CN114519311A

CN114519311A - Prediction method, system, storage medium and application of total harbor basin wave effective wave height

Info

Publication number: CN114519311A
Application number: CN202210418364.6A
Authority: CN
Inventors: 解翠; 满腾浩; 刘修栋; 董军宇
Original assignee: Ocean University of China
Current assignee: Ocean University of China
Priority date: 2022-04-21
Filing date: 2022-04-21
Publication date: 2022-05-20
Anticipated expiration: 2042-04-21
Also published as: CN114519311B

Abstract

The invention belongs to the technical field of wave information prediction, and discloses a prediction method, a prediction system, a storage medium and an application of the effective wave height of waves in a full harbor pool. The method for predicting the effective wave height of the waves in the full harbor basin comprises the following steps: firstly, constructing and preprocessing a data set, processing and normalizing missing values of waves inside and outside a harbor pool and wind field data outside the harbor pool, carrying out multi-scale feature coding on time information corresponding to the data, constructing conditions after dividing the data set to generate an antagonistic network, inputting wind field and wave features subjected to multi-scale time feature splicing and fusion processing into a neural network, continuously iteratively training and updating and adjusting parameters of the neural network and hyper-parameters of a model, and finally establishing an optimal end-to-end deep learning prediction model. The prediction model of the invention is very simple and easy to use, and the effective wave height of the waves in the whole harbour pool can be predicted quickly by only providing the preprocessed wind and wave data outside the harbour pool and multi-scale time coding information to input into the established prediction model.

Description

Prediction method, system, storage medium and application of total harbor basin wave effective wave height

Technical Field

The invention belongs to the technical field of wave information prediction, and particularly relates to a prediction method of the effective wave height of waves in a full harbor pool, a prediction system of the effective wave height of waves in a harbor pool, a storage medium for receiving a user input program, and application of the prediction method in the harbor pool, hydraulic engineering and near-shore engineering wave prediction.

Background

The importance of the rapid prediction and accurate prediction of the full pool wave is as follows: waves have been an important research topic in harbour basins, hydraulic engineering and near-shore engineering. Waves not only have an impact on the safe berthing, operation and navigation of ships, but also may damage coastal buildings, and thus a reliable and efficient method for predicting waves is important. First, the prediction of the wave conditions within a harbour site is crucial to controlling and regulating the safe operation of the harbour site. In particular, if severe rough seas are anticipated at those locations in the harbor basin, the moored vessel should be transferred to a safe location to avoid severe damage, and rapid predictions can be made to allow personnel to respond promptly to the emergency and thus quickly transfer the vessel to a safe location. Secondly, the waves from the deep water area of the open sea to the shallow water area of the offshore area are influenced by many factors such as offshore terrain, water flow, water bottom friction, breakwater and the like, so that the accurate prediction of the wave height in the harbor basin under the complex environmental influence factors is very important and challenging.

The traditional method for predicting the wave height in the harbor pool comprises two types of methods which adopt a physical model and a numerical model:

(1) the physical model method is a reduced version for constructing an actual harbor basin, and the propagation condition of waves in the harbor basin is reproduced, which is one of necessary methods in harbor projects of various countries.

(2) In terms of numerical models, many hydrodynamic models for near shore, harbour basin, water conservancy and the like are developed internationally, such as Delft3D at Delft university of dalfort, netherlands, MIKE21 series at danish hydrodynamic research institute (DHI), SMS series at american hydraulic engineering laboratories and the like, and the models are widely applied to near shore ocean engineering, such as wave propagation at harbour basin, harbour basin oscillation simulation, wave prediction, tsunami modeling and the like. The MIKE21 BW model can simulate a plurality of physical processes, the model is accurate and fits well with the actual effect, but the use is complex, a large amount of external parameters are needed, the running time is long, and the requirement on the performance of a computer is high; the MIKE21 SW model uses an unstructured network, is flexible and simple to use, neglects wave phase, and has relatively long running simulation time. The non-hydrostatic model SWASH developed by dalfovery much aims at predicting surface wave transitions and rapidly changing shallow water flows in coastal waters, while also requiring some external files and longer simulation times. In general, the wave numerical simulation method requires complicated process steps for construction and use, requires more external parameters and longer simulation calculation time, and results are unstable due to the influence of the parameters and are difficult to meet the real-time requirement.

With the rapid development of artificial intelligence, machine learning methods are being applied to wave height prediction. For example, kumquat et al attempts to predict the wave height using a Support Vector Machine (SVM). The results show that the SVM model has acceptable accuracy and requires a short computation time. Deo proposes a feed-forward network to predict the wave height in real time. Compared with a numerical calculation method, the method has stronger universality, flexibility and adaptability. The positions of 12 points selected in a harbor basin are quickly predicted by using an artificial neural network ANN model, so that the prediction by using the neural network is proved to be much faster than that by using a numerical model, and the reliability of the prediction is verified. The results of various methods of research on the prediction effect of the wave effective wave height of the Sulil lake by the Mahjoobi show that the model tree method has higher precision than a feedforward back propagation neural network. Therefore, the method has feasibility for rapidly and accurately predicting the wave height by using the neural network.

Through the above analysis, the problems and defects of the prior art are as follows:

(1) the method for numerically simulating the wave height of sea waves of the harbor basin depends on field knowledge and the experience of researchers, and meanwhile, the environmental factors such as the terrain of a target harbor basin, seabed friction, wave breaking, breakwater setting and the like need to be fully considered, for example, sufficient water depth data and measured data of an engineering area are needed, and if no measured data exists, the set conditions can directly influence the accuracy of a simulation result according to the experience;

(2) because the construction of the numerical simulation model involves a large number of external parameters, a large amount of time is spent to adjust the parameters to better simulate the real wave motion behavior, the process is quite complex, the calculation time cost is high, and the requirement of real-time prediction cannot be met;

(3) at present, most of the existing wave height prediction models based on the neural network only consider the wave height information outside the port, and the wave height in the port pool is obviously influenced by various related factors, for example, the wave height influenced by changes of tide and season shows obvious periodic changes, an external wind field also has direct influence on the wave height, and the like, so that the important factors influencing the behavior characteristics of the wave can influence the accuracy of model prediction by neglecting.

(4) The wave height prediction method in the harbor pool is limited by limited monitoring equipment in the harbor pool, most of the existing wave height prediction methods in the harbor pool only predict the wave height of a certain specific position, and the wave height prediction of each position of the whole harbor pool is not realized, so that the method brings adverse effects on the efficient and accurate construction and operation of the harbor.

The difficulty in solving the above problems and defects is:

(1) constructing a high-quality data set, wherein a deep learning model needs a large amount of data for training, the quality of the data set influences the quality of the model, and particularly, the construction of the high-quality data set under the condition of limited monitoring data is difficult;

(2) factors influencing the height of the waves are complex and changeable, and important relevant factors are selected and effectively processed to better represent the behavior characteristics (such as periodicity and the like) of the waves, so that the model can accurately learn various characteristics, which is a challenge;

(3) at present, only one or a few buoys are often arranged in a harbor basin to monitor the wave condition, so the obtained monitoring data are often sparse, and the actual need is to predict wave height information of each position of the harbor basin, which also has certain challenges.

The significance of solving the problems and the defects is as follows:

(1) under the condition that insufficient monitoring data can be used, a basic wave height data set is generated by using a numerical model, and high-quality data sets are generated by using valuable monitoring data on hand, so that an important basic role is played in developing deep learning model research in the aspect of ocean atmosphere;

(2) once the deep learning model is built, a user can quickly realize high-quality prediction of effective wave height of waves in the port only by using wind and wave information outside the port pool and multi-scale coding information of time of the wind and wave information, the user does not need to master relevant field knowledge and experience such as numerical modeling and the like, and does not need to wait for long resolving time, and the method is simple, efficient and easy to use.

(3) Through the multi-scale time information coding of the embodiment, the deep learning model can learn richer time information characteristics and learn the change rule of the harbor pool wave height under different time scales, so that the prediction result is more accurate.

(4) Based on buoy monitoring data of a small number of positions, wave height prediction of each position in the harbor basin is realized quickly, the method can be further expanded to prediction of other wave elements (period and direction) except for wave height, a user can know dynamic change conditions of waves in the harbor more comprehensively, proper berthing operation positions, proper berthing operation arrangement operation and the like are set, risks are reduced, and economic benefits are brought.

Disclosure of Invention

In order to overcome the problems in the related art, the disclosed embodiment of the invention provides a prediction method and a prediction system for the effective wave height of the waves of the full harbor pool, a program storage medium for receiving user input, and application of the prediction method and the prediction system in the wave prediction of harbor pool, hydraulic engineering and near-shore engineering. The invention also relates to the technical fields of coastal engineering, artificial intelligence and neural networks. The technical scheme is as follows:

the method for predicting the effective wave height of the waves of the full harbor pool comprises the following steps:

data set construction and preprocessing: constructing data sets inside and outside a harbor pool, preprocessing the data sets to generate characteristic vectors of wind fields and waves, carrying out multi-scale characteristic coding on time information corresponding to the data, and finally dividing the data sets to prepare for training a deep learning model;

building and training a deep learning model: establishing a deep learning model for predicting the effective wave height of the harbor pool based on the generated countermeasure network, performing alternate iterative training, learning and model parameter adjusting optimization on the generator network and the discriminator network, and establishing an end-to-end prediction model for the effective wave height of the harbor pool;

predicting the effective wave height of the harbor pool waves: and after the wind and wave data outside the harbor pool are subjected to the same missing value and normalization processing as those of the data set construction and pretreatment module and time multi-scale coding, the obtained wind, wave and time characteristics are directly spliced and fused, and are input into a trained harbor pool wave effective wave height prediction model, so that the effective wave height of the waves in the whole harbor pool is predicted.

In one embodiment, the data set construction and preprocessing specifically includes:

1) constructing high-quality data sets inside and outside the harbor pool, and acquiring or generating wind and wave data outside the target harbor pool and wave height data in the harbor pool, wherein the high-quality data sets respectively comprise: average wind speed and average wind direction of a wind field outside the harbor basin, effective wave height, average period and average direction of waves outside the harbor basin, and effective wave height data of waves at each grid point inside the harbor basin;

the method comprises the following steps that wind and wave data inside and outside a harbor pool are acquired through a re-analysis data set disclosed on the internet or a monitoring data set recorded by a monitoring station; if no available public data set exists and monitoring data are insufficient, generating a large amount of data by adopting pattern calculation, calibrating based on the monitoring data, and taking the calibrated pattern data as a data set to prepare for building and training a deep learning model;

2) constructing data sets inside and outside a pretreatment port pool, uniformly processing the numerical value of the land part into 0, and carrying out interpolation filling on the missing value of effective wave height data in the port pool; meanwhile, normalizing the wave and wind field data outside the harbor pool respectively to generate a characteristic vector of the wave outside the harbor pool and a characteristic vector of the wind field outside the harbor pool; the characteristic vector of the wave outside the harbor basin comprises three components of the effective wave height, the average period and the average direction after normalization; the characteristic vector of the wind field outside the harbor basin comprises two components of the average wind speed and the average wind direction after normalization;

3) coding the time information in different scales, and carrying out multi-scale one-hot coding on the time information corresponding to the data to obtain a time characteristic vector with the length of 36;

4) and dividing the data set, wherein 80% of the data set is used as a training set, 10% of the data set is used as a verification set, and 10% of the data set is used as a test set, so that preparation is made for training the deep learning model.

In one embodiment, in the step 2), the wave and wind field data outside the harbor basin are respectively normalized by converting the data into decimal numbers between 0 and 1 and converting dimensional expressions into dimensionless expressions

①

In the formula

Is the mean value of all the sample data,

is the standard deviation of all sample data.

In one embodiment, the building and training of the deep learning model specifically includes:

(1) building and generating a confrontation network deep learning model, setting the number of hidden layer neural units of a neural network generator and a discriminator, initializing the weight of a neural network, and setting a loss function of the neural network model; setting a gradient descent optimizer, the number of training rounds epoch and the learning rate;

(2) deep learning model training and parameter adjustment, alternately training by a discriminator and a generator, and terminating the training when the discriminator cannot judge true and false data; in the training process, the generator and the discriminator are repeatedly adjusted in structure and super parameters by using a verification set, and the optimal network model structure parameters are finally determined according to multiple comparison experiment results;

(3) and evaluating the performance of the generator model, predicting the generator model trained in the previous step on a test set, if the prediction result is worse than the current optimal reference model, continuing to adjust the model parameters for retraining, otherwise, if the prediction result is obviously better than the current optimal reference model, obtaining a deep learning model capable of more accurately predicting the effective wave height of each position in the harbor basin.

In one embodiment, in step (1), the building and generating the confrontation network deep learning model comprises: the generator and the discriminator are built:

the construction generator part: the generator is a multi-layer perceptron with a residual error structure, wherein a main network contains 4 hidden layers, the residual error part has 2 hidden layers, and an activation function ReLU is used behind each hidden layer to increase the nonlinearity of an output result; meanwhile, the height of the waves in the harbor pool is periodically influenced by wind fields, waves, tides and seasons outside the harbor pool, so that the three characteristics of the waves, the wind field data and the multi-scale coded time information outside the harbor pool after splicing and fusing pretreatment are input into a generator model, and a residual error structure inside the generator further enhances the influence of the waves outside the harbor pool and improves the performance of network prediction;

constructing a discriminator part: the discriminator uses a multilayer perceptron, the multilayer perceptron comprises 3 hidden layers, a ReLU is used as an activation function after passing through each hidden layer, and finally the hidden layers use Sigmoid to set the certainty factor score of the generated image to be in the range of 0 to 1; the input of the discriminator is that the generated wave height data and the truth value data of the wave height in the harbor pool are spliced and fused with the wave characteristic vectors outside the harbor pool respectively except the effective wave height data and the truth value data of the wave height in the harbor pool generated by the generator, then the generated wave height data and the truth value data of the wave height in the harbor pool are input into the discriminator to solve the loss respectively, the wave information outside the harbor pool is used as an additional condition input by the discriminator, and the capability of the discriminator for discriminating the data accuracy is improved.

In one embodiment, the loss function of the generator part is shown in equation (c):

②；

wherein the content of the first and second substances,

representing the characteristics of the input model after splicing and fusion;

representing data generated by the generator model;

representing the accuracy grade value given by the discriminator model after discrimination;

indicating the number of samples input to the network in bulk;

the Loss function of the discriminator part is binary cross entropy BCE Loss, as shown in formula (c):

③；

wherein the content of the first and second substances,

true value data for the wave effective wave height;

the significant wave height data generated by the generator;

indicating the number of samples input to the network in bulk;

and representing the accuracy value of the wave effective wave height after the discrimination of the discriminator model.

In one embodiment, in step (2), the deep learning model training and parameter tuning includes:

alternately training the discriminator and the generator, terminating the training when the discriminator cannot judge true and false data, repeatedly adjusting the structure and the super parameters of the generator and the discriminator by using a verification set during the training, and finally determining the optimal network model structure parameters according to the comparison experiment results for many times; when the discriminator is trained, inputting effective wave height data and true wave height data of waves in a harbor pool generated by a generator by the discriminator, splicing and fusing the generated wave height data and the true wave height data of the harbor pool with wave characteristic vectors outside the harbor pool respectively in order to further improve the capability of the discriminator for discriminating the data accuracy, inputting the generated wave information outside the harbor pool as an additional condition input by the discriminator, then inputting the generated wave height data and the true wave height data into the discriminator for training respectively, updating network parameters of the discriminator through a reverse propagation gradient according to a loss function, and performing multi-round iterative training;

when the generator is trained, fixing network parameters of a discriminator, inputting data into the generator in batches, judging the data generated by the generator by the discriminator, giving an accuracy grade of the data generated by the generator, wherein the value is 0-1, then performing gradient descent according to a loss function, updating the generator through error back propagation, and updating the network parameters of the generator; when the discriminator can not distinguish the true data from the false data, the training is terminated; and in the training process, the structure and the super-parameter adjustment are repeatedly carried out on the generator and the discriminator by using the verification set, and the network model structure and the control model complexity are adjusted according to multiple comparison experiment results.

Another object of the present invention is to provide a system for predicting the effective wave height of a harbor pool wave according to the method for predicting the effective wave height of a full harbor pool wave, the system comprising:

the data set construction and pretreatment module is used for constructing data sets inside and outside the harbor basin, carrying out pretreatment to generate characteristic vectors of wind fields and waves, carrying out multi-scale characteristic coding on time information corresponding to the data, and finally dividing the data sets to prepare for deep learning model training;

the deep learning model establishing and training module is used for establishing a deep learning model for predicting the effective wave height of the harbor pool based on the generated confrontation network, performing alternate iterative training, learning and model parameter adjusting and optimizing on the generator network and the discriminator network, and establishing an end-to-end prediction model for the effective wave height of the harbor pool;

and the effective wave height of the waves in the whole harbor pool is predicted by the module for predicting the effective wave height of the waves in the harbor pool, wherein the wind and wave data outside the harbor pool are subjected to missing value and normalization processing which are the same as those of the data set construction and pretreatment module and time multi-scale coding, and the obtained three characteristics of the wind, the waves and the time are directly spliced and fused and input into a trained prediction model for the effective wave height of the waves in the harbor pool.

It is another object of the present invention to provide a program storage medium for receiving a user input, the stored computer program causing an electronic device to execute the method for predicting the effective wave height of the harbor-wide pool wave.

The invention also aims to provide an application of the prediction method of the total harbour basin wave effective wave height in the prediction of harbour basin, hydraulic engineering and near-shore engineering waves.

By combining all the technical schemes, the invention has the advantages and positive effects that:

firstly, the method combines the thought of generating a countermeasure network in deep learning and builds a deep learning model to quickly and accurately predict the effective wave height in the harbor pool. When the deep learning model is used for predicting the effective wave height of the waves in the harbor basin, no requirements are made on the field knowledge and the related experience of a user. And the end-to-end deep learning model enables the prediction to be very simple and easy to use when the model is actually used, and the model can quickly and accurately predict the wave condition in the whole harbour pool as long as wind and wave data outside the harbour pool are provided, so that real-time prediction is realized. Meanwhile, a certain periodicity which is shown by the influence of the season wind on the geographical position of the target harbor basin and the influence of tide on the harbor basin wave height is considered, so that the method splices and integrates the information of a wind field and the time information of multi-scale coding, leads a deep learning model to learn the periodic rule of the wave height influenced by the tide and the season and the influence of the wind factor on the harbor basin wave height change through training, and realizes more accurate prediction.

In addition, the method can not only predict the wave height attribute, but also be applied to the field of predicting directions of wave attributes such as wave periods, wave directions and the like.

Secondly, the expected income and commercial value after the technical scheme of the invention is converted are as follows: the numerical model is large in calculation amount and long in required time, and a super computer is often required to execute a calculation task, so that the calculation cost and the time cost in port pool wave height prediction can be greatly reduced by using the deep learning model disclosed by the invention to predict the port pool wave height; meanwhile, the method can accurately predict the wave height in the harbor, so that the safe operation of the ship can be positively promoted, and the risk and economic loss of loading and unloading the ship under the condition of poor wave height are reduced. The method has strong expansibility, can not only predict the effective wave height in the harbor pool, but also be applied to the field of predicting wave-related attributes such as wave period, wave direction and the like, and more comprehensively meet the requirements of harbor operation service.

Drawings

The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and together with the description, serve to explain the principles of the disclosure.

FIG. 1 is a flow chart of a method for rapidly and accurately predicting the effective wave height of a harbor basin wave provided by an embodiment of the invention;

FIG. 2 is a schematic diagram of a system for rapidly and accurately predicting the effective wave height of a harbor basin wave provided by an embodiment of the invention;

FIG. 3 is a schematic diagram of simulating the wave propagation evolution of a harbor basin by numerical solution according to an embodiment of the present invention;

FIG. 4 is a graph of the comparison of the simulated wave height of the numerical model and the actual monitoring data provided by the embodiment of the invention;

FIG. 5 is a schematic diagram of a wave data acquisition location outside a harbor basin according to an embodiment of the present invention, as shown at the location of the exterior number 1 of the harbor basin;

FIG. 6 is an overall architecture of a deep learning model provided by an embodiment of the present invention, showing the composition of a complete model participating in model training;

FIG. 7 is a model diagram of a portion of a model structure for wave height prediction using a trained generator according to an embodiment of the present invention, where no discriminator is needed for real-time prediction;

FIG. 8 is a flowchart of a training process for establishing a harbor pool wave effective wave height fast and accurate prediction model according to an embodiment of the present invention;

FIG. 9 is a flowchart for predicting effective wave height of a harbor pool wave in real time by using a trained deep learning model according to an embodiment of the present invention;

fig. 10 is a comparison graph of the prediction result of the effective wave height in the harbor basin and the calibrated mode-resolved effective wave height (true value) according to the embodiment of the present invention.

In the figure: 1. a data set construction and preprocessing module; 2. the deep learning model building and training module; 3. and a module for predicting the effective wave height of the waves in the harbor basin.

Detailed Description

In order to make the aforementioned objects, features and advantages of the present invention comprehensible, embodiments accompanied with figures are described in detail below. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present invention. This invention may, however, be embodied in many different forms than those specifically described herein, and it will be apparent to those skilled in the art that many more modifications are possible without departing from the spirit and scope of the invention.

The invention develops a prediction algorithm of the wave height in the harbor pool by utilizing the thought of generating the antagonistic neural network, realizes the prediction of the effective wave height of the waves at each position in the harbor pool, and obtains the rapid and accurate prediction effect;

the influence of a wind field and a wave field outside the harbor pool on the change of the wave height inside the harbor pool is considered at the same time, the wind speed and the wind direction information outside the harbor pool and the wave height, the speed and the direction information outside the harbor pool are input into a neural network, the internal relation between the wind field outside the harbor pool and the wave height inside the harbor is excavated and learned, and the prediction accuracy of a deep learning model is improved;

the invention also considers the characteristic that the wave height change in the harbor basin is periodic, for example, the wave height shows the regular change trend due to the periodic change of the wave height caused by tide or the alternate season, so that multi-scale time information coding is introduced, the multi-scale time information coding is embedded into a neural network, the periodic phenomenon or the regular change trend of the wave height change is learned, and more accurate prediction is realized.

The deep learning model is very efficient and easy to use when in use, and can quickly and accurately predict the effective wave height of waves in the outlet pool and realize real-time prediction only by processing and normalizing missing values of wave data and wind field data outside the port pool, splicing and fusing multi-scale characteristic coding information of time and inputting the information into a trained neural network model.

The method simultaneously considers the influence of a wind field, the periodic characteristics and the variation trend of waves under different time scales, establishes a deep learning prediction model based on the idea of generating a countermeasure network, and realizes quick and more accurate prediction of the effective wave height in the harbor basin. Firstly, constructing and preprocessing a data set, processing and normalizing missing values of waves inside and outside a harbor pool and wind field data outside the harbor pool, carrying out multi-scale feature coding on time information during data sampling, constructing conditions after dividing the data set to generate an antagonistic network, inputting wind and wave field features subjected to multi-scale time feature splicing and fusion processing into a neural network, continuously iteratively training and updating and adjusting parameters of the neural network and hyper-parameters of a model, and finally establishing an optimal end-to-end deep learning prediction model. The prediction model of the invention is very simple and easy to use, and when in actual prediction, the effective wave height of the waves in the whole harbour pool can be predicted quickly by only providing the wind and wave data preprocessed outside the harbour pool and multi-scale time coding information to input into the established prediction model without mastering complex wave professional knowledge. The method can quickly and accurately predict the effective wave height of the waves in the harbor pool, improve the safety of ship operation in the harbor and increase the economic benefit of the harbor. Meanwhile, the method can be applied to the field of wave related attribute prediction such as wave period, wave direction and the like.

The technical solution of the present invention is further described below with reference to examples.

Example 1

As shown in fig. 1, the method for predicting the effective wave height of the full pool wave provided by the invention comprises the following steps:

s101, data set construction and preprocessing: constructing data sets inside and outside a harbor pool, preprocessing the data sets to generate characteristic vectors of wind fields and waves, carrying out multi-scale characteristic coding on time information corresponding to the data, and finally dividing the data sets to prepare for training a deep learning model; wherein, the step of constructing the data sets inside and outside the harbor pool comprises directly obtaining or generating the data sets inside and outside the harbor pool;

s102, building and training a deep learning model: establishing a deep learning model for predicting the effective wave height of the harbor pool based on the generated countermeasure network, and performing alternate iterative training, learning and model parameter adjusting optimization on the generator network and the discriminator network to establish an end-to-end prediction model for the effective wave height of the harbor pool;

s103, forecasting the effective wave height of the harbor pool waves: and after the wind and wave data outside the harbor pool are subjected to the same missing value and normalization processing as those of the data set construction and pretreatment module and time multi-scale coding, the obtained wind, wave and time characteristics are directly spliced and fused, and are input into a trained harbor pool wave effective wave height prediction model, so that the effective wave height of the waves in the whole harbor pool is predicted.

As shown in fig. 2, the system for predicting the effective wave height of the full pool wave provided by the invention comprises:

and the data set constructing and preprocessing module 1 is used for constructing high-quality data sets inside and outside the harbor basin. Namely, acquiring or generating wind and wave data outside a target harbor pool and wave height data in the harbor pool, which respectively comprise: the average wind speed and the average wind direction of a wind field outside the harbor basin, the effective wave height, the average period and the average direction of waves outside the harbor basin, and the effective wave height data of the waves at each grid point inside the harbor basin. The acquisition of wind and wave related data inside and outside the harbor pool can be realized through an online public data set or a monitoring data set recorded by a monitoring site; if no public data set is available and the monitoring data is insufficient, a large amount of relevant basic data can be generated by adopting pattern calculation, calibration is carried out based on the monitoring data, and the calibrated pattern data is used as a data set so as to ensure that a data set with higher quality is constructed. Then carrying out missing value and normalization processing on the data to generate a characteristic vector (effective wave height, average period and average direction) of waves outside the harbor basin and a characteristic vector (average wind speed and average wind direction) of a wind field outside the harbor basin, then carrying out multi-scale characteristic coding on the time corresponding to the data to form a time characteristic vector, and finally dividing a data set to prepare for training a deep learning model;

the deep learning model establishing and training module 2 is used for establishing a deep learning model for predicting the effective wave height of the harbor pool based on the generated countermeasure network, performing alternate iterative training, learning and model parameter adjusting and optimizing on the generator network and the discriminator network, and establishing an end-to-end prediction model for the effective wave height of the harbor pool;

and a port pool wave effective wave height predicting module 3: wind and wave data outside the harbor basin are subjected to missing value and normalization processing which are the same as those of the data set construction and pretreatment module 1 and time multi-scale coding, the obtained wind, wave and time characteristics are directly spliced and fused, and are input into a trained harbor basin wave effective wave height prediction model, so that the effective wave height of each wave in the whole harbor basin can be rapidly and accurately predicted.

Example 2

The prediction system for the effective wave height of the waves in the full harbor pool provided by the invention comprises:

(1) data set construction and preprocessing module 1: constructing high-quality data sets inside and outside the port pool, carrying out preprocessing such as missing value processing and normalization on the data, carrying out multi-scale feature coding on time information corresponding to the data, dividing the data sets, and preparing for training a deep learning model.

Firstly, high-quality data sets inside and outside a port pool are constructed. Namely, acquiring or generating wind and wave data outside a target harbor pool and wave height data in the harbor pool, which respectively comprise: the average wind speed and the average wind direction of a wind field outside the harbor basin, the effective wave height, the average period and the average direction of waves outside the harbor basin, and the effective wave height data of the waves at each grid point inside the harbor basin. The acquisition of wind and wave related data inside and outside the harbor basin can be realized through a re-analysis data set disclosed on the network or a monitoring data set recorded by a monitoring site; if no public data set is available and the monitoring data is insufficient, a large amount of relevant data can be generated by adopting pattern calculation, calibration is carried out based on the monitoring data, and the calibrated pattern data is taken as a data set so as to ensure that a data set with higher quality is constructed.

Because the construction of the deep learning model depends on a large number of training data sets, but the target harbor pool buoy monitoring data of the example is less, and the investment cost for arranging a large number of monitoring buoys on the spot is too high, the example firstly generates a large number of wave height data of each part of the harbor pool based on a wave numerical model method of the target harbor pool, and as shown in fig. 3, the numerical model provided by the embodiment of the invention simulates the propagation and evolution diagram of the wave of the target harbor pool. And then, wave height data generated by a small amount of buoy monitoring data calibration modes are utilized, as shown in fig. 4, the simulation result of a single buoy monitoring position is compared with the real buoy monitoring result, so that a large amount of relatively high-quality wave height data are obtained, preparation is made for constructing a training data set and a testing data set of a deep learning network, and the purpose of deep learning model modeling for predicting the wave height of the port pool with high quality is met.

Next, the data is subjected to missing value and normalization processing. Uniformly processing the numerical value of the land part into 0 for the data in the harbor pool, wherein the missing value in the harbor pool can be filled according to bilinear interpolation; normalization processing is adopted for wave and wind field data at the position outside the harbor basin (the position is shown as the number 1 in figure 5), the data are converted into decimal numbers between 0 and 1, and dimensional expressions are changed into dimensionless expressions. This has two benefits: firstly, the convergence rate of the model is improved, and secondly, the precision of the model is improved. The normalization process employed in this example is as follows:

①

in the formula

Is the mean value of all the sample data,

is the standard deviation of all sample data.

And the normalized data of the waves and the wind field outside the harbor pool are respectively organized into characteristic vectors of the waves outside the harbor (effective wave height, average period and average direction) and wind fields outside the harbor (average wind speed and average wind direction).

And thirdly, carrying out multi-scale feature coding on the time information corresponding to the data to form a time feature vector. Encoding time information of different scales, and encoding current time information into a vector with the length of 36 by using one-hot, wherein the month of each month is encoded into a vector with the length of 12 by using one-hot, for example, the first dimension of the vector with the length of 12 dimensions is set as 1 in 1 month, and the positions of the rest dimensions are set as 0, namely the encoding is (1,0,0,0,0,0,0,0,0,0,0 in 2 months, and the like, the encoding of other months can be obtained; encoding the 24 hours of each day with one-hot as well as length 24 vectors; the required time information is coded into a vector with the total length of 36 by the method, the time information can be input into the deep learning network of the invention after being coded into the vector, and through continuous learning and training of the neural network, the model can learn the change of the wave height in the harbor basin caused by the tide or season influence and the periodicity of the wave height in the harbor basin along with the time.

Finally, the data sets are divided for deep learning model learning and training. This example will divide the post-processing data into 80% as the training set, 10% as the validation set, and 10% as the test set.

(2) The deep learning model establishing and training module 2 is used for establishing a deep learning model for predicting the effective wave height of the harbor pool based on the generated countermeasure network, performing alternate iterative training, learning and model parameter adjusting and optimizing on the generator network and the discriminator network, and establishing an end-to-end prediction model for the effective wave height of the harbor pool;

firstly, a generation confrontation network deep learning model (the specific structure of the model is shown in fig. 6) is built, and the generation confrontation network deep learning model comprises a generator and a discriminator, wherein: the generator network is responsible for generating wave height prediction data in the harbor pool, and the discriminator network is responsible for comparing the generated effective wave height data in the harbor pool with a true value to discriminate the credibility of the effective wave height data, so that when the discriminator cannot distinguish true from false, the prediction result of the generator network model approaches the true value. Therefore, a model with accurate prediction results is obtained through training of a large amount of data. The generator has an internal structure of a multilayer perceptron with a residual error structure, the discriminator has an internal structure of a multilayer perceptron, the multilayer perceptron is a nonlinear and self-adaptive information processing system formed by interconnection of a large number of processing units and can fit a very complex nonlinear relation, the multilayer perceptron generally has a plurality of hidden layers, and a single perceptron has certain fitting capacity, so the multilayer perceptron has stronger fitting capacity and can be used for solving the more complex problem. The specific two-part model is built as follows:

the construction generator part: the generator is a multi-layer perceptron with a residual error structure, wherein a main network contains 4 hidden layers, the residual error part contains 2 hidden layers, and an activation function ReLU is used behind each hidden layer to increase the nonlinearity of an output result; meanwhile, the wave height inside the harbor pool is considered to be periodically influenced by wind fields outside the harbor, waves, tides, seasons and the like, so that three types of characteristics (preprocessed wave and wind field data outside the harbor pool and multi-scale coded time information) are spliced and fused to be input into a generator model, the influence of the waves outside the harbor is further enhanced by a residual error structure inside the generator, and the network prediction performance is improved; the input data is mapped into m x n vectors (m and n are consistent with the grid size calculated by the wave height in the harbor basin) through a multilayer perceptron, and the vectors are combined to the actual terrain. The final part of the generator converts the vector into a matrix of m x n, and the matrix is the port pool wave height prediction data generated by the invention.

Constructing a discriminator part: the discriminator uses a multi-layer perceptron, which contains 3 hidden layers, after each hidden layer using ReLU as the activation function, and finally the hidden layers using Sigmoid to set the confidence score for the resulting image in the range of 0 to 1. The input of the discriminator is that the generated wave height data and the generated truth value data are spliced and fused with the wave characteristic vectors outside the harbor pool respectively except the effective wave height data and the truth value data in the harbor pool generated by the generator, and then the generated wave height data and the truth value data are input into the discriminator to solve the loss respectively, so that the wave information outside the harbor pool is used as an extra condition input by the discriminator to improve the capability of the discriminator for discriminating the data accuracy.

Secondly, the number of neurons of a hidden layer of a neural network generator and a discriminator model is needed to be set, the weight of the neural network is initialized, then a gradient descent optimizer and the number of training rounds epoch, the learning rate and other hyper-parameters are set, loss functions of the neural network generator and the discriminator model are respectively set as a formula II and a formula III, and the method comprises the following steps:

the loss function of the generator part is shown as formula (II):

②；

wherein the content of the first and second substances,

the features of the input model after splicing and fusion are shown,

representing the data generated by the generator model,

indicating the accuracy score value after discrimination by the discriminator model,

representing the number of samples of the batch input network;

③；

wherein the content of the first and second substances,

is the true value data of the wave effective wave height,

to generate the significant wave height data for the generator,

representing the number of samples input to the network in bulk

Thirdly, the deep learning model training and parameter adjusting process comprises the following steps:

and alternately training the discriminator and the generator, terminating the training when the discriminator cannot judge true and false data, repeatedly adjusting the structure and the super parameters of the generator and the discriminator by using a verification set in the training, and finally determining the optimal network model structure parameters according to multiple comparison experiment results. When the discriminator is trained, the discriminator inputs effective wave height data and true value data of the wave height in the harbor pool generated by the generator, in order to further improve the capability of the discriminator for discriminating the data accuracy, wave information outside the harbor pool is used as an additional condition input by the discriminator, the generated wave height data and the wave height true value data are spliced and fused with wave characteristic vectors outside the harbor pool respectively, then the wave height data and the wave height true value data are input into the discriminator respectively for training, then the network parameters of the discriminator are updated through a reverse propagation gradient according to a loss function, and the identification performance of the discriminator is continuously improved through multi-round iterative training; when a generator is trained, network parameters of a discriminator are fixed, three types of spliced and fused features (pretreated waves, wind fields and coded multi-scale time information outside a harbor pool) are input into the built input generator in batch, wave height data generated by the generator is discriminated by the discriminator, accuracy scores (0-1) of the wave height data generated by the generator are given, the score is closer to 1, the score represents that the harbor pool wave height data generated by a model is closer to a true value input in the case, then gradient descent is performed according to a loss function, the generator is updated through error back propagation, and network parameters of the generator are updated, so that the generator generates more accurate wave height data; when the discriminator can not distinguish the true data from the false data, the training is terminated; in the training process, the structure and the super-parameter adjustment are repeatedly carried out on the generator and the discriminator by using the verification set, namely, the structure of the network model and the complexity of the control model are adjusted according to multiple comparison experiment results.

And finally, establishing an end-to-end high-quality port pool wave effective wave height prediction model, and also needing to evaluate the performance of the generator model, predicting the generator model trained in the last step on a test set, if the prediction result is worse than the current optimal reference model, continuing to adjust model parameters for retraining, otherwise, if the prediction result is obviously better than the current optimal reference model, obtaining a deep learning model capable of accurately predicting the effective wave height of each position in the port pool.

(3) The effective wave height of the harbor pool waves is predicted by the module 3, when the prediction model of the trained harbor pool effective wave height is used for prediction, only the trained generator network model part is needed to be used for predicting the effective wave height in the harbor pool, the specific model is shown in fig. 7, wind and wave data outside the harbor pool are subjected to the same missing value and normalization processing as those of the data set construction and the preprocessing module 1 and time multi-scale coding, the obtained three characteristics of wind, wave and time are directly spliced and fused, input into the trained generator model and subjected to forward propagation through the generator network, and the effective wave height of the waves in the whole harbor pool can be predicted quickly and accurately.

Example 3

Based on the above embodiment 2, preferably, the wind field data (average wind speed and average wind direction) and the wave data (effective wave height, average period, average direction) outside the harbor basin are obtained by reanalysis or monitoring data sets disclosed by the website, or data with higher quality is generated after the numerical model is solved and the monitoring data is calibrated.

Example 4

Based on the embodiment 2, the wave height data in the harbor pool is preferably generated by the calculation of a numerical model and the calibration of monitoring data, and can also be acquired by a large number of monitoring sensing devices arranged in the harbor or acquired by public reanalysis data.

Application example:

han Bangtotang, also known as Han Bangtotang deepwater Port, is located in the Han Bangtotang prefecture, south of Si Li lan Ka. The method is a main harbor of the Steran card and plays an important role on international airlines between Asia and Europe, the wave condition in a harbor pool, particularly the wave height in the harbor, can have important influence on the loading and unloading operation of ships, and the specific embodiment of the method is to predict the effective wave height of waves in the Hanbangtota harbor pool.

In this embodiment, the present invention selects the wave field data near the entrance of the hanbantottower port, the position of which is shown as (r) in fig. 5, where the wave is the incident wave of the hanbantottower port, and we obtain the reanalysis data of the wave and the wind field from the ERA5 data set, and then predict the effective wave height of the wave generated inside the hanbantottower port under the input condition.

The positive effects of the present invention will be further described below with reference to specific experiments.

The experiment comprises the following specific steps: in this embodiment, according to the above-mentioned prediction method of the total harbor basin wave effective wave height, the method is implemented according to three modules:

1. a data set construction and preprocessing module:

(1) firstly, wind and wave data outside a port pool are obtained from reanalysis data of ERA5, effective wave height data inside the port are constructed in a generation mode because no public data which are directly available at present, a SWASH numerical model is adopted to simulate a wave propagation process as shown in figure 3, then monitoring data obtained by a small number of sensors arranged inside the port pool are used for calibrating the wave height data generated by the mode, the effective wave height of numerical solution is enabled to approach an actual monitoring value as much as possible, and the effective wave height data generated by the calibrated mode is considered to be a true value within an acceptable range (as shown in figure 4). This yields a large amount of mode-aligned significant wave height data to prepare a high quality data set for further deep learning model building and training.

(2) Next, the data is subjected to missing value and normalization processing. Uniformly processing the numerical value of the land part into 0 for the data in the harbor pool, wherein the missing value in the harbor pool can be filled according to bilinear interpolation; normalization processing is respectively carried out on wave and wind field data outside a harbor basin (the position is shown as a serial number 1 in figure 5), the data are converted into decimal numbers between 0 and 1, and dimensional expressions are changed into dimensionless expressions. This has two benefits: firstly, the convergence rate of the model is improved, and secondly, the precision of the model is improved. The normalization process employed in this example is as follows:

；

in the formula

Is the mean value of all the sample data,

is the standard deviation of all sample data. Then the normalized data of the waves and the wind field outside the harbor basin are respectively organized into wave characteristic vectors (effective wave height, average period and average direction) and wind field characteristicsEigenvectors (mean wind speed, mean wind direction).

(3) And finally, carrying out multi-scale coding on the time information corresponding to the data to obtain a time characteristic vector. Encoding time information of different scales, and encoding current time information into a vector with the length of 36 by using one-hot, wherein the month of each month is encoded into a vector with the length of 12 by using one-hot, for example, the first dimension of the vector with the length of 12 dimensions is set as 1 in 1 month, and the positions of the rest dimensions are set as 0, namely the encoding is (1,0,0,0,0,0,0,0,0,0,0 in 2 months, and the like, the encoding of other months can be obtained; encoding the 24 hours of each day with one-hot as well as length 24 vectors; the required time information is coded into a vector with the total length of 36 by the method, the time information can be input into the deep learning network of the invention after being coded into the vector, and through continuous learning and training of the neural network, the model can learn the change of the wave height in the harbor basin caused by the tide or season influence and the periodicity of the wave height in the harbor basin along with the time.

(4) Again, the partitioned data sets are used for deep learning model learning and training. This example will divide the post-processing data into 80% as the training set, 10% as the validation set, and 10% as the test set.

2. Deep learning model building and training module: based on the thought of generating an antagonistic neural network, a deep learning neural network model shown in FIG. 6 is built, a large amount of iterative training, learning and model parameter adjusting and optimizing are carried out, and an end-to-end high-quality harbor pool wave effective wave height prediction model is built.

(1) Firstly, building a generation confrontation network deep learning model, as shown in fig. 6, building the deep learning model based on the generation confrontation network idea, including two parts, namely a generator and a discriminator, specifically building two parts of models:

the construction generator part: the generator is a multi-layer perceptron with a residual error structure, wherein a main network contains 4 hidden layers, the residual error part has 2 hidden layers, and an activation function ReLU is used behind each hidden layer to increase the nonlinearity of an output result; in this embodiment, through multiple comparison experiments, the number of neural units of the optimal network hidden layer is set as 32-64-256-1024-65536 of the neural units of the generator hidden layer, and the residual part is 32-64; meanwhile, the influence of the wave height outside the harbor pool on the outside wind field, the wave, the tide, the season and the like is considered, so that three types of characteristics (the preprocessed wave and wind field data outside the harbor pool and the multi-scale coded time information) are spliced and fused and input into a generator model (the input data are mapped into 256 × 256 vectors through a multi-layer perceptron, and the size of a wave height calculation grid inside the harbor pool is kept consistent with 256 × 256), and the influence of the waves outside the harbor pool is further enhanced by a residual error structure inside the generator, and the network prediction performance is improved; in combination with the actual terrain, the invention multiplies the output vector by the terrain mask, which preserves the data in the area inside the harbor basin and sets the land part outside the harbor basin to 0. The final part of the generator converts the vector into a matrix (256 ), which is the port pool wave height prediction data generated by the invention.

Constructing a discriminator part: the discriminator uses a multi-layered perceptron, which contains 3 hidden layers, after each hidden layer using ReLU as the activation function, and finally the hidden layers using Sigmoid to set the confidence score for the resulting image in the range of 0 to 1. The number of neural units in the hidden layer of the discriminator is 65536-1024-256-64-1. The input of the discriminator is that the generated wave height data and wave height truth value data are spliced and fused with the wave characteristic vectors outside the harbor pool respectively except the effective wave height data and the wave height truth value data in the harbor pool generated by the generator, and then the wave height data and the wave height truth value data are input into the discriminator to solve the loss respectively, so that the wave information outside the harbor pool is used as an additional condition input by the discriminator to improve the capability of the discriminator in discriminating the data accuracy.

(2) Secondly, the neuron parameters are initialized, then the weight of the neural network is set, and the gradient descent optimizer, the number of training rounds epoch, the learning rate and other hyper-parameters are set. Through a large number of experiments, a better model hyper-parameter setting is selected here: the number of training rounds epoch is 500-1000, the learning rate is 0.001-0.01, Adam is used in a generator for gradient optimization, and SGD is used in a discriminator for gradient optimization; setting the loss functions of the neural network generator and the discriminator model as a formula II and a formula III respectively, wherein the formula II and the formula III comprise the following steps:

the loss function of the generator part is shown as formula (II):

②；

wherein, the first and the second end of the pipe are connected with each other,

the characteristics of the input model after splicing and fusion are shown,

representing the data generated by the generator model,

representing the number of samples input to the network in bulk;

the Loss function of the discriminator part is binary cross entropy BCE Loss, as shown in (c):

③；

wherein the content of the first and second substances,

is the true value data of the wave effective wave height,

to generate the significant wave height data for the generator,

representing the number of samples input to the network in bulk.

(3) Thirdly, the deep learning model training and parameter adjusting process comprises the following steps:

and alternately training the discriminator and the generator, terminating the training when the discriminator cannot judge true and false data, repeatedly adjusting the structure and the super parameters of the generator and the discriminator by using a verification set in the training, and finally determining the optimal network model structure parameters according to multiple comparison experiment results.

FIG. 6 illustrates the composition of the complete model that participates in the model training process. When the discriminator is trained, wave data outside a harbor pool is used as an additional condition input by the discriminator to improve the discrimination accuracy of the discriminator under the wave condition, wave height data and truth value data generated by a generator are respectively spliced and fused with wave characteristic vectors outside the harbor pool, then the wave height data and the truth value data are respectively input into the discriminator to be trained, then network parameters of the discriminator are updated through a reverse propagation gradient according to a loss function, and the identification performance of the discriminator is continuously improved through multiple rounds of iterative training; when a generator is trained, network parameters of a discriminator are fixed, three types of characteristics of splicing and fusion (the tensor shapes of wave and wind field input outside a harbor pool after pretreatment are all (batch _ size, 3), the time characteristic tensor size is (batch _ size, 32), the spliced and fused characteristics are input into a built input generator in batch, the generator is multiplied by a harbor pool terrain mask (the land part is set to be 0) when outputting, data in the harbor pool are reserved to obtain the wave height data of the harbor pool, the wave height data generated by the generator is discriminated by the discriminator to give an accuracy score (0-1) of the wave height data generated by the generator, the closer the score is to 1, the closer the harbor pool wave height data generated by a model is to a true value input in the case, then gradient descent is performed according to a loss function, the generator is updated through error reverse propagation, the network parameters of the generator are updated, so that the generator generates more accurate wave height data.

In the training process, the structure and the super-parameter adjustment are repeatedly carried out on the generator and the discriminator by using the verification set, namely, the network model structure and the control model complexity are adjusted according to multiple comparison experiment results, and when the discriminator cannot distinguish true data from false data, the training is stopped. Earlystopping is set here for terminating the model training in advance, when training the deep learning model, the probability of the authenticity of the generated data judged by the discriminator is about 0.5, namely the true and false data can not be distinguished to reach Nash equilibrium, and the model training is terminated when 20 epochs have no loss reduction.

(4) And finally, establishing an end-to-end high-quality harbor pool wave effective wave height prediction model, and needing to evaluate the performance of a generator model, predicting the generator model trained in the last step on a test set, if the prediction result is worse than the current optimal reference model, continuing to adjust model parameters for retraining, otherwise, if the prediction result is obviously better than the current optimal reference model, obtaining a deep learning model capable of accurately predicting the effective wave height of each position in the harbor pool.

3. And a module for performing effective wave height fast and accurate prediction by using a model: after the training is finished, the wave heights of all positions of the harbor basin are finally and accurately predicted, in the testing and using stage, firstly, wave data and wind field data outside the harbor basin are constructed according to a data set, preprocessed by a preprocessing module and subjected to multi-scale time coding, and then directly spliced and fused to be input into a trained harbor basin wave effective wave height prediction model (namely, as shown in fig. 7, a generator network model part of a trained deep learning model), so that the high-quality prediction of the wave effective wave heights of all positions in the harbor basin can be automatically and quickly realized.

As shown in fig. 8, the overall training process of the deep learning model provided by the embodiment of the present invention includes: firstly, a data set is constructed, when a large amount of publicly available data exists, public monitoring or reanalysis data are obtained, and when a large amount of available data does not exist, generated data of numerical model calibration is adopted, namely the data solved by the numerical model is corrected by using the monitoring data to adjust the parameters of the numerical model so as to enable the numerical model to generate more accurate data; then, data missing value processing and normalization preprocessing are carried out, multi-scale coding of time information is carried out, then, splicing and fusion wind and wave characteristics and multi-scale time characteristics are input into a neural network for training, a generator can generate intra-port wave height prediction data, a discriminator judges the accuracy of the generated wave height data, when the discriminator judges that the data generated by the generator is not true enough, the generator needs to be trained continuously, and when the discriminator judges that the generated data is true enough and has small errors, model training is finished;

fig. 9 is a flow of predicting the height of a harbor basin wave by using the deep learning model provided in the embodiment of the present invention, which specifically includes: after model training is completed, when the wave height in a target harbor basin is predicted, firstly, data processing of S101 is carried out on wave data outside the harbor basin and wind field data, then time information codes are spliced and fused, the wave data and the wind field data are input into a trained generator network model for forward propagation, and the generator model outputs a predicted value of effective wave height in the harbor basin.

In order to verify the effectiveness and accuracy of the model, in hanbangtatookang, the following two evaluation indexes MSE and R-Squared are used in the invention as shown in table 1 by applying the experimental result pair of the example model and other machine learning models, and the experimental result proves that the method has higher prediction accuracy.

Table 1: comparison of prediction results of the model of the embodiment with those of other machine learning models

The invention verifies the performance of the model by using a test set, selects the calibrated wave height data (true value) at a specific point to compare with the predicted wave height data of the embodiment of the invention, and shows that the method has higher prediction accuracy as shown in a comparison graph of the prediction result of the effective wave height in the harbor basin and the mode-resolved effective wave height (true value) after calibration provided by the embodiment of the invention in figure 10.

In summary, the thought of generating the anti-neural network by combining deep learning is adopted, and a deep learning model is built to predict the effective wave height in the harbor basin. When the deep learning model is used for predicting the wave height inside the harbor basin, no special requirements are required for the field knowledge and the related experience of a user. In addition, the end-to-end deep learning model is very simple and easy to use, and the effective wave height of the waves in the departure pool can be quickly and accurately predicted by only providing wind field data outside the port pool and wave representative information or wave direction spectrum information such as wave height, wave period and wave direction.

Compared with a numerical model prediction method, the construction of a numerical simulation model requires higher domain knowledge and accurate information of actual topography, boundary conditions and the like of a target harbor basin, and meanwhile, numerical simulation also takes a large amount of calculation time to simulate the propagation phenomenon of waves in the harbor basin. The method is based on data driving, once a deep learning model is constructed and trained, the high-quality porthole pool wave height prediction can be rapidly and efficiently carried out, and therefore, for a user, the user does not need to master too much knowledge in related fields.

Meanwhile, the prediction method also considers the characteristics of the waves, such as obvious regularity of the change of the wave height influenced by the seasonal wind, and the wave height is periodically changed due to tide or seasonal alternation, so that when the deep learning model is constructed, the wind speed and wind direction information outside the harbor basin and the multi-scale coded time information are simultaneously input into the deep learning model, the correlation between the wind field and the wave height and the periodicity of the wave height under different time scales are automatically learned, and the prediction accuracy of the deep learning model is obviously improved.

This application is intended to cover any variations, uses, or adaptations of the disclosure following, in general, the principles of the disclosure and including such departures from the present disclosure as come within known or customary practice within the art to which the disclosure pertains. It is intended that the specification and examples be considered as exemplary only, with a true scope and spirit of the disclosure being indicated by the following claims. It will be understood that the present disclosure is not limited to the precise arrangements described above and shown in the drawings and that various modifications and changes may be made without departing from the scope thereof.

Claims

1. A prediction method for the effective wave height of waves in a full harbor pool is characterized by comprising the following steps:

data set construction and preprocessing: constructing data sets inside and outside the harbor basin, preprocessing the data sets to generate characteristic vectors of wind fields and waves, carrying out multi-scale characteristic coding on time information corresponding to the data, and finally dividing the data sets to prepare for deep learning model training;

predicting the effective wave height of the harbor pool waves: and (3) after wind and wave data outside the harbor basin are subjected to the same deficiency value and normalization processing as those of the data set construction and pretreatment module (1) and time multi-scale coding, the obtained three characteristics of wind, wave and time are directly spliced and fused, and the obtained three characteristics are input into a trained harbor basin wave effective wave height prediction model to predict the effective wave height of waves in the whole harbor basin.

2. The method for predicting the wave height of the hong kong pool wave according to claim 1, wherein the data set construction and preprocessing specifically comprises:

3. The method for predicting the effective wave height of waves in the harbor basin as claimed in claim 2, wherein in the step 2), the data of the waves outside the harbor basin and the wind field are respectively normalized by converting the data into decimal numbers between 0 and 1 and converting dimensional expressions into dimensionless expressions as shown in the following formula (I)

①

In the formula

Is the mean value of all the sample data,

is the standard deviation of all sample data.

4. The method for predicting the wave height of the hong kong pool wave according to claim 1, wherein the building and training of the deep learning model specifically comprises:

5. The method for predicting the wave height of the hong Kong pool waves according to claim 4, wherein in the step (1), the building and generating a confrontation network deep learning model comprises the following steps: the generator and the discriminator are built:

6. The method for predicting the wave height of the full-harbor pool wave according to claim 5, wherein the loss function of the generator part is represented by the formula (II):

②；

representing the characteristics of the input model after splicing and fusion;

representing data generated by the generator model;

indicating the number of samples input to the network in bulk;

③；

wherein the content of the first and second substances,

true value data for the wave effective wave height;

the significant wave height data generated by the generator;

indicating the number of samples input to the network in bulk;

7. The method for predicting the wave height of the hong Kong pool wave according to claim 4, wherein in the step (2), the deep learning model training and parameter adjustment comprises:

alternately training the discriminator and the generator, stopping training when the discriminator cannot judge true and false data, repeatedly adjusting the structure and the hyper-parameters of the generator and the discriminator by using a verification set during the training, and finally determining the optimal network model structure parameters according to the results of multiple comparison experiments; when the discriminator is trained, inputting effective wave height data and true wave height data of waves in a harbor pool generated by a generator by the discriminator, splicing and fusing the generated wave height data and the true wave height data of the harbor pool with wave characteristic vectors outside the harbor pool respectively in order to further improve the capability of the discriminator for discriminating the data accuracy, inputting the generated wave information outside the harbor pool as an additional condition input by the discriminator, then inputting the generated wave height data and the true wave height data into the discriminator for training respectively, updating network parameters of the discriminator through a reverse propagation gradient according to a loss function, and performing multi-round iterative training;

when the generator is trained, fixing network parameters of a discriminator, inputting data into the generator in batches, giving an accuracy grade of the data generated by the generator after the data generated by the generator is discriminated by the discriminator, wherein the score is 0-1, then performing gradient descent according to a loss function, updating the generator through error back propagation, and updating the network parameters of the generator; when the discriminator can not distinguish the true data from the false data, the training is terminated; and in the training process, the structure and the super-parameter adjustment are repeatedly carried out on the generator and the discriminator by using the verification set, and the network model structure and the control model complexity are adjusted according to multiple comparison experiment results.

8. A prediction system of the effective wave height of the harbor pool wave according to the prediction method of the effective wave height of the full harbor pool wave of any one of claims 1 to 7, which is characterized by comprising:

the data set construction and pretreatment module (1) is used for constructing data sets inside and outside the harbor basin, carrying out pretreatment to generate characteristic vectors of a wind field and waves, carrying out multi-scale characteristic coding on time information corresponding to the data, and finally dividing the data set to prepare for deep learning model training;

the deep learning model establishing and training module (2) is used for establishing a deep learning model for predicting the effective wave height of the harbor pool based on the generated countermeasure network, performing alternate iterative training, learning and model parameter adjusting and optimizing on the generator network and the discriminator network, and establishing an end-to-end prediction model for the effective wave height of the harbor pool;

and the effective wave height of the waves in the harbor pool is predicted by a module (3) for predicting the effective wave height of the waves in the harbor pool, wind and wave data outside the harbor pool are subjected to missing value and normalization processing which are the same as those of the data set construction and pretreatment module (1) and time multi-scale coding, and the obtained three characteristics of the wind, the waves and the time are directly spliced and fused and input into a trained prediction model of the effective wave height of the waves in the harbor pool.

9. A program storage medium for receiving user input, wherein the stored computer program causes an electronic device to execute the method for predicting full pool wave effective wave height according to any one of claims 1 to 7.

10. Use of the method for predicting the effective wave height of the full harbour basin wave according to any one of claims 1 to 7 in the prediction of harbour basin waves, hydraulic engineering waves and near-shore engineering waves.