Wave height prediction model and apparatus

Abstract

The application discloses a wave height prediction model, which comprises a first LGE module, a feature encoder, an encoder, a second LGE module and a decoder. The first LGE module is used for encoding a marine multi-element time sequence related to wave height. The input of the encoder is connected with the output of the first LGE module, and the encoder is used for outputting a high-dimensional feature of the marine multi-element. The second LGE module is used for encoding a matrix sequence spliced by a latter half sequence of the marine multi-element time sequence and a zero matrix sequence. The decoder takes the output of the second LGE module and the high-dimensional feature output by the encoder as input, and outputs a prediction result of the wave height. The encoder is composed of N (N=4) encoding layers. Each encoding layer comprises a first hollow causal convolution self-attention layer, a first residual connection and a normalization layer, a first forward propagation layer and a second residual connection and a normalization layer in sequence.

Description

Technical Field

[0001] This invention belongs to the field of marine resource development technology, and specifically relates to a wave height prediction model and device. Background Technology

[0002] As ocean waves propagate toward the nearshore area, they are significantly influenced by factors such as temperature, wind, seabed topography, shoreline boundaries, and environmental currents (such as coastal currents and tidal currents). They exhibit more complex evolutionary patterns and faster spatiotemporal changes than those in the deep sea and open continental shelf waters, and our research and understanding of them are still not very mature.

[0003] Nearshore waves are one of the most important dynamic factors in the nearshore marine environment, threatening the safety and stability of nearshore structures and causing coastal sediment movement, coastal changes, and nearshore water exchange. The calculation of nearshore waves is of great significance for coastal engineering design, shallow-sea production operations, and nearshore environmental protection. With the continuous development of the coastal economy, human activities in coastal areas are becoming increasingly frequent, the number of coastal engineering projects is increasing, the scale of investment is growing, and the risks of these projects are attracting increasing attention. All of these factors place higher demands on the accurate prediction of nearshore waves and other marine environmental factors. Therefore, providing accurate and practical wave height prediction methods has become an urgent task for coastal engineering, marine engineering, marine and coastal resource research, and military activities in recent years.

[0004] Currently, methods for predicting ocean wave motion can be broadly categorized into two types:

[0005] The first category utilizes the physical properties of wave propagation to simulate and predict the wave propagation process based on wave numerical models, such as the third-generation wave model (WAM) and the nearshore wave simulation method based on WAM (Simulating Wave Nearshore, SWAN). [1-3] .

[0006] The second category is prediction of significant wave height (SWH) based on traditional time series models, such as autoregressive (AR) models, autoregressive moving average (ARMA) models, and autoregressive integrated moving average (ARIMA) models. [4-6] Agrawal [7] Researchers used the ARIMA model to achieve online SWH prediction in different prediction intervals.

[0007] With the development of machine learning, artificial neural networks (ANNs) and support vector machines (SVMs) have begun to be applied to SHW prediction. et al. proposed a method for real-time prediction of SWH based on ANN. Experimental results show that ANN performs well in terms of accuracy and consistency. (Cornejo-bueno) [9,10] Researchers used feedforward neural networks and extreme learning machines (ELM) to predict SWH, achieving strong generalization ability and fast solution speed.

[0008] SVM, based on a well-developed mathematical theory, has a wide range of applications. (Mahjoobi)

[11] Studies by Malekmohamadi et al. have shown that SVM outperforms ANN in certain cases during SWH prediction.

[12] The prediction performance of SVM, ANN, Bayesian networks, and the Adaptive Neural Fuzzy Inference System (ANFIS) was investigated in detail. The results show that the prediction results of ANN, ANFIS, and SVM are all within acceptable ranges, while the prediction results of Bayesian networks are relatively unreliable.

[0009] With the development of Natural Language Processing (NLP), researchers have found that the sequence structure of Recurrent Neural Networks (RNNs) is very suitable for sequence prediction tasks such as weather forecasting. Researchers used LSTM to predict SWH at multiple nearshore sites. Experimental results showed that LSTM can effectively capture the temporal correlation of SWH, and the prediction accuracy is significantly improved compared with numerical models.

[0010] The references involved in this disclosure include:

[0011] [1]GROUP T W. The WAM Model—A Third Generation Ocean Wave PredictionModel[J]. Journal of Physical Oceanography, 1988, 18(12): 1775-1810.

[0012] [2]BOOIJ N, RIS R C, HOLTHUIJSEN L H. A third-generation wave modelfor coastal regions: 1. Model description and validation[J]. Journal ofGeophysical Research: Oceans, 1999, 104(C4): 7649-7666.

[0013] [3]TOLMAN H L. User manual and system documentation of WAVEWATCH IIITM version 3.14[J]. Technical note, 2009, 276:.220

[0014] [4]BOLLERSLEV T. Generalized autoregressive conditionalheteroskedasticity[J]. Journal of Econometrics, 1986, 31(3): 307-327.

[0015] [5]SAID S E, DICKEY D A. Testing for unit roots in autoregressive-moving average models of unknown order[J]. Biometrika, 1984, 71(3): 599-607.

[0016] [6]BOX G E P, PIERCE D A. Distribution of Residual Autocorrelationsin Autoregressive-Integrated Moving Average Time Series Models[J]. Journal ofthe American Statistical Association, 1970, 65(332): 1509-1526.

[0017] [7]AGRAWAL J D, DEO M C. On-line wave prediction[J]. MarineStructures, 2002, 15(1): 57-74.

[0018] [8]DEO M C, JHA A, CHAPHEKAR A S, et al. Neural networks for waveforecasting[J]. Ocean Engineering, 2001, 28(7): 889-898.

[0019] [9]CORNEJO-BUENO L, NIETO-BORGE J C, GARCÍA-DÍAZ P, et al.Significant wave height and energy flux prediction for marine energyapplications: A grouping genetic algorithm – Extreme Learning Machineapproach[J]. Renewable Energy, 2016, 97: 380-389.

[0020]

[10] KUMAR N K, SAVITHA R, AL MAMUN A. Ocean wave height predictionusing ensemble of Extreme Learning Machine[J]. Hierarchical Extreme LearningMachines, 2018, 277: 12-20.

[0021]

[11] MAHJOOBI J, ADELI MOSABBEB E. Prediction of significant waveheight using regressive support vector machines[J]. Ocean Engineering, 2009,36(5): 339-347.

[0022]

[12] MALEKMOHAMADI I, BAZARGAN-LARI MR, KERACHIAN R, et al. Evaluating the efficacy of SVMs, BNs, ANNs and ANFIS in wave heightprediction[J]. Ocean Engineering, 2011, 38(2): 487-497.

[0023]

[13] FAN S, XIAO N, DONG S. A novel model to predict significant waveheight based on long short-term memory network[J]. Ocean Engineering, 2020,205: 107298. Summary of the Invention

[0024] One embodiment of the present invention provides an effective wave height prediction method based on the correlation of local and global features of multiple ocean elements. This method predicts ocean wave height using a wave height prediction model. The model includes:

[0025] The first LGE module performs feature encoding on the time series of multiple ocean elements related to wave height.

[0026] The encoder, whose input is connected to the output of the first LGE module, is used to output high-dimensional features of multiple ocean elements;

[0027] The second LGE module encodes the matrix sequence formed by splicing the latter half of the ocean multi-element time series with the zero matrix sequence;

[0028] The decoder takes the output of the second LGE module and the high-dimensional features of the encoder output as input, and outputs the prediction result for the wave height.

[0029] Both the encoder and decoder include a dilated causal convolutional self-attention layer, a normalization layer, and a forward propagation layer.

[0030] The effective wave height prediction model based on local-global feature correlation of multiple ocean elements in this invention combines ocean elements closely related to wave activity to more effectively learn the changing characteristics of wave height, and uses an extended causal convolutional self-attention mechanism to capture the long-term correlation of ocean multi-element sequences. By introducing temporal information, the model's learning ability is enhanced, achieving accurate prediction of effective wave height. Compared with the latest deep learning models and traditional models, this model has advantages such as high accuracy and high efficiency. Attached Figure Description

[0031] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent from the following detailed description taken in conjunction with the accompanying drawings. Several embodiments of the invention are illustrated in the drawings by way of example and not limitation, wherein:

[0032] Figure 1 A general framework diagram of a prediction model according to one embodiment of the present invention.

[0033] Figure 2 A framework diagram of an LGE module according to one embodiment of the present invention.

[0034] Figure 3 A comparative schematic diagram of different attention mechanisms according to one embodiment of the present invention.

[0035] Figure 4 A schematic diagram of a decoder module according to one embodiment of the present invention.

[0036] Figure 5 A comparison curve of prediction results according to one embodiment of the present invention. Detailed Implementation

[0037] Existing numerical models typically require driving factors such as offshore incident wave conditions for prediction, and their computational efficiency is relatively low. These driving factors include various oceanic elements involved in the propagation from the open sea (seaward side) towards the coast, such as wind field elements and the angle of incidence. Waves are influenced by complex oceanic factors during propagation, but ARIMA only applies to single-factor time series prediction and cannot effectively represent the impact of complex and diverse oceanic factors. Furthermore, the ARIMA model is highly dependent on the stationarity assumption. Therefore, current deep learning-based models cannot capture the long-term correlation of sequences, making it difficult to guarantee model accuracy.

[0038] In the task of predicting ocean wave height, it is necessary to consider both the long-term periodic changes of ocean waves and the short-term changes caused by various environmental and meteorological factors. Therefore, the traditional Transformer has shortcomings when applied to the task of predicting ocean wave height.

[0039] According to one or more embodiments, a wave height prediction model is provided, which consists of a Local-Global Embedding (LGE) module, an encoder module, and a decoder module. The basic units in both the encoder and decoder consist of a dilated causal convolutional self-attention layer (DConv Self-Attention), a residual connection and normalization layer (Add&Norm), and a feed forward layer. The LGE module is used to encode features of multi-element ocean sequences. The dilated convolutional self-attention mechanism in the encoder module can capture the long-term correlation of the sequences and local variation trends. The decoder module is used to decode the high-dimensional features output by the encoder module, and finally connects them to a fully connected layer to obtain the prediction result. The overall model framework is as follows: Figure 1 As shown.

[0040] The LGE module in the wave height prediction model encodes time series data, enabling the model to capture local features among multiple ocean elements and effectively utilize temporal information. The LGE local-global encoding module framework is as follows: Figure 2 As shown. The LGE module implements the following functions:

[0041] 1) Marine multi-element feature encoding: A one-dimensional convolutional neural network is used to encode marine multi-element sequences. Encoding is performed to capture the local correlation features between multiple ocean elements. The encoded output sequence is: .

[0042] 2) Temporal Information Encoding: Attention mechanisms cannot recognize the positional order of the input sequence. Therefore, positional encoding (Position Embedding) is usually needed to artificially add positional relationships to the sequence. Positional encoding can only retain the relative positions of elements in the sequence and cannot utilize the temporal information of the time series. However, for time series problems, the sequence will exhibit a periodic trend as time changes, and in extreme cases, it may also exhibit aperiodic behavior. Therefore, making full use of temporal information plays an important role in improving the performance of time series prediction models. To this end, this disclosure uses Time2Vec to extend positional encoding to continuous time series, adaptively capturing the periodic and aperiodic patterns of the time series. This paper will use Time2vec to sample time... To perform modeling, define as Its expression is:

[0043] (1)

[0044] in, for The Middle One element, It is a periodic activation function; in this paper, a sine function is used. , The frequency and offset of the sine function are both learnable hyperparameters. In Time2Vec, the sine function can be used to capture periodic patterns, and the linear term can be used to capture aperiodic patterns. It's equivalent to using the sine function as a fully connected layer to sample the time series. Mapped to In a 3D space.

[0045] In encoders, traditional self-attention mechanisms calculate the relevance between queries and keys point-by-point (e.g., ...). Figure 3 (as shown in (a)). Therefore, the self-attention module cannot learn local information of the time series. Convolutional neural networks are commonly used in the image domain to capture local features of images. Applying convolutional neural networks to time series prediction tasks can achieve good results in capturing local features.

[0046] When a causal convolutional self-attention mechanism combines a convolutional neural network with a self-attention mechanism, that is, replacing fully connected layers with a causal convolutional neural network to generate the attention mechanism required... and K (e.g.) Figure 3 (b) shows that causal convolutional neural networks (CNNs) enhance the model's ability to capture local features. However, causal convolutional neural networks can only review historical data of a linear scale, and due to their limited receptivity, the model's ability to utilize historical information is insufficient. Although the size of the receptive field in causal convolutional neural networks can be increased by continuously adding layers, this leads to an increase in the number of parameters.

[0047] To address the aforementioned issues, embodiments of this disclosure employ a dilated causal convolutional neural network to generate... and (like Figure 3 (c) shows that, compared to causal convolutional neural networks ( Figure 3 (b) , and its perception increases exponentially with the number of convolutional layers. The convolution kernel in a dilated causal convolutional layer will affect... Perform convolution operations on the elements at the specified positions.

[0048] The dilated causal convolution designed in this disclosure maintains the causal nature of the relationship. The value of a time series at a certain moment is only convolved with the values ​​at previous moments, ensuring that the network does not violate the basic sequential dependencies in wave height prediction during construction and training. At the same time, dilated convolution can better utilize the historical information of the sequence. When calculating the similarity of data at a certain time step, the contextual relationships (such as local trends) of that time step can be used for calculation, which helps to improve the accuracy of prediction.

[0049] The decoder module adopts a similar structure to the encoder module. A mask matrix is ​​added to the causal dilated convolutional self-attention mechanism of the decoder. All its upper triangular elements are set to Used to conceal The matrix contains the part of future information, when the matrix After calculation using the Softmax function, the values ​​of all upper triangular elements become 0, so that future information does not participate in the calculation of attention scores, thus avoiding autoregression.

[0050] Traditional multi-step prediction methods employ iterative prediction, which increases the computational cost of the model and leads to error accumulation as the prediction length increases. Therefore, this paper adopts a generative prediction method, enabling the model to obtain multi-step prediction results with only one decoding operation. The decoder module is as follows: Figure 4 As shown.

[0051] The second half of the ocean multi-element sequence is input using the Encoder module as the starting sequence segment and concatenated with a zero matrix of the same length as the target sequence, as shown in formula (3).

[0052] (3)

[0053] in, Using the initial sequence segment, in the 24-hour and 48-hour prediction tasks, The lengths taken are 6 and 12 respectively. It is a zero matrix with the same length as the target predicted sequence. Predict the timing of the target sequence. Input the concatenated result into the LGE module for encoding, and then output the encoded result... The input is fed into the Decoder module for decoding. During the decoding process, it first passes through the DConv Self-Attention layer. Feature learning is performed, and Add&Norm layers are added to obtain the matrix. Next, MHA is used to calculate the attention matrix between the high-dimensional features output by the Encoder module and matrix Z, and this matrix is ​​then input into the feedforward layer. Finally, a fully connected layer is connected after the Decoder module to linearly map the result of decoding the high-dimensional features from the Encoder module. The latter half of the mapping vector (the target prediction length) is the prediction result.

[0054] This disclosure proposes a Marine multi-elements Local and Global Correlation for Wave Height Prediction (MLG-SWH) model that combines local and global features of marine multi-elements, enabling high-precision prediction of 24-hour and 48-hour SWH. The MLG-SWH model takes significant wave height and related marine multi-elements time series as input, and is based on an encoder-decoder network structure. It emphasizes the temporal information embedding capability of marine multi-elements features through Local-Global Embedding, expands the sensitivity of the self-attention model through dilated causal convolution, and utilizes more historical information to predict the future.

[0055] The technical features and beneficial effects of this invention include:

[0056] 1. Combining local-global feature correlation of multiple ocean elements. Statistical methods are used to screen ocean elements with strong correlation to significant wave height for prediction of significant wave height. Convolutional neural networks are used to capture local correlations between multiple elements, and dilated causal convolutional self-attention mechanism is used to capture long-term correlations and local change trends of sequences.

[0057] 2. Utilizing temporal information. The prediction model uses the Time2Vec method to encode time instead of location encoding as temporal information for the self-attention mechanism. This can effectively learn the changing characteristics of multi-factor wave sequences and improve prediction performance.

[0058] 3. A generative prediction method was adopted to achieve multi-step prediction of SWH, which effectively reduced the impact of error accumulation in point-by-point iterative multi-step prediction.

[0059] Therefore, compared with deep learning networks such as ARIMA, LSTM and improved Transformer, this invention can not only better capture the long-term correlation of SWH, but also effectively extract the local correlation between SWH and multiple marine elements.

[0060] To illustrate the practical effect of the wave height prediction model of this invention, an example is given below. The operating environment requirements of the model are shown in Table 1.

[0061]

[0062] The data used in the experiment came from the website of the National Oceanic and Atmospheric Administration (NOAA). This paper selected site 42019 in the Gulf of Mexico (27.910°N, 95.345°W) and site 41025 in Pamlico Bay (35.010°N, 75.454°W) as the research objects, both of which are surface buoy sites. Site 42019, located in the Gulf of Mexico, has a water depth of 83.5 meters. Situated in the tropical and subtropical zone, it experiences high temperatures and abundant rainfall, and experiences a hurricane season annually, resulting in significant fluctuations in wave height. Site 41025, located in Pamlico Bay, has a water depth of 59.4 meters and is the largest lagoon on the east coast of the United States, where wave fluctuations are relatively stable. Data collection for both sites spanned from January 1, 2018 to December 31, 2020, with a sampling interval of 1 hour. The data included wind speed (WSPD), mean wave period (APD), water temperature (WT), and air temperature (AT).

[0063]

[0064] The formation of ocean waves is influenced by a variety of complex geographical and natural factors. From a physical perspective, the growth state of wind-driven waves is closely related to three elements: wind speed, wind duration, and wind region. Swells generally have long periods; nearshore waves are also affected by seabed topography and coastline. This paper uses a data-driven approach to study the prediction of significant wave height (SWH), selecting factors related to SWH as much as possible within the existing data range (topographic data is unavailable). Therefore, statistical methods are used to assist in verifying the correlation between SWH and various factors. While ensuring significance levels, the Pearson correlation coefficient between each factor and SWH is calculated to verify the correlation between each factor and SWH. The calculation formula is as follows.

[0065] (4)

[0066] In the formula, It is an SWH sequence. For the first A sequence of elements, For SWH and the first The covariance of the element sequence, Let SWH be the variance. For the first The variance of each element sequence.

[0067] This paper selects the values ​​of each element from 4000 consecutive hours in the dataset of site 42019 for calculation, and calculates the Pearson correlation coefficient. And the significant difference value At this point, it can be considered that this factor has a significant and strong correlation with SWH. After calculation, the correlation coefficients between wind speed (WSPD), maximum hourly wind speed (GST), and average wave period (APD) and SWH are 0.726, 0.774, and 0.789, respectively; the significance values ​​are all much smaller than [value missing]. The correlation coefficients between wind speed (WSPD) and maximum hourly wind speed (GST) are relatively high, indicating a strong correlation. However, the correlation coefficients between wind speed (WSPD) and maximum hourly wind speed (GST) and mean wave period (APD) are relatively low, at only 0.226 and 0.239, respectively.

[0068] This paper treats SWH prediction as a supervised learning task. To generate the training and testing data required for model training, a fixed-length sliding window approach is used to process the time-series data, dividing the training, validation, and testing sets in a 7:1:2 ratio. Furthermore, since different oceanic elements have varying value ranges, z-score standardization is used to process the data, ensuring it follows a standard normal distribution.

[0069] This model was used to predict wave heights for 24 hours and 48 hours at two stations with different variation characteristics. The experimental results are shown in Table 3.

[0070]

[0071] To more intuitively demonstrate the model's performance in long-term prediction tasks, an experiment was designed and specific prediction results were presented. 168 hours of historical data were randomly selected from the test set data at site 41025 and input into the LogTrans and Informer models, which showed better performance in the comparison, to obtain predicted values ​​for the next 48 hours.

[0072] It should be understood that in the embodiments of the present invention, the term "and / or" is merely a description of the relationship between associated objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Furthermore, the character " / " in this document generally indicates that the preceding and following associated objects have an "or" relationship.

[0073] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0074] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0075] In the embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the couplings or direct couplings or communication connections shown or discussed may be indirect couplings or communication connections through some interfaces, apparatuses, or units, or they may be electrical, mechanical, or other forms of connection.

[0076] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0077] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A wave height prediction system, characterized by, The system comprises, a first local-global encoding module for encoding features of a time series of ocean multi-element related to wave height, an encoder, an input of which is connected with an output of the first local-global encoding module, for outputting high-dimensional features of the ocean multi-element; a second local-global encoding module for encoding a matrix sequence spliced by a latter half sequence of the time series of the ocean multi-element and a zero matrix sequence; a decoder, an input of which is connected with an output of the second local-global encoding module and the high-dimensional features output by the encoder, for outputting a prediction result for the wave height.

2. The system of claim 1, wherein, The encoder is composed of N groups of encoding layers, each of which comprises, in sequence, a first masked causal convolution self-attention layer, a first residual connection and normalization layer, a first forward propagation layer, a second residual connection and normalization layer.

3. The system of claim 1, wherein, The decoder is composed of N groups of decoding layers, each of which comprises, in sequence, a first masked causal convolution self-attention layer, a third residual connection and normalization layer, a first multi-head attention layer, a fourth residual connection and normalization layer, a second forward propagation layer, a fifth residual connection and normalization layer.

4. The system of claim 3, wherein, The decoder is connected with a full connection layer to give the prediction result.

5. The system of claim 1, wherein, The first local-global encoding LGE module comprises encoding of ocean multi-element features. One-dimensional convolutional neural networks are used to analyze ocean multi-element sequences. Encoding is performed to capture the local correlation features between multiple ocean elements. The encoded output sequence is: .

6. The system of claim 5, wherein, The first local-global encoding LGE module comprises encoding of time series information of the ocean multi-element.

7. The system of claim 6, wherein, For the time series information encoding of marine multi-elements, the sampling time is modeled using Time2vec, defined as with the expression . （1） wherein, is in the first element, is a periodic activation function, , is the frequency and offset of the sine function.

Images