Effective wave height time sequence intelligent prediction method based on deep learning

Abstract

The invention discloses a significant wave height time sequence intelligent prediction method based on deep learning, and the method comprises the steps: S1, obtaining a wave data set, and carrying out the preprocessing of the wave data set; s2, performing modal decomposition on a time sequence in the processed wave data set by adopting an empirical mode decomposition algorithm, and splicing a modal decomposition result with the processed wave data set to obtain an enhanced time sequence feature; s3, constructing a significant wave height time sequence prediction model, and performing focused learning on frequency components, which have significant influence on the significant wave height, in the enhanced time sequence features through an adaptive fusion mechanism to obtain a trained significant wave height time sequence prediction model; and S4, predicting the significant wave height of the time point needing to be predicted. According to the method, the significant wave height time sequence prediction model can simultaneously capture local details (such as wind speed abrupt change) and a global trend (such as tidal cycle) of wave generation by obtaining enhanced time sequence characteristics; by learning the frequency components which have obvious influence on the significant wave height in the enhanced time sequence characteristics, the complex dynamic state of the marine environment can be modeled more accurately, so that the prediction precision of the significant wave height is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of significant wave height prediction, and particularly to an intelligent prediction method for significant wave height time series based on deep learning. Background Art

[0002] Since the sea surface waves are actually an irregular combination of various waves with different wave heights, periods and traveling directions, the wave height value of a single wave is not representative. Arranging the wave heights in a given wave train from largest to smallest, the average value of the largest 1 / 3 part of the wave heights is called the "Significant Wave Height (SWH)". The significant wave height is one of the most important parameters of ocean waves, which represents the visual average level of the fluctuations and is close to the visually estimated wave height value. The accurate prediction of SWH is of great significance in the fields of ocean engineering, ocean meteorology, navigation safety, exploitation and utilization of ocean resources, and ocean disaster warning, mainly reflected in the following aspects:

[0003] (1) Ocean engineering design and planning: Ocean engineering projects, such as offshore wind farms, coastal protection projects, port and ship designs, etc., need to accurately evaluate the environmental characteristics of the area where they are located. SWH is one of the important indicators for evaluating the impact of ocean waves on these engineering structures, and it directly affects the design parameters of these structures, such as the height of the protective wall and the design strength of offshore wind towers. Therefore, accurate prediction of SWH is of great significance for the design, construction and operation of ocean engineering projects.

[0004] (2) Ocean meteorological prediction and navigation safety: Ocean waves are an important part of ocean meteorological prediction, affecting navigation safety and offshore operations. Especially for ships such as fishing boats, large passenger liners and ocean-going cargo ships, navigation safety is of utmost importance. Accurate prediction of SWH can help ships plan their routes and avoid bad sea conditions, reducing the occurrence of navigation accidents.

[0005] (3) Exploitation and utilization of ocean resources: Industries such as ocean energy and fishery, which exploit and utilize ocean resources, need to consider the impact of ocean waves on their operations. Accurate SWH prediction helps to reasonably arrange operation plans, improve resource utilization efficiency and reduce losses. Especially for ocean energy development projects such as wave energy and offshore wind energy, accurate SWH prediction helps to optimize equipment design and project location selection to achieve the maximum economic benefits.

[0006] (4) Ocean disaster warning: Extreme meteorological conditions such as typhoons, hurricanes, tsunamis and storm surges, which are large-scale ocean meteorological events, may pose a serious threat to coastal areas and ocean engineering structures. Accurate prediction of SWH can early warn of ocean disaster risks, take effective preventive and response measures, and protect the lives and property safety of people in coastal areas.

[0007] The SWH intelligent prediction method based on deep learning has the advantages of strong non-linear learning ability, low computational cost, fast computational speed, etc., and has received extensive attention from researchers in recent years. However, there are still some problems in the existing wave prediction technology: First, traditional models (such as ARIMA, LSTM) directly process the original time series data, are sensitive to non-stationary signals (such as sudden fluctuations in wave data), are vulnerable to noise interference, and it is difficult to effectively capture the multi-scale characteristics of wave generation; Second, models such as RNN / LSTM can only capture unidirectional or bidirectional time dependencies, and it is difficult to model long-term cross-cycle associations (such as the influence of lunar phases). Although Transformer captures long-distance dependencies through the self-attention mechanism, its computational complexity is high, and its sensitivity to local periodic patterns is insufficient; In addition, existing models usually treat all features equally, unable to distinguish the differences caused by different frequency components, resulting in limited prediction accuracy; Finally, traditional methods often rely on artificial experience to select features, lack the combination of domain knowledge, and are prone to introducing redundant noise, affecting the generalization performance of the model. Summary of the Invention

[0008] The present invention provides an effective wave height time series intelligent prediction method based on deep learning to overcome the technical problem of limited prediction accuracy of effective wave height in the prior art.

[0009] In order to achieve the above object, the technical solution of the present invention is:

[0010] S1: Obtain a wave data set, and preprocess the wave data set to obtain a processed wave data set, where the wave data set includes significant wave height, wind speed, wind direction, and wave direction data;

[0011] S2: Use the empirical mode decomposition algorithm to perform mode decomposition on the time series in the processed wave data set, and splice the mode decomposition result with the processed wave data set to obtain enhanced time series features including wave local details and global trends;

[0012] S3: Based on the time series neural network model TimesNet, construct an effective wave height time series prediction model, input the enhanced time series features into the effective wave height time series prediction model, and the effective wave height time series prediction model focuses on learning the frequency components in the enhanced time series features that have a significant impact on the effective wave height through an adaptive fusion mechanism to obtain a trained effective wave height time series prediction model;

[0013] S4: Based on the trained effective wave height time series prediction model, predict the significant wave height at the required prediction time point.

[0014] Furthermore, the empirical mode decomposition algorithm is used to perform mode decomposition on the time series in the processed wave dataset, and the mode decomposition result is spliced with the processed wave dataset to obtain enhanced time series features, including:

[0015] S21: Let the time series be x(t), and the local maximum and minimum values of the time series x(t) are obtained by comparison;

[0016] S22: Based on the local maximum and minimum values of the time series x(t) and using the cubic spline difference method, the upper and lower envelope lines of the time series x(t) are obtained, and the average envelope line of the time series x(t) is calculated according to the upper and lower envelope lines. The calculation formula is:

[0017]

[0018] where, e + (t) represents the upper envelope line of the time series, e - (t) represents the lower envelope line of the time series, and m1(t) represents the average envelope line;

[0019] S23: Subtract the average envelope line from the time series x(t) to obtain a new signal with low-frequency components removed It is expressed as:

[0020]

[0021] S24: Judge whether it meets the definition conditions of IMF. If it meets, execute S25; otherwise, obtain the next time series as the new x(t) until k that meets the definition conditions of IMF is obtained. k is the number of cycles;

[0022] The definition conditions of the IMF are that the difference between the number of extreme points and the number of zero-crossing points of the time series x(t) is not greater than 1, and the mean value of the upper and lower envelopes of the time series x(t) at any time is 0;

[0023] S25: Set the first-order IMF composite component of the time series x(t) as:

[0024]

[0025] where, c1(t) represents the first-order IMF composite component of the time series x(t);

[0026] S26: Subtract c1(t) from the time series x(t) to obtain the residual component r1(t):

[0027] r1(t) = x(t) - c1(t) (14)

[0028] S27: Determine whether c1(t) or r1(t) satisfies any of the following conditions: c1(t) is less than the set termination threshold, or the residual component r1(t) is less than the set termination threshold, or the residual component r1(t) is a monotonic function, or the residual component r1(t) is a constant. If any condition is satisfied, execute S28; otherwise, continue to obtain the next time series and use it as the new x(t) until any of the following conditions are met: the nth-order IMF composite component c n (t) is less than the set termination threshold, or the residual component r n (t) is less than the set termination threshold, or the residual component r n (t) is a monotonic function, or the residual component r n (t) is a constant;

[0029] S28: Obtain the decomposition result of the time series x(t), expressed as:

[0030]

[0031] where r n (t) is the residual term, reflecting the average trend or average value of the signal; c i (t) is a data sequence with the characteristics of an intrinsic mode function;

[0032] S29: Concatenate the data sequence with the characteristics of the intrinsic mode function and the processed wave data set in the feature dimension to obtain enhanced time series features.

[0033] Furthermore, the effective wave height time series prediction model constructed based on the time series neural network model TimesNet includes the processing procedures of S31 - S34, and the processing procedures of S31 - S34 are as follows:

[0034] S31: Convert the enhanced time series features into a two-dimensional tensor and obtain the peak frequency;

[0035] S32: Use the two-dimensional convolution module Inception to extract the two-dimensional time series features in the two-dimensional tensor , expressed as:

[0036]

[0037] S33: Return the two-dimensional time series features to one-dimensional space to obtain a one-dimensional representation for information aggregation, expressed as:

[0038]

[0039] where Trunc(·) represents the operation of removing the supplemented 0;

[0040] S34: Use an adaptive fusion mechanism to perform dynamic weighted summation on the one-dimensional representation and its matched peak frequency intensity to obtain the significant wave height. The formula is:

[0041]

[0042] where, represents the one-dimensional representation corresponding to the peak frequency intensity, represents the finally output significant wave height.

[0043] Furthermore, in S31, converting the one-dimensional tensor into a two-dimensional tensor includes:

[0044] S311: Extract the frequency and period of the enhanced temporal features, including

[0045] Using the fast Fourier transform method to transform each one-dimensional time series feature from the time domain to the frequency domain to obtain the frequency spectrum;

[0046] Extract the peak frequency from the frequency spectrum, and the period of the signal corresponding to the peak frequency;

[0047] S312: According to the period, divide the one-dimensional tensor by the period length to obtain several segmented segments, and adjust the sequence lengths of the several segments through zero-padding method to make them adapt to the number of rows and columns of the two-dimensional structure, so as to obtain the two-dimensional tensor

[0048] Furthermore, the preprocessing of the wave dataset includes:

[0049] Eliminate the outliers in the wave dataset according to the set standard;

[0050] Use the standard deviation normalization method to perform normalization processing on the wave dataset after eliminating outliers. The formula for normalization processing is:

[0051]

[0052] where x is the wave dataset after eliminating outliers, x' is the data after normalization, μ is the mean of the wave dataset after eliminating outliers, and σ is the standard deviation of the wave dataset after eliminating outliers.

[0053] Beneficial effects: The combination of empirical mode decomposition (EMD) and the effective wave height time series prediction model in the present invention can achieve accurate prediction of the effective wave height. The empirical mode decomposition algorithm is used to perform mode decomposition on the time series in the processed wave dataset without manual intervention, thereby improving the processing efficiency of time series data. The mode decomposition results are spliced with the processed wave dataset to obtain enhanced time series features, enabling the effective wave height time series prediction model to simultaneously capture the local details (such as sudden changes in wind speed) and global trends (such as tidal cycles) of wave generation. Through the adaptive fusion mechanism of the effective wave height time series prediction model, it focuses on learning the frequency components that have a significant impact on the effective wave height in the enhanced time series features, and can more accurately model the complex dynamics of the marine environment, thereby improving the prediction accuracy of the effective wave height. Description of the Drawings

[0054] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0055] Figure 1 It is a flowchart of an intelligent prediction method for the effective wave height time series based on deep learning in the present invention;

[0056] Figure 2 It is a comparison chart of the prediction results of each model in the embodiments of the present invention. Detailed Embodiments

[0057] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0058] This embodiment provides an intelligent prediction method for the effective wave height time series based on deep learning, as Figure 1 shown, and the specific steps include:

[0059] S1: Obtain a wave dataset, and preprocess the wave dataset to obtain a processed wave dataset, where the wave dataset includes effective wave height, wind speed, wind direction, and wave direction data;

[0060] S2: Use the empirical mode decomposition algorithm to perform mode decomposition on the time series in the processed wave dataset, and splice the mode decomposition result with the processed wave dataset to obtain enhanced time series features that include wave local details and global trends.

[0061] S3: Based on the time series neural network model TimesNet, construct an effective wave height time series prediction model, and input the enhanced time series features into the effective wave height time series prediction model. The effective wave height time series prediction model focuses on learning the frequency components in the enhanced time series features that have a significant impact on the effective wave height through an adaptive fusion mechanism to obtain a trained effective wave height time series prediction model.

[0062] S4: Based on the trained effective wave height time series prediction model, predict the effective wave height at the required prediction time point.

[0063] In a specific embodiment, the preprocessing of the wave dataset includes:

[0064] Eliminate the outliers in the wave dataset according to a set standard.

[0065] Specifically, in this embodiment, the data points that deviate from the normal statistical distribution, that is, the values outside the range of the mean ± 3 times the standard deviation or the unreasonable values in terms of physical meaning, are regarded as outliers.

[0066] Use the standard deviation normalization method, that is, the Z-score normalization method, to normalize the wave dataset after eliminating outliers. This method is a method of converting the feature values of the dataset into a standard normal distribution (mean of 0 and standard deviation of 1).

[0067] Specifically, use the standard deviation normalization method to perform normalization processing on the wave dataset after eliminating outliers. The formula for the normalization processing is:

[0068]

[0069] where x is the wave dataset after eliminating outliers, x′ is the data after normalization, μ is the mean of the wave dataset after eliminating outliers, and σ is the standard deviation of the wave dataset after eliminating outliers.

[0070] In a specific embodiment, use the empirical mode decomposition algorithm EMD to perform mode decomposition on the time series in the processed wave dataset. The empirical mode decomposition algorithm is a signal processing method designed for non-linear and non-stationary time series. By performing mode decomposition on the time series through EMD, the original signal is decomposed into the sum of IMFs determined by the local feature time scales of the signal, and the mode decomposition result is spliced with the processed wave dataset to obtain enhanced time series features including:

[0071] S21: Let the time series be x(t), and obtain the local maximum and minimum values of x(t) by comparison;

[0072] S22: Based on the local maximum and minimum values of the time series x(t) and using the cubic spline difference method, obtain the upper and lower envelope lines of the time series x(t), and calculate the average envelope line of the time series x(t) according to the upper and lower envelope lines. The calculation formula is:

[0073]

[0074] where, e + (t) represents the upper envelope line of the time series, e - (t) represents the lower envelope line of the time series, and m1(t) represents the average envelope line;

[0075] Specifically, perform cubic spline interpolation on the local maximum points to generate a smooth curve covering all local maximum points, which is the upper envelope line; perform cubic spline interpolation on the local minimum points to generate another smooth curve covering all local minimum points, which is the lower envelope line, and take the average of the corresponding point values of the upper envelope line and the lower envelope line to obtain the average envelope line.

[0076] S23: Subtract the average envelope line from the time series x(t) to obtain a new signal with low-frequency components removed It is expressed as:

[0077]

[0078] S24: Judge whether it meets the definition conditions of IMF. If it meets, execute S25; otherwise, obtain the next time series as the new x(t) until obtaining k is the number of loop times;

[0079] The definition conditions of the IMF are that the difference between the number of extreme points and the number of zero-crossing points of the time series x(t) is not greater than 1, and the mean value of the upper and lower envelopes of the time series x(t) at any moment is 0;

[0080] S25: Set the first-order IMF combined component of the time series x(t) as:

[0081]

[0082] where, c1(t) represents the first-order IMF combined component of the time series x(t);

[0083] S26: Subtract c1(t) from the time series x(T) to obtain the residual component r1(t):

[0084] r1(t) = x(t) - c1(t) (25)

[0085] S27: Determine whether c1(t) or r1(t) satisfies any of the following conditions: c1(t) is less than the set termination threshold, or the residual component r1(t) is less than the set termination threshold, or the residual component r1(t) is a monotonic function, or the residual component r1(t) is a constant. If any condition is satisfied, execute S28; otherwise, continue to obtain the next time series and use it as the new x(t) until any of the following conditions is satisfied: the nth-order IMF combined component c n (t) is less than the set termination threshold, or the residual component r n (t) is less than the set termination threshold, or the residual component r n (t) is a monotonic function, or the residual component r n (t) is a constant;

[0086] S28: Obtain the decomposition result of the time series x(t), expressed as:

[0087]

[0088] where r n (t) is the residual term, reflecting the average trend or average value of the signal; c i (t) is a data sequence with the characteristics of the intrinsic mode function;

[0089] S29: Concatenate the data sequence with the characteristics of the intrinsic mode function and the processed wave data set in the feature dimension to obtain enhanced time series features.

[0090] Specifically, decompose the time series x(t) through EMD to obtain multi-scale intrinsic mode function IMF components and a residual term. These IMFs and residual terms jointly characterize the different frequency components and trend information of the original signal. These decomposition results are used to enhance the representation ability of time series data and provide richer input features for the subsequent TimesNet deep learning model.

[0091] In a specific embodiment, the effective wave height time series prediction model constructed based on the time series neural network model TimesNet includes the S31 - S34 processing procedures, and the S31 - S34 processing procedures are as follows:

[0092] S31: Since the enhanced time series feature is a one-dimensional tensor, it is necessary to convert the one-dimensional tensor into a two-dimensional tensor and obtain the peak frequency;

[0093] Specifically, the enhanced temporal features, that is, the original one-dimensional time series structure can only represent the changes between adjacent time points. To solve this problem, TimesNet expands the time series changes into a two-dimensional structure, which can capture the changes within and between periods, thus having more advantages in terms of representation ability and facilitating subsequent representation learning.

[0094] S32: Use the two-dimensional convolutional module Inception to extract the two-dimensional tensor in the two-dimensional time series features, expressed as:

[0095]

[0096] S33: Return the two-dimensional time series features to one-dimensional space to obtain a one-dimensional representation for information aggregation, expressed as:

[0097]

[0098] Among them, Trunc(·) represents the operation of removing the supplemented 0s;

[0099] S34: Use an adaptive fusion mechanism to weightedly sum the one-dimensional representation

[0100]

[0101] Among them, represents the one-dimensional representation corresponding to the peak frequency intensity, represents the finally output significant wave height.

[0102] Through the above design, TimesNet completes the time-varying modeling process of extracting two-dimensional time series features from multiple periods and then performing adaptive fusion.

[0103] In a specific embodiment, in S31, converting the one-dimensional tensor into a two-dimensional tensor includes:

[0104] S311: Extract the frequency and period of the enhanced temporal features, including:

[0105] Use the fast Fourier transform method to convert each one-dimensional time series feature from the time domain to the frequency domain to obtain a spectrum, with the formula:

[0106] A = Avg(Amp(FFT(X 1D )))

[0107] f1,...,f k = argTopk(A)

[0108]

[0109] Among them, FFT(·) and Amp(·) represent the FFT operation and the calculation operation on the amplitude value. A represents the amplitude of each frequency, which is the average value in the C-dimensional data calculated by Avg(·). Considering the sparsity of the frequency domain, to avoid the noise brought by meaningless high frequencies, only k amplitude values are selected to obtain the most significant frequencies {f1,…,f k}, and the amplitude of the unnormalized amplitude is Among them, k is a hyperparameter, and the selected frequencies also correspond to k cycle lengths. Due to the conjugacy of the frequency domain, only the frequencies within are considered here. The above formula can be summarized as:

[0110] A,{f1,…,f k},{p1,…,p k}=Period(X 1D )

[0111] Extract the peak frequencies from the said spectrum, and the periods of the signals corresponding to the peak frequencies;

[0112] According to the selected peak frequencies {f1,…,f k} and the corresponding period lengths {p1,…,p k}, reshape the one-dimensional time series X 1D into multiple two-dimensional tensors through the following equation

[0113]

[0114] where Padding(·) is to extend the time series by zeros along the time dimension to make it compatible with , p i and f i represent the number of rows and columns of the transformed two-dimensional tensor respectively.

[0115] Specifically, TimesNet is a stack composed of multiple TimesBlocks. In each TimesBlock, first find the different periods in the data through FTT, and then reshape the one-dimensional tensor into a two-dimensional tensor according to the periods and transmit it to the two-dimensional convolutional module Inception.

[0116] S312: According to the said period, divide the one-dimensional tensor by the period length to obtain several segmented segments, and adjust the sequence lengths of the several segmented segments through the zero-extension method to make it adapt to the number of rows and columns of the two-dimensional structure, so as to obtain a two-dimensional tensor

[0117] For example, if the period is TT, the time series is divided into one row every TT time steps to form a two-dimensional tensor with the number of rows × the number of columns. The number of rows is determined by the number of periods, and the number of columns is determined by the period length.

[0118] In this embodiment, buoy wave data is obtained and wave elements are screened. Appropriate prediction factors are selected to construct an input wave data set. The buoy wave data used in the present invention is collected from the National Data Buoy Center (NDBC) of the United States. The following four data features are selected as the input for the significant wave height time series prediction model: significant wave height (WVHT), wind speed (WSPD), wind direction (WDIR), and wave direction (MWD). Table 1 details the definitions of these wave features and the sampling intervals.

[0119] Table 1:

[0120] Variable Name Definition Unit Sampling Interval WVHT Significant Wave Height m 1 hour WSPD Wind Speed m / s 10 minutes WDIR Wind Direction Degree 10 minutes MWD Wave Direction Degree 1 hour

[0121] Several key parameters of the significant wave height time series prediction model are set. The specific hyperparameter settings are shown in Table 2.

[0122] Table 2:

[0123]

[0124] To verify the effectiveness of the method proposed in this embodiment, the performance of the significant wave height time series prediction model EMD-TimesNet in this embodiment is compared with that of other models. Table 3 lists the specific values of MAE, MAPE, RMSE, and CC at buoy No. 41010 for each model.

[0125] Table 3:

[0126]

[0127] Such as Figure 2As shown, from the trends of various indicators, the prediction performance of each model is as follows: EMD-TimesNet > TimesNet > Autoformer > Transformer ≈ CNN-BiLSTM-Attention. From the specific values, as shown in Table 3, when the prediction time is 1h, the prediction performance of each model is very excellent. Although the accuracy of CNN-BiLSTM-attention is the worst, the CC can still reach 0.9408, showing a relatively high correlation. Considering the four evaluation indicators comprehensively, EMD-TimesNet performs the best, and all its evaluation indicators are the best among the five models. The MAE is 0.0348m, the RMSE is 0.0494m, the MAPE is 2.51%, and the CC is 0.9936, showing a very strong correlation and excellent prediction performance. In the 1h prediction, the MAE of EMD-TimesNet is 0.0348m, which is 21.26%, 50.36%, 52.72%, and 65.68% lower than that of TimesNet, Autoformer, Transformer, and CNN-BiLSTM-Attention respectively; the RMSE of EMD-TimesNet is 0.0494m, which is 27.46%, 48.54%, 58.10%, and 67.15% lower than that of TimesNet, Autoformer, Transformer, and CNN-BiLSTM-Attention respectively; the CC of EMD-TimesNet is 0.9936, which is 0.59%, 1.81%, 3.11%, and 5.61% higher than that of TimesNet, Autoformer, Transformer, and CNN-BiLSTM-Attention respectively; when the prediction time is 3h, the performance of EMD-TimesNet is still the best, and all its evaluation indicator values are also the best among the five models. Its MAE is 0.0653m, the RMSE is 0.0982m, the MAPE reaches 4.58%, and the CC is 0.9747. The indicators of Transformer and CNN-BiLSTM-Attention are similar, and the prediction accuracy of these two models for SWH is basically the same; when the prediction time is 6h, EMD-TimesNet still performs the best, and all its evaluation indicators are better than the other four models. The MAE is 0.1026m, the RMSE is 0.1573m, the MAPE is 7.12%, and the CC is 0.9352. The indicators of Transformer and CNN-BiLSTM-Attention are very close, and their prediction accuracies for the significant wave height are roughly the same.

[0128] The effective wave height time series prediction model proposed in this embodiment has more advantages than the prior art in predicting the effective wave height, mainly due to its unique technical characteristics and high adaptability to ocean time series data. First, Empirical Mode Decomposition (EMD) can adaptively decompose non-stationary wave data into multiple Intrinsic Mode Functions (IMFs) and a residual term. Each IMF represents the fluctuation characteristics at different time scales (such as high-frequency short-period fluctuations, low-frequency long-period trends), and the residual term reflects the overall trend. This decomposition not only effectively removes noise but also retains the physical meaning of the signal. At the same time, by splicing the IMFs with the processed wave data sets (such as wind speed, wave direction), multi-scale input features are formed, enabling the model to capture both the local details of wave generation (such as sudden changes in wind speed) and the global trends (such as tidal cycles). In contrast, traditional models (such as ARIMA, LSTM) directly process the original time series data, are sensitive to non-stationary signals, and are vulnerable to noise interference. The decomposition ability of EMD significantly improves the physical interpretability and stability of the features. Second, the TimesNet framework extracts significant periods from one-dimensional time series through the Fast Fourier Transform (FFT), reshapes the data into a two-dimensional tensor according to the period length, and uses two-dimensional convolution (such as Inception module) to capture the correlations between adjacent time series points and similar points within the period simultaneously, thus more efficiently modeling the periodic fluctuations of waves (such as diurnal variations, tidal patterns). This multi-period modeling ability is superior to the unidirectional time dependence of RNN / LSTM and the high computational complexity of Transformer. In addition, this model dynamically weights the one-dimensional representations corresponding to different frequency components (extracted by FFT) through an adaptive fusion mechanism, highlighting the contributions of key frequencies, ensuring that the model focuses on the frequency components that significantly affect the prediction target (effective wave height). Existing models (such as XGBoost, ordinary CNN) usually treat all features equally and cannot distinguish the contribution differences of different frequency components. The above characteristics enable this model to more accurately model the complex dynamics of the ocean environment in the effective wave height prediction task. Compared with traditional ocean wave numerical models and statistical predictions, the present invention has significant advantages in dealing with complex time series relationships, automatic feature extraction, model adaptability, and processing large-scale data, and is significantly superior to the prior art in terms of prediction accuracy and generalization performance.

[0129] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some or all of the technical features. However, such modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent prediction of effective wave height time series based on deep learning, characterized in that: The specific steps include: S1: Acquire a wave data set, and preprocess the wave data set to obtain a processed wave data set, wherein the wave data set includes significant wave height, wind speed, wind direction, and wave direction data; S2: using an empirical mode decomposition algorithm to perform mode decomposition on the time series in the processed wave data set, and splicing the mode decomposition result with the processed wave data set to obtain enhanced time series features including local details and global trends of the waves; S3: constructing a significant wave height time series prediction model based on the time series neural network model TimesNet, inputting the enhanced time series features into the significant wave height time series prediction model, and the significant wave height time series prediction model focuses on learning the frequency components in the enhanced time series features that have a significant impact on the significant wave height through an adaptive fusion mechanism, so as to obtain a trained significant wave height time series prediction model; S4: Based on the trained effective wave height time series prediction model, predict the effective wave height at the required prediction time point.

2. The method for intelligent prediction of effective wave height time series based on deep learning according to claim 1 is characterized in that: The empirical mode decomposition algorithm is used to perform mode decomposition on the time series in the processed wave data set, and the mode decomposition result is spliced ​​with the processed wave data set to obtain enhanced time series features including: S21: Let the time series be x(t), and obtain the local maximum and minimum values ​​of the time series x(t) by comparison; S22: Based on the local maximum and minimum values ​​of the time series x(t), the upper and lower envelopes of the time series x(t) are obtained using the cubic spline difference method, and the average envelope of the time series x(t) is calculated according to the upper and lower envelopes. The calculation formula is: Among them, e + (t) represents the upper envelope of the time series, e - (t) represents the lower envelope of the time series, and m1(t) represents the mean envelope; S23: Subtract the average envelope from the time series x(t) to obtain a new signal with low-frequency components removed It is expressed as: S24: Determine h1 1 (t) whether it meets the definition conditions of IMF. If so, execute S25. Otherwise, obtain the next time series as the new x(t) until a time series that meets the definition conditions of IMF is obtained. k is the number of cycles; The definition condition of the IMF is that the difference between the number of extreme points and the number of zero-crossing points of the time series x(t) is not greater than 1, and the mean of the upper and lower envelopes of the time series x(t) at any time is 0; S25: Set the first-order IMF composite component of the time series x(t) to: Where c1(t) represents the first-order IMF composite component of the time series x(t); S26: Subtract c1(t) from the time series x(t) to obtain the residual component r1(t): r1(t)=x(t)-c1(t) (4) S27: Determine whether c1(t) or r1(t) satisfies any of the following conditions: c1(t) is less than the set termination threshold or the residual component r1(t) is less than the set termination threshold or the residual component r1(t) is a monotonic function or the residual component r1(t) is a constant. If any of the conditions is met, execute S28; otherwise, continue to obtain the next time series and use it as the new x(t) until any of the following conditions is met: the nth-order IMF combination component c n (t) is less than the set termination threshold or residual component r n (t) is less than the set termination threshold or residual component r n (t) is a monotonic function or residual component r n (t) is a constant; S28: The decomposition result of the time series x(t) is obtained, which is expressed as: Among them, r n (t) is the residual term, reflecting the average trend or average value of the signal; c i (t) is a data sequence with intrinsic mode function characteristics; S29: splicing the data sequence of the intrinsic mode function characteristics with the processed wave data set in the feature dimension to obtain enhanced time series characteristics.

3. The method for intelligent prediction of effective wave height time series based on deep learning according to claim 2 is characterized in that: The effective wave height time series prediction model built based on the time series neural network model TimesNet includes the S31-S34 processing process, and the S31-S34 processing process is: S31: Convert the enhanced time series features into a two-dimensional tensor And get the peak frequency; S32: Use the two-dimensional convolution module Inception to extract the two-dimensional tensor The two-dimensional time series features in are expressed as: S33: Return the two-dimensional time series features to a one-dimensional space to obtain a one-dimensional representation for information aggregation, which is expressed as: in, Trunc(·) means removing the added 0; S34: Adopting an adaptive fusion mechanism to represent the one-dimensional And the peak frequency intensity that matches it is dynamically weighted and summed to obtain the effective wave height, the formula is: in, represents the one-dimensional representation corresponding to the peak frequency intensity, Indicates the effective wave height of the final output.

4. The method for intelligent prediction of effective wave height time series based on deep learning according to claim 3 is characterized in that: In S31, convert a one-dimensional tensor into a two-dimensional tensor include: S311: extracting the frequency and period of the enhanced timing feature, including The fast Fourier transform method is used to transform each one-dimensional time series feature from the time domain to the frequency domain to obtain the spectrum; Extracting the peak frequency and the period of the signal corresponding to the peak frequency from the frequency spectrum; S312: According to the period, the one-dimensional tensor is divided according to the period length to obtain a plurality of divided segments, and the sequence lengths of the plurality of segments are adjusted by a zero extension method to adapt the number of rows and columns of the two-dimensional structure to obtain a two-dimensional tensor 5. The method for intelligent prediction of effective wave height time series based on deep learning according to claim 1 is characterized in that: The preprocessing of the wave data set comprises: Eliminate outliers in the wave data set according to set criteria; The standard deviation standardization method is used to standardize the wave data set that removes outliers. The formula for standardization is: Among them, x is the wave data set with outliers removed, x′ is the standardized data, μ is the mean of the wave data set with outliers removed, and σ is the standard deviation of the wave data set with outliers removed.

Images