay method. Under extreme weather conditions (typhoons, storm surges, etc.), the improvement in prediction accuracy is even more significant, with the mean absolute error reduced by 18.5% and the root mean square error reduced by 17.2%. In terms of spatiotemporal data fusion efficiency, the multi-scale spatiotemporal registration matrix of this invention increases the data integrity rate from 82.4% of the traditional method to 96.5%, an improvement of 17.1%. Under complex sea conditions with a nonlinear coupling strength parameter greater than 0.75, this invention improves the prediction accuracy by 15.7% compared to the linear overlay method by adaptively initiating the nonlinear modeling process. In terms of model training efficiency, the meta-learning-based optimization algorithm improves the convergence speed by 19.3% and shortens the training time by 14.6%. These technological advancements fully verify the significant advantages of this invention in solving the problems of multimodal data fusion and coastal water level prediction accuracy, as shown in Table 2 and the appendix. Figures 3 to 6 As shown.

[0123] Table 2 Performance Comparison of the Invention and Existing Technologies

[0124]

[0125] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A coastal water level prediction method based on multimodal observation data, characterized in that, include: Collect multimodal ocean observation data from different sensors; A multi-scale spatiotemporal registration matrix was established to align observation data with different sampling frequencies along the time axis. Bilinear interpolation was used to unify data with different spatial resolutions into the same spatial grid system. A marine dynamic process feature extractor was constructed to extract astronomical tide features, storm surge features, and wave features, and nonlinear coupling strength parameters and nonstationarity indices were calculated. An adaptive spatiotemporal fusion algorithm was used to fuse multimodal marine observation data, generating a fused spatiotemporally consistent dataset. A marine dynamic coupling strength discrimination model was established, and a nonlinear modeling process was initiated based on the nonlinear coupling strength parameters. An extreme weather condition processing module was added based on the nonstationarity index. A sparse matrix compressed sensing reconstruction mechanism is constructed, which reconstructs a high-dimensional original matrix from a low-dimensional observation matrix by utilizing the sparsity assumption of a spatiotemporally consistent dataset; an eigenvalue sensitivity analysis mechanism is designed to calculate the sensitivity of the eigenvalues ​​of the multi-scale spatiotemporal registration matrix to changes in the nonlinear coupling strength parameter and adjust the dynamic weight allocation coefficient control strategy accordingly. A multi-level prediction framework is constructed, and the corresponding prediction algorithm and parameter configuration are automatically selected according to the prediction duration. A preliminary prediction result set is generated by using a high-dimensional original matrix and a dynamic weight allocation coefficient control strategy. The preliminary prediction results are iteratively optimized using an ocean dynamic response optimization model to output the final prediction results; based on the final prediction results, a multi-timescale coastal water level prediction product is generated.

2. The method according to claim 1, characterized in that, The multi-scale spatiotemporal registration matrix achieves a unified representation of heterogeneous data by establishing a time interpolation function and a spatial mapping relationship.

3. The method according to claim 2, characterized in that, The marine dynamic process feature extractor is specifically an algorithm module used to identify and quantify various physical processes in the ocean.

4. The method according to claim 3, characterized in that, The astronomical tide characteristics are specifically the periodic water level changes caused by the gravitational influence of celestial bodies, including the amplitude and phase information of semi-diurnal and diurnal tides. The storm surge characteristics are specifically the statistical characteristics of abnormal sea level rise and fall caused by meteorological factors, including the degree of influence of wind speed, air pressure, and wind direction on water level. The wave characteristics are specifically the statistical descriptive parameters of sea surface fluctuations, including physical quantities such as significant wave height, period, and wave direction.

5. The method according to claim 4, characterized in that, The nonlinear coupling strength parameter is specifically a dimensionless index that measures the intensity of interaction between different ocean dynamic factors.

6. The method according to claim 5, characterized in that, The adaptive spatiotemporal fusion algorithm is a data processing method that dynamically adjusts the fusion weights based on data quality and reliability.

7. The method according to claim 6, characterized in that, The spatiotemporal consistency dataset is specifically a multimodal data set with a unified spatiotemporal reference formed after processing with a multi-scale spatiotemporal registration matrix.

8. The method according to claim 7, characterized in that, The extreme weather conditions processing module is specifically designed as a water level prediction algorithm enhancement module for extreme weather conditions such as typhoons and cold waves. It includes an extreme weather identification unit, a nonlinear dynamic amplification unit, and a prediction error compensation unit. The extreme weather identification unit judges the current marine environment state based on the non-stationarity index, the nonlinear dynamic amplification unit performs nonlinear coupling calculations on astronomical tide characteristics, storm surge characteristics, and wave characteristics, and the prediction error compensation unit corrects the prediction results based on the statistical law of prediction deviations of historical extreme weather events.

9. The method according to claim 8, characterized in that, The data flow of the extreme weather condition processing module is as follows: it receives astronomical tide features, storm surge features, wave features and non-stationarity index from the ocean dynamic process feature extractor, and after internal processing, it passes the enhanced feature data to the adaptive spatiotemporal fusion algorithm. At the same time, the extreme weather identification results are passed to the ocean dynamic coupling strength discrimination model to determine whether to start the nonlinear modeling process.

10. The method according to claim 9, characterized in that, The sparse matrix compressed sensing reconstruction mechanism is specifically a method that uses the principle of signal sparsity to reconstruct the original high-dimensional data from a small amount of observation data.

Images