CN115659828A

CN115659828A - Wave height prediction model and device

Info

Publication number: CN115659828A
Application number: CN202211382301.6A
Authority: CN
Inventors: 宋巍; 赵勐; 徐慧芳; 贺琪; 张明华; 包晓光
Original assignee: Shanghai Ocean University
Current assignee: Shanghai Ocean University
Priority date: 2022-11-07
Filing date: 2022-11-07
Publication date: 2023-01-31
Also published as: NL2034208A

Abstract

The invention discloses a wave height prediction model, which comprises a first LGE module, an encoder and a second LGE module, wherein the first LGE module is used for carrying out characteristic coding on marine multi-element time series related to wave height; the second LGE module is used for coding a matrix sequence formed by splicing the second half-segment sequence of the ocean multi-element time sequence and a zero matrix sequence; a decoder outputting a prediction result for a wave height using an output of the second LGE block and a high-dimensional feature of an output of the encoder as inputs. The encoder is composed of N (N = 4) groups of encoding layers, and each encoding layer sequentially comprises a first hole causal convolution attention layer, a first residual connection and normalization layer, a first forward propagation layer and a second residual connection and normalization layer.

Description

Wave height prediction model and device

Technical Field

The invention belongs to the technical field of ocean resource development, and particularly relates to a wave height prediction model and a wave height prediction device.

Background

In the process of wave propagation to the near shore, the wave is obviously influenced by factors such as temperature, wind, submarine topography, shore boundary and environmental flow (coastal flow, tidal current and the like), has more complex evolution rules and faster time-space change than deep sea and open land frame sea areas, and is not mature at present.

The near-shore wave is one of the most important power factors in the near-shore marine environment, threatens the safety and stability of near-shore buildings, and causes the movement of coastal silt, the transition of coastal and near-shore water body exchange. The calculation of the offshore waves has very important significance on the aspects of coastal engineering design, shallow sea production operation, offshore environment protection and the like. With the continuous development of social economy in coastal areas, the activities of human beings in coastal zones are more and more frequent, the quantity of coastal engineering projects is more and more, the investment scale is larger and larger, the risk of the engineering projects is more and more aroused high attention of people, and the requirements for accurate prediction of offshore environment elements such as near-shore waves are higher. Therefore, providing an accurate and practical wave height prediction method is an urgent task in recent years for coastal engineering, ocean and coastal zone resource research and military activities.

At present, the prediction methods for ocean wave motion can be roughly divided into two types:

the first type is to simulate and predict the Wave propagation process based on a Wave numerical Model by using the physical characteristics of Wave propagation, such as a third generation Wave Model (WAM) and a WAM-based near-shore Wave simulation method (SWAN) ^[1-3] 。

The second type is the prediction of the effective wave height SWH (SWH) based on the conventional time series model, such as Autoregressive model (AR), autoregressive Moving Average (ARMA), and Autoregressive Integrated Moving Average (ARIMA), etc ^[4-6] 。Agrawal ^[7] And others use ARIMA models to achieve online SWH prediction over different prediction intervals.

With the development of Machine learning, artificial Neural Network (ANN) and Support Vector Machine (SVM) and the like are beginning to be applied to prediction of SHW, deo ^[8] Et al propose real-time prediction of SWH based on ANN. Experimental results show that the ANN has better performance in the aspects of accuracy and consistency. Cornejo-bueno ^[9,10] The others predict the SWH by using an Extreme Learning Machine (ELM) of a feedforward neural network to obtain stronger generalization ability and faster solving speed.

SVM based on well-established mathematical theory is very widely applied. Mahjoobi ^[11] Et al showed that SVM has better performance than ANN in SWH prediction in some cases. Malekmohamadi ^[12] Intensive research is carried out on the prediction effects of SVM, ANN, bayesian networks and Adaptive Neural Fuzzy Inference Systems (ANFIS). The results show that the prediction results of ANN, ANFIS and SVM are in acceptable range, and the prediction results of the Bayesian network are relatively unreliable.

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. Fan ^[13] The LSTM is used for predicting the SWH of the offshore multi-site, and experimental results show that the LSTM can effectively capture the time correlation of the SWH, and the prediction accuracy is remarkably improved compared with a numerical model.

Among others, the present disclosure relates to references:

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

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

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

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

[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.

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

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

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

[9]CORNEJO-BUENO L,NIETO-BORGE J C,

P,et al.Significant wave height and energy flux prediction for marine energy applications:A grouping genetic algorithm–Extreme Learning Machine approach[J].Renewable Energy,2016,97:380-389.

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

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

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

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

disclosure of Invention

In one embodiment of the invention, the ocean wave height is predicted by a wave height prediction model. The model is composed of a model of a plurality of models,

a first LGE block that feature encodes a marine multi-element time series related to wave height,

an encoder, an input of which is connected to an output of the first LGE module, for outputting marine multi-element high-dimensional features;

the second LGE module is used for coding a matrix sequence formed by splicing the second half-segment sequence of the marine multi-element time sequence and a zero matrix sequence;

a decoder outputting a prediction result for a wave height using an output of the second LGE block and a high-dimensional feature of an output of the encoder as inputs.

The encoder and decoder each include a hole causal convolution self-attention layer, a normalization layer, and a forward propagation layer.

The ocean multi-element local-global feature associated effective wave height prediction model disclosed by the embodiment of the invention combines the ocean elements closely related to the wave activity, more effectively learns the change characteristic of the wave height of the ocean, and captures the long-time correlation of the ocean multi-element sequence by using an expanded causal convolution self-attention mechanism. By introducing time information, the learning capacity of the model is enhanced, the accurate prediction of the effective wave height is realized, and compared with the latest deep learning model and the traditional model, the model has the advantages of high accuracy, high efficiency and the like.

Drawings

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

FIG. 1 is a general framework diagram of a predictive model according to one embodiment of the invention.

Fig. 2 is a block diagram of an LGE module according to one embodiment of the present invention.

FIG. 3 is a schematic diagram comparing different attention mechanisms according to one embodiment of the present invention.

FIG. 4 is a block diagram of a decoder according to one embodiment of the present invention.

FIG. 5 is a comparison graph of predicted results according to one embodiment of the present invention.

Detailed Description

The existing numerical model prediction usually needs driving factors such as open sea incident sea wave conditions and the like, and the calculation efficiency is low. Here, the driving factors such as the open sea incident wave conditions include various sea elements involved in propagation from the open sea (to the sea side) to the coast side, for example, wind field elements, incident angles, and the like. Sea waves can be influenced by complex ocean elements in the process of propagation, ARIMA is only used for time series prediction of single elements and cannot effectively represent influences brought by complex and various ocean elements, and an ARIMA model extremely depends on conditions of stable assumptions. Therefore, the current deep learning-based model cannot capture the long-time correlation of the sequence, and the model precision is difficult to guarantee.

In the wave height prediction task of the sea wave, not only the long-term periodic change of the sea wave is considered, but also the short-term change caused by various environmental and meteorological factors is adapted, so that the application of the traditional Transformer to the wave height prediction task of the sea wave has the defects.

According to one or more embodiments, a wave height prediction model is composed of a Local-Global encoding (LGE) module, an Encoder (Encoder) module, and a Decoder (Decoder) module. The basic units in the Encoder and the Decoder are composed of a hole causal convolution Self-Attention layer (DConv Self-Attention), a residual connection and normalization layer (Add & Norm) and a Forward propagation layer (Feed Forward). The LGE module is used for carrying out feature coding on the ocean multi-element sequence, and a hole convolution self-attention mechanism in the Encoder module can capture long-time correlation and local variation trend of the sequence. The Decoder module is used for decoding the high-dimensional characteristics output by the Encode module and finally connecting the full-connection layer to obtain a prediction result. The overall framework of the model is shown in fig. 1.

The LGE blocks in the wave height prediction model encode time series, so that the model has the capability of capturing local features among multiple marine elements and effectively utilizing time information, and an LGE local-global coding block framework is shown in FIG. 2. The functions performed by the LGE modules include,

1) Marine multi-element feature coding: ocean multi-element sequence by adopting one-dimensional convolution neural network

Encoding is carried out, local association characteristics among ocean multiple elements are captured, and the output sequence after encoding is

2) Time information encoding: note that the mechanism cannot recognize the Position order of the input sequence, and therefore, position encoding (Position Embedding) is usually required to artificially add a Position relationship to the sequence. Position coding can only preserve the relative positions of elements in a sequence and cannot utilize time information of a time sequence, but for the problem of time sequence, the sequence tends to change periodically with time and in extreme cases, the sequence also tends to change aperiodically. Therefore, the full utilization of the time information plays an important role in improving the performance of the time series prediction model. Therefore, the embodiment of the disclosure adopts Time2Vec to generalize position coding to continuous Time sequences, and adaptively captures periodic and aperiodic patterns of the Time sequences. The sampling Time T will be modeled herein using Time2vec, defined as T2v (T), whose expression is:

wherein T2v (T) [ i ]]Is the ith element in T2v (T),

is an activation function with periodicity, here a sinusoidal function, w _i ，

The frequency and the offset of the sine function are both learnable hyper-parameters. The sine function in Time2Vec can be used to capture periodic patterns and the linear term can be used to capture non-periodic patterns, which is equivalent to mapping the sample Time sequence T to d using the sine function as a fully connected layer _r In the space of the dimension.

In the encoder, the correlation between the query and the key is computed point by point due to the conventional self-attention mechanism (as shown in fig. 3 (a)). Therefore, the self-attention module cannot learn time-series local information. The convolutional neural network is generally used in the field of images and is used for capturing local features of the images, and the convolutional neural network is applied to a time series prediction task and can achieve a good effect in capturing the local features.

When the causal convolution self-attention mechanism combines a convolutional neural network with a self-attention mechanism, i.e., the causal convolutional neural network replaces the fully connected layer to generate the Q and K required for the attention mechanism (as shown in fig. 3 (b)), the ability of the model to capture local features is enhanced. However, causal convolutional neural networks can only review linear-scale historical data, and the model has insufficient ability to utilize historical information due to limited perceptions. Although the causal convolutional neural network can increase the size of the receptive field by stacking layers one upon another, it will result in an increase in the number of parameters.

To solve as aboveIn view of the above, embodiments of the present disclosure design a hole causal convolutional neural network to generate Q and K (as shown in fig. 3 (c)), which experience also increases exponentially as the number of layers of convolution increases, compared to the causal convolutional neural network (fig. 3 (b)). Convolution kernel pair in the l-th hole causal convolution layer (2) ^(l-1) -1) convolution operation of the elements at the positions.

The cavity causal convolution designed by the method maintains the property of causal relationship, the value of a time sequence at a certain moment is only convoluted with the value of the previous moment, and the basic front-back dependency relationship on the time sequence in wave height prediction is not violated in the process of constructing and training the network. Meanwhile, the hole convolution can better utilize the historical information of the sequence, and when the similarity of data at a certain time step is calculated, the context relationship (such as local trend) of the time step can be utilized for calculation, so that the accuracy of prediction can be improved.

The decoder module adopts a similar structure as the encoder module. Masking matrix added in causal hole convolution self-attention mechanism of decoder

All upper triangle elements are set to- ∞formasking

When the matrix M is calculated through a Softmax function, the values of all upper triangular elements are changed into 0, so that the future information does not participate in calculation of the attention score, and autoregressive is avoided. The traditional multi-step prediction method adopts iterative prediction, which causes the calculated amount of a model to increase, and the error accumulation condition can occur along with the increase of the prediction length. Therefore, the generative prediction method is adopted here, so that the model can obtain multi-step prediction results only by decoding once, and a decoder module is shown in fig. 4.

And (3) inputting the second half of the ocean multi-element sequence as an initial sequence segment by adopting an Encoder module, and splicing the initial sequence segment and a zero matrix with the same length as the target sequence, as shown in a formula (3).

Wherein, Y _token For the initial sequence segment, Y in a prediction task of 24, 48 hours _token Are 6, 12, respectively. X ₀ Is a zero matrix of the same length as the target prediction sequence,

the time of the sequence is predicted for the target. And inputting the spliced result into an LGE module for encoding, and inputting the encoded result D into a Decoder module for decoding. In the decoding process, firstly, the feature learning is carried out on D through a DConv Self-authorization layer, and Add&Norm layer obtaining matrix

And secondly, calculating an attention matrix between the high-dimensional characteristics output by the Encoder module and the matrix Z by using the MHA and inputting the attention matrix into a feedforward layer. And finally, connecting a full connection layer behind the Decoder module, and performing linear mapping on the result of decoding the high-dimensional characteristics of the Encode module in the Decoder module. The second half (target prediction length) of the mapping vector is the prediction result.

The embodiment of the disclosure provides an SWH Prediction model (MLG-SWH) combining Marine multi-element Local and Global features, and can realize high-precision Prediction of SWH in 24 hours and 48 hours. The MLG-SWH model takes the effective wave height and relevant marine multi-element time series as input, and based on an encoder-decoder network structure, emphasizes the time information Embedding capacity of marine multi-element characteristics through Local-Global encoding (Local-Global encoding), enlarges the feeling of a self-attention model through cavity causal convolution, and predicts the future by using more historical information.

The technical characteristics and the beneficial effects of the invention comprise:

1. and combining the marine multi-element local-global feature association. The method comprises the steps of screening out ocean elements with strong correlation with effective wave height through a statistical method to predict the effective wave height, capturing local correlation among multiple elements by using a convolutional neural network, and capturing long-time correlation and local variation trend of a sequence by using an expansion causal convolution self-attention mechanism.

2. Time information is utilized. The prediction model uses the Time2Vec method to encode Time and replaces position encoding as Time information of a self-attention mechanism, so that the change characteristics of a multi-element sea wave sequence can be effectively learned, and the prediction effect is improved.

3. A generative prediction mode is adopted to realize one-time multi-step prediction of SWH, and the influence of error accumulation in point-by-point iterative multi-step prediction is effectively reduced.

Therefore, compared with deep learning networks such as ARIMA, LSTM and improved Transformer, the method provided by the invention can better capture the long-time correlation of SWH, and can effectively extract the local correlation between SWH and ocean multiple elements.

An example is given below to illustrate the practical effects of the wave height prediction model of the present invention. The operating environment requirements of the model are shown in table 1.

TABLE 1 System operating Environment

The data used in the experiments originated from the national oceanic and atmospheric administration website, which selected as research subjects the bay No. 42019 (27.910 ° N,95.345 ° W) in mexico and the bay No. 41025 in pamlicoco (35.010 ° N,75.454 ° W) in each case as surface buoy stations. Wherein, the water depth of the No. 42019 station positioned in the gulf of Mexico is 83.5 meters, the station is in tropical zone and subtropical zone, the high temperature is rainy, the precipitation is more, and there are hurricane seasons every year, so the wave height fluctuation range of the sea wave is larger; 41025 station is located in the Bay of Pamliko, and has a water depth of 59.4 m, which is the largest lagoon on the east coast of the United states, and the wave amplitude is relatively stable. The time span of data acquisition of the two stations is from 1 month and 1 day in 2018 to 31 days in 12 months 2020, the data sampling interval is 1 hour, and the data set comprises Wind Speed (WSPD), sea wave Average Period (APD), water Temperature (WT), air Temperature (AT) and other elements.

TABLE 1 selected site information

Tab.1 Details of selected stations

The formation of ocean waves is influenced by a variety of complex geographical and natural factors. The growth state of the stormy waves is closely related to three factors of wind speed, wind time and wind area in terms of physical mechanism; swell generally has the characteristic of long cycle; the offshore waves are also affected by the sea floor topography and shoreline. The prediction of the effective wave height of the sea wave is researched based on a data-driven method, and factors related to SWH are selected as far as possible in the existing data range (terrain data is not available). Therefore, the correlation between the SWH and each element is verified by performing an auxiliary verification of the correlation by using a statistical method, and the correlation between each element and the SWH is verified by calculating a Pearson correlation coefficient (Pearson correlation coefficient) between each element and the SWH while ensuring the significance level, and the calculation formula is as follows.

Wherein A is SWH sequence, B _i Is the i-th element sequence, cov (A, B) _i ) Is the covariance, σ, of SWH and the ith element sequence _A Is the variance of the SWH and is,

is the variance of the ith element sequence.

The continuous 4000-hour element value in the 42019 station data set is selected for calculation, and when the Pearson correlation coefficient is calculated

And the difference value p-value of the significance<At 0.05, this element can be considered to have a significantly strong correlation with SWH. Through calculation, windCorrelation coefficients between speed (WSPD), maximum wind speed per hour (GST), and average period of ocean waves (APD) and SWH are 0.726,0.774, and 0.789, respectively; the values of the significant differences are far less than 0.001, the correlation coefficients between the Wind Speed (WSPD) and the maximum wind speed per hour (GST) are high and have strong correlation, and the correlation coefficients between the Wind Speed (WSPD) and the maximum wind speed per hour (GST) and the average period of sea waves (APD) are low and are only 0.226 and 0.239.

SWH prediction is considered herein as a supervised learning task. In order to generate training data and test data required by model training, the time sequence data is processed in a fixed-length sliding window mode, and a training set, a verification set and a test set are divided according to a ratio of 7. In addition, since the value ranges of different elements of the ocean are different, the data is processed by using z-score standardization so as to be in accordance with standard normal distribution.

The model carries out wave height prediction for 24 hours and 48 hours on two sites with different change characteristics, and the experimental results are shown in table 1.

TABLE 1 comparison of Performance of different prediction methods

Note: bolding indicates the best result

In order to more intuitively represent the performance of the model in a long-term prediction task, an experiment is designed and a specific prediction result is given. 168 hours of historical data are randomly extracted from the test set data of the 41025 station, and the data are input into LogTrans and Informmer with better performance in a comparison model for prediction, so that predicted values in the future of 48 hours are respectively obtained.

It should be understood that, in the embodiment of the present invention, the term "and/or" is only one kind of association relationship describing an association object, and indicates that three kinds of relationships may exist. For example, a and/or B, may represent: a exists alone, A and B exist simultaneously, and B exists alone. In addition, the character "/" herein generally indicates that the former and latter related objects are in an "or" relationship.

Those of ordinary skill in the art will appreciate that the various illustrative components and algorithm steps described in connection with the embodiments disclosed herein may be implemented as electronic hardware, computer software, or combinations of both, and that the components and steps of the various examples have been described above generally in terms of their functionality in order to clearly illustrate this interchangeability of hardware and software. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the implementation. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present invention.

It can be clearly understood by those skilled in the art that, for convenience and simplicity of description, the specific working processes of the above-described systems, apparatuses and units may refer to the corresponding processes in the foregoing method embodiments, and are not described herein again.

In the several embodiments provided in the present application, it should be understood that the disclosed system, apparatus and method may be implemented in other ways. For example, the above-described apparatus embodiments are merely illustrative, and for example, the division of the units is only one logical division, and other divisions may be realized in practice, for example, a plurality of units or components may be combined or integrated into another system, or some features may be omitted, or not executed. In addition, the shown or discussed mutual coupling or direct coupling or communication connection may be an indirect coupling or communication connection through some interfaces, devices or units, and may also be an electric, mechanical or other form of connection.

The integrated unit, if implemented in the form of a software functional unit and sold or used as a stand-alone product, may be stored in a computer readable storage medium. Based on such understanding, the technical solution of the present invention essentially or partly contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device) to execute all or part of the steps of the method according to the embodiments of the present invention. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), a magnetic disk or an optical disk, and other various media capable of storing program codes.

While the invention has been described with reference to specific embodiments, the invention is not limited thereto, and various equivalent modifications and substitutions can be easily made by those skilled in the art within the technical scope of the invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A wave height prediction model, characterized in that the model comprises,

an encoder having an input connected to an output of the first LGE module for outputting marine multi-element high-dimensional features;

and a decoder which takes the output of the second LGE block and the high-dimensional characteristics output by the encoder as input and outputs a prediction result for the wave height.

2. The wave height prediction model according to claim 1, wherein the encoder is composed of N (N = 4) sets of coding layers, each coding layer comprising, in order, a first hole causal convolution self-attention layer, a first residual concatenation and normalization layer, a first forward propagation layer, a second residual concatenation and normalization layer.

3. The wave height prediction model of claim 1, wherein the decoder comprises N (N = 4) sets of decoding layers, each decoding layer comprising, in order, a first mask hole causal convolution self-attention layer, a third residual concatenation and normalization layer, a first multi-head attention layer, a fourth residual concatenation and normalization layer, a second forward propagation layer, a fifth residual concatenation and normalization layer.

4. The wave height prediction model according to claim 3, wherein the decoder concatenates a fully concatenated layer to give the prediction result.

5. The wave height prediction model of claim 1, wherein the first LGE block comprises, for marine multi-element feature coding:

ocean multi-element sequence using one-dimensional convolutional neural network

6. The wave height prediction model of claim 5 wherein the first LGE block includes time series information coding for ocean multi-elements.

7. The wave height prediction model according to claim 6, wherein for marine multi-element Time series information coding, the sampling Time T is modeled by using Time2vec, defined as T2v (T), and expressed as:

wherein T2v (T) i is the ith element in T2v (T),

is an activation function with a periodicity that is,

w _i ，

is the frequency and offset of the sine function.

8. An apparatus for wave height prediction, characterized in that the apparatus comprises a memory; and

a processor coupled to the memory, the processor configured to execute instructions stored in the memory, the processor to:

from the obtained sea parameters, the wave height is predicted using the model of claim 1.

9. A storage medium on which a computer program is stored which, when being executed by a processor, carries out a prediction model according to any one of claims 1 to 7.