A method and system for accurate prediction of regional wave height using multi-source data fusion and dynamic calibration

Abstract

This invention relates to the field of marine environmental forecasting and marine engineering safety technology, and discloses a method and system for accurate prediction of regional wave height using multi-source data fusion and dynamic calibration. The method includes: S1, acquiring time-synchronized multi-source observation data of the target area, performing three-level quality control processing on the multi-source observation data to obtain a spatiotemporally consistent dataset; S2, calculating multi-dimensional physical features based on the dataset, constructing an input feature matrix after standardization, and dividing the input feature matrix into training, validation, and test sets; S3, constructing a deep spatiotemporal network model, extracting spatial correlation and temporal dependence features from the input feature matrix; S4, supervising the training of the deep spatiotemporal network model and dynamically updating it to obtain a prediction model; S5, performing destandardization processing on the output of the prediction model, generating regional grid wave height prediction data through spatial interpolation, and outputting it. This method improves the accuracy of wave height spatiotemporal evolution prediction.

Description

Technical Field

[0001] This invention relates to the field of marine environmental forecasting and marine engineering safety technology, specifically to a method and system for accurate prediction of regional wave height using dynamic calibration based on multi-source data fusion. Background Technology

[0002] Regional wave height prediction is a core component of marine environmental forecasting and marine engineering safety assurance, and its accuracy directly impacts offshore operation planning, disaster risk assessment, and emergency response decisions. Currently, the mainstream wave height prediction methods in the industry are mainly divided into two categories: physical numerical models and data-driven models. Both types of methods have significant technical limitations and are difficult to meet the high-precision forecasting requirements under complex sea conditions.

[0003] Numerical physical models (such as the SWAN model and the WAVEWATCH-III model) are built based on hydrodynamic equations and require extensive manual parameter tuning (such as bottom friction coefficient and drag coefficient) to adapt to different sea area characteristics. When dealing with the modulation effect of ocean currents on waves, these models use simplified parameterized formulas (such as assuming that ocean currents only linearly affect wave speed along the wave direction), failing to accurately characterize the nonlinear effects of "anti-current wave amplification and downstream wave reduction" in strong current regions, resulting in significant prediction biases in strong current areas. Data-driven models often rely on time series analysis (such as the ARIMA model) or traditional machine learning methods (such as random forests and support vector machines). These models generally suffer from the problem of limited input features—they only consider the statistical correlation between wind speed, wind direction, and wave height, completely ignoring the crucial influencing factor of ocean currents. Furthermore, traditional data-driven models lack the ability to jointly model spatiotemporal dimensions: either they only construct time series models based on single-point observation data, failing to capture the spatial correlation of wave height within a region; or they use static spatial interpolation methods to process regional data, failing to dynamically characterize the evolution of wave height over time, leading to unstable accuracy in short-term forecasts (1-24 hours ahead).

[0004] With the development of marine observation technology, my country has established a three-dimensional observation network covering nearshore waters, capable of acquiring multi-source heterogeneous data such as wind, ocean currents, waves, and air pressure in real time (e.g., the sampling frequency of buoy ADCP equipment reaches 1 hour / time, and ocean current profile observations cover water depths of 0-50m). However, in existing technologies, traditional data assimilation methods have high computational complexity, requiring several hours of computing resources for a single assimilation, making it impossible to achieve real-time fusion of multi-source data; moreover, the assimilation process requires manual design of observation operators, which is difficult to adapt to the complex nonlinear coupling relationship among wind, current, and waves, resulting in the value of multi-source data not being fully explored.

[0005] In summary, existing technologies have significant shortcomings in three aspects: adaptability to complex sea conditions, efficiency of multi-source data fusion, and ability to capture spatiotemporal correlations. There is an urgent need for a wave height prediction technology that can deeply integrate multi-source data, automatically learn spatiotemporal coupling mechanisms, and has real-time performance, in order to break through the industry bottleneck. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the present invention aims to provide a method and system for accurate prediction of regional wave height using multi-source data fusion and dynamic calibration, which can solve the aforementioned technical problems.

[0007] To achieve the above objectives, as a first aspect of the present invention, a method for accurate prediction of regional wave height using multi-source data fusion and dynamic calibration is proposed, comprising: S1, acquire time-synchronized multi-source observation data of the target area, the multi-source observation data includes wave data, wind data, ocean current data and auxiliary air pressure data, perform three-level quality control processing on the multi-source observation data to obtain a spatiotemporally consistent dataset; S2, calculate multi-dimensional physical features based on the dataset, standardize the multi-dimensional physical features, construct an input feature matrix, and divide the input feature matrix into a training set, a validation set, and a test set; S3, Construct a deep spatiotemporal network model, which includes a convolutional neural network submodule and a long short-term memory network submodule. The spatial correlation features of the input feature matrix are extracted through the convolutional neural network submodule, and the temporal dependence features are extracted through the long short-term memory network submodule. S4, supervise the training of the deep spatiotemporal network model using the training set, adjust the model parameters based on the validation set, and dynamically update the model using newly added observation data in real time to obtain the prediction model; S5 performs denormalization on the output of the prediction model, generates regional grid wave height prediction data through spatial interpolation, and outputs it.

[0008] In one possible implementation, the wave data includes wave height sequence, wave direction, and wave period, acquired using an acoustic Doppler current profiler; the wind data includes wind speed and wind direction sequences at a height of 10 meters above the sea surface, acquired by a wind sensor; and the ocean current data includes surface current velocity and direction.

[0009] In one possible implementation, step S1 specifically includes: S11. If the multi-source observation data value exceeds the range of [μ-3σ,μ+3σ], it is judged as an outlier and removed, where μ is the data mean and σ is the standard deviation. S12: Align the multi-source observation data to the whole-point timestamp, use linear interpolation to complete the data with a missing rate ≤5%, use random forest interpolation to complete the data with a missing rate of 5%-20%, and remove the time slice data with a missing rate >20% as a whole. S13, calculate the Pearson correlation coefficient of the concurrent multi-source observation data of adjacent buoy observation points. If the Pearson correlation coefficient is <0.5, troubleshoot the equipment failure and re-collect the data to ensure the spatial continuity of the regional data.

[0010] In one possible implementation, step S2 includes at least one of the following: Wind data features are obtained by decomposing wind speed into eastward and northward wind speed components and performing trigonometric function calculations on wind direction to obtain wind direction sequence features. The ocean current data features are obtained by decomposing the current velocity into eastward and northward velocity components and calculating the current direction sequence features by performing trigonometric function calculations on the current direction. The characteristics of the air pressure data are air pressure sequences P; The angle between wind and waves is used to quantify the influence of wind on the direction of wave propulsion. Wind stress characteristics, calculated using the formula τ=ρ a ×C_d×U 10 2 , where ρ a Where C is the air density, C_d is the drag coefficient, and U is the air density. 10 The wind speed at 10 meters; Wave age characteristics, calculated using the following formula: / U 10 ,in , The acceleration due to gravity is taken as 9.8 m / s². 2 T is the wave period; Wave data features include the features obtained by trigonometric function calculation of wave direction, wave period sequence, and wave height sequence as the prediction target.

[0011] In one possible implementation, when calculating wind stress characteristics, the drag coefficient C_d is determined according to the rule that when the wind speed U... 10 When the speed is ≤10 m / s, C_d = 1.2 × 10 -3 ; when U 10 When the speed is >10 m / s, C_d = 1.0 × 10 -3 .

[0012] In one possible implementation, step S2 involves standardization using Z-score normalization, calculated as: x'=(x-μ) x ) / σ x , where x is the original eigenvalue, μx Let σ be the mean of this characteristic sequence. x denoted as the standard deviation of this characteristic sequence.

[0013] In another aspect, the present invention proposes a regional wave height-accuracy prediction system based on multi-source data fusion and dynamic calibration, for implementing the regional wave height-accuracy prediction method described in the first aspect, comprising: The data acquisition module is used to collect time-synchronized multi-source observation data of the target area, including wave data, wind data, ocean current data, and auxiliary air pressure data. The data quality control module is used to perform three-level quality control processing on the multi-source observation data to obtain a spatiotemporally consistent dataset; The feature engineering module is used to calculate multi-dimensional physical features based on the spatiotemporally consistent dataset, standardize the multi-dimensional physical features, construct an input feature matrix, and divide the dataset into training set, validation set, and test set. The model building module is used to build a deep spatiotemporal network model, which includes a convolutional neural network submodule and a long short-term memory network submodule. The convolutional neural network submodule is used to extract spatial correlation features of the input feature matrix. The convolutional neural network submodule is a one-dimensional convolutional neural network submodule. The long short-term memory network submodule is used to extract time-dependent features. The training and calibration module is used to supervise the training of the deep spatiotemporal network model using the training set, adjust the model parameters based on the validation set, and dynamically update the model by adding new observation data in real time. The prediction output module is used to perform denormalization processing on the output results of the deep spatiotemporal network model, generate regional grid wave height prediction data through spatial interpolation, and output it.

[0014] In one possible implementation, the wave data includes wave height sequence, wave direction, and wave period, acquired using an acoustic Doppler current profiler; the wind data includes wind speed and wind direction sequences at a height of 10 meters above the sea surface, acquired by a wind sensor; and the ocean current data includes surface current velocity and direction.

[0015] In one possible implementation, the feature engineering module includes a feature calculation unit, a normalization unit, and a matrix construction unit: The feature calculation unit is used to calculate multi-dimensional physical features, which include at least one of the following: Wind data features are obtained by decomposing wind speed into eastward and northward wind speed components and performing trigonometric function calculations on wind direction to obtain wind direction sequence features. The ocean current data features are obtained by decomposing the current velocity into eastward and northward velocity components and calculating the current direction sequence features by performing trigonometric function calculations on the current direction. The characteristics of the air pressure data are air pressure sequences P; The angle between wind and waves is used to quantify the influence of wind on the direction of wave propulsion. Wind stress characteristics, calculated using the formula τ=ρ a ×C_d×U 10 2 , where ρ a Where C is the air density, C_d is the drag coefficient, and U is the air density. 10 The wind speed at 10 meters; Wave age characteristics, calculated using the following formula: / U 10 ,in , The acceleration due to gravity is taken as 9.8 m / s². 2 T is the wave period; Wave data features include features obtained by trigonometric function calculation of wave direction, wave period sequence, and wave height sequence as the prediction target; The standardization unit is used to standardize the multi-dimensional physical features using Z-score normalization. The matrix construction unit is used to organize the standardized features into multivariate time series samples according to an hourly time interval, construct the input feature matrix, and divide the training set, validation set, and test set.

[0016] In one possible implementation, the training calibration module is dynamically updated by periodically or in real-time inputting new observation data into the training set after system deployment, thereby fine-tuning the parameters of the deep spatiotemporal network model.

[0017] Compared with the prior art, the beneficial effects of the present invention are: 1. By using time-synchronized observation data of wind, ocean currents, waves, and air pressure as the core input for wave height prediction, this method overcomes the limitations of traditional data-driven models that only consider the single relationship between wind and waves. It allows the model to fully learn the complex nonlinear coupling relationship between wind, current, and waves, accurately characterizing the nonlinear effect of "wave amplification against the current and wave reduction with the current" in strong current areas. This solves the industry pain point of large prediction deviation in strong current areas by traditional methods, and significantly improves the wave height prediction accuracy under complex sea conditions compared to traditional models that ignore ocean current characteristics. 2. By setting up three-level quality control rules, abnormal data is accurately removed, effective data is supplemented in a targeted manner, equipment failures are investigated in a timely manner, and data is re-collected. This effectively avoids the interference of low-quality and inconsistent data on model training, making the wave height spatiotemporal evolution law learned by the model more consistent with the real sea conditions. This not only effectively reduces the prediction error caused by data noise, but also improves the convergence speed of model training. 3. A deep spatiotemporal network architecture was constructed to achieve joint modeling of the spatial correlation and temporal dependence of wave height, which greatly improved the spatiotemporal representation capability of the model. A three-layer one-dimensional CNN was used to accurately learn the spatial influence relationship between the wave height of a single observation point and the wind field and flow field of the surrounding area. A two-layer LSTM was used to effectively remember the temporal driving effect of long-term wind and flow changes on wave height. The coupled design of the two abandons the limitation of the single spatiotemporal modeling of traditional models, making the spatiotemporal evolution prediction of regional wave height more accurate, improving the accuracy and stability of short-term forecasts, and solving the technical problems of single-point prediction without spatial correlation and static interpolation without temporal evolution in traditional models. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall process of the regional wave height accurate prediction method of multi-source data fusion dynamic calibration according to the present invention; Figure 2 This is a deep spatiotemporal network model architecture diagram of the regional wave height accurate prediction method of multi-source data fusion dynamic calibration in this invention. Figure 3 This is an architecture diagram of the regional wave height accurate prediction system of the multi-source data fusion dynamic calibration system of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0021] Existing technologies struggle to accurately describe the modulation effect of ocean currents on wave height (such as wave amplification by countercurrents and wave reduction by downstreams) in terms of physical models, and are overly parameterized and simplified when dealing with this effect. In terms of data models, most models only consider wind-wave relationships, have single input features, and ignore the key contribution of ocean currents. The models have simple structures, cannot effectively process spatiotemporal sequence data, and have limited predictive capabilities. In terms of data assimilation, traditional assimilation methods are complex and computationally intensive, making it difficult to achieve real-time, efficient, and deep fusion of observational data with models.

[0022] Therefore, please see Figure 1 A preferred embodiment of this invention provides a method for accurate prediction of regional wave height using multi-source data fusion and dynamic calibration. This method comprehensively integrates historical and real-time observational data from multiple sources, including wind, ocean currents, waves, and air pressure. By constructing a deep spatiotemporal network architecture, it automatically learns the nonlinear spatiotemporal coupling mechanism of wind, current, and waves, eliminating the need for manual design of physical parameters or observation operators. It enables dynamic updating and online learning of the prediction model, ensuring high prediction accuracy even during environmental changes (such as seasonal transitions or typhoon passage). It meets the real-time requirements of short-term forecasts, with a single prediction time controlled within 1 second, providing rapid decision support for maritime emergency response. The method includes the following steps: S1, acquire time-synchronized multi-source observation data of the target area, the multi-source observation data includes wave data, wind data, ocean current data and auxiliary air pressure data, perform three-level quality control processing on the multi-source observation data to obtain a spatiotemporally consistent dataset; S2, calculate multi-dimensional physical features based on the dataset, standardize the multi-dimensional physical features, construct an input feature matrix, and divide the input feature matrix into a training set, a validation set, and a test set; S3, Construct a deep spatiotemporal network model, which includes a convolutional neural network submodule and a long short-term memory network submodule. The spatial correlation features of the input feature matrix are extracted through the convolutional neural network submodule, and the temporal dependence features are extracted through the long short-term memory network submodule. S4, supervised training of the deep spatiotemporal network model using the training set, adjusting model parameters based on the validation set, and dynamically updating the model using newly added observation data in real time to obtain a prediction model; dynamic updating means periodically or in real time inputting new observation data into the training set to fine-tune the model parameters; S5 performs denormalization on the output of the prediction model, generates regional grid wave height prediction data through spatial interpolation, and outputs it.

[0023] In one possible implementation, the wave data includes wave height sequence, wave direction, and wave period, acquired using an acoustic Doppler current profiler; the wind data includes wind speed and wind direction sequences at a height of 10 meters above the sea surface, acquired by a wind sensor; and the ocean current data includes surface current velocity and direction.

[0024] In one possible implementation, step S1 specifically includes: S11, if the multi-source observation data value exceeds [μ-3σ, If the value falls outside the range, it is considered an outlier and removed, where μ is the data mean and σ is the standard deviation. S12: Align the multi-source observation data to the whole-point timestamp, use linear interpolation to complete the data with a missing rate ≤5%, use random forest interpolation to complete the data with a missing rate of 5%-20%, and remove the time slice data with a missing rate >20% as a whole. S13, calculate the Pearson correlation coefficient of the concurrent multi-source observation data of adjacent buoy observation points. If the Pearson correlation coefficient is <0.5, troubleshoot the equipment failure and re-collect the data to ensure the spatial continuity of the regional data.

[0025] In one possible implementation, step S2 includes at least one of the following: Wind data features are obtained by decomposing wind speed into eastward and northward wind speed components and performing trigonometric function calculations on wind direction to obtain wind direction sequence features. The ocean current data features are obtained by decomposing the current velocity into eastward and northward velocity components and calculating the current direction sequence features by performing trigonometric function calculations on the current direction. The characteristics of the air pressure data are air pressure sequences P; The angle between wind and waves is used to quantify the influence of wind on the direction of wave propulsion. Wind stress characteristics, calculated using the formula τ=ρ a ×C_d×U 10 2 , where ρ a Where C is the air density, C_d is the drag coefficient, and U is the air density. 10 The wind speed at 10 meters; Wave age characteristics, calculated using the following formula: / U 10 ,in , The acceleration due to gravity is taken as 9.8 m / s². 2 T is the wave period; Wave data features include the features obtained by trigonometric function calculation of wave direction, wave period sequence, and wave height sequence as the prediction target.

[0026] In one possible implementation, the drag coefficient C_d is determined according to the rule that when the wind speed U... 10 When the speed is ≤10 m / s, C_d = 1.2 × 10 -3 ; when U 10 When the speed is >10 m / s, C_d = 1.0 × 10 -3 .

[0027] In one possible implementation, step S2 involves standardization using Z-score normalization, calculated as: x'=(x-μ) x ) / σ x , where x is the original eigenvalue, μ x Let σ be the mean of this characteristic sequence. x denoted as the standard deviation of this characteristic sequence.

[0028] In this embodiment of the invention, the construction of the multi-source spatiotemporal dataset specifically involves collecting time-synchronized multi-source data from the ocean observation network of the target area: 1. Wave Data: Wave data is provided by ADCP on the buoy, wave height sequence , wave direction Wave period, sampling frequency 1 hour 2. Wind data: Wind data is provided by wind sensors on the buoy, which can measure wind speed sequences at a height of 10 meters above the sea surface. Wind direction sequence , 3. Ocean current data: Surface current velocity Flow direction , 4. Auxiliary data: air pressure data P.

[0029] A three-level quality control process is employed for the collected data to ensure data reliability. The first level of quality control involves outlier detection and removal, based on the 3σ criterion (data values ​​exceeding the range [μ-3σ, μ+3σ] are considered outliers and removed, where μ is the data mean and σ is the standard deviation). The second level involves time alignment and missing value completion, aligning all data to the exact timestamp, using linear interpolation for data with a missing rate ≤5%, random forest interpolation for data with a missing rate between 5% and 20%, and removing time slices with a missing rate >20%. The third level involves spatial consistency verification, calculating the correlation between contemporaneous data from adjacent buoy observation points (e.g., Pearson correlation coefficient for wind speed and wave height). If the correlation coefficient is <0.5, equipment malfunctions are investigated and data is recollected to ensure the spatial continuity of the regional data.

[0030] Multi-dimensional physical features are calculated based on the dataset, and the multi-dimensional physical features specifically include: (1) Wind data characteristics: The wind speed is decomposed into eastward and northward directions to obtain the wind speed characteristics: The wind direction is calculated using trigonometric functions to obtain the characteristics of the wind direction sequence: (2) Characteristics of ocean current data: The current velocity series was decomposed into eastward and northward directions to obtain: The flow sequence is calculated using trigonometric functions to obtain: .

[0031] (3) Characteristics of air pressure data sequence: P.

[0032] (4) Angle between wind and waves It is used to quantify the influence of wind on the direction of wave drive.

[0033] (5) Wind stress: ,in Let the air density be denoted as . , This is the drag coefficient, whose magnitude is related to wind speed. When the wind speed... , Values ,when , Values , The wind speed at 10 meters is expressed in units of... , (6) Wave Age: Wave age reflects the maturity of a wave and is a key indicator for determining whether a wave is in its growth stage. The calculation formula is: ,in ( The acceleration due to gravity is taken as 9.8 m / s². 2 ).

[0034] (7) Wave data: Trigonometric function calculation of wave direction Wave periodic sequence , wave height (Also serving as a prediction target).

[0035] The calculation principle for wind direction and current direction trigonometric functions is similar. The core method for wave direction trigonometric function calculation is to convert the geographical azimuth of the wave direction into orthogonal north and east components, realizing a numerical quantitative expression of the wave direction angle to meet the feature learning needs of deep spatiotemporal network models. The specific calculation process is as follows: The wave direction adopts the geographical azimuth measurement rule commonly used in the marine field: with true north as 0° (360°), the azimuth angle of the wave direction is measured in a clockwise direction (denoted as ). The value range is 0°≤ (<360°), this benchmark is completely consistent with the trigonometric function calculation benchmark for wind direction and flow direction, ensuring the consistency of multi-source directional characteristics.

[0036] Due to the requirements of trigonometric function calculations, the angular value of the wave direction is first converted into a radian value. The formula is: = ; The wave direction radian value is decomposed into a northward component using sine and cosine functions (denoted as ). ) and the eastward component (denoted as The two orthogonal numerical characteristics are the core of the wave-direction trigonometric function calculation, and the formula is: , Both components have values ​​ranging from [-1, 1] to accurately represent the geographic direction attribute of the wave direction.

[0037] In this embodiment of the invention, the multi-dimensional physical features are standardized. Specifically, Z-score normalization is used to eliminate the influence of dimensions. Z-score normalization involves calculating the mean and standard deviation for each feature sequence and then scaling it. The calculation formula is as follows: , where μ is the mean and σ is the standard deviation.

[0038] In this embodiment of the invention, the multi-dimensional physical features are standardized to construct an input feature matrix. Specifically, for each time step, all data features are organized into multivariate time series samples according to one-hour time intervals to construct the input features. Input features , construct training set, validation set and test set.

[0039] In this embodiment of the invention, a deep spatiotemporal network model is constructed, such as... Figure 2 As shown, firstly, multi-dimensional features are processed by convolution, and a CNN is used to learn how the wave height at a certain point is affected by the wind field and flow field at its surrounding points. Then, a one-dimensional convolutional neural network is used to extract local correlation patterns from the features at each time point.

[0040] The calculation formula is as follows: ,in: Convolutional layer number ; The number of input channels for the i-th layer (e.g., the first layer) (Set the number of input features to 16). Kernel size (1 / 3 layer) , 2nd layer); For the first Layer The first input channel, the first Each convolutional kernel parameter; For the first Layer Each channel, location The input feature values; For the first Layer The bias term for each output channel; The activation function is given by the formula: , used to introduce nonlinear feature representation.

[0041] Specifically, a 3-layer one-dimensional convolutional network (1D-CNN) is used, with the following parameters for each layer: First layer: The number of input channels equals the number of features (such as wind u / v, ocean current u / v, historical wave height, etc.), the output channels are 64, the kernel size is 3, the stride is 1, the padding is 1, and the activation function is ReLU.

[0042] Second layer: 64 input channels, 128 output channels, kernel size 5, stride 1, padding 2, activation function ReLU.

[0043] Third layer: 128 input channels, 64 output channels, kernel size 3, stride 1, padding 1, activation function ReLU.

[0044] After the convolution calculation is complete, batch normalization (BatchNorm) is added after each layer to improve training stability. The calculation formula is as follows: in, This is the output of the convolutional layer. This is the average value of the current batch. This represents the variance of the current batch. To prevent division by zero errors, the value is set to... The learnable scaling factor is initialized to 1. B is the learnable bias term, initially set to 0, and B is the batch size, i.e., the number of samples trained each time.

[0045] For temporal feature processing, a Long Short-Term Memory (LSTM) network is used to capture long-term temporal dependencies in the sequence data. LSTM can remember the impact of wind and current changes on the current wave height over a long period of time.

[0046] Specifically, the convolutional feature data is used as the feature input of the LSTM layer to train the neural network and automatically learn sequence features. The LSTM consists of cell states and gate mechanisms, and captures long-term dependencies through cell state updates. The cell state can update long-term memory, as shown in the following formula:

[0047] First, the multidimensional features are processed by convolution and then input into the forget gate (which determines the historical information to be discarded): , Filter the current input The process of generating new and valid information to prepare for updating cell states involves two steps: screening and generation. First, candidate information is generated using the following formula. Then update the cell state. Finally, update the cell state. .

[0048] Output gate (determines the current information to be output): Generates the output gate signal: , Generate current hidden state update:

[0049] Where is the time step (1≤t≤24). The output features of the CNN at time step t (dimension 64). Let be the LSTM hidden state (dimension 128) at time step t-1. The LSTM cell state at time step t-1 (dimension 128). This is the weight matrix. For bias terms; This is the sigmoid activation function (outputs 0~1, controlling the gate switch). The hyperbolic tangent activation function (output -1 to 1, which are normalized state values); Weight initialization: Orthogonal initialization is used to ensure the weight matrix... satisfy (Identity matrix), to avoid gradient explosion / vanishing in the early stages of training, the formula is:

[0050] Where is the input dimension (e.g., the first layer of LSTM) and is the output dimension (128).

[0051] Forget gate bias initialization: Zero initialization is used, i.e. (Dimension 128) to avoid excessive forgetting of historical information in the early stages of training.

[0052] In this embodiment of the invention, a two-layer LSTM output layer is constructed: LSTM layer 1: Input dimension 64 (CNN output), number of hidden units 128, output is the hidden state at 24 time steps (dimension 24×128). LSTM Layer 2: Input dimension 128 (output of LSTM Layer 1), number of hidden units 128, output is the hidden state of the last time step (dimension 1×128, i.e. global temporal features).

[0053] Use a 2-layer LSTM with 128 hidden units, a dropout rate of 0.2, and a time step set to the length of the historical sequence (e.g., 24 hours).

[0054] The output of the spatiotemporal feature extraction module is finally connected to a fully connected layer. The output layer uses a linear activation function to output the predicted wave height for several future time steps.

[0055] In this embodiment of the invention, the model training and dynamic calibration process involves optimizing model parameters through supervised training and achieving "online evolution" of the model based on measured data to adapt to changes in sea state.

[0056] Supervised training of the deep spatiotemporal network was performed using historical test datasets, and the loss function was minimized using the Adam optimizer. Basic optimizer parameters: learning rate. Momentum parameters Second-order momentum parameters Numerical stability parameters ; The parameter update formula is as follows: , in, For model parameters, The gradient of the loss function with respect to the parameters. , These are first-order and second-order momentum estimates, respectively. , This is the momentum estimate after bias correction.

[0057] The loss function uses mean squared error (MSE) and mean absolute error (MAE) to predict wave height and actual observed wave height.

[0058] The basic loss function is the mean squared error (MSE): suitable for scenarios with relatively uniform wave height distribution, the formula is:

[0059] Mean Absolute Error (MAE): Suitable for scenarios with extreme wave heights (such as during typhoon season), it is more robust to outliers. The formula is:

[0060] in, For the sample size, To predict the time step, For the first The first sample The true wave height at the time step, This corresponds to the predicted value.

[0061] After the system is deployed, new observation data are added to the training set periodically or in real time to fine-tune the model so that it can adapt to changes in the environment and maintain its predictive performance.

[0062] The output of the prediction model is denormalized, and regional grid wave height prediction data is generated and output through spatial interpolation.

[0063] Destandardization: Since the input features have already been Z-score standardized, the predicted output needs to be destandardized to restore the true wave height. The formula is: The formula is: in, These are the mean and standard deviation of the "true wave height sequence" in the training set, respectively.

[0064] The output results from each station are weighted according to the inverse distance weight to generate regional grid data with a spatial resolution of 0.1°×0.1°.

[0065] This invention also provides a regional wave high-precision prediction system with multi-source data fusion and dynamic calibration, comprising: The data acquisition module is used to collect time-synchronized multi-source observation data of the target area, including wave data, wind data, ocean current data, and auxiliary air pressure data. The data quality control module is used to perform three-level quality control processing on the multi-source observation data to obtain a spatiotemporally consistent dataset; The feature engineering module is used to calculate multi-dimensional physical features based on the spatiotemporally consistent dataset, standardize the multi-dimensional physical features, construct an input feature matrix, and divide the dataset into training set, validation set, and test set. The model building module is used to build a deep spatiotemporal network model, which includes a convolutional neural network submodule and a long short-term memory network submodule. The convolutional neural network submodule is used to extract spatial correlation features of the input feature matrix, and the long short-term memory network submodule is used to extract temporal dependency features. The training and calibration module is used to supervise the training of the deep spatiotemporal network model using the training set, adjust the model parameters based on the validation set, and dynamically update the model by adding new observation data in real time. The prediction output module is used to perform denormalization processing on the output results of the deep spatiotemporal network model, generate regional grid wave height prediction data through spatial interpolation, and output it.

[0066] In one possible implementation, the wave data includes wave height sequence, wave direction, and wave period, acquired using an acoustic Doppler current profiler; the wind data includes wind speed and wind direction sequences at a height of 10 meters above the sea surface, acquired by a wind sensor; and the ocean current data includes surface current velocity and direction.

[0067] This invention is based on the deep fusion of multi-source heterogeneous marine observation data, and uses a deep spatiotemporal network coupled with convolutional neural networks and long short-term memory networks as the core modeling carrier. Through the collaborative operation of various modules, it achieves end-to-end automated processing from marine observation data acquisition to refined prediction of regional wave height. Its core working principle is based on a modular process design, which standardizes and performs quality control and extracts physical features from multi-source data such as wind, ocean currents, waves, and air pressure. The convolutional neural network submodule captures the spatial correlation features between wave height and wind and current fields in the region. The long short-term memory network submodule explores the dynamic evolution of wave height over time. Then, through supervised training and real-time dynamic calibration, the model parameters are optimized so that the model can adapt to the dynamic changes in sea state. Finally, through destandardization and spatial interpolation, the model prediction results are restored to regional grid wave height data that fits the actual application, realizing accurate and real-time prediction of regional wave height under complex sea conditions.

[0068] The specific working process of this system is carried out sequentially according to the execution order of the modules. The output results of each module serve as the core input of the subsequent modules, forming a closed-loop and iterative prediction process. The specific process is as follows: First, the data acquisition module synchronously collects wave data, wind data, ocean current data, and barometric pressure data from devices such as buoys and current meters based on the marine three-dimensional observation network deployed in the target area, ensuring that the timestamps of various types of data are consistent, forming an original multi-source observation dataset, and then transmitting the dataset to the data quality control module in real time. Then, after receiving the original multi-source observation dataset, the data quality control module performs three-level quality control processing. First, it detects and removes outliers based on the 3σ criterion. Then, it performs hourly time alignment on the data. It uses linear interpolation and random forest interpolation to fill in missing data or remove invalid time slice data according to the missing rate gradient. Finally, it completes the spatial consistency test by calculating the correlation of data from adjacent observation points at the same time. It also checks for equipment failures and collects valid data to obtain a spatiotemporally consistent, high-quality standardized dataset, which is then transmitted to the feature engineering module. Next, the feature engineering module calculates multi-dimensional features with physical meaning, such as wind / current direction decomposition, wind-wave angle, wind stress, and wave age, based on a spatiotemporally consistent dataset. Z-score normalization is used to standardize all features to eliminate the influence of dimensions. Then, the standardized features are organized into multivariate time series samples at 1-hour intervals to construct the input feature matrix. The matrix is ​​then divided into training set, validation set, and test set according to a preset ratio. The divided feature matrix is ​​then transmitted to the model building module and the training and calibration module. Furthermore, after receiving the input feature matrix, the model building module constructs a deep spatiotemporal network model containing a convolutional neural network submodule and a long short-term memory network submodule. The convolutional neural network submodule performs convolution and batch normalization on the feature matrix to extract the spatial correlation features between wave height and wind field and flow field. The spatial feature results are then transmitted to the long short-term memory network submodule, which mines the long-term time dependencies in the feature sequence to complete the joint extraction of spatiotemporal features. The constructed deep spatiotemporal network model will serve as the core prediction model and be connected to the training and calibration module. Subsequently, the training and calibration module first uses the training set to supervise the training of the deep spatiotemporal network model, and uses a preset optimizer and loss function to minimize the model prediction error. Based on the validation set, it monitors the model training effect in real time and adjusts the model parameters to obtain the initial optimal prediction model. After the system is deployed, this module continuously receives the observation data added in real time by the data acquisition module. After quality control and feature engineering, it fine-tunes and dynamically updates the parameters of the initial optimal model to adapt the model to changes in sea conditions such as seasonal transitions and typhoons, and obtains the optimal model that can respond to sea conditions in real time, which is then transmitted to the prediction output module. Finally, the prediction output module calls the optimal deep spatiotemporal network model obtained from the training and calibration module, inputs the feature matrix to be predicted into the model to obtain the standardized wave height prediction result, and then performs destandardization processing on the prediction result based on the mean and standard deviation of the true wave height in the training set to restore it to the true wave height scale. Finally, the inverse distance weighted interpolation method is used to perform spatial interpolation on the single-point prediction result to generate regional grid wave height prediction data with a preset resolution, and completes the final result output in the form of numerical files and visualization charts.

[0069] Meanwhile, the various modules of this system can realize real-time data interaction and iterative optimization of the model. The dynamic update results of the training and calibration module can be fed back to the model building module and the feature engineering module to continuously optimize the model structure and feature extraction rules, thereby continuously improving the prediction accuracy of the system.

[0070] In one possible implementation, the feature engineering module includes a feature calculation unit, a normalization unit, and a matrix construction unit: The feature calculation unit is used to calculate multi-dimensional physical features, which include at least one of the following: Wind data features are obtained by decomposing wind speed into eastward and northward wind speed components and performing trigonometric function calculations on wind direction to obtain wind direction sequence features. The ocean current data features are obtained by decomposing the current velocity into eastward and northward velocity components and calculating the current direction sequence features by performing trigonometric function calculations on the current direction. The characteristics of the air pressure data are air pressure sequences P; The angle between wind and waves is used to quantify the influence of wind on the direction of wave propulsion. Wind stress characteristics, calculated using the formula τ=ρ a ×C_d×U 10 2 , where ρ a Where C is the air density, C_d is the drag coefficient, and U is the air density. 10 The wind speed at 10 meters; Wave age characteristics, calculated using the following formula: / U 10 ,in , The acceleration due to gravity is taken as 9.8 m / s². 2 T is the wave period; Wave data features include features obtained by trigonometric function calculation of wave direction, wave period sequence, and wave height sequence as the prediction target; The standardization unit is used to standardize the multi-dimensional physical features using Z-score normalization. The matrix construction unit is used to organize the standardized features into multivariate time series samples according to 1-hour time intervals, construct the input feature matrix, and divide the training set, validation set, and test set.

[0071] In one possible implementation, the training calibration module is dynamically updated by periodically or in real-time inputting new observation data into the training set after system deployment, thereby fine-tuning the parameters of the deep spatiotemporal network model.

[0072] The training and calibration module employs a dynamic update approach that periodically or in real-time integrates new observational data into the training set and fine-tunes the parameters of the deep spatiotemporal network model. This enables the model to learn online and evolve adaptively in line with changes in sea state, effectively addressing the technical drawbacks of traditional wave height prediction models, which rely on long-term use after a single training phase and are prone to continuous decline in prediction accuracy due to dynamic changes in sea state. On one hand, for routine sea state changes such as seasonal transitions and slow adjustments in ocean current direction, periodically integrating new data to fine-tune parameters allows the model to gradually learn the progressive patterns of sea state changes, continuously conforming to the characteristics of the real marine environment and ensuring stable prediction accuracy throughout long-term deployment. On the other hand, for sudden sea state changes such as typhoon passage and sudden increases in current velocity in strong current areas, real-time integration of new observational data allows for rapid fine-tuning of model parameters, enabling the model to adapt to the abrupt changes in sea state and avoiding prediction failures caused by a disconnect between sea state data and model training data. Meanwhile, this update method only makes minor adjustments to the model parameters and does not require retraining of the deep spatiotemporal network model, which greatly reduces the computational resource consumption and time cost of model updates. Moreover, the entire process does not require manual intervention to adjust physical parameters or model structure, which reduces the cost of manual operation and allows the model to always complete wave height prediction in the optimal state, providing continuous model performance assurance for long-term and accurate short-term forecasts of regional wave heights.

[0073] This invention incorporates ocean current observation data as a core input feature alongside wind and historical wave data into the wave height prediction model. It employs a deep spatiotemporal network architecture combining CNN and LSTM, rather than a single model, to simultaneously capture spatial dependence and temporal evolution patterns. An end-to-end multi-source data fusion learning mechanism allows the model to automatically discover the complex physical relationships between wind, current, and waves from the data, rather than relying on manual settings. By introducing ocean currents and utilizing deep learning to capture complex nonlinear relationships, the prediction accuracy in complex sea conditions such as strong current areas far exceeds that of traditional physical models and simple data models that ignore ocean currents. The CNN-LSTM architecture can simultaneously learn spatial correlations and temporal dynamics, resulting in more accurate predictions of the spatiotemporal evolution of wave height.

[0074] The above description is a detailed description of the preferred embodiments of the present invention. However, the embodiments are not intended to limit the scope of the patent application of the present invention. All equivalent changes or modifications made under the technical spirit of the present invention should fall within the patent scope covered by the present invention.

Claims

1. A method for accurate prediction of regional wave height using multi-source data fusion and dynamic calibration, characterized in that: include: S1, acquire time-synchronized multi-source observation data of the target area, the multi-source observation data includes wave data, wind data, ocean current data and auxiliary air pressure data, perform three-level quality control processing on the multi-source observation data to obtain a spatiotemporally consistent dataset; S2, calculate multi-dimensional physical features based on the dataset, standardize the multi-dimensional physical features, construct an input feature matrix, and divide the input feature matrix into a training set, a validation set, and a test set; S3, Construct a deep spatiotemporal network model, which includes a convolutional neural network submodule and a long short-term memory network submodule. The spatial correlation features of the input feature matrix are extracted through the convolutional neural network submodule, and the temporal dependence features are extracted through the long short-term memory network submodule. S4, supervise the training of the deep spatiotemporal network model using the training set, adjust the model parameters based on the validation set, and dynamically update the model using new observation data to obtain the prediction model; S5 performs denormalization on the output of the prediction model, generates regional grid wave height prediction data through spatial interpolation, and outputs it.

2. The method for accurate prediction of regional wave height according to claim 1, characterized in that, The wave data includes wave height sequence, wave direction and wave period, acquired by an acoustic Doppler current profiler; the wind data includes wind speed sequence and wind direction sequence at a height of 10 meters above the sea surface, acquired by a wind sensor; the ocean current data includes surface current velocity and direction.

3. The method for accurate prediction of regional wave height according to claim 1, characterized in that, In S1, the three-level quality control process specifically includes: S11, if the multi-source observation data values ​​exceed If the value falls outside the range, it is considered an outlier and removed, where μ is the data mean and σ is the standard deviation. S12: Align the multi-source observation data to the whole-point timestamp, use linear interpolation to complete the data with a missing rate ≤5%, use random forest interpolation to complete the data with a missing rate of 5%-20%, and remove the time slice data with a missing rate >20% as a whole. S13, calculate the Pearson correlation coefficient of the concurrent multi-source observation data of adjacent buoy observation points. If the Pearson correlation coefficient is <0.5, troubleshoot the equipment failure and re-collect the data to ensure the spatial continuity of the regional data.

4. The method for accurate prediction of regional wave height according to claim 1, characterized in that, In S2, the multi-dimensional physical features include at least one of the following: Wind data features are obtained by decomposing wind speed into eastward and northward wind speed components and performing trigonometric function calculations on wind direction to obtain wind direction sequence features. The ocean current data features are obtained by decomposing the current velocity into eastward and northward velocity components and calculating the current direction sequence features by performing trigonometric function calculations on the current direction. The characteristics of the air pressure data are air pressure sequences P; The angle between wind and waves is used to quantify the influence of wind on the direction of wave propulsion. Wind stress characteristics, calculated using the formula τ=ρ a ×C_d×U 10 2 , where ρ a Where C is the air density, C_d is the drag coefficient, and U is the air density. 10 The wind speed at 10 meters; Wave age characteristics, calculated using the following formula: / U 10 ,in , The acceleration due to gravity is taken as 9.8 m / s². 2 T is the wave period; Wave data features include the characteristics obtained by trigonometric function calculations on wave direction, wave period sequence, and wave height sequence.

5. The method for accurate prediction of regional wave height according to claim 4, characterized in that, The drag coefficient C_d is determined according to the rule that when the wind speed U 10 When the speed is ≤10 m / s, C_d = 1.2 × 10 -3 ; when U 10 When the speed is >10 m / s, C_d = 1.0 × 10 -3 .

6. The method for accurate prediction of regional wave height according to claim 1, characterized in that, In S2, the spatial interpolation adopts the inverse distance weighted interpolation method.

7. The method for accurate prediction of regional wave height according to claim 1, characterized in that, The standardization process employs Z-score normalization.

8. A regional wave height prediction system with multi-source data fusion and dynamic calibration, characterized in that, For implementing the regional wave height accurate prediction method as described in any one of claims 1-7, comprising: The data acquisition module is used to collect time-synchronized multi-source observation data of the target area, including wave data, wind data, ocean current data, and auxiliary air pressure data. The data quality control module is used to perform three-level quality control processing on the multi-source observation data to obtain a spatiotemporally consistent dataset; The feature engineering module is used to calculate multi-dimensional physical features based on the spatiotemporally consistent dataset, standardize the multi-dimensional physical features, construct an input feature matrix, and divide the dataset into training set, validation set, and test set. The model building module is used to build a deep spatiotemporal network model, which includes a convolutional neural network submodule and a long short-term memory network submodule. The convolutional neural network submodule is used to extract spatial correlation features of the input feature matrix, and the long short-term memory network submodule is used to extract temporal dependency features. The training and calibration module is used to supervise the training of the deep spatiotemporal network model using the training set, adjust the model parameters based on the validation set, and dynamically update the model by adding new observation data. The prediction output module is used to perform denormalization processing on the output results of the deep spatiotemporal network model, generate regional grid wave height prediction data through spatial interpolation, and output it.

9. The regional wave height accurate prediction system according to claim 8, characterized in that, The feature engineering module includes a feature calculation unit, a normalization unit, and a matrix construction unit; The feature calculation unit is used to calculate multi-dimensional physical features, which include at least one of the following: Wind data features are obtained by decomposing wind speed into eastward and northward wind speed components and performing trigonometric function calculations on wind direction to obtain wind direction sequence features. The ocean current data features are obtained by decomposing the current velocity into eastward and northward velocity components and calculating the current direction sequence features by performing trigonometric function calculations on the current direction. The characteristics of the air pressure data are air pressure sequences P; The angle between wind and waves is used to quantify the influence of wind on the direction of wave propulsion. Wind stress characteristics, calculated using the formula τ=ρ a ×C_d×U 10 2 , where ρ a Where C is the air density, C_d is the drag coefficient, and U is the air density. 10 The wind speed at 10 meters; Wave age characteristics, calculated using the following formula: / U 10 ,in , The acceleration due to gravity is taken as 9.8 m / s². 2 T is the wave period; Wave data features include features obtained by trigonometric function calculations on wave direction, wave period sequence, and wave height sequence; The standardization unit is used to standardize the multi-dimensional physical features using Z-score normalization. The matrix construction unit is used to organize the standardized features into multivariate time series samples according to an hourly time interval, construct the input feature matrix, and divide the training set, validation set, and test set.

10. The regional wave height accurate prediction system according to claim 8, characterized in that, The training and calibration module is dynamically updated by periodically or in real-time inputting new observation data into the training set to fine-tune the parameters of the deep spatiotemporal network model.

Images