CN116340868A - A method and device for ensemble forecasting of wave features based on machine learning - Google Patents

Abstract

The invention discloses a wave characteristic set forecasting method and device based on machine learning, wherein the method comprises the steps of constructing a set forecasting model of wave height, wave period and wave direction through a multi-layer perceptron MLP and a decision tree DT machine learning model; taking historical wave height, wave period and wave direction data and parameters of several days of future wind as input of the set forecasting model; and the set forecasting model outputs future wave height, wave period and wave direction forecasting results. The invention builds a wave height, wave period and wave direction set prediction model by using a multi-layer perceptron (MLP) and a Decision Tree (DT) machine learning model, wherein the model builds a machine learning prediction model by learning the nonlinear relation between wind and the analysis historical data set disclosed by wave elements (including wave height, wave period and wave direction), and selects a prediction result with relatively high precision, thereby accurately predicting wave characteristics.

Description

A method and device for ensemble forecasting of wave features based on machine learning

technical field

The invention relates to ocean wave prediction technology, in particular to a machine learning-based wave feature set prediction method and device.

Background technique

At present, most of the methods for predicting ocean waves based on artificial intelligence only focus on one element of wave height, and the prediction results of wave period and wave direction cannot be obtained, and the purpose of predicting wave height is only achieved by learning wave height time series data sets, which cannot be generated through wave generation. physical mechanisms to explain the predicted results.

The physical mechanism of ocean wave generation is relatively mature at present, and the third-generation wave mathematical model based on the wave energy balance equation is also widely used. The current mainstream models mainly include WAM, WWIII and SWAN, which are based on the wave energy balance equation:

In the formula, N is the abbreviation of N(x,y,t; σ,θ), where σ is the frequency and θ is the direction; C _gx and C _gy are the propagation speeds of wave energy in the x and y directions; C _gσ and C _gθ is the propagation velocity in frequency and direction; S is the source term that includes different physical processes, mainly including deep water and shallow water processes, including physical processes such as wind and wave growth process, wave interaction, seabed friction, white hat process, and wave breaking .

The current forecast model only considers the complex nonlinear relationship between wind and waves, including wave-wave interaction, wind-wave growth and other mechanisms, but factors such as wave-current interaction, seabed friction, and the influence of terrain on waves are still not considered.

Contents of the invention

Aiming at the problem that the existing forecasting model only considers the complex nonlinear relationship between wind and sea waves, the present invention provides a wave feature set forecasting method and device based on machine learning.

For realizing the above object, technical scheme of the present invention is:

In the first aspect, the present invention provides a method for forecasting wave feature sets based on machine learning, including:

Construct an ensemble prediction model of wave height, wave period, and wave direction through multi-layer perceptron MLP and decision tree DT machine learning model;

Using historical wave height, wave period and wave direction data and the parameters of the wind in the next few days as the input of the ensemble forecast model;

The ensemble forecasting model outputs forecast results of future wave height, wave period and wave direction.

Further, the ensemble prediction model is constructed in the following way:

Obtain reanalysis historical data sets of wind and wave elements to generate training data set samples required for machine learning model learning; said wave elements include wave height, wave period, and wave direction

Train the MLP model and DT model based on the training data set samples;

According to the judgment factor, the combination of optimal forecast samples and models with different forecast lead times is selected to obtain the ensemble forecast model.

Further, the judgment factors include coefficient of determination R ² , mean absolute error MAE and mean absolute percentage error MAPE.

其中

Further, the determination coefficient

in

Further, the mean absolute error

Further, the average relative error

Further, the reanalysis historical data set of the wind and wave elements adopts model reanalysis data, or adopts a public reanalysis data set.

Further, the forecast lead time is an integer multiple of the training data set sample time interval.

In a second aspect, the present invention provides a machine learning-based wave feature set forecasting device, including a memory, a processor, and a computer program stored in the memory and operable on the processor, and the processor executes the When the above-mentioned computer program realizes the steps of any one of the above-mentioned methods.

In a third aspect, the present invention provides a computer-readable storage medium, the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of any one of the methods described above are implemented

Compared with the prior art, the present invention has the beneficial effects of:

The present invention constructs the ensemble prediction model of wave height, wave period and wave direction through multi-layer perceptron (MLP) and decision tree (DT) machine learning model, and this model is by learning wind and wave element (comprising wave height, wave period and wave direction) ) has been published to reanalyze the nonlinear relationship between historical data sets, establish a machine learning forecast model, and compare the pros and cons of the model at different forecast lead times through the coefficient of determination R ² , the average absolute error MAE and the average absolute percentage error MAPE, The forecast results with relatively high accuracy are selected so that the wave characteristics can be accurately forecasted.

Description of drawings

Fig. 1 is the flow chart of the method for forecasting wave feature sets based on machine learning provided by Embodiment 1 of the present invention;

Fig. 2 is the training set composition of the prediction model of significant wave height;

Figure 3 is a statistical chart of the machine learning model score and mean absolute error for significant wave height prediction;

Figure 4 is a statistical diagram of the average relative error of the significant wave height forecasting machine learning model;

Fig. 5 is the average wave cycle forecasting machine learning model score and the average absolute error statistics;

Fig. 6 is the average relative error statistical diagram of the average period forecast machine learning model;

Fig. 7 is the average wave direction prediction machine learning model score and the statistical graph of the average absolute error;

Fig. 8 is the average relative error statistical diagram of the average wave direction forecasting machine learning model;

Fig. 9 is a statistical diagram of prediction result accuracy;

Fig. 10 is a schematic diagram of the composition of the machine learning-based wave feature set forecasting device provided by Embodiment 2 of the invention.

Detailed ways

The content of the present invention will be described in further detail below in conjunction with the accompanying drawings and specific embodiments.

Example 1:

Compared with traditional numerical models based on physical relationships, machine learning generally has two advantages in forecasting waves: 1. Less computing resources are required for forecasting, and the time required for forecasting is short; 2. High forecasting accuracy effectively solves the time limit for ocean wave forecasting problem.

The current forecast model only considers the complex nonlinear relationship between wind and waves, including wave-wave interaction, wind-wave growth and other mechanisms, but factors such as wave-current interaction, seabed friction, and the influence of terrain on waves are still not considered. Among them, terrain factors and bottom friction generally do not change with time, and their influence is reflected in the historical data set to a certain extent, and has little impact on the forecast results; ocean currents change with time, and should be used as input conditions to train the model like wind, so the size of the training data set Increase, the computer memory required for the training process of the machine learning model is greatly increased.

For this reason, the present invention constructs the ensemble prediction model of wave height, wave period and wave direction through multi-layer perceptron (MLP) and decision tree (DT) machine learning model, and this ensemble prediction model is by learning wind and wave element (comprising wave height, Wave period and wave direction) have been published to reanalyze the nonlinear relationship between historical data sets, establish a machine learning forecasting model, and compare the models in different forecasting advances through the coefficient of determination R ² , the mean absolute error MAE and the mean absolute percentage error MAPE According to the advantages and disadvantages of time, choose the forecast result with relatively high accuracy.

The multi-layer perceptron model is developed based on the neural model of the human brain. It is a forward-structured artificial neural network that maps a set of input vectors to a set of output vectors. An MLP can be viewed as a directed graph consisting of multiple layers of nodes, each fully connected to the next layer. Except for the input node, each node is a neuron (or processing unit) with a nonlinear activation function. A supervised learning method known as the backpropagation algorithm is often used to train MLPs. A multilayer perceptron follows the principles of the human nervous system, learning and making predictions from data. It first learns, then stores the data with weights, and uses algorithms to adjust the weights and reduce bias, the error between actual and predicted values, during training. The main strength is its ability to quickly solve complex problems. The basic structure of multi-layer perception consists of three layers: the first input layer, the middle hidden layer and the final output layer. The product of input elements and weights is fed to the summation node with neuron bias. The main advantage lies in its fast resolution of complex ability to problem. It can learn non-linear function samples to identify transactions with n feature vectors.

The decision tree regression model is also a kind of machine learning method integrated in this forecast model. The regression decision tree mainly refers to the CART (classification and regression tree) algorithm. The values of internal node features are "yes" and "no", which is a binary tree structure . The so-called regression is to determine the corresponding output value according to the feature vector. The regression tree is to divide the feature space into several units, and each division unit has a specific output. Because each node is a "yes" and "no" judgment, the boundaries of the division are parallel to the coordinate axes.

Referring to Fig. 1, the machine learning-based wave feature set prediction method provided in this embodiment mainly includes the following steps:

101. Construct an ensemble prediction model of wave height, wave period, and wave direction through multi-layer perceptron MLP and decision tree DT machine learning model;

102. Using historical wave height, wave period and wave direction data and wind parameters in the next few days as the input of the ensemble forecast model;

103. The ensemble forecast model outputs future forecast results of wave height, wave period and wave direction.

In a specific embodiment, the above-mentioned ensemble prediction model is constructed in the following way:

Obtain reanalysis historical datasets of wind and wave elements (e.g. ERA5 dataset, see https://www.ecmwf.int/en/forecasts/datasets/reanalysis-datasets/era5) to generate the data needed for machine learning model learning Training dataset; the wave features include wave height, wave period, and wave direction.

The MLP model and DT model are trained based on the samples of the training data set; taking the forecast model of significant wave height as an example, the composition of the training data set is shown in Figure 2:

Figure 2 includes initial wave conditions: significant wave height (SWH), mean period (MWP) and mean wave direction (MWD) at time T; initial wind field and wind field of forecast step (n-1); the output is 1 The wave characteristic variable at +n time, if the effective wave height is trained, it is the corresponding effective wave height value.

According to the judgment factor, the combination of optimal forecast samples and models with different forecast lead times is selected to obtain the ensemble forecast model.

Specifically, the above judgment factors include coefficient of determination R ² , mean absolute error MAE and mean absolute percentage error MAPE.

其中/>

The coefficient of determination

where />

The mean absolute error

The mean relative error

In addition, in order to improve the calculation timeliness and accuracy of prediction, the reanalysis historical data set available for learning can use the model reanalysis data, or use the public reanalysis data set, the present invention uses the hourly data of ERA5 for verification; for training The time interval of the reanalysis data set has a decisive effect on the forecast lead time, and the forecast lead time of the present invention is an integer multiple of the time interval of the training samples; to predict different wave elements, it is necessary to change the output variable of the training data set, and different output variables correspond to different Forecast of wave elements.

Therefore, compared with the prior art, the present invention has the following technical advantages:

1. Occupies less computing resources and takes less time:

The main technical content of the present invention can be divided into two steps: 1 training model; 2 model prediction

Computer cpu type: Intel(R) Xeon(R) Gold 6226R CPU@2.90GHz

The time required to train the model varies depending on the number of samples. The most commonly used model for forecasting 8 steps is about 3600 seconds (one hour), and the time required for prediction is about 5 seconds. It may take 4-5 hours to use the same level of cpu and the same forecast area, greatly improving the calculation time.

2. High forecast accuracy:

Here, era5 is used as an example, as shown in Figure 3-8, which shows the forecast accuracy of different forecast lead times:

For the significant wave height SWH, when forecasting 1 hour in advance, MAE≈0.102m, MAPE≈0.04; when forecasting 1 day in advance, MAE≈0.26m, MAPE≈0.127; when forecasting 4 days in advance, MAE≈0.1037m, MAPE ≈0.189;

For the average period MWP, when forecasting 1 hour in advance, MAE≈0.188 seconds, MAPE≈0.02; when forecasting 1 day in advance, MAE≈0.487 seconds, MAPE≈0.071; when forecasting 4 days in advance, MAE≈0.848 seconds, MAPE ≈0.108;

For the average period MWD, when forecasting 1 hour in advance, MAE≈6.185 degrees, MAPE≈0.159; when forecasting 1 day in advance, MAE≈12.34 degrees, MAPE≈0.178; when forecasting 4 days in advance, MAE≈17.06 degrees, MAPE ≈0.27;

The technical effect of the present invention is further described in conjunction with an experimental example below:

The sea waves in the Pearl River Estuary [112°E-117°E, 18°N-23°N] area are predicted by designing experiments. The training set data uses the reanalysis data set of wind and waves from 2012 to 2021 in ERA5 for 11 years. And analyze the feasibility, advantages and disadvantages of this method. This calculation example selects the radial and latitudinal components of the wind for a total of 87,672 hours in 11 years from 2012 to 2021, as well as the effective wave height SWH, average wave period MWP, and average wave direction MWD as the input variables of the training data set. Arranged in a certain order; in order to increase the forecast duration, the present invention extracts four kinds of samples from the original samples according to 1 hour, four intervals of 3 hours, 6 hours and 12 hours for calculation, and each kind of sample predicts 8 steps in the future (one Step size and sample interval time are equal), the combination of different samples and forecast step size can get different training sample size and forecast lead time, details can be seen in the table below.

Table training data for each time interval, forecast lead time and corresponding training data array size

The accuracy of the training model for various combinations is as follows:

Table B.2: MAE for SWH

Table B.3: MAPE for SWH

Table B.5: MAE for MWP

Table B.6: MAPE for MWP

Table B.8: MAE for MWD

Table B.9: MAPE for MWD

Using the trained model, input historical data and wind parameters in the next few days, it can be used to predict future wave height, wave period and wave direction.

Analysis of forecast results of wave elements during typhoons:

The wind and wave data during 74 typhoons during ERA52001-2011 were extracted, and the pre-trained models were used to make predictions. The accuracy statistics of the prediction results of the two models are shown in Figure 9, and the forecast results are basically consistent with the actual wind and wave data.

Example 2:

Referring to Fig. 3, the machine learning-based wave feature collection forecasting device provided in this embodiment includes a processor 101, a memory 102, and a computer program 103 stored in the memory 102 and operable on the processor 101, such as An ensemble forecasting program for wave features based on machine learning. When the processor 101 executes the computer program 103, the steps in the first embodiment above are implemented, for example, the steps shown in FIG. 1 .

Exemplarily, the computer program 103 may be divided into one or more modules/units, and the one or more modules/units are stored in the memory 102 and executed by the processor 101 to complete this invention. The one or more modules/units may be a series of computer program instruction segments capable of accomplishing specific functions, and the instruction segments are used to describe the execution process of the computer program 103 in the machine learning-based wave feature set forecasting device .

The machine learning-based wave feature set forecasting device can be computing devices such as desktop computers, notebooks, palmtop computers, and cloud servers. The apparatus for predicting wave feature sets based on machine learning may include, but is not limited to, a processor 101 and a memory 102 . Those skilled in the art can understand that Figure 5 is only an example of a machine learning-based wave feature set forecasting device, and does not constitute a limitation of the machine learning-based wave feature set forecasting device, and may include more or less components than those shown in the illustration , or combine certain components, or different components, for example, the machine learning-based wave feature set forecasting apparatus may also include input and output devices, network access devices, buses, and the like.

The so-called processor 101 may be a central processing unit (Central Processing Unit, CPU), and may also be other general-purpose processors, a digital signal processor (Digital Signal Processor, DSP), an application specific integrated circuit (Application Specific Integrated Circuit, ASIC), Off-the-shelf programmable gate array (Field Programmable Gate Array, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor, or the processor may be any conventional processor, or the like.

The memory 102 may be an internal storage element of the device for forecasting a set of wave features based on machine learning, such as a hard disk or memory of the device for forecasting a set of wave features based on machine learning. The memory 102 can also be an external storage device of the machine learning-based wave feature set forecasting device, such as a plug-in hard disk equipped on the machine learning-based wave feature set forecasting device, a smart memory card (SmartMedia Card, SMC), Secure Digital (Secure Digital, SD) card, Flash Card (Flash Card), etc. Further, the memory 102 may also include both an internal storage unit of the machine learning-based wave feature set forecasting apparatus and an external storage device. The memory 102 is used to store the computer program and other programs and data required by the machine learning-based wave feature set forecasting device. The memory 102 can also be used to temporarily store data that has been output or will be output.

Example 3:

This embodiment provides a computer-readable storage medium, where the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the method described in Embodiment 1 are implemented.

The illustrated computer readable medium can be any means that can contain, store, communicate, propagate or transport the program for use by or in connection with an instruction execution system, apparatus or device. More specific examples (non-exhaustive list) of computer-readable media include the following: electrical connection with one or more wires (electronic device), portable computer disk case (magnetic device), random access memory (RAM), Read Only Memory (ROM), Erasable and Editable Read Only Memory (EPROM or Flash Memory), Fiber Optic Devices, and Portable Compact Disc Read Only Memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program can be printed, for example by optically scanning the paper or other medium, followed by editing, interpretation or other suitable means if necessary The process then obtains the program electronically and stores it in the computer memory.

The above-mentioned embodiments are only to illustrate the technical concept and characteristics of the present invention, and its purpose is to enable those of ordinary skill in the art to understand the content of the present invention and implement it accordingly, and cannot limit the protection scope of the present invention. All equivalent changes or modifications made according to the essence of the content of the present invention shall fall within the protection scope of the present invention.

Claims (10)

1. A machine learning based wave feature set forecasting method, comprising:

constructing a set forecasting model of wave height, wave period and wave direction through a multi-layer perceptron MLP and a decision tree DT machine learning model;

taking historical wave height, wave period and wave direction data and parameters of several days of future wind as input of the set forecasting model;

and the set forecasting model outputs future wave height, wave period and wave direction forecasting results.

2. The machine learning based wave feature set forecasting method of claim 1, wherein the set forecasting model is constructed by:

acquiring an analysis historical data set of wind and wave elements to generate training data set samples required by machine learning model learning; the wave elements include wave height, wave period and wave direction;

training an MLP model and a DT model based on the training dataset samples;

and selecting the combination of the optimal prediction samples and the models with different prediction advance time according to the judgment factors to obtain the set prediction model.

3. The machine learning based wave feature set forecasting method of claim 2, wherein the decision factor includes a decision coefficient R ² Average absolute error MAE and average absolute percent error MAPE.

4. The machine learning based wave feature set forecasting method of claim 3, wherein said decision coefficients

Wherein->

5. The machine learning based wave feature set forecasting method of claim 1, wherein said average absolute error

6. The machine learning based wave feature set forecasting method of claim 1, wherein said average relative error

7. The machine learning based wave feature set forecasting method of claim 2, wherein the analysis-history data set of wind and wave elements is employed with pattern analysis data or with a published analysis data set.

8. The machine learning based wave feature set forecasting method of claim 1 or 2, wherein the forecasting lead time is an integer multiple of the training data set sample time interval.

9. A machine learning based wave feature set forecasting device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that the processor implements the steps of the method according to any one of claims 1 to 8 when executing the computer program.

10. A computer readable storage medium storing a computer program, characterized in that the computer program when executed by a processor implements the steps of the method according to any one of claims 1 to 8.

Images