A Method for Predicting Irregular Ocean Wave Height Based on Numerical Pool and Deep Learning Model

Abstract

This invention discloses a method for predicting the wave height of irregular ocean waves based on numerical water tank simulation and a deep learning model. The method includes using a numerical water tank simulation combined with extreme weather conditions to generate a relatively complete dataset of irregular ocean waves; and constructing a deep learning model, TA-DCLSTM, for end-to-end training and prediction, effectively improving the prediction accuracy of wave height. This invention employs numerical water tank simulation to accurately reproduce wave dynamics, while incorporating extreme weather conditions to generate an irregular wave dataset. After preprocessing, the dataset is input into a deep learning model, proposing a time-series prediction model, TA-DCLSTM, based on a time attention mechanism and dilated convolution. Dilated convolution expands the receptive field, capturing multi-scale temporal features in the wave data, while the time attention mechanism assigns different weights to each time step, focusing on key features to achieve accurate prediction of the wave height of irregular ocean waves.

Description

Technical Field

[0001] This invention belongs to the fields of oceanography and deep learning technology, specifically relating to a method for predicting ocean wave height based on numerical water tanks and deep learning models. Background Technology

[0002] With the deepening of marine scientific research, wave height prediction has gradually become a research focus in the fields of marine exploration and offshore operations, possessing significant safety value and scientific significance. Wave data is characterized by high complexity, strong time-series dependence, and non-stationarity. Complexity stems from the interaction of various physical factors such as wind speed, wind direction, and tides; time-series dependence is reflected in the autocorrelation characteristics of wave data, where current wave height is significantly influenced by past ocean conditions; and non-stationarity causes the statistical properties of wave data to change over time. These characteristics of wave data limit the predictive models' ability to cope with complex marine environments, making it difficult to meet the requirements of high-precision prediction.

[0003] In existing technologies, such as the deep learning-based nearshore single-point wave height forecasting method disclosed in announcement number CN113283588A, the following steps are included: specifying a regional marine hydrological and meteorological dataset; cleaning and processing the data to obtain effective data and dividing it into training and testing sets; training the model to obtain a trained wave height prediction model; verifying the accuracy of the wave height prediction model and predicting the wave height results. Using a self-developed deep learning model to predict single-point wave height improves the accuracy of wave height prediction, and the trained model can be directly used in future nearshore single-point wave height prediction applications. In recent years, machine learning methods have shown significant advantages in wave height prediction tasks. To address conventional prediction tasks, by introducing techniques such as recurrent neural networks (RNNs), long short-term memory networks (LSTMs), and gated recurrent units (GRUs), the model can automatically learn features from wave height data, effectively removing noise and capturing the autocorrelation and long-term trend changes of time series. Under sufficient training data, such models have strong generalization ability and can handle wave height prediction tasks in various complex scenarios. However, in order to accurately predict the wave height in complex environments, it is crucial to reduce the strong non-stationarity and complex physical interaction processes in wave data. Conventional time series prediction models have low accuracy in predicting the wave height of irregular waves. Therefore, this application proposes a method for predicting the wave height of irregular waves based on numerical pools and deep learning models to improve the accuracy of wave height prediction. Summary of the Invention

[0004] The purpose of this invention is to provide a method for predicting the wave height of irregular ocean waves based on a numerical water tank and a deep learning model in order to solve at least one of the above-mentioned technical problems.

[0005] This invention achieves the above objective through the following technical solution: a method for predicting the wave height of irregular ocean waves based on a numerical water tank and a deep learning model, comprising the following steps:

[0006] S1. The numerical water tank method is used to simulate the pusher wave tank. Combined with extreme weather conditions, the corresponding wave equation is calculated to output the pusher motion parameters. By adjusting the pusher motion parameters, the precise control and generation of irregular waves can be achieved. The generated irregular wave data is used for deep learning model training after preprocessing.

[0007] S2. Construct a deep learning model TA-DCLSTM. The TA-DCLSTM deep learning model combines a time attention mechanism with dilated convolution. The deep learning model extracts long-term dependency features of ocean wave data through the LSTM module. The deep learning model extracts local features at different time scales through dilated convolution. The deep learning model uses the attention mechanism to assign different weights to each time step, focusing on the key moments that have a greater impact on the prediction results. Accurate prediction of irregular ocean wave height is achieved through end-to-end training.

[0008] As a further aspect of the present invention: In S1, the generation process of irregular ocean waves is verified and controlled by simulating a pusher wave-generating tank using a numerical water tank method; wherein, the vertically arranged pusher wave-generating tank is composed of a closed elongated tank containing liquid, and perforated wave-damping blocks are symmetrically arranged at both ends of the tank, with both sides of the wave-damping blocks designed as slopes; a groove is provided at the bottom of the tank, and a vertical pusher plate is horizontally installed in the groove, with the pusher plate slidingly connected to the side wall of the groove, and the length of the pusher plate being consistent with the width of the tank, and the pusher plate is driven to generate waves in the tank through reciprocating up and down motion.

[0009] As a further aspect of the present invention: the wave equation is calculated based on the target wave parameters, and corresponding vertical pusher motion data is generated; specifically including:

[0010] Let the thickness of the pusher plate be Z. After time dt, the distance the pusher plate moves vertically is dx. The volume change of the groove should be equal to the volume change of the liquid at the bottom of the tank, dV. That is:

[0011] dV=Zdx

[0012] Let the velocity of the vertical push plate be v(t), then we have:

[0013] dV = Zdx = Zvdt

[0014] Assuming that the fluid volume change caused by the vertical pusher is entirely converted into water surface ripples without loss, we have:

[0015] Zvdt=2Cη(t)dt

[0016] Therefore, the velocity relationship between the vertical pusher and the wave surface can be derived as follows:

[0017] v(t)=2Cη(t)ξ / Z

[0018] Where C is the wave propagation speed, η is the free surface waveform, and ξ is the wave height amplification factor;

[0019] Substituting different waveform functions into the formula yields the velocity equation for the vertical pusher plate; substituting the preset pusher plate size and target wave parameters into the corresponding pusher plate velocity equation calculates the motion parameters of the pusher plate, and the expected wave is generated by controlling the size and motion of the pusher plate.

[0020] As a further aspect of the present invention: during the wave-generating process, the wave data generated by the pusher plate wave-generating tank is compared with the target wave; if the error between the actual wave information and the target wave exceeds 5%, it is corrected by adjusting the wave height amplification factor ξ, and the error of the wave period is corrected by the target wave period until the wave parameters meet the requirements; finally, the dimensions and motion data of the vertical pusher plate are saved.

[0021] When a vertical pusher is used to generate a regular wave, the waveform function of the regular wave is:

[0022]

[0023] Where S is the wave height and T is the wave period, substituting the regular wave expression into the equation for the vertical pusher motion of the regular wave yields the following equation:

[0024]

[0025] When generating irregular waves using the vertical pusher wave generation method, it is necessary to set the target wave spectrum, then decompose the irregular wave into multiple different regular waves in the frequency domain, and sum them to obtain the motion equation of the vertical pusher:

[0026]

[0027] Among them, S i Represents wave height, T i For the wave period, ε i It is the random initial phase of the wave component, ξ. i This is the wave height amplification factor;

[0028] In reality, airflow over the sea surface affects wave generation, and changes in air velocity should be taken into account. This is based on the Monin-Obukhov similarity theory for describing the ocean surface:

[0029]

[0030] Where u(z) is the average wind speed at sea level z, u0 is the sea surface current velocity, and u * Let Ψ be the friction velocity, k be the Karman constant, z0 represent the surface roughness length, and Ψ be the friction velocity. m Describes the stability function. This represents the dimensionless stability parameter, where L is the Monin-Obukhov length;

[0031] The stability function characterizes the sea surface temperature T sea and air temperature T air The relationship when T sea >T air When this occurs, it is considered an unstable state, and there is When T air >T sea Sometimes, Dimensionless stability parameters Similarly, stable and unstable states also exist, which can be represented by the Bulk Richardson number (Ri). b Further description: Its expression is as follows:

[0032]

[0033] In the formula, △T=T air -T sea g is the acceleration due to gravity, z represents the air height, and U 10 The wind speed at a height of 10m, when T sea >T air At that time, there are dimensional stability parameters. When T air >T sea Sometimes,

[0034] As a further aspect of the present invention: In S2, the constructed deep learning model TA-DCLSTM specifically includes:

[0035] The irregular wave data generated by S1 is normalized to standardize the wave height values ​​to the range of [0,1]. The normalized dataset is divided into two parts in a 7:3 ratio. 90% of the data is used as the training set to allow the model to fully learn the characteristics of irregular waves, and the remaining 10% is used as the validation set to evaluate the model performance.

[0036] As a further step in this invention: the deep learning model TA-DCLSTM performs one-dimensional dilated convolution processing on the input irregular ocean wave data. By introducing a gap between the convolution kernels, the receptive field is expanded to cover a larger time range without increasing the number of convolution kernel parameters. Assuming the convolution kernel size is k, the dilation rate is d, and the input data is x, the output features after dilated convolution processing can be expressed as:

[0037]

[0038] As a further aspect of the present invention: when the deep learning model TA-DCLSTM processes the extracted features through the LSTM module, it specifically includes:

[0039] The LSTM module learns and maintains the long-term dependencies of ocean wave data through its internal memory cells, capturing the trends and periodic characteristics of ocean waves over time. Because ocean wave data has strong time dependencies, the LSTM module can effectively control the flow of information through gating mechanisms such as input gates, forget gates, and output gates, avoiding gradient vanishing or exploding phenomena in long-term sequences. The forget gate processing is as follows:

[0040] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0041] Among them, h t-1 It is the hidden state of the previous time step, x t This is the current input, W f and b f These are the weights and configuration options for the forget gate;

[0042] The input gate processing procedure is as follows:

[0043] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0044] Furthermore, the update process for memory units is as follows:

[0045]

[0046] Memory unit C t It preserves the dependencies in the long-term sequence, combining the memory of the previous moment with the updated information of the current input;

[0047] The output gate processing procedure is as follows:

[0048] ot =σ(W o ·[h t-1 ,x t ]+b o )

[0049] How does the output gate control the hidden state at the current moment affect the model's output?

[0050] The final output after LSTM processing is:

[0051] h t =o t ·tanh(C t )

[0052] Hidden state h t It is the output of the current time step, which combines long-term dependency information and features of the current input. Through this mechanism, LSTM can effectively capture the long-term trend and periodic change features of irregular wave data, and make full use of the influence of historical data to generate more accurate future prediction results.

[0053] As a further aspect of this invention: Deep learning models using a time attention mechanism for data processing enable the models to pay more attention to key time steps when processing time-series data, specifically including:

[0054] By dynamically assigning different weights to each time step, the model can focus on moments with greater impact on the final prediction result in complex time series data, especially when facing irregular waves and extreme wave characteristics, effectively improving the model's accuracy. The time attention mechanism first assigns different weights to each time step through a weight matrix. Based on the input features, the weights are dynamically adjusted, and the model finally calculates a weighted sum based on these weights, enabling the model to more flexibly capture complex temporal features. Let the input time series be X = [x1, x2, ..., x...]. T ], where T is the number of time steps, x t This represents the input feature at time step t. Temporal attention first calculates the attention score for each time step:

[0055] score t =tanh(W·x) t +b)

[0056] Where W is a trainable weight matrix and b is a bias term;

[0057] Then, the attention scores are converted into weights using the Softmax function:

[0058]

[0059] These weights are then applied to the original input features, i.e.:

[0060]

[0061] in It is the weighted output sequence, that is, the weighted feature representation; in this way, the model can dynamically focus on key time steps according to different input features, thereby effectively improving prediction accuracy.

[0062] The beneficial effects of this invention are as follows: Based on wave data obtained from actual engineering projects, this invention addresses the problems of poor controllability and repeatability, low completeness and coverage, and high acquisition costs by employing numerical water tank technology to simulate waves, accurately reproducing wave dynamics. Simultaneously, extreme weather conditions are introduced to generate an irregular wave dataset, which is then preprocessed and input into a deep learning model. Secondly, to solve the problems of insufficient long-term dependency capture and low sensitivity to local short-wave fluctuations exhibited by traditional LSTM models in predicting complex wave heights, a time series prediction model based on temporal attention and dilated convolution, TA-DCLSTM, is proposed. Dilated convolution expands the receptive field, capturing multi-scale temporal features in the wave data, while the temporal attention mechanism assigns different weights to each time step, focusing on key features to achieve accurate prediction of irregular wave heights. Attached Figure Description

[0063] Figure 1 This is a schematic diagram of the overall process of predicting irregular ocean wave height using a numerical water tank and a deep learning model in this invention.

[0064] Figure 2 This is a schematic diagram of the process of generating irregular ocean wave data using numerical water tank and pusher plate wave generation technology according to the present invention.

[0065] Figure 3 This is a schematic diagram of the deep learning model for predicting irregular ocean wave heights according to the present invention.

[0066] Figure 4 This is an example diagram of irregular ocean wave data from the present invention;

[0067] Figure 5 This is an example diagram showing the prediction results of irregular ocean wave height according to the present invention. Detailed Implementation

[0068] 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.

[0069] Example 1

[0070] like Figures 1 to 5 As shown, a method for predicting the wave height of irregular ocean waves based on a numerical water tank and a deep learning model includes the following steps:

[0071] S1. The numerical water tank method is used to simulate the pusher wave tank. Combined with extreme weather conditions, the corresponding wave equation is calculated to output the pusher motion parameters. By adjusting the pusher motion parameters, the precise control and generation of irregular waves can be achieved. The generated irregular wave data is used for deep learning model training after preprocessing.

[0072] S2. Construct the deep learning model TA-DCLSTM. The TA-DCLSTM deep learning model combines a temporal attention mechanism with dilated convolution. The deep learning model extracts long-term dependency features of wave data through the LSTM module to handle long-term dependencies and periodic changes in wave data. The deep learning model extracts local features at different time scales through dilated convolution and captures short-term fluctuations in the sequence by expanding the receptive field. The deep learning model uses the attention mechanism to assign different weights to each time step, focusing on the key moments that have a greater impact on the prediction results. Through end-to-end training, it achieves accurate prediction of irregular wave heights, focusing on the key time steps that have a greater impact on the prediction results, thereby optimizing the weight allocation and significantly improving the prediction accuracy.

[0073] Example 2

[0074] In addition to all the technical features included in Embodiment 1, this embodiment also includes:

[0075] The generation process of irregular ocean waves was verified and controlled by simulating a pusher wave tank using a numerical water tank method. The vertically arranged pusher wave tank consists of a closed, elongated tank filled with liquid. Porous wave-damping blocks are symmetrically arranged at both ends of the tank, and both sides of the wave-damping blocks are designed with slopes. There is a groove at the bottom of the tank, and a vertical pusher plate is horizontally installed in the groove. The pusher plate is slidably connected to the side wall of the groove. The length of the pusher plate is the same as the width of the tank. The pusher plate generates waves in the tank by reciprocating up and down.

[0076] Furthermore, the wave equation is calculated based on the target wave parameters, and corresponding vertical pusher motion data is generated; specifically including:

[0077] Let the thickness of the pusher plate be Z. After time dt, the distance the pusher plate moves vertically is dx. The volume change of the groove should be equal to the volume change of the liquid at the bottom of the tank, dV. That is:

[0078] dV=Zdx

[0079] Let the velocity of the vertical push plate be v(t), then we have:

[0080] dV = Zdx = Zvdt

[0081] Assuming that the fluid volume change caused by the vertical pusher is converted into water surface ripples without loss, we have:

[0082] Zvdt=2Cη(t)dt

[0083] Therefore, the velocity relationship between the vertical pusher and the wave surface can be derived as follows:

[0084] v(t)=2Cη(t)ξ / Z

[0085] Where C is the wave propagation speed, η is the free surface waveform, and ξ is the wave height amplification factor;

[0086] Substituting different waveform functions into the formula yields the velocity equation for the vertical pusher plate; substituting the preset pusher plate size and target wave parameters into the corresponding pusher plate velocity equation calculates the motion parameters of the pusher plate, and the expected wave is generated by controlling the size and motion of the pusher plate.

[0087] Furthermore, during the wave-generating process, the pusher plate wave-generating tank compares the generated wave data with the target wave. If the error between the actual wave information and the target wave exceeds 5%, it is corrected by adjusting the wave height amplification factor ξ. The error in the wave period is corrected by the target wave period until the wave parameters meet the requirements. Finally, the dimensions and motion data of the vertical pusher plate are saved.

[0088] The regular waveform function generated by the vertical pusher is:

[0089]

[0090] Where S is the wave height and T is the wave period, substituting the regular wave expression into the equation for the vertical pusher motion of the regular wave yields the following equation:

[0091]

[0092] When generating irregular waves using the vertical pusher wave generation method, it is necessary to set the target wave spectrum, then decompose the irregular wave into multiple different regular waves in the frequency domain, and sum them to obtain the motion equation of the vertical pusher:

[0093]

[0094] Among them, S i T represents the wave height of a certain component of a wave. i For its period, ε i It is the random initial phase of the wave component, ξ. i This is the wave height amplification factor;

[0095] In reality, airflow over the sea surface affects wave generation, and changes in air velocity should be taken into account. This is based on the Monin-Obukhov similarity theory for describing the ocean surface:

[0096]

[0097] Where u(z) is the average wind speed at sea level z, u0 is the sea surface current velocity, and u * Let Ψ be the friction velocity, k be the Karman constant, z0 represent the surface roughness length, and Ψ be the friction velocity. m Describes the stability function. This represents the dimensionless stability parameter, where L is the Monin-Obukhov length;

[0098] The stability function characterizes the sea surface temperature T sea and air temperature T air The relationship when T sea >T air When this occurs, it is considered an unstable state, and there is When T air >T sea Sometimes, Dimensionless stability parameters Similarly, stable and unstable states also exist, which can be represented by the Bulk Richardson number (Ri). b Further description: Its expression is as follows:

[0099]

[0100] In the formula, △T=T air -T sea g is the acceleration due to gravity, z represents the air height, and U 10 The wind speed at a height of 10m, when T sea >T air At that time, there are dimensional stability parameters. When T air >T sea Sometimes,

[0101] Example 3

[0102] In addition to all the technical features included in Embodiment 1, this embodiment also includes:

[0103] The constructed deep learning model TA-DCLSTM specifically includes:

[0104] The irregular wave data generated by S1 is normalized to standardize the wave height values ​​to the range of [0,1]. Normalization can prevent the model from being affected by excessively large or small wave height feature values ​​during training, thereby improving the training effect and convergence speed, and facilitating model training and analysis. The normalized dataset is divided into two parts in a 7:3 ratio, with 90% of the data used as the training set to allow the model to fully learn the characteristic patterns of irregular waves, and the remaining 10% used as the validation set to evaluate the model performance.

[0105] Furthermore, the deep learning model TA-DCLSTM performs one-dimensional dilated convolution processing on the input irregular ocean wave data. By introducing a gap between the convolution kernels, the receptive field is expanded to cover a larger time range without increasing the number of kernel parameters, effectively capturing multi-scale features in irregular ocean waves and providing richer temporal information for subsequent processing. Assuming the kernel size is k, the dilation rate is d, and the input data is x, the output features after dilated convolution processing can be expressed as:

[0106]

[0107] Furthermore, when the deep learning model TA-DCLSTM processes the extracted features through the LSTM module, it specifically includes:

[0108] The LSTM module learns and maintains the long-term dependencies of ocean wave data through its internal memory cells, capturing the trends and periodic characteristics of ocean waves over time. Because ocean wave data has strong time dependencies, the LSTM module can effectively control the flow of information through gating mechanisms such as input gates, forget gates, and output gates, avoiding gradient vanishing or exploding phenomena in long-term sequences. The forget gate processing is as follows:

[0109] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0110] Where h t-1 It is the hidden state of the previous time step, x t This is the current input, W f and b f These are the weights and configuration options for the forget gate;

[0111] The input gate processing procedure is as follows:

[0112] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0113] Furthermore, the update process for memory units is as follows:

[0114]

[0115] Memory unit C t It preserves the dependencies in the long-term sequence, combining the memory of the previous moment with the updated information of the current input;

[0116] The output gate processing procedure is as follows:

[0117] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0118] How does the output gate control the hidden state at the current moment affect the model's output?

[0119] The final output after LSTM processing is:

[0120] h t =o t ·tanh(C t )

[0121] Hidden state h t It is the output of the current time step, which combines long-term dependency information and features of the current input. Through this mechanism, LSTM can effectively capture the long-term trend and periodic change features of irregular wave data, and make full use of the influence of historical data to generate more accurate future prediction results.

[0122] Furthermore, deep learning models, through time attention mechanisms, can focus more on key time steps when processing time-series data. Specifically, this includes:

[0123] By dynamically assigning different weights to each time step, the model can focus on moments with greater impact on the final prediction result in complex time series data, especially when facing irregular waves and extreme wave characteristics, effectively improving the model's accuracy. The time attention mechanism first assigns different weights to each time step through a weight matrix. Based on the input features, the weights are dynamically adjusted, and the model finally calculates a weighted sum based on these weights, enabling the model to more flexibly capture complex temporal features. Let the input time series be X = [x1, x2, ..., x...]. T ], where T is the number of time steps, x t This represents the input feature at time step t. Temporal attention first calculates the attention score for each time step:

[0124] scoret =tanh(W·x) t +b)

[0125] Where W is the trainable weight matrix and b is the bias term;

[0126] Then, the attention scores are converted into weights using the Softmax function:

[0127]

[0128] These weights are then applied to the original input features, i.e.:

[0129]

[0130] in It is the weighted output sequence, that is, the weighted feature representation; in this way, the model can dynamically focus on key time steps according to different input features, thereby effectively improving prediction accuracy.

[0131] like Figure 4 and Figure 5 As shown, an example of wave height prediction based on numerical water tank and deep learning model is as follows:

[0132] This paper presents irregular ocean wave data generated using numerical pooling and pusher wave generation techniques. As shown in the figure, irregular ocean wave data exhibits significant randomness and complexity, with parameters such as wave height, wave period, and wave direction typically displaying high variability and irregularity. It reflects the random characteristics of natural ocean waves, containing multiple wave components of different frequencies and amplitudes; therefore, its time series data often exhibits non-periodicity and long-term dependence. Furthermore, irregular ocean wave data may be affected by various factors such as wind speed, tides, and seabed topography, leading to significant changes in wave amplitude and frequency over time and space. Therefore, processing this type of data requires modeling that considers both local characteristics and long-term trends. This invention proposes key technologies for predicting the wave height of irregular ocean waves, effectively improving the accuracy of wave height prediction.

[0133] Application Step 1: Input wave data containing wave height information into the network's input. This wave data is first processed by a one-dimensional dilated convolutional layer. Dilated convolution can effectively expand the receptive field without adding extra parameters. By introducing holes in the convolutional kernel, it skips parts of the input and captures wave features over a larger time range. This not only preserves the local information of the wave data but also allows for the extraction of wave fluctuation patterns over longer time scales, such as periodicity and trend changes.

[0134] Step 2: The data processed by dilated convolution will be input into the LSTM layer. The LSTM (Long Short-Term Memory) layer has powerful memory capabilities, enabling it to handle long-term dependencies in wave data. It can capture the changing trends and periodic characteristics of wave data, which is crucial for processing complex time-series data such as wave height. By maintaining sensitivity to dependencies over time, LSTM ensures that the network can learn the long-term changing patterns and short-term fluctuation patterns of wave data.

[0135] Step 3: After processing by the LSTM layer, the network uses a temporal attention mechanism to further process the data. This mechanism assigns different weights to each time step, allowing the model to focus on key moments that have a greater impact on the prediction results. This allows the model to dynamically adjust its attention allocation based on the importance of data at different time points. Especially when dealing with irregular waves and extreme weather events, the attention mechanism helps the model focus more on moments that significantly affect wave height changes, thereby improving prediction accuracy. Furthermore, the temporal attention mechanism can effectively handle noise in long-term series, preventing the model from being interfered with by unimportant or irrelevant data.

[0136] After processing by the aforementioned modules, the network ultimately generates a prediction of wave height. Through this combination of network architectures, the model can simultaneously capture short-term fluctuations, long-term trends, and key moment features in wave data, thereby achieving high-precision prediction of wave height. This process fully leverages the advantages of dilated convolution, LSTM, and temporal attention mechanisms, ensuring that the model possesses both broad feature capture capabilities and efficient processing of long-term dependencies and local features in complex time-series data.

[0137] Working principle: Based on the numerical water tank method, combined with extreme weather conditions (such as wind speed, air pressure, etc.) for simulation, a relatively complete irregular wave dataset is generated; then, a deep learning model TA-DCLSTM is constructed for end-to-end training and prediction, which effectively improves the prediction accuracy of wave height.

[0138] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0139] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for predicting the wave height of irregular ocean waves based on a numerical water tank and a deep learning model, characterized in that, The method for predicting the height of irregular ocean waves includes the following steps: S1. The numerical water tank method is used to simulate the pusher wave tank. Combined with extreme weather conditions, the corresponding wave equation is calculated to output the pusher motion parameters. By adjusting the pusher motion parameters, the precise control and generation of irregular waves can be achieved. The generated irregular wave data is used for deep learning model training after preprocessing. S2. Construct a deep learning model TA-DCLSTM. The deep learning model TA-DCLSTM combines a time attention mechanism with dilated convolution. The deep learning model extracts long-term dependency features of ocean wave data through the LSTM module. The deep learning model extracts local features at different time scales through dilated convolution. The deep learning model uses the attention mechanism to assign different weights to each time step, focusing on the key moments that have a greater impact on the prediction results. Accurate prediction of irregular ocean wave height is achieved through end-to-end training. In step S1, the generation process of irregular ocean waves is verified and controlled by simulating a pusher wave tank using a numerical water tank method. The vertically arranged pusher wave tank consists of a closed, elongated tank filled with liquid. Porous wave-damping blocks are symmetrically arranged at both ends of the tank, and both sides of the wave-damping blocks are designed with slopes. A groove is provided at the bottom of the tank, and a vertical pusher plate is horizontally installed in the groove. The pusher plate is slidably connected to the side wall of the groove, and the length of the pusher plate is the same as the width of the tank. The pusher plate is driven to generate waves in the tank by reciprocating up and down motion. The wave equation is calculated based on the target wave parameters and generates corresponding vertical pusher motion data; specifically including: Let the thickness of the pusher plate be Z. After time dt, the vertical distance the pusher plate moves is dx. The volume change of the groove should be equal to the volume change of the liquid at the bottom of the tank, dV. That is: dV=Zdx Let the velocity of the vertical push plate be v(t), then we have: dV = Zdx = Zvdt Assuming that the fluid volume change caused by the vertical pusher is entirely converted into water surface ripples without loss, we have: Zvdt=2Cη(t)dt Therefore, the relationship between the vertical pusher and the wave surface can be derived as follows: v(t)=2Cη(t)ξ / Z Where C is the wave propagation speed, η is the free surface waveform, and ξ is the wave height amplification factor; Substituting different waveform functions into the formula yields the velocity equation of the vertical pusher plate; substituting the preset pusher plate size and target wave parameters into the corresponding pusher plate velocity equation calculates the motion parameters of the pusher plate, and generates the expected wave by controlling the size and motion of the pusher plate. During the wave-generating process, the wave data generated by the pusher plate wave-generating tank is compared with the target wave. If the error between the actual wave information and the target wave exceeds 5%, it is corrected by adjusting the wave height amplification factor ξ. The error of the wave period is corrected by the target wave period until the wave parameters meet the requirements. Finally, the dimensions and motion data of the vertical pusher plate are saved. The regular waveform function generated by the vertical pusher is: Where S is the wave height and T is the wave period, substituting the regular wave expression into the equation for the vertical pusher motion of the regular wave yields the following equation: When generating irregular waves using the vertical pusher wave generation method, it is necessary to set the target wave spectrum, then decompose the irregular wave into multiple different regular waves in the frequency domain, and sum them to obtain the motion equation of the vertical pusher: Among them, S i Represents wave height, T i For the wave period, ε i It is the random initial phase that makes up the wave, ξ i This is the wave height amplification factor; In reality, airflow over the sea surface affects wave generation, and changes in air velocity should be taken into account. This is based on the Monin-Obukhov similarity theory for describing the ocean surface: Where u(z) is the average wind speed at sea level z, u0 is the sea surface current velocity, and u * Let Ψ be the friction velocity, k be the Karman constant, z0 represent the surface roughness length, and Ψ be the friction velocity. m Describes the stability function. This represents the dimensionless stability parameter, where L is the Monin-Obukhov length; The stability function characterizes the sea surface temperature T sea and air temperature T air The relationship when T sea >T air When this occurs, it is considered an unstable state, and there is When T air >T sea Sometimes, Dimensionless stability parameters Similarly, stable and unstable states also exist, which can be represented by the Bulk Richardson number (Ri). b Further description: Its expression is as follows: In the formula, △T=T air -T sea g is the acceleration due to gravity, z represents the air height, and U 10 The wind speed at a height of 10m, when T sea >T air At that time, there are dimensional stability parameters. When T air >T sea Sometimes, 2. The method for predicting the wave height of irregular ocean waves according to claim 1, characterized in that: In S2, the constructed deep learning model TA-DCLSTM specifically includes: The irregular wave data generated by S1 is normalized to standardize the wave height values ​​to the range of [0,1]. The normalized dataset is divided into two parts in a 7:3 ratio. 90% of the data is used as the training set to allow the model to fully learn the characteristics of irregular waves, and the remaining 10% is used as the validation set to evaluate the model performance.

3. The method for predicting the wave height of irregular ocean waves according to claim 1, characterized in that: The deep learning model TA-DCLSTM performs one-dimensional dilated convolution processing on the input irregular wave data. By introducing a gap between the convolution kernels, the receptive field is expanded to cover a larger time range without increasing the number of kernel parameters. Assuming the kernel size is k, the dilation rate is d, and the input data is x, the output feature y(t) after dilated convolution processing is expressed as:

4. The method for predicting the wave height of irregular ocean waves according to claim 3, characterized in that: When the deep learning model TA-DCLSTM processes the extracted features through the LSTM module, it specifically includes: The LSTM module learns and maintains the long-term dependencies of ocean wave data through its internal memory cells, capturing the trends and periodic characteristics of ocean waves over time. Because ocean wave data has strong time dependencies, the LSTM module can effectively control the flow of information through gating mechanisms such as input gates, forget gates, and output gates, avoiding gradient vanishing or exploding phenomena in long-term sequences. The forget gate processing is as follows: f t =σ(W f ·[h t-1 ,x t ]+b f ) Among them, h t-1 It is the hidden state of the previous time step, x t This is the current input, W f and b f These are the weights and configuration options for the forget gate; The input gate processing procedure is as follows: i t =σ(W i ·[h t-1 ,x t ]+b i ) Furthermore, the update process for memory units is as follows: Memory unit C t It preserves the dependencies in the long-term sequence, combining the memory of the previous moment with the updated information of the current input; The output gate processing procedure is as follows: the t =σ(W o ·[h t-1 ,x t ]+b o ) How does the output gate control the hidden state at the current moment affect the model's output? The final output after LSTM processing is: h t =o t ·tanh(C t ) Hidden state h t It is the output of the current time step, which combines long-term dependency information and features of the current input. Through this mechanism, LSTM can effectively capture the long-term trend and periodic change features of irregular wave data, and make full use of the influence of historical data to generate more accurate future prediction results.

5. The method for predicting the wave height of irregular ocean waves according to claim 4, characterized in that: The deep learning model uses a time attention mechanism for data processing, enabling it to focus more on key time steps when processing time-series data. Specifically, this includes: By dynamically assigning different weights to each time step, the model can focus on moments with greater impact on the final prediction result in complex time series data. The time attention mechanism first assigns different weights to each time step through a weight matrix. Based on the input features, the weights are dynamically adjusted, and the model finally calculates a weighted sum based on these weights, enabling the model to more flexibly capture complex temporal features. Let the input time series be X = [x1, x2, ..., x...]. T ], where T is the number of time steps, x t This represents the input feature at time step t. Temporal attention first calculates the attention score for each time step: score t =tanh(W·x t +b) Where W is a trainable weight matrix and b is a bias term; Then, the attention scores are converted into weights using the Softmax function: These weights are then applied to the original input features, i.e.: in, It is the weighted output sequence, that is, the weighted feature representation; This method allows the model to dynamically focus on key time steps based on different input features, thereby effectively improving prediction accuracy.

Images