Marine significant wave height intelligent prediction method, device, equipment, medium and product

Abstract

The invention discloses a marine significant wave height intelligent prediction method, device, equipment, medium and product, and relates to the field of marine environment monitoring and forecasting, and the method comprises the steps: obtaining a marine meteorological hydrological environment time sequence signal; performing discrete wavelet transform on the marine meteorological and hydrological environment time sequence signal to obtain a de-noised time sequence signal; according to the de-noised time sequence signal, an effective wave height prediction model is adopted to determine the ocean effective wave height at the future moment; the significant wave height prediction model comprises a long short-term memory network layer, a time step attention layer and a full connection layer which are connected in sequence; the long-short-term memory network layer learns the de-noised time sequence signal to capture a long-term dependency relationship in the de-noised time sequence signal and outputs a hidden state vector sequence; the time step attention layer dynamically allocates weights for hidden states in the hidden state vector sequence and performs weighted summation on the hidden states to obtain context vectors; and the full connection layer maps the context vector into the ocean significant wave height at the future moment. According to the invention, the prediction precision of the ocean significant wave height is improved.

Description

Technical Field

[0001] This application relates to the field of marine environmental monitoring and forecasting, and in particular to a method, device, equipment, medium and product for intelligent prediction of marine significant wave height. Background Technology

[0002] The complexity and randomness of marine environmental systems make accurate forecasting a persistent research challenge and frontier. Current mainstream forecasting approaches mainly fall into two categories: (i) Physical numerical prediction: Based on the equations of fluid mechanics and ocean dynamics, simulation is performed through discretization and numerical solution. The mechanism of this method is clear, but it has significant limitations: huge computational resource consumption, dependence on the accuracy of initial field and boundary conditions, insufficient ability to characterize sudden processes at small and medium scales, and difficulty in debugging due to the complexity of the model.

[0003] (II) Statistical and Machine Learning Forecasting: With the accumulation of ocean observation data, data-driven methods are receiving increasing attention. Early methods, such as autoregressive models (AR) and support vector machines (SVM), struggle to effectively handle long-term dependencies. In recent years, recurrent neural networks (RNNs) and their improved forms, such as long short-term memory (LSTM) networks and gated recurrent units (GRUs), have demonstrated outstanding performance in time series forecasting, capable of capturing long-term dependencies.

[0004] However, direct application in marine scenarios still faces three major challenges: First, the challenge of data quality. Raw observational data inevitably contains instrument noise, transmission errors, and high-frequency environmental disturbances, which can interfere with the model's ability to learn the true physical evolution. Second, the challenge of information focus. Traditional LSTM processes information from all historical time steps equally, failing to distinguish the differences in the contribution of different moments to the current prediction. For example, the state of the ocean at a more recent moment is usually more valuable than its state in the distant past, and information from key moments in specific weather events is also more important. Third, the challenge of predicting extreme events. Extreme wave (such as typhoon) samples are scarce but extremely destructive. General models often fit normal data well, but their ability to predict extreme values ​​is weak, which is precisely the part that disaster prevention and early warning are most concerned about.

[0005] While some studies have attempted to combine filtering methods with neural networks or introduce attention mechanisms in related technologies, they often lack systematic design. Typically, only simple moving average or Fourier filtering is used for noise reduction, which is insufficient for time-frequency localization analysis; or attention mechanisms are simply mechanically superimposed on the network without fully considering their adaptability to ocean sequence characteristics. Therefore, developing a method that can synergistically optimize data quality and dynamically focus on key information has urgent practical needs and significant technological value. Summary of the Invention

[0006] The purpose of this application is to provide a method, device, equipment, medium, and product for intelligent prediction of ocean significant wave height, which can improve the prediction accuracy of ocean significant wave height.

[0007] To achieve the above objectives, this application provides the following solution: Firstly, this application provides an intelligent prediction method for ocean significant wave height, including: Acquire time-series signals of marine meteorological and hydrological environment; Discrete wavelet transform is performed on the marine meteorological and hydrological environment time series signal to obtain a denoised time series signal; Based on the denoised time series signal, the ocean significant wave height at future times is determined using a significant wave height prediction model; The significant wave height prediction model is obtained by training a training sample set in advance. Each training sample in the training sample set includes the sample marine meteorological and hydrological environment time series signal and the corresponding future marine significant wave height. The significant wave height prediction model comprises a long short-term memory network layer, a time-step attention layer, and a fully connected layer connected in sequence. The long short-term memory network layer learns the denoised time-series signal to capture long-term dependencies and outputs a sequence of hidden state vectors. The time-step attention layer dynamically assigns weights to the hidden states in the sequence of hidden state vectors and performs a weighted summation of the hidden states to obtain a context vector. The fully connected layer maps the context vector to the significant wave height of the ocean.

[0008] Secondly, this application provides an intelligent prediction device for ocean significant wave height, comprising: The signal acquisition module is used to acquire time-series signals of marine meteorological and hydrological environment; The signal denoising module is used to perform discrete wavelet transform on the marine meteorological and hydrological environment time series signal to obtain a denoised time series signal. The ocean significant wave height prediction module is used to determine the ocean significant wave height at future times based on the denoised time series signal using a significant wave height prediction model; The significant wave height prediction model is obtained by training a training sample set in advance. Each training sample in the training sample set includes the sample marine meteorological and hydrological environment time series signal and the corresponding future marine significant wave height. The significant wave height prediction model comprises a long short-term memory network layer, a time-step attention layer, and a fully connected layer connected in sequence. The long short-term memory network layer learns the denoised time-series signal to capture long-term dependencies and outputs a sequence of hidden state vectors. The time-step attention layer dynamically assigns weights to the hidden states in the sequence of hidden state vectors and performs a weighted summation of the hidden states to obtain a context vector. The fully connected layer maps the context vector to the significant wave height of the ocean.

[0009] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described intelligent prediction method for ocean effective wave height.

[0010] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described intelligent prediction method for ocean effective wave height.

[0011] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the above-described intelligent prediction method for ocean effective wave height.

[0012] According to the specific embodiments provided in this application, this application achieves the following technical effects: It introduces discrete wavelet transform for adaptive denoising at the data level, separating effective signals from noise in the time-frequency domain, significantly improving the quality of input data and laying a clean and stable data foundation for subsequent modeling. Furthermore, it embeds a time-step attention mechanism into the Long Short-Term Memory network, enabling the significant wave height prediction model to autonomously learn and dynamically assign differentiated importance weights to different time steps in the historical sequence, enhancing the model's ability to capture key forecasting factors and thus improving the prediction accuracy of ocean significant wave height. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a flowchart illustrating an intelligent prediction method for ocean significant wave height provided in an embodiment of this application.

[0015] Figure 2 This is a flowchart illustrating the discrete wavelet transform denoising process in one embodiment of this application.

[0016] Figure 3 This is a schematic diagram of the effective wave height prediction model in one embodiment of this application.

[0017] Figure 4 This is a schematic diagram of the training and prediction process of the effective wave height prediction model in one embodiment of this application.

[0018] Figure 5 This is a comparison diagram of the effective wave height before and after discrete wavelet transform denoising in one embodiment of this application.

[0019] Figure 6 This is a comparison chart of the predicted and actual values ​​of the effective wave height prediction model on the test set in one embodiment of this application.

[0020] Figure 7 This is a schematic diagram of the functional modules of an intelligent prediction device for ocean significant wave height provided in an embodiment of this application. Detailed Implementation

[0021] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] The purpose of this application is to provide a machine learning hybrid prediction model for high-noise, nonlinear marine meteorological and hydrological environmental time-series data, namely, a significant wave height prediction model. This model deeply integrates signal processing technology and deep learning algorithms, aiming to achieve real-time, high-precision prediction of key marine dynamic parameters (significant wave height), providing reliable technical support for marine disaster prevention and mitigation, shipping safety, and resource development.

[0023] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0024] In one exemplary embodiment, such as Figure 1 As shown, an intelligent prediction method for ocean significant wave height is provided. This method is executed by computer equipment, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method includes steps 101 to 103.

[0025] Step 101: Obtain marine meteorological and hydrological environmental time-series signals.

[0026] The marine meteorological and hydrological environmental time-series signal includes marine meteorological and hydrological environmental data for multiple consecutive time steps. Each time step's data includes wind direction, average wind speed, peak gust speed, significant wave height, dominant wave period, average wave period, dominant wave direction, sea level pressure, air temperature, and sea surface temperature. The peak gust speed is the 5-second or 8-second peak gust speed.

[0027] Step 102: Perform Discrete Wavelet Transform (DWT) on the marine meteorological and hydrological environmental time series signal to obtain a denoised time series signal. The process is as follows: Figure 2 As shown.

[0028] The purpose of discrete wavelet transform is to suppress high-frequency random noise to the greatest extent possible while preserving the main evolution trend and abrupt change characteristics of the effective wave height. In a specific application example, step 102 includes steps 21 to 24.

[0029] Step 21: Perform multi-level wavelet decomposition on the marine meteorological and hydrological environmental time-series signal to obtain a coefficient set. The coefficient set includes approximate coefficients and detail coefficient vectors for each level. The detail coefficient vector includes multiple detail coefficients. The approximate coefficients represent the low-frequency main signal, while the detail coefficient vector represents the high-frequency components, mainly containing noise and detail features.

[0030] Specifically, considering the non-stationary characteristics of ocean wave height sequences, wavelet basis functions with good time-frequency localization properties are selected. Decomposition level. L The data sampling frequency and the dominant noise frequency band are adaptively determined, typically decomposed to layers 3 through 5. Then, the marine meteorological and hydrological environmental time series signals are analyzed. conduct L Layer wavelet decomposition yields a set of coefficients: .in, For the first L Approximation coefficients of the layer For the first i Layer detail coefficient vector, The length of the detail coefficient vector is equal to the length of the marine meteorological and hydrological environmental time series signal. N Divide by Such as the detail coefficient vector of layer 1 . to This is the detail factor.

[0031] Step 22: Calculate the global threshold based on the detail coefficient vector of the first layer.

[0032] This application is based on the detail coefficient vector of the first layer. The robust median absolute deviation method is used to robustly estimate the noise standard deviation. A general threshold criterion is adopted. Calculate the global threshold. The global threshold ensures that, under the Gaussian white noise assumption, signal coefficients are significantly preserved while noise coefficients are suppressed. Wherein, TH The global threshold. The standard deviation of noise. N The length of the marine meteorological and hydrological environment monitoring signal. This is the detail coefficient vector for the first layer. This is the median-value function.

[0033] Step 23: Based on the global threshold, perform soft thresholding on the detail coefficient vector of each layer to obtain the detail coefficient vector after thresholding of each layer.

[0034] Specifically, the detail coefficient vector of each layer is soft-thresholded using the following formula: ; in, For the first i The first detail coefficient vector after layer thresholding j A detailed coefficient, For the first i The first layer in the detail coefficient vector j A detailed coefficient, For the sign function, To find the maximum value function, For the first i The dynamic threshold of the layer, , TH As a global threshold, the soft threshold function produces a more continuous estimate and avoids the pseudo-Gibbs phenomenon that may be caused by the hard threshold function.

[0035] Step 24: Perform wavelet inverse transform on the detail coefficient vector after thresholding of each layer and the approximation coefficient of each layer to obtain the denoised time series signal.

[0036] This application performs discrete wavelet transform on ocean meteorological and hydrological environmental time series signals to obtain denoised time series signals, which have the following characteristics: (1) Wavelet transform has time-frequency localization capability: Wavelet transform can decompose the signal into different scales (frequency bands) and locate it on the time axis. Noise and real signals exhibit significant amplitude separation characteristics in the wavelet domain: Real physical processes such as sudden changes in wind speed and wave height response usually have large amplitude wavelet coefficients even at high frequencies; while sensor noise, transmission errors, etc., are characterized by small amplitude and randomly distributed coefficients.

[0037] (2) Threshold selection mechanism guarantee: general threshold criterion Based on noise variance estimation, its design principle is to retain large-amplitude coefficients (signal) and suppress small-amplitude coefficients (noise). Wavelet coefficient amplitudes corresponding to synchronous high-frequency pulsations under extreme weather conditions (such as sudden increases in wind speed or wave height) far exceed the threshold and will not be filtered out. Independent denoising filters out sensor noise, transmission errors, and environmental high-frequency jitter that are unrelated to the signal, rather than the physically driven high-frequency signal.

[0038] (3) The hysteresis of the physical response of ocean waves: There is a physical hysteresis of several hours in the response of significant wave height to wind speed (depending on wind area, wind time, etc.). The high-frequency random fluctuations of wind speed (the period is shorter than the wave growth time) do not directly drive the synchronous high-frequency fluctuations of wave height. Therefore, the independent denoising mainly filters out physically unrelated high-frequency random noise, rather than the key driving signal.

[0039] In this application, independent denoising does not compromise covariance or the model's ability to predict extreme events.

[0040] First, covariance will not be violated: (1) The physical covariance between wind speed and wave height is mainly reflected in low-frequency trends (weather processes, seasonal variations) and mid-frequency energy transfer (wind and wave growth on an hourly scale). These components correspond to approximate coefficients in wavelet decomposition. The mid-level detail coefficients, with their large amplitude, are fully preserved.

[0041] (2) Independent denoising removes high-frequency random noise that is independent of each variable. This noise is physically uncorrelated. Removing it can enhance the true covariance between variables. Empirical evidence shows that the wind speed-wave height cross-correlation curve after denoising is smoother, the peak lag is more in line with marine meteorological theory (6-12 hours), and spurious fluctuations are significantly reduced.

[0042] Secondly, the ability to predict extreme events remains unaffected: The sudden changes in wind speed and the abrupt rise in wave height under extreme weather conditions are high-frequency components with large amplitude and structural characteristics, and their wavelet coefficients are much higher than the threshold.

[0043] Therefore, independent wavelet denoising is feasible and safe in this application. It does not destroy the physical covariance between variables, nor does it filter out key high-frequency components of extreme events; instead, it improves the model's ability to capture real physical laws and the accuracy of its predictions of extreme events by suppressing random noise. This method has been validated in multiple marine meteorological and hydrological research cases and has a solid theoretical and empirical foundation.

[0044] Step 103: Based on the denoised time series signal, the ocean significant wave height at future times is determined using the significant wave height prediction model.

[0045] Among them, such as Figure 3 As shown, the effective wave height prediction model includes a long short-term memory network layer, a time step attention layer, and a fully connected layer connected in sequence.

[0046] The Long Short-Term Memory (LSTM) network layer learns the denoised time-series signal to capture long-term dependencies and outputs a sequence of hidden state vectors.

[0047] Long Short-Term Memory (LSTM) network layers are composed of multiple stacked LSTM units, with the input being a denoised sequential signal. ,in, For the first t Multiple feature vectors at each time step. The LSTM layer outputs a sequence of hidden state vectors at each time step. , ,in, For the first t Hidden states at each time step d The hidden state dimension. LSTM effectively captures and transmits long-term dependent information through its internal gating mechanisms (forget gate, input gate, output gate) and cell states.

[0048] The forget gate controls the forgetting or discarding of corresponding information at a certain time. It uses different information weight values ​​set in the sigmoid activation function to retain the necessary information and discard other information. For information to be forgotten, the activation function's information weight parameter is set to 0; for information to be retained, the activation function's information weight parameter is set to 1. The input value in the forget gate is the result of the current time step input and the hidden state of the previous time step, while the output value is a real number representing the forgetting rate, ranging from [0,1]. The relevant calculation formula is as follows: ;in, It is the first t The forgetting value of the forgetting gate at each time step. It is a layer update control for the corresponding object. It is the first t Hidden states at -1 time step, It is the first t The bias value corresponding to the forget gate at each time step. It is the first t The corresponding weights of the forget gates at each time step.

[0049] Input gates control whether corresponding information can enter the cellular state, i.e., long-term memory. In this approach, it is achieved using element-wise dot product and the sigmoid activation function. The calculation formula is: ; ;in, It is the first t The output of each time-step input gate can control long-term memory. Whether to update the current cell state information, and the specific value after the update is controlled by the tanh function. It is a layer update control for the corresponding object. It is the bias value corresponding to the input gate. These are the corresponding weights of the input gate. It is the weight value of long-term memory. It is the bias value of long-term memory. It is the unit state corresponding to the memory unit of the current neuron.

[0050] The output gate controls how much information the current LSTM neuron passes forward for output. The output vector is determined by the hidden state of the previous time step and the input of the current time step. The calculation formula is: ; ;in, It is the first t The output of the time step output gate It is the weight matrix of the output gate. It is the bias term of the output gate. It represents the cell state at the current time step, and tanh is the hyperbolic tangent function.

[0051] Updating the cell state involves updating the forgotten value. State of past time steps After multiplying, the resulting value is compared with the current state of the memory cell. Input gate The parameters are multiplied and then added together. The formula for calculating cell state is: .

[0052] The time-step attention layer is used to dynamically assign weights to the hidden states in the hidden state vector sequence and perform a weighted summation of the hidden states to obtain the context vector.

[0053] Specifically, firstly, based on the hidden state vector sequence, the formula is used... Calculate the unnormalized attention score for each time step, reflecting the importance of that time step. For the first t Unnormalized attention score at each time step For the first t Hidden states at each time step W This is the weight matrix. b For bias vectors, v For attention weight vectors, W , b and v All were obtained by training using a training sample set in advance. , , The superscript T indicates the transpose operation.

[0054] Then, the Softmax function is used to convert the score into a normalized attention weight distribution, i.e., using the formula... The unnormalized attention score at each time step is converted into a normalized attention weight distribution; where, For the first t Normalized attention weight distribution at each time step T The total number of time steps. e k For the first k The unnormalized attention score for each time step.

[0055] Then, the original hidden state sequence is weighted and summed based on attention weights to generate a context vector that incorporates global key information, i.e., using the formula... The hidden states are weighted and summed to obtain the context vector; where, c As a context vector, it condenses the most critical information from the historical sequence for the current prediction.

[0056] The fully connected layer maps the context vector to the significant wave height of the ocean at future times. Specifically, the fully connected layer maps the context vector and ultimately outputs the predicted significant wave height for the next hour. During the training phase, a Dropout layer is added after the fully connected layer to prevent overfitting.

[0057] The significant wave height prediction model is pre-trained using a training sample set, where each training sample includes a time-series signal of the marine meteorological and hydrological environment and the corresponding future significant wave height. The significant wave height prediction model can automatically learn and focus on the historical moments that have the greatest influence on predicting future wave heights.

[0058] The loss function used during training of the effective wave height prediction model is the mean squared error loss function, which has good properties for regression tasks. The formula is as follows: ;in, M The total number of training samples. For the first m The true values ​​of ocean significant wave height at future times corresponding to each sample of marine meteorological and hydrological environmental time series signals. This is the predicted value output by the effective wave height prediction model.

[0059] The optimization process employs an adaptive moment estimation (Adam) optimizer. This optimizer accumulates historical gradient information by calculating the first-order moment estimate (momentum term) of the gradient, thereby accelerating convergence. Simultaneously, it adaptively assigns learning rates to different parameters by calculating the second-order moment estimate of the gradient to address gradient sparsity or scale sensitivity issues. Based on this mechanism, the optimizer can achieve fast and stable convergence during training iterations.

[0060] The core idea of ​​this application is to construct an intelligent prediction system encompassing "preprocessing - feature extraction - dynamic attention - accurate output." It achieves precise separation of signal and noise in the time-frequency domain through Discrete Wavelet Transform (DWT) to purify the input; leverages the powerful sequence modeling capabilities of the LSTM network to extract deep temporal features; innovatively introduces a time-step-oriented attention mechanism, endowing the model with the ability to dynamically evaluate and weight key historical moment information; and finally outputs high-precision prediction results through a fully connected layer. Specifically, at the data level, this application introduces Discrete Wavelet Transform for adaptive denoising, separating effective signals and noise in the time-frequency domain, significantly improving the quality of input data and laying a clean and stable data foundation for subsequent modeling. At the model structure level, a time-step attention mechanism is embedded in the LSTM network, enabling the effective wave height prediction model to autonomously learn and dynamically assign differentiated importance weights to different time steps in the historical sequence, enhancing the model's ability to capture key forecasting factors. In addition to focusing on overall prediction accuracy, it also specifically designs a performance evaluation for extreme wave height events, enhancing the relevance and reliability of the effective wave height prediction model in practical applications such as disaster prevention and mitigation.

[0061] like Figure 4 As shown below, the detailed prediction process is illustrated by implementing the intelligent prediction method for ocean significant wave height provided in this application using NDBC 44013 buoy data to predict the ocean significant wave height for the next hour.

[0062] (1) Data: Hourly measured marine meteorological and hydrological environmental time series signals were acquired. This application uses hourly data from NDBC buoy 44013 from 2013 to 2019. Ten parameters were selected as features: wind direction, average wind speed, peak gust speed, ocean significant wave height, main wave period, average wave period, dominant wave direction, sea level pressure, air temperature, and sea surface temperature. Ocean significant wave height was used as the prediction target.

[0063] (2) Denoising: First, outliers and missing values ​​in the marine meteorological and hydrological environmental time series signal are processed. Then, DWT denoising is performed independently on each parameter sequence, especially the marine significant wave height sequence (using “db4” wavelet, L=4 layers). Taking a typhoon process of buoy 44013 as an example: the wind speed jumps from 8m / s to 21m / s in 5 hours. The corresponding wavelet coefficient is about 10-12 times the noise estimate, which is significantly higher than the threshold. It will be completely preserved and will not be misjudged as noise and filtered out.

[0064] (3) Dataset Construction: The training set, validation set, and test set are divided into three parts in a 6:2:2 ratio according to time sequence. The minimum and maximum values ​​of each feature in the training set are calculated, and these values ​​are used to normalize all datasets. .in, These are standardized values. These are the original values. It is the minimum value of that attribute in the dataset. It is the maximum value of this attribute in the dataset.

[0065] (4) Sequence construction: The sliding window width is set to 24 (historical 24 hours), and the step size is 1. The input of each sample is a 24×10 matrix, and the output is the ocean significant wave height value of the 25th hour.

[0066] (5) Model configuration: Construct a two-layer LSTM with 64 and 64 hidden units respectively; the dimension of the attention layer is consistent with the dimension of the LSTM output; the number of neurons in the fully connected layer is 32 and 1; the dropout rate is 0.2.

[0067] (6) Training: The Adam optimizer (initial learning rate 0.001) is used, the batch size is 64, the training is conducted for a maximum of 50 epochs, and an early stopping strategy is adopted.

[0068] (7) Evaluation: Make predictions on the test set, then inversely normalize the results to verify the accuracy of the model, and calculate the MAE and Specifically, extreme samples with wave heights exceeding a certain quantile are selected from the test set, and their prediction errors are calculated separately to evaluate the model's ability to capture extreme events. Figure 5 and Figure 6 As shown, the evaluation results are: Overall performance: MAE=0.04m, R 2 =0.981. Extreme event prediction performance: MAE=0.15m, an improvement of approximately 25% compared to the baseline model (standard LSTM).

[0069] The above results fully demonstrate the effectiveness, advancement, and practical value of the method proposed in this application for ocean significant wave height prediction.

[0070] In summary, the beneficial effects of this application include at least the following: (1) Significantly improved prediction accuracy. Data quality is improved from the source through DWT denoising, enabling the effective wave height prediction model to learn more closely to the actual physical laws; the attention mechanism allows the effective wave height prediction model to focus on historical segments strongly correlated with the prediction target. Examples show that the MAE of this application on the test set can reach 0.04 meters, and R... 2 With an accuracy exceeding 0.98, its overall accuracy surpasses that of a single LSTM model and a hybrid model without denoising.

[0071] (2) The model has strong robustness and generalization ability. It is more robust to noise and missing data, and the adaptive thresholding of wavelet transform can cope with noise environments of different intensities. The effective wave height prediction model learns the robust features after denoising and shows good transfer potential on buoy data in different sea areas or at different time periods.

[0072] (3) Improved predictive ability for extreme events. Through targeted evaluation and model design (such as attention mechanisms to capture abrupt change points), the effective wave height prediction model has improved the accuracy of predicting high-wave and other extreme ocean conditions, which has crucial application value for disaster early warning such as typhoon waves and storm surges. The MAE of extreme event prediction in this application on the test set is 0.15m, which is about 12% better than the model without denoising, proving that independent denoising does not filter out key high-frequency components, but instead improves the predictability of extreme events.

[0073] (4) Balance between computational efficiency and practicality. Compared with complex numerical models, once the training is completed, the computational overhead in the prediction stage of this application is minimal, enabling near real-time forecasting. The entire process is highly automated and easy to deploy for operational use.

[0074] Based on the same inventive concept, this application also provides an apparatus for implementing the method described above. The solution provided by this apparatus is similar to the solution described in the above method; therefore, specific limitations in one or more apparatus embodiments provided below can be found in the limitations of the method described above, and will not be repeated here.

[0075] In one exemplary embodiment, such as Figure 7 As shown, an intelligent prediction device for ocean significant wave height is provided, including: a signal acquisition module 701, a signal denoising module 702, and an ocean significant wave height prediction module 703.

[0076] The signal acquisition module 701 is used to acquire marine meteorological and hydrological environmental time-series signals.

[0077] The signal denoising module 702 is used to perform discrete wavelet transform on the marine meteorological and hydrological environmental time series signal to obtain a denoised time series signal.

[0078] The ocean significant wave height prediction module 703 is used to determine the ocean significant wave height at future times based on the denoised time series signal and an significant wave height prediction model.

[0079] The significant wave height prediction model is trained in advance using a training sample set. Each training sample in the training sample set includes the sample marine meteorological and hydrological environment time series signal and the corresponding future significant wave height.

[0080] The significant wave height prediction model comprises a Long Short-Term Memory (LSTM) network layer, a time-step attention layer, and a fully connected layer connected in sequence. The LSM network layer learns from the denoised time-series signal to capture long-term dependencies and outputs a sequence of hidden state vectors. The time-step attention layer dynamically assigns weights to the hidden states in the hidden state vector sequence and performs a weighted summation of the hidden states to obtain a context vector. The fully connected layer maps the context vector to significant wave heights in the ocean.

[0081] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0082] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0083] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0084] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0085] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods.

[0086] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0087] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A smart prediction method for ocean significant wave height, characterized in that, The intelligent prediction method for ocean significant wave height includes: Acquire time-series signals of marine meteorological and hydrological environment; Discrete wavelet transform is performed on the marine meteorological and hydrological environment time series signal to obtain a denoised time series signal; Based on the denoised time series signal, the ocean significant wave height at future times is determined using a significant wave height prediction model; The significant wave height prediction model is obtained by training a training sample set in advance. Each training sample in the training sample set includes the sample marine meteorological and hydrological environment time series signal and the corresponding future marine significant wave height. The significant wave height prediction model comprises a long short-term memory network layer, a time-step attention layer, and a fully connected layer connected in sequence. The long short-term memory network layer learns the denoised time-series signal to capture long-term dependencies and outputs a sequence of hidden state vectors. The time-step attention layer dynamically assigns weights to the hidden states in the sequence of hidden state vectors and performs a weighted summation of the hidden states to obtain a context vector. The fully connected layer maps the context vector to the significant wave height of the ocean at future times.

2. The intelligent prediction method for ocean significant wave height according to claim 1, characterized in that, The marine meteorological and hydrological environment time series signal includes marine meteorological and hydrological environment data for multiple consecutive time steps; the marine meteorological and hydrological environment data for each time step include wind direction, average wind speed, peak gust speed, ocean significant wave height, main wave period, average wave period, dominant wave direction, sea level pressure, air temperature, and sea surface temperature.

3. The intelligent prediction method for ocean significant wave height according to claim 1, characterized in that, The marine meteorological and hydrological environmental time series signal is subjected to discrete wavelet transform to obtain a denoised time series signal, including: The marine meteorological and hydrological environmental time series signal is subjected to multi-level wavelet decomposition to obtain a coefficient set; the coefficient set includes approximate coefficients and detail coefficient vectors for each level; the detail coefficient vector includes multiple detail coefficients. Calculate the global threshold based on the detail coefficient vector of the first layer; Based on the global threshold, soft thresholding is performed on the detail coefficient vector of each layer to obtain the detail coefficient vector after thresholding of each layer. The detailed coefficient vector after thresholding at each layer and the approximation coefficients at each layer are subjected to inverse wavelet transform to obtain the denoised time series signal.

4. The intelligent prediction method for ocean significant wave height according to claim 3, characterized in that, The global threshold is calculated using the following formula: ； ； in, TH The global threshold. The standard deviation of noise. N The length of the marine meteorological and hydrological environment monitoring signal. This is the detail coefficient vector for the first layer. This is the median-value function.

5. The intelligent prediction method for ocean significant wave height according to claim 3, characterized in that, The following formula is used to perform soft thresholding on the detail coefficient vector of each layer: ； in, For the first i The first detail coefficient vector after layer thresholding j A detailed coefficient, For the first i The first layer in the detail coefficient vector j A detailed coefficient, For the sign function, To find the maximum value function, For the first i The dynamic threshold of the layer, , TH This is the global threshold.

6. The intelligent prediction method for ocean significant wave height according to claim 1, characterized in that, The process by which the time-step attention layer dynamically assigns weights to the hidden states in the hidden state vector sequence and performs a weighted summation of the hidden states to obtain the context vector includes: Based on the hidden state vector sequence, the formula is used. Calculate the unnormalized attention score at each time step; where, For the first t Unnormalized attention score at each time step For the first t Hidden states at each time step W This is the weight matrix. b For bias vectors, v For attention weight vectors, W , b and v All of them were obtained in advance using a training sample set, and the superscript T indicates the transpose operation; Using formula The unnormalized attention score at each time step is converted into a normalized attention weight distribution; where, For the first t Normalized attention weight distribution at each time step T The total number of time steps. e k For the first k Unnormalized attention score at each time step; Using formula The hidden states are weighted and summed to obtain the context vector; where, c This is the context vector.

7. A smart ocean significant wave height prediction device, applied to the smart ocean significant wave height prediction method according to any one of claims 1-6, characterized in that, The intelligent prediction device for ocean significant wave height includes: The signal acquisition module is used to acquire time-series signals of marine meteorological and hydrological environment; The signal denoising module is used to perform discrete wavelet transform on the marine meteorological and hydrological environment time series signal to obtain a denoised time series signal. The ocean significant wave height prediction module is used to determine the ocean significant wave height at future times based on the denoised time series signal using a significant wave height prediction model; The significant wave height prediction model is obtained by training a training sample set in advance. Each training sample in the training sample set includes the sample marine meteorological and hydrological environment time series signal and the corresponding future marine significant wave height. The significant wave height prediction model comprises a long short-term memory network layer, a time-step attention layer, and a fully connected layer connected in sequence. The long short-term memory network layer learns the denoised time-series signal to capture long-term dependencies and outputs a sequence of hidden state vectors. The time-step attention layer dynamically assigns weights to the hidden states in the sequence of hidden state vectors and performs a weighted summation of the hidden states to obtain a context vector. The fully connected layer maps the context vector to the significant wave height of the ocean.

8. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the intelligent prediction method for ocean significant wave height according to any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the intelligent prediction method for ocean effective wave height as described in any one of claims 1-6.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the intelligent prediction method for ocean effective wave height as described in any one of claims 1-6.

Images