CN-121995548-A - Intelligent prediction method, device, equipment, medium and product for ocean effective wave height

CN121995548ACN 121995548 ACN121995548 ACN 121995548ACN-121995548-A

Abstract

The application discloses a method, a device, equipment, a medium and a product for intelligently predicting ocean effective wave height, which relate to the field of ocean environment monitoring and forecasting; the method comprises the steps of performing discrete wavelet transformation on marine meteorological hydrologic environment time sequence signals to obtain denoising time sequence signals, determining marine effective wave heights at future time according to the denoising time sequence signals by adopting an effective wave height prediction model, wherein the effective wave height prediction model comprises a long-period memory network layer, a time step attention layer and a full-connection layer which are sequentially connected, the long-period memory network layer learns the denoising time sequence signals to capture long-term dependency relations and output hidden state vector sequences, the time step attention layer dynamically distributes weights for hidden states in the hidden state vector sequences and performs weighted summation on the hidden states to obtain context vectors, and the full-connection layer maps the context vectors to the marine effective wave heights at the future time. The application improves the prediction precision of the ocean effective wave height.

Inventors

XU ZHAOYUE
AN HAOWEN
LIU YUANYING
LI XIAOXU
HU JINGUO
LI YAWEN

Assignees

国家海洋技术中心

Dates

Publication Date: 20260508
Application Date: 20260409

Claims (10)

1. The intelligent prediction method for the ocean effective wave height is characterized by comprising the following steps of: Acquiring a marine meteorological hydrologic environment time sequence signal; performing discrete wavelet transformation on the marine meteorological hydrologic environment time sequence signal to obtain a denoising time sequence signal; according to the denoising time sequence signal, determining the ocean effective wave height at the future moment by adopting an effective wave height prediction model; The method comprises the steps that an effective wave height prediction model is obtained by training a training sample set in advance, and each training sample in the training sample set comprises a sample marine meteorological hydrologic environment time sequence signal and a corresponding marine effective wave height at a future moment; The effective wave height prediction model comprises a long-period memory network layer, a time step attention layer and a full connection layer which are sequentially connected, wherein the long-period memory network layer learns the denoising time sequence signal to capture long-period dependency relationship therein and output a hidden state vector sequence, the time step attention layer is used for dynamically distributing weights for hidden states in the hidden state vector sequence and carrying out weighted summation on the hidden states to obtain a context vector, and the full connection layer is used for mapping the context vector into ocean effective wave heights at future time.
2. The method of intelligent prediction of ocean effective wave height according to claim 1, wherein the marine meteorological hydrodynamic environment time sequence signal comprises marine meteorological hydrodynamic environment data of a plurality of continuous time steps, wherein each time step of marine meteorological hydrodynamic environment data comprises wind direction, average wind speed, peak gust speed, ocean effective wave height, main wave period, average wave period, dominant wave direction, sea level pressure, air temperature and sea surface temperature.
3. The method of intelligent prediction of ocean significant wave height according to claim 1, wherein performing discrete wavelet transform on the marine meteorological hydrologic environment timing signal to obtain a denoising timing signal comprises: Performing multi-layer wavelet decomposition on the marine meteorological hydrologic environment time sequence signal to obtain a coefficient set, wherein the coefficient set comprises an approximate coefficient and a detail coefficient vector of each layer; Calculating a global threshold according to the detail coefficient vector of the first layer; according to the global threshold, soft threshold processing is carried out on the detail coefficient vector of each layer respectively, and the detail coefficient vector after threshold processing of each layer is obtained; And carrying out wavelet inverse transformation on the detail coefficient vector after each layer of thresholding and the approximate coefficient of each layer to obtain a denoising time sequence signal.
4. A method of intelligent prediction of ocean wave height according to claim 3, wherein the global threshold is calculated using the formula: ; ; where TH is the global threshold, Is the standard deviation of noise, N is the length of the marine meteorological hydrologic environment monitoring signal, For the detail coefficient vector of the first layer, To take the median function.
5. A method of intelligent prediction of ocean significant wave height according to claim 3, wherein the detail coefficient vector for each layer is soft thresholded using the following formula: ; Wherein, the For the j-th detail coefficient in the i-th layer thresholded detail coefficient vector, For the j-th detail coefficient in the detail coefficient vector of the i-th layer, In order to take the sign function of the symbol, In order to take the function of the maximum value, For the dynamic threshold of the i-th layer, TH is a global threshold.
6. The method of claim 1, wherein the dynamically assigning weights to hidden states in the sequence of hidden state vectors and weighting and summing the hidden states to obtain the context vector by the time-step attention layer comprises: According to the hidden state vector sequence, adopting a formula Calculating an unnormalized attention score for each time step, wherein, For an unnormalized attention score for the t-th time step, For the hidden state of the T-th time step, W is a weight matrix, b is a bias vector, v is an attention weight vector, W, b and v are obtained by training in advance by adopting a training sample set, and the superscript T represents transposition operation; Using the formula The non-normalized attention score for each time step is converted into a normalized attention weight distribution, respectively, wherein, For a normalized attention weight distribution for the T-th time step, T is the total number of time steps, e k is the un-normalized attention score for the k-th time step; Using the formula And carrying out weighted summation on the hidden states to obtain a context vector, wherein c is the context vector.
7. An intelligent prediction device for ocean effective wave height, which is applied to the intelligent prediction method for ocean effective wave height according to any one of claims 1-6, and is characterized in that the intelligent prediction device for ocean effective wave height comprises: the signal acquisition module is used for acquiring marine meteorological hydrologic environment time sequence signals; The signal denoising module is used for performing discrete wavelet transformation on the marine meteorological hydrologic environment time sequence signal to obtain a denoising time sequence signal; The ocean effective wave height prediction module is used for determining the ocean effective wave height at the future moment by adopting an effective wave height prediction model according to the denoising time sequence signal; The method comprises the steps that an effective wave height prediction model is obtained by training a training sample set in advance, and each training sample in the training sample set comprises a sample marine meteorological hydrologic environment time sequence signal and a corresponding marine effective wave height at a future moment; The effective wave height prediction model comprises a long-period memory network layer, a time step attention layer and a full connection layer which are sequentially connected, wherein the long-period memory network layer learns the denoising time sequence signal to capture long-period dependency relationship therein and output a hidden state vector sequence, the time step attention layer is used for dynamically distributing weights for hidden states in the hidden state vector sequence and carrying out weighted summation on the hidden states to obtain a context vector, and the full connection layer is used for mapping the context vector into ocean effective wave heights.
8. A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the processor executes the computer program to implement the ocean wave height intelligent prediction method according to any one of claims 1-6.
9. A computer readable storage medium having stored thereon a computer program, which when executed by a processor implements the ocean wave height intelligent prediction method according to any one of claims 1-6.
10. A computer program product comprising a computer program which, when executed by a processor, implements the ocean wave height intelligent prediction method of any one of claims 1-6.

Description

Intelligent prediction method, device, equipment, medium and product for ocean effective wave height Technical Field The application relates to the field of marine environment monitoring and forecasting, in particular to a marine effective wave height intelligent forecasting method, device, equipment, medium and product. Background The complexity and randomness of marine environmental systems have made their accurate predictions a constant research difficulty and forefront. The current mainstream forecasting approach mainly comprises the following two types: And (one) physical numerical forecasting, namely carrying out simulation by discretization and numerical solution based on a hydrodynamic and marine dynamics equation set. The method has definite mechanism, but has obvious limitations that the calculation resource consumption is huge, the accuracy depending on the initial field and the boundary condition is high, the depicting capability of the medium-small scale burstiness process is insufficient, and the debugging is difficult due to complex mode. And secondly, statistics and machine learning forecast, wherein as marine observation data are accumulated, a data driving method is increasingly paid attention to. Early methods such as autoregressive models (Autoregressive model, AR), support vector machines (Support Vector Machine, SVM), etc., have been difficult to handle long sequence dependencies efficiently. In recent years, recurrent neural networks (Recurrent Neural Network, RNN) and their modified Long Short-Term Memory (LSTM) and gated loop units (Gated Recurrent Unit, GRU) have emerged in timing predictions, enabling capture of Long-Term dependencies. However, direct application in marine scenarios still faces three challenges, one being data quality challenges. The raw observation data inevitably contains instrument noise, transmission errors and environmental high-frequency disturbances, which can interfere with model learning of the true physical evolution law. And secondly, information focusing challenges. Traditional LSTM processes all historical time-step information equally, failing to distinguish between differences in contribution to the current prediction at different times. For example, near-time ocean conditions are often more reference than far past conditions, and critical point-in-time information for a particular weather process to occur is also more important. And thirdly, extreme event prediction challenges. Samples of extreme sea waves (such as typhoons) are rare but very damaging. The general model is often well fitted to normal data, but has weak prediction capability to extreme values, which is the most concerned part of disaster prevention early warning. In the related art, there are many attempts to combine filtering methods with neural networks or draw attention to mechanisms, but there is a lack of systematic design. Typically, only simple moving average or fourier filtering denoising is performed, which is insufficient in time-frequency localization analysis capability, or only attention mechanisms are mechanically superimposed on the network, and the suitability of the attention mechanisms to marine sequence characteristics is not fully considered. Therefore, the development of the dynamic focusing key information capable of cooperatively optimizing the data quality has urgent practical requirements and important technical values. Disclosure of Invention The application aims to provide an intelligent prediction method, device, equipment, medium and product for ocean effective wave height, which can improve the prediction precision of ocean effective wave height. In order to achieve the above object, the present application provides the following solutions: In a first aspect, the application provides an intelligent prediction method for ocean effective wave height, comprising the following steps: Acquiring a marine meteorological hydrologic environment time sequence signal; performing discrete wavelet transformation on the marine meteorological hydrologic environment time sequence signal to obtain a denoising time sequence signal; according to the denoising time sequence signal, determining the ocean effective wave height at the future moment by adopting an effective wave height prediction model; The method comprises the steps that an effective wave height prediction model is obtained by training a training sample set in advance, and each training sample in the training sample set comprises a sample marine meteorological hydrologic environment time sequence signal and a corresponding marine effective wave height at a future moment; The effective wave height prediction model comprises a long-period memory network layer, a time step attention layer and a full connection layer which are sequentially connected, wherein the long-period memory network layer learns the denoising time sequence signal to capture long-period dependency relationship therein and output a hidden sta