Method and system for predicting sea wave parameters

Abstract

The invention relates to an improved sea wave prediction method comprising the following steps: obtaining training data from a sea wave data set, the training data comprising a plurality of sea wave variables in a spatial domain; processing the training data by a normalization process; using the processed data to train a machine learning model for sea wave parameter prediction; and generating a plurality of sea wave parameter predictions in the spatial domain using the trained machine learning model. The invention aims to provide more accurate and real-time sea wave prediction data so as to meet the requirements in the fields of ocean navigation, offshore operation, coastal construction, disaster prevention and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of ocean wave prediction, in particular to an ocean wave prediction system and method for marine environmental monitoring and marine activity safety management. Background Art

[0002] The ocean, as the largest component of the Earth's surface, its weather system and ocean wave activities have a profound impact on various human activities. In particular, the prediction of ocean wave height is of crucial significance for maritime shipping safety, ocean engineering design, coastal tourism activities, etc. In these activities, the instability and unpredictability of ocean waves often lead to safety accidents, posing a major threat to personnel safety and economic interests.

[0003] Currently, some new methods and technologies have emerged in the field of ocean wave prediction, including physics-based simulation models, analytical models, and machine learning / deep learning models.

[0004] Traditional ocean wave prediction methods relying on empirical formulas and complex physical models usually not only rely on huge computing power but also require a large number of ocean and atmospheric parameters, such as wind speed, wind direction, flow velocity, etc., to predict the generation, propagation, and attenuation of ocean waves. Although these methods are effective within a certain time and space range, they have obvious limitations in dealing with the complexity of the marine environment. For example, the high nonlinearity of the marine environment and the spatio-temporal variability of parameters make it difficult for traditional models to accurately capture the rapid changes in ocean wave height, especially under extreme weather conditions. In addition, when facing large-scale prediction or the need for real-time feedback, traditional methods are often difficult to meet the requirements due to high computational complexity.

[0005] For example, physical models such as SWAN (Simulating Waves Nearshore) are widely used for generating global ocean wave data, but their computational cost is high and they are not suitable for real-time updates. Analytical models predict ocean waves by interpolating observed data, but their estimation error increases as the distance from the observation point increases. In addition, although machine learning and deep learning models have shown strong performance in multiple tasks, current applications mainly focus on the duration prediction of a single ocean wave parameter, and most studies are limited to internal data verification in specific regions.

[0006] The prior art also faces other problems when conducting regional sea wave prediction. First is the problem of data collection. The observation data of the ocean environment is often sparse and unevenly distributed, resulting in the lack of geographical and temporal continuity of the input data for the model. Second is the problem of the model's generalization ability. Traditional models perform well in specific regions, but when extended to new regions, due to the lack of adaptability, their prediction accuracy often drops significantly. Third is the problem of computational efficiency. The ocean environment monitoring and early warning system needs to respond quickly, while existing models often take a long time in the calculation process and cannot provide real-time prediction. Therefore, in order to effectively improve the accuracy and real-time performance of sea wave prediction, it is imperative to develop new prediction models and technologies.

[0007] These challenges have prompted researchers and engineers to seek more efficient and accurate sea wave prediction methods. Among them, the application of machine learning and deep learning technologies to sea wave prediction, especially in high-resolution spatial dimensions, has become a research hotspot in recent years. These methods can not only handle the ocean environment changes that are difficult to capture by traditional models, but also overcome the problems of data sparsity and insufficient model generalization ability to a certain extent. However, current research mostly focuses on the prediction of specific sea areas or specific sea wave parameters, lacking research on the comprehensive prediction of multiple parameters such as sea wave height, period, and direction. In addition, existing research mostly uses data from local areas for training and validation, and the global applicability and universality of the sea wave prediction model still need to be improved.

[0008] In view of this, currently, some units involved in offshore operations such as offshore wind farms / coastal nuclear power plants still rely on local meteorological and ocean bureaus or public weather forecasts to predict wave information such as wave height and wave direction. Regarding the prediction of the local meteorological bureau, the response is not always immediate. And the public weather forecast (from physical simulation models) can only provide hourly 72-hour predictions at a large grid level of 0.5 * 0.5 degrees (50 kilometers by 50 kilometers). This is not precise enough for some on-site technicians, so it is difficult to know the exact wave conditions at the operation location at a specific moment. In addition, since the wave forecast does not have the ability to be updated in real time, once the operation starts, without additional special assistance, it can only rely on human observation and judgment, and it is difficult to detect emergencies / rapid changes.

[0009] Therefore, it is necessary to develop a new type of sea wave prediction method that can effectively integrate and analyze multi-source data, improve the accuracy and real-time performance of prediction, and at the same time can adapt to different geographical regions and ocean environments to better meet the safety and efficiency requirements of ocean activities. Summary of the Invention

[0010] The object of the present invention is to overcome the defects and problems existing in the background technology.

[0011] For this purpose, according to one aspect of the present invention, the present invention proposes a method for predicting ocean wave parameters, including the following steps:

[0012] · Obtain training data from an ocean wave dataset, where the training data includes various ocean wave variables in the spatial domain;

[0013] · Process the training data through a normalization process;

[0014] · Use the processed data to train a machine learning model for ocean wave parameter prediction; and

[0015] · Use the trained machine learning model to generate predictions of various ocean wave parameters in the spatial domain.

[0016] By utilizing machine learning techniques, the present invention can accurately and timely predict the key parameters of ocean waves, thereby improving the safety and efficiency of ocean activities.

[0017] Among them, various ocean wave parameters and parameter predictions at least include parameters such as ocean wave height, ocean wave direction, and ocean wave period. Including but not limited to the following parameters:

[0018] 1) Time: corresponding to the start time

[0019] 2) Apparent wave height (m)

[0020] 3) Main wave direction (°)

[0021] 4) Wave amplitude (°)

[0022] 5) Water depth depth (m)

[0023] 6) Wind direction and wind speed

[0024] 8) Wave period

[0025] 9) Average wave period

[0026] 10) Peak wave period

[0027] Optionally, the ocean wave dataset can be sourced from actual ocean observation data or generated by a physics-based ocean wave simulation model, providing accurate input information for ocean wave prediction.

[0028] Optionally, the machine learning model is a long short-term memory (LSTM) model. The long short-term memory model has excellent performance in dealing with time series data, such as the change of ocean wave height over time.

[0029] Preferably, the LSTM model includes at least four layers, including at least one input layer, at least two hidden layers, and an output layer. This multi-layer structure allows the model to capture the complex change patterns of ocean wave parameters over time.

[0030] Preferably, the range of the hidden layer is selected from 64 to 128, which can effectively handle the complexity of the sea wave data while maintaining the computing efficiency of the model.

[0031] Optionally, the performance of the machine learning sea wave model is evaluated using metrics including R-squared, mean absolute error (MAE), mean squared error (MSE), and root mean squared error (RMSE), and the model is optimized according to these metrics. The model is optimized based on these metrics to achieve higher prediction accuracy.

[0032] Optionally, the method according to the present invention further includes a transfer learning step to adapt the trained machine learning model to different geographical regions, thereby enhancing the generality and application scope of the model.

[0033] Optionally, the method according to the present invention further includes a warning step of sending a warning signal when the predicted sea wave parameters exceed a preset threshold.

[0034] According to another aspect of the present invention, the present invention provides a system for predicting sea wave parameters, including:

[0035] · A data acquisition module configured to obtain training data from a sea wave data set, wherein the training data includes various sea wave variables in the spatial domain;

[0036] · A data processing module configured to normalize the training data;

[0037] · A training module for training a machine learning model for sea wave parameter prediction using the processed data; and

[0038] · A prediction module configured to generate predictions of various sea wave parameters in the spatial domain using the trained machine learning model.

[0039] This system can support efficient and accurate sea wave parameter prediction.

[0040] Optionally, the machine learning model of the training module is a long short-term memory (LSTM) model trained to predict sea wave parameters based on the processed training data, and includes at least four layers, including at least one input layer, at least two hidden layers, and an output layer.

[0041] According to yet another aspect of the present invention, the present invention provides a computer-readable medium containing instructions that, when executed by a computer, cause the computer to perform the steps of the method according to the present invention. This enables the present invention to be easily deployed through computer program updates, providing users with a flexible application method.

[0042] In summary, the present invention provides an efficient and accurate method for predicting wave parameters through an integrated machine learning algorithm combined with the normalization processing of wave data. Compared with the prior art, the present invention has technical advantages such as standardized data processing, optimized model structure, comprehensive performance evaluation, and transfer learning ability. Through these innovative points, the present invention not only improves the accuracy of wave prediction, but also enhances the adaptability and application scope of the model, especially in the application of different geographical regions, bringing significant improvements in safety and efficiency to the fields related to marine activities.

[0043] The features and advantages of other aspects of the present invention will be discussed in the following detailed implementation. Those skilled in the art can clearly understand the content of the present invention and the obtained technical effects based on the following embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] It should be understood that in the present invention, except for obvious contradictions or incompatibilities, all features, deformation methods, and / or specific embodiments can be combined in various combinations.

[0045] By reading the following specific embodiments as non-limiting descriptions and in combination with the drawings, other features and advantages of the present invention will be obvious, in the drawings:

[0046] - Figure 1 is a flowchart of an embodiment according to the method of the present invention;

[0047] - Figure 2 is Figure 1 an example of the specific steps included in step 2 in the shown flowchart;

[0048] - Figure 3 is Figure 1 an example of the specific steps included in step 4 in the shown flowchart. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0049] The following are exemplary embodiments according to the present invention. The relevant definitions below are only used to describe the exemplary embodiments and are not intended to limit the scope of the present invention. Since the embodiments described here are exemplary, they can also be extended to modifications related to the functions, purposes, and / or structures of the present invention.

[0050] Figure 1 An exemplary embodiment according to the method of the present invention is shown, which includes the following steps:

[0051] Step S1: Obtain training data

[0052] First, obtain training data from the wave dataset for subsequent machine learning training of the model. The wave dataset can be actual historical data or a dataset generated based on a physical model. The wave dataset contains parameters such as wave height, wave direction, and wave period, which include time and space dimensions.

[0053] For a dataset generated based on a physical model, for example, the ANEMOC - China dataset can be used. This dataset is calculated by a simulation model. ANEMOC data is a simulated wave dataset, which is verified against buoy and satellite observation data using measurement methods such as normalized bias (NB) and symmetric root mean square error (HH). Under normal and extreme weather conditions, the observation performance of this dataset is good, with the overall normalized bias (NB) being less than 0.04 and the normalized bias of the mean period (MP) variable being less than 0.03.

[0054] As an example, the dataset can contain the following data:

[0055] The data type is datetime for time and float for the rest.

[0056] The data variables include:

[0057] 1) Time: Corresponding to the start time

[0058] 2) Significant wave height (m)

[0059] 3) Main wave direction (°)

[0060] 4) Wave spread (°)

[0061] 5) Depth (m): Water depth

[0062] 6) windx (m / s): Wind speed in the x - direction

[0063] 7) windy (m / s): Wind speed in the y - direction

[0064] 8) TMOY (s): Wave period

[0065] 9) TM01: Mean wave period

[0066] 10) TM02: Mean wave period

[0067] 11) TP: Peak wave period

[0068] Among them, TM01: Average wave period using the 0th and 1st spectral moments = m0 / m1, where

[0069] m0 is the 0th spectral moment, which represents the total energy of the wave energy spectrum. Physically, it is related to the average height of the waves because energy is proportional to the square of the wave height.

[0070] m1 is the 1st spectral moment, which is related to the frequency distribution of wave energy and can be associated with the average period of the wave system.

[0071] For example, TM01: Absolute average wave period (in seconds), defined as

[0072]

[0073] E(ω,θ) is the variance density spectrum, and ω is the absolute radian frequency determined by the dispersion relation of the Doppler shift frequency.

[0074] TM02: Among them

[0075] m0 is the 0th spectral moment, which represents the total energy of the wave energy spectrum. Physically, it is related to the average height of the waves because energy is proportional to the square of the wave height.

[0076] m2 is the 2nd spectral moment, which involves the square of the wave frequency and is related to the shape parameter of the wave, such as the steepness of the wave.

[0077] For example, TM02: Absolute average wave period (in seconds), defined as

[0078]

[0079] E(ω,θ) is the variance density spectrum, and ω is the absolute radian frequency determined by the dispersion relation of the Doppler shift frequency.

[0080] The following is a set of exemplary ANEMOC - China datasets

[0081]

[0082] Step S2: Data preprocessing

[0083] Next, preprocess the obtained training data, especially normalize the training data, which is used to adjust the numerical values in the dataset to have a unified scale or distribution, and is very important for improving the performance and stability of the model. Of course, this step can include Figure 2 the specific steps exemplified in

[0084] Step S2.1: Clean data (or data screening)

[0085] In data preprocessing, first, nearshore and inshore data points in a selected area can be deleted from a dataset containing historical data generated by a physical simulation model. For example, a data integrity threshold is used to exclude missing or corrupted data points. Another example is to screen out duplicate records based on timestamps in wave data.

[0086] Step S2.2: Normalize data

[0087] Next, the cleaned data is normalized using different preprocessing strategies. The present invention can adopt various normalization methods, such as including Min - Max Normalization, Z - Score Normalization (also known as standardization), decimal scaling, logarithmic transformation, etc.

[0088] For example, the input value of wind direction is "m", in units of angle (0 - 360 degrees). To help the model better learn this variable, "m" is decomposed into c = cosine(m) and s = sine(m), where the spatial range of c and s is (-1, 1). "c" and "s" will be used as two variables for predicting the wave direction and input into the subsequent learning model.

[0089] In other words, this is a transformation from polar coordinates (r, φ) to rectangular coordinates (x, y). Among them, r is the radius = 1, φ is the angle = 180 degrees = ∏, and polar coordinates (1, ∏) = rectangular coordinates (-1, 0). Since this processing technology is reversible, an inverse transformation will be applied in post - processing to convert the numerical values back from rectangular coordinates to polar coordinates.

[0090] The variable of wind direction is converted into a format that is more easily processed by the subsequent learning model. This conversion helps to improve the model's understanding and prediction ability of wind direction data.

[0091] For other variables such as wave height and wave period, they can also be standardized during preprocessing, for example, converting the data to have a mean of 0 and a standard deviation of 1.

[0092] Step S2.3: Format the dataset

[0093] After completing the normalization operation on the dataset, optionally, the normalized dataset is formatted, especially formatting the time series of the dataset.

[0094] For example, first, ensure that the data is arranged in chronological order, which usually means sorting the data based on a timestamp or a date column. Next, process the timestamps to ensure their format is consistent and suitable for model processing, such as converting date and time to the standard ISO format. Additionally, extra features can be extracted from the timestamps, such as hour, day of the week, month, etc., and these features are helpful for the model's prediction.

[0095] The data needs to be split into "windows" for training subsequent time series related models. This means defining a time window of a fixed size and grouping consecutive data points within this window.

[0096] Furthermore, if there are time gaps or missing values in the data, a decision needs to be made on how to handle this missing data. Possible methods include interpolation, forward filling, backward filling, or simply removing the missing data.

[0097] Step S2.4: Divide the training and test data sets

[0098] After completing the above data preprocessing steps, the data set needs to be divided into a training data set and a test data set for subsequent machine learning.

[0099] Correctly dividing the training data and the test data is crucial for building an effective model later. In this embodiment, most of the data is allocated as training data, and the rest is used as test data. The specific data volume ratio, such as 80 (training data) / 20 (test data) to 60 (training data) / 40 (test data), depends on the total amount of data and its characteristics. To ensure that both the training and test data can represent the characteristics of the entire data set, random sampling can be used for division. Additionally, for data sets with important categories or categories with a small number of samples, the stratified sampling method can ensure that these categories are reasonably represented in both the training set and the test set. This method is particularly suitable for unbalanced data sets, where the number of instances in some categories may be much less than that in other categories. Of course, when dealing with time series data, the division method can also be based on chronological order to maintain the time series nature of the data. For example, the first 80% of the data set can be selected as training data, while the remaining 20% is used as test data. This can ensure that the test data is temporally after the training data, thus simulating the situation where the model faces future data in the real world.

[0100] Generally speaking, data division should consider the characteristics of the data set and the requirements of the model to ensure that the performance of the model can be evaluated fairly and effectively during the training and test processes.

[0101] For example, in an embodiment, the data used for training can be selected from a specific sea area A in the ANEMOC-China dataset. Spatially, the entire area A is divided into an M×N grid system, and each grid represents a sub-region. Each sub-region consists of a small l×l grid, and the corner points of each small grid are used as training points, while the middle point is used as a test point. Temporally, the data covers a period from a certain start time point to a certain end time point. During this period, a timestamp is assigned to each position point to form, for example, hourly data. Thus, these data contain spatial and temporal dimensions. In other words, these data contain (simulated) wave parameters (such as height, direction, and period) at specific space and time. In addition, when preprocessing the data, in addition to normalization, missing value imputation and outlier handling are also performed to ensure the quality of the data. Finally, these data are divided into training data and test data according to the steps described above.

[0102] Step S3: Train the machine learning model for wave prediction

[0103] After completing the preprocessing step of the training data including normalization, according to the embodiments of the present invention, the next step is to use the processed data to train the machine learning model for wave prediction.

[0104] Figure 3 An exemplary model training embodiment according to the present invention is shown, including the following steps.

[0105] Step S3.1: Establish the model and definition

[0106] Before training, a machine learning model needs to be established. In this process, defining the configuration of the model is a key step, which involves multiple decision points.

[0107] Specifically, first, the type of input variables needs to be selected. It can be single-variable input, that is, the model makes predictions based on only one input feature, which is suitable for simple tasks or when there is only one important predictor; or it can be multi-variable input, that is, the model uses multiple input features for prediction, which is suitable for more complex tasks and can capture the relationships between features. In the present invention, multiple wave variables in the spatial domain (at least including wave direction, height, and period) are adopted.

[0108] Next is the selection of the output. It can be single-position output, which makes predictions for a specific position; or it can be multi-position output, enabling the model to simultaneously predict the results of multiple positions, which is suitable for scenarios where multiple regions need to be monitored.

[0109] The determination of the step size is also an important configuration, which affects the fineness of the model's processing of spatial data. In the present invention, the range can be between 50 kilometers and 5 kilometers.

[0110] In addition, the selection of the training scope is another key decision point. In the present invention, local training or global training can be selected. Local training means that the model is trained only using data from a specific area, which may provide higher accuracy in that area; while global training uses data from multiple areas and is suitable for models that need to be generalized to different areas.

[0111] Finally, it is also necessary to determine the maximum number of epochs for model training, which is an important hyperparameter that affects the length of the training process.

[0112] In the present invention, these settings can be flexibly adjusted according to specific data characteristics, requirements, and expected results. For example, single-variable input plus local training may be suitable for high-precision requirements in a specific area, while multi-output settings are more suitable for scenarios that widely monitor multiple locations. When defining the model configuration, it is necessary to comprehensively consider the characteristics of the data, the nature of the prediction task, and the desired accuracy and generalization ability.

[0113] Step S3.2: Train the model

[0114] After the model is established and configured, the machine learning model for wave prediction is trained.

[0115] For example, in the machine learning process, the model iteratively traverses all position points in the training area A and uses the hourly data corresponding to these points to update its parameters to gradually learn and grasp the patterns and characteristics of the data. Specifically, if it is desired to predict information such as the wave height and direction at a certain position at a specific time point, then a training set containing a series of timestamp data before and after that time point is first selected. This training set includes not only the data of the target position point during this period but also the data of several other key position points. At each timestamp, the data of these key position points are provided as inputs to the model, while the actual observed value of the target position point is used as the true value of the output.

[0116] In addition, there are various models for processing sequential data that can be selected for training. In this example, a long short-term memory (LSTM) network model is selected.

[0117] The long short-term memory (LSTM) model is a special type of recurrent neural network (RNN) specifically designed to handle the long-term dependence problem of sequential data. In a standard recurrent neural network, the connections between the hidden layers of the network form a circular structure, enabling the network to transfer information from one time step to the next. However, standard RNNs encounter problems of vanishing gradients or exploding gradients when dealing with long sequences, making it difficult for them to capture long-term dependencies.

[0118] The LSTM model solves this problem by introducing a complex gating mechanism. These gating units include an input gate, a forget gate, and an output gate, which control the storage, update, and extraction of information. The forget gate determines which information should be discarded from the unit's state, the input gate controls the entry of new information, and the output gate determines the information that should be included in the next hidden state.

[0119] The inventors have found that these properties of the LSTM model make it very suitable for performing tasks such as wave prediction that require considering long-term temporal dependencies in time series data.

[0120] Specifically, this embodiment uses an LSTM network to perform spatio-temporal modeling and prediction of a large number of variables in order to predict meteorological and oceanographic phenomena in a specific area.

[0121] The LSTM network of this embodiment is composed of, for example, four hidden layers, each layer containing 64 / 128 neurons, and is connected to a fully connected output layer. The output size is determined according to the grid size in area A, usually the total number of grid points minus the number of training points at the four corners. Of course, in actual training, more hidden layers and neurons can also be used, which also depends on factors such as training accuracy and computing power.

[0122] The training process uses a time series segmentation method to ensure the temporal continuity of the training set and the test set and prevent future data from leaking into the training phase. In each training cycle, the model tests the data by traversing the corner points of each small grid and the corresponding intermediate points.

[0123] The model uses the mean squared error (MSE) as the loss function during training. This function measures the difference between the model output and the actual observed values and is optimized by accumulating the losses of all training and test position pairs in area A in each training cycle. To minimize this loss, stochastic gradient descent (SGD) is used as the optimizer. At the same time, momentum and learning rate decay strategies are applied to accelerate the training process and improve the convergence speed.

[0124] In addition, to improve training efficiency and prevent the model from overfitting the training data, an early stopping mechanism can be set. The working principle of this mechanism is that if the loss on the validation set does not improve in consecutive training cycles (e.g., 20 cycles), then the training will be terminated early. When the loss of the model stabilizes within 20 cycles, it can be considered that the model has converged at the current stage, so it is reasonable to stop training.

[0125] Of course, in addition to the mean squared error (MSE) and stochastic gradient descent (SGD) mentioned above, a variety of other loss functions, optimizers, and overfitting prevention strategies can also be selected. For example, for the loss function, cross-entropy loss, Huber loss, logarithmic loss, Hinge loss, etc. can be adopted; for the optimizer, Adam optimizer, RMSprop optimizer, Adagrad optimizer, Adadelta optimizer, etc. can be used; for the overfitting strategy, data augmentation, Dropout, etc. can be employed.

[0126] Ultimately, throughout the entire learning process, the goal of the model is to understand the complex patterns of various wave parameter changes within the entire region A, so as to accurately predict the wave height at future time points. Through this method, the model can be effectively applied to the wave height prediction task for the entire region.

[0127] Step S3.2: Result post-processing

[0128] Next, post-process the obtained results to restore the variable values in the original data space. This is the reverse step of the data preprocessing in Step S2. Taking the same preprocessing as an example, the wind direction variable "m" (0 - 360 degrees) has been decomposed into "s" and "c" (ranging from [-1, 1]) as the input of the model. Therefore, correspondingly, the model will predict "s'" and "c'" as its output. To convert "s'" and "m'" into the metric value of "m'" (0 - 360 degrees), post-processing using arccos(c') and arcsin(s') is required to find the corresponding "m'". In order to convert the results of the model into meaningful outputs in practical applications, such as the height, direction, and period of ocean waves at a specific location, etc.

[0129] Step S3.4: Model validation

[0130] Then, the model is validated, including the following two aspects:

[0131] Step S3.4.1: Find the best model configuration

[0132] Use methods such as R-squared, MAE, MSE, or RMSE to evaluate the results of the test dataset. If the results are not satisfactory, return to Step S3.2 and run the model with other configuration settings until the best results are found, and save the best model for each target variable.

[0133] Step S3.4.2: Validate the performance and stability of the model under various real-world scenarios

[0134] Evaluate the performance of the method model using methods such as MSE, RMSE, R - squared or MAE. For example, test in three typical time periods: normal days, typhoon days and winter rainstorm days, and test in two different regions to verify the ability of the model against the results of the simulation model.

[0135] Although these two verification steps focus on different aspects, they are both part of the model verification process, aiming to ensure that the final model can perform well on a specific dataset and maintain its performance and accuracy in diverse and challenging real - world situations.

[0136] Step S4: Predict wave parameters

[0137] Once the wave prediction model passes the verification phase, indicating that its performance on the given dataset meets expectations, the next step is to deploy the model into a practical application environment. This may involve integrating the model into wave business processes, systems or applications.

[0138] After the model is deployed, the model can start to be used for actual prediction tasks. This may be real - time prediction of new data (daily or hourly rolling prediction), or batch processing of a batch of data. According to an embodiment of the present invention, during the prediction process, the model outputs multiple wave prediction parameters to predict the wave height, direction and period of a sea area at a specific location and time.

[0139] In addition, after the model starts to run in the actual environment, it is important to continuously monitor its performance. This includes tracking accuracy, response time and other key metrics. If the performance of the model starts to decline, adjustments or retraining may be required. Since machine learning is an iterative process. Depending on the performance of the model in the production environment, it may be necessary to periodically re - evaluate and adjust the model, including retraining with new data.

[0140] Of course, the prediction results can be presented in multiple ways, either visualized as an interactive line chart for a specific location, showing hourly trends and values, or as a heat map, showing the predicted variable values updated hourly for a selected area.

[0141] The multiple wave parameters obtained from the prediction can be used for a variety of applications. Figure 1 An exemplary application for issuing dangerous wave warnings is given. Among them, when the parameters predicted by the model, such as wave height and / or wave frequency, reach a certain predetermined threshold, the warning system will be triggered, for example, an alarm will be sent to the user, and the results will be sent to maritime personnel and offices. The following is a detailed description of this process.

[0142] The warning trigger mechanism is based on the condition that the model prediction result exceeds the normal range. For example, if the wave height predicted by the model exceeds the 95% confidence interval calculated based on historical data, the system will issue a warning. This means that the warning will only be triggered when the wave height is very likely to exceed the normal variation range. This strategy aims to reduce false alarms and ensure that alerts are only issued when there is a real potential danger.

[0143] Among them, the setting of the threshold is based on in-depth analysis of historical data and consideration of safety standards. Statistical methods can be used, such as calculating the distribution of historical wave height data and determining the 95% confidence interval accordingly. At the same time, the opinions of marine safety experts can also be referred to for qualitative analysis to ensure that the threshold is both scientific and meets the actual safety requirements. This comprehensive method helps to ensure that the warning system is both reliable and practical.

[0144] Once the warning is triggered, the system will immediately notify relevant personnel or systems via email, text message, or other instant messaging tools. The notification logic includes classification of warning priorities at different levels. For example, an emergency warning, such as a rapid increase in wave height to a level that may cause serious impacts, will directly notify the management and the maritime emergency response team so that they can take prompt actions. While a general warning, such as a slight increase in wave height but not reaching the dangerous level, is recorded in the system log for subsequent analysis. Such classification not only ensures that emergencies can be quickly responded to but also avoids unnecessary interference with daily operations.

[0145] In this way, the wave height warning system can timely and accurately issue alerts to relevant personnel, improving the response speed and efficiency to potential marine hazards, and at the same time ensuring the safety of marine activities.

[0146] According to another embodiment of the present invention, the present invention can be implemented through a software application, combining the powerful computing capabilities of the cloud and the convenience of mobile applications to provide effective wave prediction services.

[0147] Specifically, the learning and computing parts of the model are deployed in the cloud, making full use of cloud computing resources to process and store a large amount of marine data. This cloud deployment method ensures that the model can be updated in real time and always reflects the latest data and trends.

[0148] On the user side, a mobile application can be developed to enable users to easily input the location and time information they want to query. This application is designed to be intuitive and easy to use, and users can send prediction requests to the cloud with just simple operations. After the cloud server receives the request, it will call the trained model to process these queries and quickly predict wave parameters such as wave height and direction.

[0149] After the prediction is completed, the results will be quickly transmitted back from the cloud to the mobile application, where users can directly view them. To provide more flexibility in usage, the prediction results can also be downloaded to local devices for users to conduct further analysis or save. This function is particularly suitable for users who need to long-term track and analyze the wave changes at specific locations.

[0150] In addition, to handle emergencies, an automatic warning notification function is integrated into the system. When the wave height prediction exceeds the safety threshold, the system will immediately notify the user to ensure a quick response at critical moments. This timely feedback mechanism improves the practicality and efficiency of early warnings.

[0151] In summary, the present invention proposes a new wave prediction model aimed at providing higher accuracy and practicality than existing methods. The structural design of this model is simple and lightweight, enabling it to be easily embedded into applications and run locally on mobile devices such as mobile phones. The model in the example consists of four layers of long short-term memory networks (LSTMs), and the hyperparameters can be customized according to different requirements and hardware conditions. The number of neurons in the hidden layer is set between 64 and 128, and a suitable optimizer is selected from, such as Stochastic Gradient Descent (SGD), Adam, and RMSProp. The learning rate is set between 0 and 0.1, and a learning rate scheduler is introduced to improve the learning efficiency of the model. At the same time, multiple loss functions are provided for selection, including Mean Squared Error (MSE), Mean Absolute Error (MAE), Root Mean Square Error (RMSE), or Mean Absolute Percentage Error (MAPE).

[0152] In the initial training stage of the model, historical or physics-based wave parameters are used to enable the model to learn seasonal variations and wave characteristics. Obviously, the more data, the better the performance of the model. It is worth mentioning that the model supports transfer learning, which means that the knowledge learned in one research area can be adapted to the characteristics of another area through fine-tuning. Such a design enables the model to require less training data when continuously updated and also improves the efficiency.

[0153] In addition, the present invention also proposes a unique spatio-temporal training method and a data-driven design, which not only enhances the adaptability of the model to a wider area but also maintains good prediction accuracy. In particular, the present invention performs data preprocessing of reversible angular transformation on periodic data variables, such as the average wave period, significantly improving the prediction accuracy. It is also proposed to use a rich dataset with a long time span and a wide regional scope, namely the "ANEMOC-China" database, which contains data under different water depths and various weather conditions and can predict multiple wave parameters, such as significant wave height, wave direction, and period.

[0154] Finally, to evaluate the performance of the model, the present invention was verified under three typical weather conditions: normal days, typhoon days, and winter storm days. These test results show that the method according to the present invention can maintain a high level of prediction accuracy under different conditions, thus verifying its practicability and reliability.

[0155] Generally speaking, the present invention combines advanced machine learning techniques with modern software engineering to provide a comprehensive, flexible, and user-friendly ocean wave height prediction solution. Whether for maritime workers who need immediate wave height information or scientists who conduct long-term research on ocean data, the present invention can provide strong support.

[0156] Those skilled in the art have mastered multiple embodiments, various deformations, and improvements. In particular, it should be clear that, except for obvious contradictions or incompatibilities, the features, deformation methods, and / or specific embodiments described in the present invention can be combined with each other. All these embodiments, deformations, and improvements fall within the protection scope of the present invention.

Claims

1. A method for predicting wave parameters, comprising the following steps: · Training data is obtained from a sea wave dataset, where the training data includes various sea wave variables in the spatial domain; · Process the training data through a normalization process; · Train a machine learning model for predicting ocean wave parameters using the processed data; and · Generate predictions of various ocean wave parameters in the spatial domain using a trained machine learning model.

2. The method according to claim 1, wherein The wave dataset is generated by a physics-based wave simulation model.

3. The method according to claim 1, wherein The machine learning model is a long short-term memory (LSTM) model.

4. The method according to claim 3, wherein the long short-term memory model comprises at least four layers, including at least one input layer, at least two hidden layers, and an output layer.

5. The method according to claim 4, wherein The selection range of the hidden layer is from 64 to 128.

6. The method according to claim 1, wherein The performance of the machine learning wave model is evaluated using metrics including R-squared, mean absolute error (MAE), mean squared error (MSE), and root mean squared error (RMSE), and the model is optimized according to these metrics.

7. The method according to claim 1, characterized in that It further comprises a transfer learning step to adapt the trained machine learning model to different geographical regions.

8. A system for predicting wave parameters, comprising: · A data acquisition module, configured to obtain training data from a data set, wherein the training data includes multiple ocean wave variables in the spatial domain; · A data processing module, configured to normalize the training data; · A training module configured to train a machine learning model for predicting ocean wave parameters using the processed data; and · A prediction module configured to generate predictions of multiple ocean wave parameters in the spatial domain using a trained machine learning model.

9. The system according to claim 8, wherein The machine learning model of the training module is a long short-term memory (LSTM) model trained to predict wave parameters based on the processed training data, and comprises at least four layers, including at least one input layer, at least two hidden layers, and an output layer.

10. A computer-readable medium containing instructions that, when executed by a computer, cause the computer to perform the steps of the method according to claim 1.

Images