Sea wave spectrum prediction method based on deep learning

Abstract

The invention provides a sea wave spectrum prediction method based on deep learning, and belongs to the technical field of sea wave spectrum prediction, and the method comprises the steps: calculating an accumulated energy function and relative accumulated energy based on sea wave spectrum data, and fixedly dividing a frequency domain into a plurality of subintervals according to a preset energy percentile threshold value; for each sub-interval, extracting at least one physical feature of segment energy, center frequency, peak frequency and spectral width of the interval, and combining the original spectral vector of the interval with the extracted physical features to form an input feature vector of the sub-interval; an independent neural network model is constructed and trained for each subinterval, the input feature vector of the corresponding subinterval is used as input, and the predicted spectrum vector of the subinterval is output; and splicing the prediction spectrum vectors output by the neural network of each subinterval according to a frequency sequence, performing weighted average processing at the junction of the subintervals, and performing reconstruction to obtain a complete sea wave spectrum prediction result.

Description

Technical Field

[0001] This invention relates to a deep learning-based method for predicting ocean wave spectrum, belonging to the field of ocean wave spectrum prediction technology. Background Technology

[0002] Current intelligent forecasting methods in this field have evolved from traditional numerical model post-processing to rapid prediction using machine learning or deep learning. However, research still focuses on regression modeling of a few spectral elements such as significant wave height, characteristic period, and peak frequency. While these methods can improve the computational efficiency and fitting accuracy of numerical indicators in short-term forecasts of point time series or regional grid points, the output is limited to a few scalars, making it difficult to reflect the overall energy distribution in the frequency domain and the dynamic evolution of the spectral shape over time, resulting in incomplete information representation. For applications such as ship motion forecasting and offshore platform load assessment, which use the complete wave spectrum as input, an information gap exists between single-element forecasting and spectrum-driven applications, leading to incomplete engineering linkages or amplified accumulated errors.

[0003] On the other hand, existing spectrum prediction methods generally use a single network to directly model and output the complete spectrum. Although this is relatively simple to implement, the prediction results are prone to deviations in the overall spatial distribution and are not accurate enough in characterizing the main energy concentration areas. In particular, there are shortcomings in the reproduction of high-energy band positions and intensities, resulting in limited spatial consistency and energy intensity restoration of the predicted spectrum, which affects the reliability and practicality of the results. Summary of the Invention

[0004] This patent overcomes the shortcomings of existing methods that suffer from insufficient prediction accuracy and interpretability due to overall modeling of the wave spectrum. It proposes a deep learning-based wave spectrum prediction method. By conducting long-term statistical analysis of the wave spectrum, the boundary frequencies corresponding to the cumulative energy percentiles are determined, dividing the frequency domain into low-frequency, mid-frequency, and high-frequency bands. An independent neural network is then constructed for each frequency band to model and predict the wave spectrum. The prediction results are then concatenated to reconstruct the complete spectrum. This method achieves targeted learning of the evolutionary characteristics of different frequency bands, effectively improving prediction accuracy.

[0005] The points to be protected in this patent include: a deep learning-based method for predicting ocean wave spectrum, which divides the frequency domain into multiple sub-intervals and then splices the predictions from each sub-interval to output a complete spectrum distribution; and a network input construction mechanism that combines physical characteristics such as segmented energy, center frequency, peak frequency, and spectral width. This mechanism enhances the physical interpretability and engineering applicability of the prediction results, distinguishes it from traditional non-segmented black-box prediction methods, and forms a core technological advantage with independent innovation.

[0006] A deep learning-based method for predicting ocean wave spectra includes the following steps: S1. Based on the wave spectrum data, calculate the cumulative energy function and relative cumulative energy, and divide the frequency domain into multiple sub-intervals according to the preset energy percentile threshold. S2. For each sub-interval, extract at least one physical feature from the segmented energy, center frequency, peak frequency and spectral width of the sub-interval, and merge the original spectrum vector of the sub-interval with the extracted physical features to form the input feature vector of the sub-interval. S3. Construct and train an independent neural network model for each sub-interval, input the input feature vector of the corresponding sub-interval into the neural network model, and output the predicted spectrum vector of the sub-interval. S4. The predicted spectrum vectors output from each sub-interval are concatenated in frequency order, and a weighted average is performed at the boundary of the sub-intervals to reconstruct the complete wave spectrum prediction result.

[0007] Preferably, in step S1, a time window is set. scalar spectrum on for: ; Introducing the cumulative energy function : ; Where f is the effective frequency, and These are the lowest and highest effective frequencies, respectively, and v represents the indefinite integral variable; Define relative cumulative energy for: ; For all sample time windows Calculate the average to obtain the relative cumulative average energy. : .

[0008] Preferably, based on the physical characteristics of the spectral energy distribution, the boundary frequencies are set as the frequency points f1 and f2 corresponding to the cumulative energy percentile: ; Therefore, the spectrum is fixedly divided into three segments: Low frequency band: ; Mid-frequency band: ; High frequency band: ; Preferably, the low-frequency band corresponds to the swell component, the high-frequency band corresponds to the wind-sea component, and the mid-frequency band reflects the transition region.

[0009] Preferably, in step S2, in each segment Internally, construct segmented energy, center frequency, peak frequency, and spectral width: Segmented energy: ; Center frequency: ; Peak frequency: ; Spectral width: ; in, For the energy of segment b, f b It is the frequency of segment b, f b-1 It is the frequency of segment b-1. is the center frequency of segment b.

[0010] Preferably, the segmented energy, center frequency, peak frequency, and spectral width are compared with the original segmented spectral vector. The merger constitutes the first Segment input features : .

[0011] Preferably, in step S4, after the three independent networks output three prediction results respectively, the spectrum is spliced ​​and reconstructed: ; At the boundary frequency point, a weighted average of adjacent segments is used: ; in, For the first The spectrum obtained from segment prediction; Finally, the complete predicted spectrum is obtained. This enables segmented modeling and full-spectrum prediction under a fixed three-segment division.

[0012] 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 using a loss function, and the optimizer is Adam.

[0013] Compared with the prior art, the present invention has the following beneficial effects: The first advantage of this invention lies in its proposed wave spectrum modeling method based on fixed-frequency segmentation. By dividing the frequency domain into three segments—low-frequency, mid-frequency, and high-frequency—and constructing an independent neural network for each segment to perform predictions, the entire spectrum is then stitched together to obtain the complete spectrum. This method can specifically capture the evolutionary characteristics of different frequency bands. Compared to the overall modeling approach without segmentation, it significantly improves the accuracy and stability of spectral level prediction.

[0014] On the other hand, the present invention incorporates physical quantities such as segmented energy, center frequency, peak frequency, and spectral width into the feature construction, thereby enhancing the physical interpretability of the model. For applications such as ship motion and marine engineering structure response that rely on the spectrum as input, the present invention can provide more reliable and detailed environmental forecast information, and has significant engineering application value and promotional significance. Attached Figure Description

[0015] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0016] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the scope of protection of the present invention.

[0017] like Figure 1 The diagram shows a flowchart of the deep learning-based wave spectrum prediction method of the present invention, which includes the following steps: (1) Fixed-segment modeling and feature construction of spectrum To achieve intelligent prediction of ocean wave spectra, this invention first divides the frequency domain into several fixed intervals, allowing the spectrum to be represented segment by segment and modeled segment by segment. A time window is then defined. scalar spectrum on for: ; Where f is the effective frequency, and These are the lowest and highest effective frequencies, respectively.

[0018] To ensure the physical consistency and statistical stability of the segments, a cumulative energy function is introduced. : ; And define the relative cumulative energy as : ; In the long-term statistical sense, for all sample time windows Calculate the average to obtain the relative cumulative average energy. : ; Based on the physical characteristics of the spectral energy distribution, the boundary frequencies are set as points f1 and f2 corresponding to the cumulative energy percentiles: ; Therefore, the spectrum is fixedly divided into three segments: Low frequency band: ; Mid-frequency band: ; High frequency band: ; This division ensures that the low-frequency band mainly corresponds to swell components, the high-frequency band mainly corresponds to wind and sea components, and the mid-frequency band reflects the transition region between the two, thus providing a clear physical interpretation.

[0019] (2) Segmented feature extraction and network-driven expression In each segment Within the model, multiple features are constructed to drive the prediction model, including: piecewise energy, center frequency, peak frequency, and spectral width.

[0020] Segmented energy: ; For the energy of segment b, f b It is the frequency of segment b, f b-1 It is the frequency of segment b-1.

[0021] Center frequency: ; is the center frequency of segment b.

[0022] Peak frequency: ; The peak frequency of segment b Spectral width: ; Let be the spectral width of segment b.

[0023] The above features and the original segmented spectral vector The merger constitutes the first Segment input features : ; Subsequently, an independent neural network model is constructed for each sub-interval. The neural network models are completely independent of each other, each predicting three frequency spectrum segments separately. The neural network model structure includes an input layer, hidden layers, and an output layer. The input layer receives the spectrum vector and its features for each sub-interval; the input dimension depends on the frequency band length. The hidden layers consist of three fully connected layers with 128, 64, and 32 nodes respectively, all using the ReLU activation function. The output layer corresponds to the predicted spectrum vector for that segment. , The predicted time number, dimension, and original piecewise spectral vector. The activation function is the same as above, and the loss function chosen for training is mean squared error. The optimizer is Adam. The initial learning rate is 0.001, and an exponential decay strategy is used (decreasing by a factor of 0.9 every 20 epochs). The batch size is 32, and the training run is 100 epochs.

[0024] (3) Spectral segment splicing and reconstruction After the three independent networks each output three prediction segments, the spectrum is spliced ​​and reconstructed: ; in, For the first The spectrum obtained from segment prediction.

[0025] To eliminate numerical discontinuities at splicing boundaries, a weighted average of adjacent segments is used at the dividing frequency points: ; in, The spectrum at the boundary between the two segments. For the spectrum at the open boundary, This represents the spectrum at the closed boundary.

[0026] Finally, the complete predicted spectrum is obtained. This enables segmented modeling and full-spectrum prediction under a fixed three-segment division.

[0027] Table 1

[0028] To compare the advantages of this invention, a neural network was constructed to directly predict the wave spectrum without segmented modeling. The results are shown in Table 1. The results show that, compared with the non-segmented prediction method, the segmented prediction method of this invention has significant advantages in overall performance: its root mean square error of the predicted spectrum is significantly reduced, and the correlation coefficient is significantly improved, verifying the effectiveness and superiority of this method in intelligent prediction of the wave spectrum.

[0029] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A deep learning-based method for predicting ocean wave spectrum, characterized in that, Includes the following steps: S1. Based on the wave spectrum data, calculate the cumulative energy function and relative cumulative energy, and divide the frequency domain into multiple sub-intervals according to the preset energy percentile threshold. S2. For each sub-interval, extract at least one physical feature from the segmented energy, center frequency, peak frequency and spectral width of the sub-interval, and merge the original spectrum vector of the sub-interval with the extracted physical features to form the input feature vector of the sub-interval. S3. Construct and train an independent neural network model for each sub-interval, input the input feature vector of the corresponding sub-interval into the neural network model, and output the predicted spectrum vector of the sub-interval. S4. The predicted spectrum vectors output from each sub-interval are concatenated in frequency order, and a weighted average is performed at the boundary of the sub-intervals to reconstruct the complete wave spectrum prediction result.

2. The deep learning-based wave spectrum prediction method according to claim 1, characterized in that, In step S1, a time window is set. scalar spectrum on for: ； Introducing the cumulative energy function : ； Where f is the effective frequency, and These are the lowest and highest effective frequencies, respectively, and v represents the indefinite integral variable; Define relative cumulative energy for: ； For all sample time windows Calculate the average to obtain the relative cumulative average energy. : 。 3. The deep learning-based wave spectrum prediction method according to claim 2, characterized in that, Based on the physical characteristics of the spectral energy distribution, the boundary frequencies are set as the frequency points f1 and f2 corresponding to the cumulative energy percentile: ； Therefore, the spectrum is fixedly divided into three segments: Low frequency band: ; Mid-frequency band: ; High frequency band: .

4. The deep learning-based wave spectrum prediction method according to claim 3, characterized in that, The low-frequency band corresponds to the swell component, the high-frequency band corresponds to the wind and sea component, and the mid-frequency band reflects the transition region.

5. The deep learning-based wave spectrum prediction method according to claim 1, characterized in that, In step S2, in each segment Internally, construct segmented energy, center frequency, peak frequency, and spectral width: Segmented energy: ； Center frequency: ； Peak frequency: ； Spectral width: ； in, For the energy of segment b, f b It is the frequency of segment b, f b-1 It is the frequency of segment b-1. is the center frequency of segment b.

6. The deep learning-based wave spectrum prediction method according to claim 5, characterized in that, The segmented energy, center frequency, peak frequency, and spectral width are compared with the original segmented spectral vector. The merger constitutes the first Segment input features : 。 7. The deep learning-based wave spectrum prediction method according to claim 6, characterized in that, In step S4, after the three independent networks output three prediction segments respectively, the spectrum is spliced ​​and reconstructed: ； At the boundary frequency point, a weighted average of adjacent segments is used: ； in, For the first The spectrum obtained from segment prediction; Finally, the complete predicted spectrum is obtained. This enables segmented modeling and full-spectrum prediction under a fixed three-segment division.

8. The deep learning-based wave spectrum prediction method according to claim 7, characterized in that, 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 using a loss function, and the optimizer is Adam.

Images