A ship motion prediction method based on AR-OLSSVR model

By combining iterative learning algorithms and autoregressive models with online least squares support vector regression models, the high precision and time cost problems of online ship motion forecasting are solved, and efficient ship motion forecasting is achieved.

CN114564694BActive Publication Date: 2024-11-19DALIAN MARITIME UNIVERSITY
View PDF 6 Cites 0 Cited by

Patent Information

Application Number
CN202210108766.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-01-28
Publication Date
2024-11-19
Estimated Expiration
2042-01-28

AI Technical Summary

Technical Problem

The prior art is difficult to achieve high-precision online forecast of ship motion, especially in nonlinear non-stationary characteristic data. A single forecast model is difficult to fully capture feature information and has a high time cost.

Method used

The combination of iterative learning algorithm, autoregression (AR) model and online least squares support vector regression (OLSSVR) model is used to determine the order of the AR model through the Akaike information criterion, establish the AR model and iteratively learn it, and regression is carried out in combination with the OLSSVR model. Error forecasting, realize online forecasting of dual models.

Benefits of technology

The accuracy and computing speed of ship motion forecasting are improved, and online forecasting is realized, which avoids the insufficient capture of nonlinear non-stationary data feature information by a single model, and reduces the model operation time.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114564694B_ABST
    Figure CN114564694B_ABST
Patent Text Reader

Abstract

The present invention discloses a ship motion prediction method based on an AR-OLSSVR model, comprising the following steps: establishing and training an AR model; establishing a regression equation of ship motion sample data and establishing a ship motion prediction model based on the AR model; establishing an OLSSVR model; and adding the ship motion prediction value obtained by the ship motion prediction model and the AR model regression error prediction value obtained by the OLSSVR model to each other as the final ship motion prediction value. The present invention adopts an iterative learning strategy to update the parameters of the ship motion prediction model and the regression error prediction model in real time. Each time the parameters are updated, only the last result needs to be used to calculate in an iterative manner, and there is no need to re-use a large amount of sample data for training, thereby improving the calculation speed of the prediction model and realizing the online prediction of ship motion. The present invention adopts dual-model prediction to improve the accuracy of ship motion prediction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of ship and ocean engineering, and in particular to a ship motion prediction method which integrates an iterative learning algorithm, an autoregressive (AR) model and an online least squares support vector regression (OLSSVR) model. Background Art

[0002] As a kind of marine transfer equipment, the offshore corridor is equivalent to a marine passage, which facilitates the transfer between the host ship and the target ship or offshore platform. When transferring at sea, the ship will produce six degrees of freedom movement of sway, surge, heave, roll, pitch and bow pitch under the influence of the marine environment such as wind, waves and current, which will cause safety hazards to the transfer of personnel at sea. The sway, surge and bow pitch of the ship can be compensated by the dynamic positioning system, while the roll, pitch and heave movement need to be compensated by the offshore corridor so that the corridor can always remain stable, just like on land, to provide protection for the safe transfer of personnel at sea. In order to improve the wave compensation control accuracy of the offshore corridor and realize the synchronous compensation of the disturbance of the offshore corridor caused by the ship's roll, pitch and heave movement caused by waves and other marine environments to the offshore corridor, it is necessary to make real-time and extremely short-term forecasts of ship movement.

[0003] Chinese invention patent CN103926932A (examination) discloses a method for intelligent prediction of ship motion posture decomposition domain, which realizes ship motion prediction by combining data preprocessing method with neural network. Chinese invention patent CN111709140A (examination) discloses a method for ship motion prediction based on intrinsic plasticity echo state network, which realizes ship motion prediction by applying new echo state network. However, none of the above patents can realize online prediction. Chinese invention patent CN113156815A (examination) discloses a data-driven real-time prediction method for marine ship motion posture, which realizes online prediction by combining sliding window and neural network, but the time cost of this method is large. In addition, the above patents all use a single prediction model for training and prediction. It is difficult to fully capture the characteristic information of ship motion data with nonlinear and non-stationary characteristics, and it is difficult to meet the actual needs of high-precision prediction. Summary of the invention

[0004] In order to solve the above problems, the present invention proposes a ship motion prediction method based on the AR-OLSSVR model to realize online prediction of ship motion, reduce model calculation time and improve prediction accuracy.

[0005] In order to achieve the above object, the technical solution of the present invention is as follows:

[0006] A ship motion prediction method based on the AR-OLSSVR model comprises the following steps:

[0007] A. Collect ship motion data with T as the sampling period, and obtain n sample data x k , k = 1, 2, ..., n; the ship motion includes rolling motion, pitching motion and heave motion;

[0008] B. Build and train the AR model;

[0009] B1. Establish the AR model of the ship motion sample data as follows:

[0010]

[0011] In the formula, p is the undetermined order of the AR model, a i is the undetermined coefficient of the AR model, let a=[a1,a2,…,a p ] T is the undetermined coefficient vector of the AR model, e k is the error of the AR model;

[0012] B2. Determine the value of p according to the Akaike information criterion;

[0013] First, the Akaike information criterion function AIC(·) is defined for the AR model as follows:

[0014] AIC(p)=nln(ξ 2 )+2p (2)

[0015]

[0016] In the formula, ξ 2 represents the error variance of the AR model, is the estimated value of the undetermined coefficient vector of the AR model, ln(·) represents the natural logarithm; assuming that the maximum order of the AR model is p max , and 0<p≤p max <n, calculate AIC(1),…,AIC(p max ) values, if the minimum of these values ​​is AIC(p * ), then the order of the AR model is determined to be p = p * ;

[0017] B3. Yes The following online iterative learning strategy is proposed:

[0018]

[0019] Where t is the current sampling time, Q is the p×p design constant matrix, X(t-1)=[x t-2 ,x t-3 ,…,x t-1-p ] T , I is the p×p identity matrix, G(t-1) is the p-dimensional gain vector, C(t-1) is the p×p covariance matrix, and R is the design constant;

[0020] C. Based on the AR model, the regression equation of the ship motion sample data is established:

[0021]

[0022] In the formula, For sample data x k The regression value of .

[0023] The np sample data before the current time t are regressed by formula (5), and the sample data are compared with their regression values ​​to obtain np AR model regression errors:

[0024]

[0025] In the formula, e reg,k is the regression error of the AR model.

[0026] D. Based on the AR model, the ship motion prediction model is established as follows:

[0027]

[0028] Where H is the number of forecast steps, is the forecast value for the next HT seconds;

[0029] E. Establish the OLSSVR model as follows:

[0030]

[0031] In the formula, e t =[e reg,p+m+1 ,e reg,p+m+2 ,…,e reg,n ], m is the embedding dimension, 2≤m<np, ω j ,j=p+1,···,nm, is the model coefficient of OLSSVR, M=nmp, ω=[ω p+1 ,ω p+2 ,···,ω j ,···,ω n-m ] is the coefficient vector of the OLSSVR model, b is the bias term, and b = [b,…,b] is the M-dimensional bias term vector.

[0032]

[0033] in, is the radial basis kernel function, E j =[e reg,j ,e reg,j+1 ,…,e reg,j+m-1 ],

[0034]

[0035] The embedding dimension m is calculated as follows:

[0036] First, let m = 2 and calculate the correlation dimension D(m) of the regression error:

[0037]

[0038] In the formula, d(E u ,E v ) is the vector E u With E v The Euclidean distance between them, r is the Euclidean distance in m-dimensional Euclidean space. u is the radius of the spherical region Ω centered, and h(·) is the unit step function;

[0039] Then, let m = m + 1, and continue to calculate the correlation dimension D(m) of the regression error according to formula (9) until the difference between the two correlation dimensions is less than 10 -3 So far, m is the embedding dimension.

[0040] According to the principle of least squares algorithm, the coefficient vector of OLSSVR model is designed and the bias term The iteration formula is as follows:

[0041]

[0042]

[0043] Where σ1=[1,1,…,1] T is an M-dimensional column vector, U(t)=U(t-1)+r1(t)r1(t) T z1(t), where:

[0044]

[0045] F(t-1)=[K(E t-m ,E t-m-1 )+1 / C,…,K(E t-M-m+2 ,E t-m-1 )] T , f(t-1)=K(E t-m ,Et-m ) / C, C is the penalty factor, r1(t)=[V(t) T W(t-1) -1 ,-1] T , V(t)=[K(E t-m+1 ,E t-M-m+2 ),…,K(E t-m+1 ,E t-m )] T ,v(t)=K(E t-m+1 ,E t-m+1 )+1 / C;

[0046] Using the np AR model regression errors obtained in step C, we can calculate and According to the OLSSVR model, the AR model regression error e is established reg,k The forecast model:

[0047]

[0048] Where H is the number of forecast steps, is the forecast value for the next HT seconds.

[0049] F. Add the ship motion prediction value obtained by the ship motion prediction model (7) and the AR model regression error prediction value obtained by the OLSSVR model to obtain:

[0050]

[0051] In the formula, is the corrected ship motion forecast value, which is used as the final ship motion forecast value.

[0052] G. Collect new ship motion data, set t=t+1, go to step B3, and continue the forecast for the next HT second.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] 1. The present invention adopts an iterative learning strategy to update the parameters of the ship motion prediction model and the regression error prediction model in real time. Each time the parameters are updated, only the previous results need to be used to calculate them in an iterative manner, and there is no need to re-use a large amount of sample data for training, which improves the calculation speed of the prediction model and realizes the online prediction of ship motion.

[0055] 2. The present invention adopts the concept of dual-model forecasting, avoiding the drawback that a single forecasting model is difficult to fully capture the characteristic information of ship motion data with nonlinear and non-stationary characteristics, so that the characteristic information is effectively utilized and the accuracy of ship motion forecasting is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 This is a flow chart of a ship motion prediction method based on the AR-OLSSVR model proposed by the present invention.

[0057] Figure 2 The curve of the change of the average absolute percentage error of the predicted value versus the prediction time when the method proposed by the present invention and other methods are used to predict the ship's rolling motion when the sampling period T is 200ms. DETAILED DESCRIPTION

[0058] The present invention will be further described below in conjunction with the accompanying drawings. Figure 1 In the process shown in FIG. 1 , taking the ship rolling motion as an example, when the sampling period T = 200 ms and the number of sample data n = 500, the AR model, the OLSSVR model, the EMD-AR model and the AR-OLSSVR model proposed in the present invention are used to predict the ship rolling motion, and the prediction time is 1 s, 3 s and 5 s respectively. The change curve of the average absolute percentage error of the obtained prediction value with respect to the prediction time is shown in FIG. Figure 2 As shown in the figure, it can be seen from the curve that the method proposed by the present invention has the highest prediction accuracy.

[0059] The present invention is not limited to this embodiment, and any equivalent concepts or changes within the technical scope disclosed by the present invention are included in the protection scope of the present invention.

Claims

1. A ship motion prediction method based on AR-OLSSVR model, characterized by: The following steps are involved: A. Collect ship motion data with T as the sampling period, and obtain n sample data x k , k = 1, 2, ..., n; the ship motion includes rolling motion, pitching motion and heave motion; B. Build and train the AR model; B1. Establish the AR model of the ship motion sample data as follows: In the formula, p is the undetermined order of the AR model, a i is the undetermined coefficient of the AR model, let a=[a1,a2,…,a p ] T is the undetermined coefficient vector of the AR model, e k is the error of the AR model; B2. Determine the value of p according to the Akaike information criterion; First, the Akaike information criterion function AIC(·) is defined for the AR model as follows: AIC(p)=nln(ξ 2 )+2p (2) In the formula, ξ 2 represents the error variance of the AR model, is the estimated value of the undetermined coefficient vector of the AR model, ln(·) represents the natural logarithm; assuming that the maximum order of the AR model is p max , and 0<p≤p max <n, calculate AIC(1),…,AIC(p max ) values, if the minimum of these values ​​is AIC(p * ), then the order of the AR model is determined to be p = p * ; B3. Yes The following online iterative learning strategy is proposed: Where t is the current sampling time, Q is the p×p design constant matrix, X(t-1)=[x t-2 ,x t-3 ,…,x t-1-p ] T , I is the p×p identity matrix, G(t-1) is the p-dimensional gain vector, C(t-1) is the p×p covariance matrix, and R is the design constant; C. Based on the AR model, the regression equation of the ship motion sample data is established: In the formula, For sample data x k The regression value of The np sample data before the current time t are regressed by formula (5), and the sample data are compared with their regression values ​​to obtain np AR model regression errors: In the formula, e reg,k is the regression error of the AR model; D. Based on the AR model, the ship motion prediction model is established as follows: Where H is the number of forecast steps, is the forecast value for the next HT seconds; E. Establish the OLSSVR model as follows: In the formula, e t =[e reg,p+m+1 ,e reg,p+m+2 ,…,e reg,n ], m is the embedding dimension, 2≤m<np, ω j ,j=p+1,…,nm, is the model coefficient of OLSSVR, M=nmp, ω=[ω p+1 ,ω p+2 ,…,ω j ,…,ω n-m ] is the coefficient vector of the OLSSVR model, b is the bias term, and b = [b,…,b] is the M-dimensional bias term vector. in, is the radial basis kernel function, E j =[e reg,j ,e reg,j+1 ,…,e reg,j+m-1 ], The embedding dimension m is calculated as follows: First, let m = 2 and calculate the correlation dimension D(m) of the regression error: In the formula, d(E u ,E v ) is the vector E u With E v The Euclidean distance between them, r is the Euclidean distance in m-dimensional Euclidean space. u is the radius of the spherical region Ω centered, and h(·) is the unit step function; Then, let m = m + 1, and continue to calculate the correlation dimension D(m) of the regression error according to formula (9) until the difference between the two correlation dimensions is less than 10 -3 So far, m is the embedding dimension; According to the principle of least squares algorithm, the coefficient vector of OLSSVR model is designed and the bias term The iteration formula is as follows: Where σ1=[1,1,…,1] T is an M-dimensional column vector, U(t)=U(t-1)+r1(t)r1(t) T z1(t), where: F(t-1)=[K(E t-m ,E t-m-1 )+1 / C,…,K(E t-M-m+2 ,E t-m-1 )] T , f(t-1)=K(E t-m ,E t-m ) / C, C is the penalty factor, r1(t)=[V(t) T W(t-1) -1 ,-1] T , V(t)=[K(E t-m+1 ,E t-M-m+2 ),…,K(E t-m+1 ,E t-m )] T ,v(t)=K(E t-m+1 ,E t-m+1 )+1 / C; Using the np AR model regression errors obtained in step C, we can calculate and According to the OLSSVR model, the AR model regression error e is established reg,k The forecast model: Where H is the number of forecast steps, is the forecast value for the next HT seconds; F. Add the ship motion prediction value obtained by the ship motion prediction model (7) and the AR model regression error prediction value obtained by the OLSSVR model to obtain: In the formula, is the corrected ship motion forecast value, which is used as the final ship motion forecast value; G. Collect new ship motion data, set t=t+1, go to step B3, and continue the forecast for the next HT second.

Citation Information

Patent Citations

  • Intelligent ship moving posture decomposition field forecasting method

    CN103926932A

  • Ship motion forecasting method based on intrinsic plasticity echo state network

    CN111709140A

  • Data-driven marine ship motion attitude real-time forecasting method

    CN113156815A

  • Method for estimating rolling motion response function of unmanned ship by recursive least square method

    CN111046485A

  • Ship motion forecasting method based on improved EMD-AR model

    CN113378387A