A deep learning-based sea wave spectrum prediction method

Abstract

The application provides a sea wave spectrum prediction method based on deep learning, and belongs to the technical field of sea wave spectrum prediction. Based on sea wave spectrum data, a cumulative energy function and relative cumulative energy are calculated, and a frequency domain is fixedly divided into multiple subintervals according to a preset energy percentile threshold. For each subinterval, at least one physical feature of segmented energy, center frequency, peak frequency and spectral width of the interval is extracted, and an original spectrum vector of the interval is combined with the extracted physical feature to form an input feature vector of the subinterval. An independent neural network model is constructed and trained for each subinterval, and the input feature vector of the corresponding subinterval is taken as input, respectively, to output a predicted spectrum vector of the subinterval. The predicted spectrum vectors output by the neural networks of the subintervals are spliced in frequency order, and weighted average processing is performed at the junctions of the subintervals to reconstruct a complete sea wave spectrum prediction result.

Description

TECHNICAL FIELD

[0001] The present application relates to a deep learning-based sea wave spectrum prediction method, belonging to the technical field of sea wave spectrum prediction. BACKGROUND

[0002] Current intelligent prediction methods in the same field have gradually developed from traditional numerical model post-processing to rapid prediction using machine learning or deep learning, but the research focus is still concentrated on the regression modeling of a few spectral elements such as significant wave height, characteristic period, peak frequency, etc. Such methods can improve the calculation efficiency and fitting accuracy of numerical indicators in point time series or regional grid short-term prediction, but the output results are limited to a few scalars, which are difficult to reflect the overall distribution of energy in the frequency domain and the dynamic evolution process of the spectral shape over time, and there is a problem of incomplete information expression. For application scenarios such as ship motion prediction and offshore platform load evaluation that take complete sea wave spectrum as input, there is an information fault between single-element prediction and spectrum-driven applications, which leads to the incompletion of the engineering link or the amplification of error accumulation.

[0003] On the other hand, existing spectrum prediction methods generally use a single network to directly model and output the complete spectrum, which is relatively simple to implement, but the prediction results are prone to deviations in overall spatial distribution, and the description of the main energy concentration area is not accurate, especially in the reproduction of high-energy band position and intensity, which limits the spatial consistency and energy intensity restoration of the predicted spectrum, thereby affecting the reliability and practicality of the results. SUMMARY

[0004] The present application breaks through the problems of insufficient prediction accuracy and insufficient explainability caused by the overall modeling of sea wave spectrum in existing methods, and proposes a deep learning-based sea wave spectrum prediction method. Through long-term statistical analysis of sea wave spectrum, the demarcation frequency corresponding to the cumulative energy percentile is determined, and the frequency domain is divided into three segments of low frequency, medium frequency and high frequency; and an independent neural network is constructed for each frequency segment for modeling and prediction, and the prediction results are spliced and reconstructed into a complete spectrum. This method realizes targeted learning of the evolution characteristics of different frequency bands, effectively improving the prediction accuracy.

[0005] The points to be protected in the present application include: a deep learning-based sea wave spectrum prediction method, which divides the frequency domain into multiple subintervals and splices the output of the complete spectrum distribution after prediction by independent neural networks; and a network input construction mechanism combining segmented energy, center frequency, peak frequency and spectral width, etc. The mechanism enhances the physical explainability and engineering applicability of the prediction results, which is different from traditional non-segmented black box prediction methods, forming a core technical advantage with independent innovation.

[0006] A deep learning-based sea wave spectrum prediction method, comprising the following steps:

[0007] S1, based on the sea wave spectrum data, calculate the cumulative energy function and the relative cumulative energy, and divide the frequency domain into multiple subintervals according to the preset energy percentile threshold;

[0008] S2, for each subinterval, extract at least one physical feature of the subinterval from the segmented energy, center frequency, peak frequency and spectral width, and merge the original spectrum vector of the subinterval with the extracted physical features to form an input feature vector of the subinterval;

[0009] S3, for each subinterval, construct and train an independent neural network model, input the input feature vector corresponding to the subinterval into the neural network model, and output the predicted spectrum vector of the subinterval;

[0010] S4, the predicted spectrum vectors output by each subinterval are spliced in frequency order, and weighted average processing is performed at the junction of the subintervals, and the complete sea wave spectrum prediction result is reconstructed.

[0011] Preferably, in step S1, the time window The scalar spectrum is:

[0012] ;

[0013] The cumulative energy function is introduced:

[0014] ;

[0015] Where f is the effective frequency, And are the lowest and highest effective frequencies, respectively, and v represents the indefinite integral variable;

[0016] The relative cumulative energy is defined as:

[0017] ;

[0018] The average of all sample time windows is obtained, and the relative cumulative average energy :

[0019] .

[0020] Preferably, according to the physical characteristics of the spectral energy distribution, the boundary frequency is set as the frequency points f1 and f2 corresponding to the cumulative energy percentile:

[0021] ;

[0022] Thus, the spectrum is fixedly divided into three segments:

[0023] Low frequency segment: ;

[0024] Mid frequency segment: ;

[0025] High frequency segment: ;

[0026] Preferably, the low frequency segment corresponds to the swell component, the high frequency segment corresponds to the wind sea component, and the mid frequency segment reflects the transition region.

[0027] Preferably, in step S2, in each segment , the segmented energy, the center frequency, the peak frequency and the spectral width are constructed:

[0028] Segmented energy:

[0029] ;

[0030] Center frequency:

[0031] ;

[0032] Peak frequency:

[0033] ;

[0034] Spectral width:

[0035] ;

[0036] wherein, is the energy of the bth segment, f b is the frequency of the bth segment, f b-1 is the frequency of the (b-1)th segment, is the center frequency of the bth segment.

[0037] Preferably, the segmented energy, the center frequency, the peak frequency and the spectral width are merged with the original segmented spectrum vector to form the input features of the bth segment:

[0038] .

[0039] Preferably, in step S4, after the three independent networks respectively output the prediction results of the three segments, the spectrum is spliced and reconstructed:

[0040] ;

[0041] The weighted average of adjacent segments is used at the boundary frequency point: ​

[0042] ;

[0043] wherein, is the spectrum predicted by the first segment;

[0044] Finally, the complete predicted spectrum is obtained, realizing the segmented modeling and full-spectrum prediction under fixed three-segment division.

[0045] Preferably, in step S3, an independent neural network model is constructed and trained for each sub-interval The parameters of the neural network models are completely independent, the neural network models are trained through a loss function, and the optimizer is Adam.

[0046] Compared with the prior art, the present application has the following beneficial effects:

[0047] The first advantage of the present application is that a sea wave spectrum modeling method based on fixed frequency segmentation is proposed. By dividing the frequency domain into three segments of low frequency, medium frequency and high frequency, and constructing an independent neural network for each segment for prediction, and then splicing to obtain the complete frequency spectrum, the evolution characteristics of different frequency bands can be captured. Compared with the whole modeling method without segmentation, the accuracy and stability of the spectrum level prediction are significantly improved.

[0048] On the other hand, the present application introduces segmented energy, center frequency, peak frequency and spectral width in feature construction, thereby enhancing the physical interpretability of the model. For ship motion, ocean engineering structure response and other applications that rely on frequency spectrum as input, the present application can provide more reliable and detailed environmental prediction information, and has significant engineering application value and popularization significance. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 is the flowchart of the present application. DETAILED DESCRIPTION

[0050] The present application will be described in detail below with specific embodiments. The following examples will help those skilled in the art to further understand the present application, but do not limit the present application in any form. It should be noted that for those skilled in the art, without departing from the concept of the present application, a number of changes and improvements can be made. These all belong to the protection scope of the present application.

[0051] As Figure 1 shown, it is the flowchart of the sea wave spectrum prediction method based on deep learning of the present application, including the following steps:

[0052] (1) Fixed segmentation modeling of spectrum and feature construction

[0053] To realize the intelligent prediction of sea wave spectrum, the present application firstly divides the frequency domain into several fixed intervals, so that the spectrum can be expressed and modeled segment by segment. Set the scalar spectrum of time window as:

[0054]

[0055] where f is the effective frequency, and fmin and fmax are the lowest and highest effective frequencies, respectively.

[0056] To ensure the physical consistency and statistical stability of the segments, the cumulative energy function is introduced:

[0057]

[0058] and the relative cumulative energy is defined as :

[0059]

[0060] In the long-term statistical sense, the relative cumulative average energy is obtained by averaging all sample time windows :

[0061]

[0062] According to the physical characteristics of the spectral energy distribution, the dividing frequencies are set as the points f1 and f2 corresponding to the cumulative energy percentiles:

[0063]

[0064] Therefore, the spectrum is divided into three segments:

[0065] Low-frequency segment:

[0066] Mid-frequency segment:

[0067] High-frequency segment:

[0068] This division can ensure that the low-frequency segment mainly corresponds to the swell component, the high-frequency segment mainly corresponds to the wind sea component, and the mid-frequency segment reflects the transition region of the two, thereby having clear physical interpretability.

[0069] (2) Segment feature extraction and network-driven expression

[0070] In each segment ​​​​​​​​​​Within, a plurality of features are constructed to drive the prediction model, including: segment energy, center frequency, peak frequency, and spectral width.

[0071] Segment Energy:

[0072] ;

[0073] is the energy of the bth segment. b is the frequency of the bth segment. b-1 is the frequency of the b-1th segment.

[0074] Center Frequency:

[0075] ;

[0076] is the center frequency of the bth segment.

[0077] Peak Frequency:

[0078] ;

[0079] is the peak frequency of the bth segment.

[0080] Spectral Width:

[0081] ;

[0082] is the spectral width of the bth segment.

[0083] The above features and the original segment spectral vector are combined to form the input features of the bth segment :

[0084] ;

[0085] Subsequently, an independent neural network model is constructed for each sub-interval , and the parameters of the neural network models are completely independent. Each model predicts the spectrum of one of the three segments. The neural network model structure includes an input layer, a hidden layer, and an output layer. The input layer receives the spectrum vector and its feature quantity of the sub-interval, and the input dimension varies with the length of the frequency segment. The hidden layer uses three fully connected layers, with node numbers of 128, 64, and 32, respectively, and all use ReLU activation functions. The output layer outputs the predicted spectrum vector , is the number of predicted time points, and the dimension is the same as the original segment spectral vector ​The same, the activation function is a linear function. The loss function selected for training is mean square error. The optimizer is Adam. The initial learning rate is 0.001, and an exponential decay strategy is adopted (decay by 0.9 times every 20 epochs). The batch size is 32, and the training is performed for 100 rounds.

[0086] (3) Spectrum splicing reconstruction

[0087] After the three independent networks output three segments of prediction results, the spectrum splicing reconstruction is performed:

[0088] ;

[0089] wherein, is the spectrum obtained by the prediction of the i-th segment.

[0090] To eliminate the numerical discontinuity of the splicing boundary, the weighted average of adjacent segments is used at the boundary frequency point:

[0091] ;

[0092] wherein, is the spectrum at the junction of two segments, is the spectrum at the open boundary, is the spectrum at the closed boundary.

[0093] Finally, the complete predicted spectrum is obtained, realizing the segmented modeling and full-spectrum prediction under the fixed three-segment division.

[0094] Table 1

[0095]

[0096] To compare the advantages of the present application, a neural network is constructed to directly predict the sea wave spectrum without segmented modeling, and the results are shown in Table 1. The results show that, compared with the non-segmented prediction method, the segmented prediction method of the present application has obvious advantages in overall performance: the root mean square error of the predicted spectrum is significantly reduced, and the correlation coefficient is significantly improved, verifying the effectiveness and superiority of the method in intelligent prediction of sea wave spectrum.

[0097] The specific embodiments of the present application are described above. It should be understood that the present application is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essential content of the present application. In the case of no conflict, the embodiments of the present application and the features in the embodiments can be arbitrarily combined with each other.​

Claims

1. A deep learning-based sea wave spectrum prediction method, characterized in that, The method comprises the following steps: S1. Based on the sea wave spectrum data, calculate the cumulative energy function and the relative cumulative energy, and divide the frequency domain into multiple subintervals according to a preset energy percentile threshold; Let S be the scalar spectrum on a time window t t (f) is: S t (f), f e [f min , f max ]; Introducing the cumulative energy function C t (f): where f is the effective frequency, f min and f max are the lowest and highest effective frequencies, respectively, and v represents the indefinite integral variable; Defining the relative cumulative energy R t (f) is: The relative cumulative average energy is averaged over all sample time windows t = 1,..., T According to the physical characteristics of the spectrum energy distribution, the dividing frequency is set as the frequency points f1 and f2 corresponding to the cumulative energy percentiles: f1 = min{f | R(f) ≥ 0.30} Thus, the spectrum is divided into three segments: Low frequency band: f e [f min , f1); Medium frequency segment: f ∈ [f1, f2); High frequency band: f e [f2, f max ]; S2. For each subinterval, extract at least one of the physical characteristics of the segmented energy, the center frequency, the peak frequency and the spectrum width of the subinterval, and merge the original spectrum vector of the subinterval with the extracted physical characteristics to form an input feature vector of the subinterval; S3. Construct and train an independent neural network model for each subinterval, input the input feature vector of the corresponding subinterval into the neural network model, and output the predicted spectrum vector of the subinterval; S4. The predicted spectrum vectors output by each subinterval are spliced in frequency order, and weighted average processing is performed at the junction of the subintervals to reconstruct the complete sea wave spectrum prediction result.

2. The deep learning-based sea wave spectrum prediction method according to claim 1, characterized in that, The low frequency segment corresponds to the swell component, the high frequency segment corresponds to the wind sea component, and the medium frequency segment reflects the transition region. 3.The deep learning-based sea wave spectrum prediction method according to claim 1, characterized in that, In step S2, within each segment b ∈ {1, 2, 3}, the segmented energy, the center frequency, the peak frequency and the spectrum width are constructed: Segmented energy: Center frequency: Peak frequency: Spectrum width: where E b (t) is the energy of the bth segment, f b is the frequency of the bth segment, f b-1 is the frequency of the b-1th segment, f c,b (t) is the center frequency of the bth segment.

4. The deep learning-based sea wave spectrum prediction method according to claim 3, characterized in that, The segmented energy, center frequency, peak frequency, and spectral width are combined with the original segmented spectral vector x b (t) = {S t (f)} to form the input features for the bth segment 5. The deep learning-based sea wave spectrum prediction method according to claim 4, characterized in that, In step S4, after the three independent networks output the prediction results of the three segments, the spectrum is spliced and reconstructed: At the dividing frequency point, the weighted average of the adjacent segments is used: wherein is the spectrum predicted for the b-th segment; Final predicted spectrum is obtained Segment modeling and full spectrum prediction are implemented under fixed three-segment division.

6. The deep learning-based sea wave spectrum prediction method according to claim 5, characterized in that, In step S3, an independent neural network model is constructed and trained for each sub-interval The parameters between the neural network models are completely independent, the neural network model is trained by a loss function, and the optimizer is Adam.

Images