Improved multi-step prediction method for ship roll motion based on NARX

Abstract

The application discloses an improved NARX ship roll motion multi-step prediction method, which comprises the following steps: step 1, obtaining a plurality of roll angle time series data as training samples; step 2, determining a fixed step length of a multi-step prediction model; step 3, constructing an improved NARX neural network multi-step prediction model; step 4, training the improved NARX neural network multi-step prediction model through the training samples and the fixed step length, and obtaining a trained improved NARX neural network multi-step prediction model; and step 5, predicting the actually collected roll angle time through the trained improved NARX neural network multi-step prediction model, and taking any prediction value in the obtained fixed step number as a prediction result. The application can obtain better prediction effect in the case that the sensor sampling period is long and the system time lag is compensated, and can ensure that the ship equipment can be stably and safely operated.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ship motion prediction, in particular to an improved NARX ship roll motion multi-step prediction method. BACKGROUND

[0002] With the improvement of science and technology, human beings increase the pace of deep-sea resource development to meet the growing demand for resources. In the complex and changeable marine environment, the roll motion of the ship during navigation will bring many adverse effects to the personnel, cargo and ship itself, will cause the crew to be unable to work normally due to the appearance of restlessness and fear psychology, will cause the shipborne instruments, equipment and other failures, and will cause the cargo on the ship to appear left and right shaking, and cannot be normally fixed.

[0003] The traditional NARX neural network performs single-step prediction on the ship roll motion, although it has high prediction accuracy, but for the slow sampling of the ship sensor and the problem of large time lag, it cannot provide the measured value in time, resulting in poor prediction effect, and the general multi-step prediction method is to construct a model according to the value at a certain time and before to predict the value at future multiple times, and once the prediction step is too long, the prediction value will gradually replace the measured value, and the effect will be worse and worse. Therefore, it is very important to provide a roll motion multi-step prediction method with high prediction accuracy. SUMMARY

[0004] The present application provides an improved NARX ship roll motion multi-step prediction method to solve the technical problem of low prediction accuracy in the prior art, which is suitable for occasions with slow sensor sampling and large time lag.

[0005] The present application provides an improved NARX ship roll motion multi-step prediction method, comprising the following steps:

[0006] Step 1: Obtain a plurality of roll angle time series data as training samples;

[0007] Step 2: Determine the fixed step length of the multi-step prediction model;

[0008] Step 3: Construct an improved NARX neural network multi-step prediction model, specifically:

[0009]

[0010] In the formula: is the prediction value; y is the training sample data value; is the prediction value obtained after correction; m is the prediction number; n is the sample data number; h is the number of historical inputs required for prediction; k is the prediction fixed step length, k << h; l is the prediction time series; J(x) is the prediction function;

[0011] Step 4: training the improved NARX neural network multi-step prediction model through the training sample and the fixed step length to obtain the trained improved NARX neural network multi-step prediction model;

[0012] Step 5: predicting the actually collected roll angle time through the trained improved NARX neural network multi-step prediction model, and taking any prediction value in the fixed step length as the prediction result.

[0013] Further, the fixed step length is determined by estimating the lag length of the sensor and the controller system of the ship in the actual application and the model running speed in step 2.

[0014] Further, the fixed step length ranges from 4 to 6.

[0015] Further, in step 3, the correction function is:

[0016]

[0017] Further, in step 5, the last step prediction value in the fixed step length is selected as the prediction result.

[0018] The beneficial effects of the present application are as follows:

[0019] The present application establishes an improved NARX neural network multi-step prediction model, so that the model can obtain better prediction effect in the case of long sensor sampling period and compensation system time lag. Thus, the control system can timely compensate the disturbance to ensure the stable and safe operation of the ship equipment.

[0020] In the model, the fixed step length is used for prediction, which can provide sufficient time for compensation. When the preliminary prediction value is obtained, the corrected prediction value is obtained by averaging the prediction values at the same time, instead of the single prediction value in the traditional method, so as to avoid error accumulation and improve the prediction accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0021] The features and advantages of the present application will be more clearly understood through the following detailed description with reference to the accompanying drawings, which are schematic and should not be understood as any limitation to the present application, and in which:

[0022] Figure 1 It is the open loop training and closed loop prediction structure of the traditional NARX neural network;

[0023] Figure 2 It is the flow chart of the specific embodiment in the present application;

[0024] Figure 3 It is the ITTC double parameter spectrum under different sea conditions;

[0025] Figure 4 Flow chart for solving ship motion time series;

[0026] Figure 5 Rolling motion time series chart for existing method under secondary sea state;

[0027] Figure 6 Rolling motion time series chart for existing method under tertiary sea state;

[0028] Figure 7 Rolling motion time series chart for existing method under quaternary sea state;

[0029] Figure 8 Rolling direction time series chart for existing method under quinary sea state;

[0030] Figure 9 Prediction effect chart for specific embodiment of the present application under secondary sea state;

[0031] Figure 10 Error chart for specific embodiment of the present application under secondary sea state;

[0032] Figure 11 Effect chart for specific embodiment of the present application under tertiary sea state;

[0033] Figure 12 Error chart for specific embodiment of the present application under tertiary sea state;

[0034] Figure 13 Effect chart for specific embodiment of the present application under quaternary sea state;

[0035] Figure 14 Error chart for specific embodiment of the present application under quaternary sea state;

[0036] Figure 15 Effect chart for specific embodiment of the present application under quinary sea state;

[0037] Figure 16 Error chart for specific embodiment of the present application under quinary sea state. Specific embodiments

[0038] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described below in connection with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the present application.

[0039] The embodiment of the present application provides an improved NARX ship rolling motion multi-step prediction method, which comprises the following steps:

[0040] Step 1: Obtain several roll angle time series data as training samples;

[0041] Suppose n roll angle time series data is collected as {y1, y2, …, y h , …, y n-1 , y n}.

[0042] Step 2: Determine the fixed step length of the multi-step prediction model;

[0043] According to the lag length of the sensor and controller system of the ship in actual application, the fixed step length is determined, and the model running speed needs to be considered. The step length of the fixed step length should not be too large, so the fixed step length is k steps, and the range is 4≤k≤6;

[0044] Step 3: Construct an improved NARX neural network multi-step prediction model, specifically:

[0045]

[0046] In the formula: The superscript 1 in y^1 indicates the first predicted value obtained by the prediction model; similarly, the superscripts 2, 3, 4, …, m indicate the second, third, fourth, …, mth predicted values at the same time; y^1, y^2, …, y^m represent the predicted values obtained after correction by the correction function; y1, y2, …, y n-h , y n-h+1 , …, y n represent the sampled training sample data; h represents the number of historical inputs required for prediction; k represents the fixed step length of prediction, generally k<<h; l represents the prediction time series; J(x) is the prediction function;

[0047] The correction function is:

[0048]

[0049] Step 4: Train the improved NARX neural network multi-step prediction model through the training samples and the fixed step length to obtain the trained improved NARX neural network multi-step prediction model; the specific training method is represented as:

[0050]

[0051] In the formula: y1, y2, …, y h , …, y n are the measured values in the training samples at different time intervals; h represents the number of historical inputs required for training; k represents the training fixed step length, which is also the prediction step length;

[0052] Step 5: The actual collected roll angle time is predicted by the improved NARX neural network multi-step prediction model after training, and the last step prediction value in the fixed step number is obtained as the prediction result.

[0053] The embodiment of the present application is verified by selecting a fixed step length k=4 when the sensor and the compensation system have a short lag time.

[0054] The improved NARX multi-step prediction method provided by the embodiment of the present application is mainly improved by the traditional NARX multi-step prediction, and the traditional NARX neural network open-loop training and closed-loop prediction structure is as shown in Figure 1 The flow chart of the specific embodiment is as shown in Figure 2 The specific embodiment includes the following processes:

[0055] The time sequence angle data of the ship roll motion is obtained, and due to the condition limitation, the ship roll motion is simulated under different sea conditions by using the Marine Systems Simulator (MSS) software platform developed by the Norwegian University of Science and Technology. The ship dynamics model is established as follows:

[0056]

[0057] In the formula: η 6 = (x y z φ θ ψ) T is the position of the ship in the north-east plane coordinate system

[0058] and the attitude angle is the linear velocity and angular velocity of the ship in the body coordinate system; τ 6 is the thrust and torque on the six degrees of freedom of the ship in the body coordinate system; ω env is the external environmental disturbance; M 6 is the inertia matrix; D 6 is the damping matrix; g (η) is the force and torque caused by gravity and buoyancy; g 0 represents the restoring force provided by the ship ballast water; J θ (η 6) is the coordinate transformation matrix; the related matrix is as follows:

[0059]

[0060]

[0061]

[0062] The sea wave model uses the ITTC double-parameter wave energy spectrum to establish the sea wave spectrum, and the specific mathematical expression is as follows:

[0063]

[0064] In the formula: H sHs represents the significant wave height; ω represents the wave angular frequency; and T1 represents the characteristic period of the wave.

[0065] The ITTC two-parameter wave energy spectrum under different sea states is obtained according to the formula as shown in Figure 3 , and the random sea waves are simulated under different sea states by frequency equalization method. The sea state grade definition is shown in Table 1 Sea State Grade Definition:

[0066]

[0067] Table 1

[0068] The response of the ship to the wave is solved by the force response amplitude operator (Force RAO) and the ship's own modeling data, and the parameters of the ship are shown in Table 2 Ship Parameters:

[0069]

[0070] Table 2

[0071] The training data of the roll angle is obtained according to the ship motion time sequence flow chart as shown in Figure 4 . The time sequence roll angle values obtained under different sea states are shown in Figure 5 , 6 , 7, 8. The sampling time under each sea state is 300s, the sampling time interval is 0.1s, and a total of 3000 groups of data, 1000 data points in the first 100s are used as the training set. 2000 data points in the last 200s are used as the test data set. The improved NARX multi-step prediction model is established, and the specific steps are as follows:

[0072] 1) The collected data is trained, and the specific fixed step length is 4 steps, and the mathematical model is trained as shown in

[0073] y h+4 , y h+3 , y h+2 , y h+1 = J (y h , y h-1 …y2, y1)

[0074] y h+5 , y h+4 , y h+3 , y h+2 = J (y h+1 , y h …y3, y2)

[0075] y h+6 , y h+5 , y h+4 , y h+3 = J (y h+2 , yh+1 , y4, y3)

[0076] y n , y n-1 , y n-2 , y n-3 = J(y n-4 , y n-5 ,... y n-h+1 , y n-h )

[0077] 2) When the training prediction error meets the initial requirements, the traditional method performs a fixed step 4-step prediction, which can be represented by the following mathematical model:

[0078]

[0079] The above mainly uses the first prediction value at the last time to replace it. If the prediction value and the measured value error is large, the feedback at this moment is used as input, which greatly reduces the multi-step prediction accuracy.

[0080] 3) The improved NARX neural network fixed step 4-step prediction model can be described by the mathematical model:

[0081]

[0082] The prediction value obtained by using the correction function is used to replace the first prediction value

[0083] 4) The correction function in the above is represented by the following mathematical model:

[0084]

[0085] 5) Finally, the last step prediction value in the fixed step is selected as the prediction result.

[0086] The parameter settings of the prediction model are shown in Table 3:

[0087]

[0088] Table 3

[0089] The improved NARX four-step prediction model is used to predict different sea conditions, and the prediction effects and prediction errors of the second, third, fourth and fifth levels are as follows Figures 9 to 16 .

[0090] In order to evaluate the invented method, based on the test set, the average error, the maximum error and the mean square error are calculated as the performance index for judgment. The prediction performance is shown in Table 4: ​

[0091]

[0092] Table 4

[0093] Table 4 is the comparison of the traditional NARX multi-step model and the improved NARX multi-step model in terms of average error, maximum error and mean square error. It can be seen from Table 4 that:

[0094] The traditional NARX multi-step prediction model depends on the accuracy of the previous predicted value, and the measured value is gradually replaced by the predicted value, resulting in error accumulation, so the effect is poor in the later prediction. The improved NARX multi-step prediction model has better prediction effect compared with the traditional NARX multi-step prediction model.

[0095] With the increase of sea state level, the average error of the improved NARX multi-step prediction model is about 6%, the maximum error is basically maintained at about 20%, and the mean square error is basically less than 0.5%. It is verified that it has high prediction accuracy in multiple sea state conditions.

[0096] In the iterative prediction of the traditional NARX model, the measured data in the early stage plays a leading role, so the prediction effect is good. With the continuous movement of time series, the predicted value gradually plays a leading role, and the prediction effect in the later stage is poor. The improved NARX reduces the error of each step of the predicted value by constructing a correction function, and the predicted value is as close as possible to the measured value, so that the later prediction effect is good.

[0097] Although the embodiments of the present application are described in combination with the drawings, various modifications and variations can be made by those skilled in the art without departing from the spirit and scope of the present application, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. An improved multi-step prediction method for NARX ship roll motion, characterized in that, Includes the following steps: Step 1: Obtain several roll angle time series data as training samples; Step 2: Determine the fixed step size for the multi-step prediction model; Step 3: Construct an improved NARX neural network multi-step prediction model, specifically as follows: In the formula: y is the predicted value; y is the value of the training sample data. The corrected predicted value is denoted as ; m is the number of predictions; n is the number of sample data; h is the number of historical inputs required for prediction; k is the fixed prediction step size, k << h; l is the prediction time series; J(x) is the prediction function. Step 4: Train the improved NARX neural network multi-step prediction model using training samples and a fixed step size to obtain the trained improved NARX neural network multi-step prediction model. Step 5: Predict the actual roll angle time using the trained and improved NARX neural network multi-step prediction model, and take any prediction value from the fixed number of steps as the prediction result.

2. The improved NARX ship roll motion multi-step prediction method as described in claim 1, characterized in that, In step 2, the fixed step size is determined by estimating the lag time of the sensor and controller systems of the ship in actual application and the running speed of the model.

3. The improved NARX ship roll motion multi-step prediction method as described in claim 1 or 2, characterized in that, The fixed step size is between 4 and 6.

4. The improved NARX ship roll motion multi-step prediction method as described in claim 1, characterized in that, In step 3, the correction function is: 。 5. The improved NARX ship roll motion multi-step prediction method as described in claim 1, characterized in that, In step 5, the predicted value of the last step in the fixed number of steps is selected as the prediction result.

Images