CEEMDAN-Transform-based offshore platform motion response forecasting method and device

Abstract

The invention discloses an offshore platform motion response forecasting method and device based on CEEMDAN-Transformer, and the method comprises the steps: obtaining motion response monitoring data of an offshore floating platform, and carrying out the preprocessing of the data, and removing abnormal values and noise; and decomposing the preprocessed motion response data by using the CEEMDAN. Dividing each sub-sequence obtained through decomposition into a training set and a test set, carrying out standardization processing, respectively building a Transform neural network, setting related structure parameters of the Transform neural network, and carrying out prediction work of the sub-sequences to obtain a trained neural network model; the test set is used to test the Transform neural network; and reconstructing the forecast results of all the subsequences to obtain the forecast result of the original motion response, and then carrying out comprehensive quantitative evaluation on the forecast result of the motion response of the offshore floating platform to realize the forecast work of the motion response of the offshore floating platform. According to the method, accurate forecasting work of the motion response of the offshore floating platform is realized, and auxiliary reference can be provided for field decision making.

Description

Technical Field

[0001] The invention belongs to the field of marine engineering technology, and in particular relates to a method and a device for predicting motion response of an offshore platform based on CEEMDAN-Transformer. Background Art

[0002] At present, offshore floating platforms face many challenges in operation, such as harsh marine environment, including strong winds and waves, complex currents and changeable marine meteorological conditions, etc. By accurately predicting the movement of the platform, corresponding safety measures can be taken in advance, such as adjusting the platform's operating parameters, strengthening the platform structure, evacuating some non-essential personnel, etc., thereby effectively reducing the risk of accidents and ensuring the safety of the people on the platform and the normal operation of the equipment.

[0003] Offshore oil and gas development operations have high requirements for platform stability. Accurate prediction of platform motion can reasonably arrange operation time and operation process, thereby improving overall operation efficiency and reducing operating costs. Accurate platform motion prediction data can provide an important basis for the design of offshore platforms. At present, the relevant motion prediction algorithms used in offshore floating platforms have the problem of poor generalization ability, making it difficult to accurately capture the nonlinear characteristics of platform motion under complex sea conditions, and unable to provide auxiliary reference for offshore operations. Summary of the invention

[0004] To this end, the present invention provides a method and device for predicting the motion response of an offshore platform based on CEEMDAN-Transformer, which solves the problems of low prediction accuracy and poor generalization ability in the traditional prediction of the motion response of an offshore floating platform.

[0005] In order to achieve the above object, the present invention provides the following technical solution: a method for predicting motion response of an offshore platform based on CEEMDAN-Transformer, comprising:

[0006] Acquiring motion response monitoring data of an offshore floating platform, and preprocessing the motion response monitoring data;

[0007] Using CEEMDAN algorithm to perform empirical mode decomposition on the preprocessed motion response monitoring data to obtain a plurality of subsequences, wherein the subsequences include a plurality of intrinsic mode functions (IMFs) and a residual sequence;

[0008] Dividing the obtained subsequences into a training set and a test set, and performing standardization processing on the data in the training set and the test set;

[0009] Constructing a Transformer neural network model, and configuring neural network training parameters and neural network structure for the Transformer neural network model; training the Transformer neural network model using the standardized training set to obtain the trained Transformer neural network model;

[0010] The trained Transformer neural network model is tested by using the standardized test set, and the prediction results of all the test sets are summarized and reconstructed to obtain the prediction results of the motion response monitoring data;

[0011] A comprehensive quantitative evaluation is conducted on the prediction results of the motion response of the offshore floating platform to achieve the prediction of the motion response of the offshore floating platform.

[0012] As a preferred solution for the offshore platform motion response prediction method based on CEEMDAN-Transformer, the motion response monitoring data of the offshore floating platform is obtained through the differential global positioning system and the motion reference unit system;

[0013] The preprocessing method for the motion response monitoring data includes filtering to remove abnormal values ​​and noise.

[0014] As a preferred solution of the offshore platform motion response prediction method based on CEEMDAN-Transformer, the steps of using the CEEMDAN algorithm to perform empirical mode decomposition on the pre-processed motion response monitoring data to obtain several subsequences include:

[0015] a1) Define operator E m (·) Carry out m-order empirical mode decomposition on the signal; add multiple Gaussian white noises w that obey the standard normal distribution N(0,1) to the original motion response signal Y(t) i (t), and obtain the motion response signal Y i (t):

[0016] Y i (t) = Y (t) + ε i w i (t),i=1,2,…I

[0017] In the formula, ε i is the noise coefficient, which is used to determine the signal-to-noise ratio of each step; I is the number of Gaussian white noise added;

[0018] a2) The obtained multiple motion response signals Y i (t) Carry out empirical mode decomposition to obtain multiple subsequences Define the first IMF as:

[0019]

[0020] The first residual signal R1(t) is:

[0021]

[0022] In the formula, is the first intrinsic mode function obtained by the CEEMDAN algorithm;

[0023] a3) For the decomposition result R1(t)+ε1E1(w i (t)) signal is subjected to empirical mode decomposition, and the second IMF is obtained as:

[0024]

[0025] In the formula, is the second intrinsic mode function obtained by CEEMDAN algorithm; is the intrinsic mode function of the second subsequence obtained by decomposing the residual R1(t) after adding multiple Gaussian white noises;

[0026] a4) Repeat step a3) to get the kth IMF:

[0027]

[0028] In the formula, is the kth intrinsic mode function obtained by CEEMDAN algorithm; is the residual R k-1 (t) After adding multiple Gaussian white noises, the second subsequence intrinsic mode function obtained by each decomposition; R k-1 (t) is the k-1th residual signal left after decomposition; ε k-1 is the noise coefficient of the k-1th step; E k-1 (w i (t)) is to carry out k-1 order empirical mode decomposition of the added Gaussian white noise;

[0029] a5) Determine the k-th order residual signal R k (t) whether it contains at least two extreme values. If it does, continue to decompose and repeat step a3) to obtain the k+1th IMF; if it does not contain at least two extreme values, the decomposition process ends; finally, the time series signal Y(t) is decomposed into all subsequence IMFs that meet the conditions and the residual R representing the signal trend K (t) and.

[0030] As a preferred solution of the offshore platform motion response prediction method based on CEEMDAN-Transformer, the formula for standardizing the data in the training set and the test set is:

[0031]

[0032] In the formula, y i is the offshore floating platform motion monitoring value, y norm is the result after standardization, y max is the maximum value of the input data, y min is the minimum value of the input data.

[0033] As a preferred solution of the offshore platform motion response prediction method based on CEEMDAN-Transformer, the neural network training parameter configuration objects of the Transformer neural network model include historical time step, prediction time step, data batch size and number of training rounds;

[0034] The neural network structure configuration objects of the Transformer neural network model include embedding layer, multi-head attention layer, normalization layer, feedforward layer, residual connection layer, and linearization layer; Softmax is selected to calculate the probability distribution of data, MSE is selected as the training loss function, and Adam optimizer is selected to update the neural network parameters.

[0035] As a preferred solution of the offshore platform motion response prediction method based on CEEMDAN-Transformer, the steps of training the Transformer neural network model by using the standardized training set to obtain the trained Transformer neural network model include:

[0036] b1) Input the data in the training set into the Transformer neural network model to generate the corresponding query vector Q, key vector K and value vector V:

[0037]

[0038] Where W Q , W K and W V are the weight matrices of query, key and value respectively, and X is the input data;

[0039] b2) Calculate the attention score using the dot product of the query vector Q and the key vector K:

[0040]

[0041] Where Attention(Q,K,V) is the attention score of the input data; d k is the dimension of the key vector K;

[0042] b3) Calculate the multi-head attention matrix:

[0043]

[0044] MultiHead(Q,K,V)=(head1,head2,…,head h )W O

[0045] In the formula, head i is the i-th attention weight; Corresponding to the weight matrices of the i-th Q, K, and V, W O is the final attention matrix after linear transformation; MultiHead(Q,K,V) is the multi-head attention score of the input data;

[0046] b4) Repeat steps b1), b2), and b3) to generate multiple query vectors Q, key vectors K, and value vectors V, and calculate the attention scores and weighted sums respectively; concatenate the results of all attention weights and obtain the final output through linear transformation to obtain the trained Transformer neural network.

[0047] As the optimal solution for the prediction method of offshore platform motion response based on CEEMDAN-Transformer, the prediction results of offshore floating platform motion response are comprehensively and quantitatively evaluated. In the process of predicting the motion response of offshore floating platform, the selected evaluation indicators are the root mean square error, mean absolute error and determination coefficient of motion.

[0048] The formula for root mean square error RMSE is:

[0049]

[0050] The formula for mean absolute error MAE is:

[0051]

[0052] Determination coefficient R 2 The formula is:

[0053]

[0054] In the formula, y i is the platform motion monitoring value, is the platform motion prediction value, is the average value of the platform motion monitoring value; n is the number of motion data points.

[0055] The present invention also provides an offshore platform motion response prediction device based on CEEMDAN-Transformer, comprising:

[0056] A motion response monitoring data processing module is used to obtain the motion response monitoring data of the offshore floating platform and pre-process the motion response monitoring data;

[0057] A monitoring data empirical mode decomposition module, used for performing empirical mode decomposition on the pre-processed motion response monitoring data using a CEEMDAN algorithm to obtain a plurality of subsequences, wherein the subsequences include a plurality of intrinsic mode functions IMFs and a residual sequence;

[0058] A subsequence division processing module, used to divide the obtained subsequence into a training set and a test set, and perform standardization processing on the data in the training set and the test set;

[0059] A Transformer neural network model construction module, used to construct a Transformer neural network model, and perform neural network training parameter and neural network structure configuration on the Transformer neural network model;

[0060] A Transformer neural network model training module, used to train the Transformer neural network model using the standardized training set to obtain the trained Transformer neural network model;

[0061] A Transformer neural network model testing module, used to test the trained Transformer neural network model using the standardized test set;

[0062] A forecast result aggregation and reconstruction module, used for aggregating and reconstructing the forecast results of all the test sets to obtain the forecast results of the motion response monitoring data;

[0063] The forecast result quantitative evaluation module is used to conduct a comprehensive quantitative evaluation of the forecast results of the motion response of the offshore floating platform, so as to realize the forecast of the motion response of the offshore floating platform.

[0064] As a preferred solution of the offshore platform motion response prediction device based on CEEMDAN-Transformer, in the motion response monitoring data processing module:

[0065] Acquire the motion response monitoring data of offshore floating platforms through differential global positioning system and motion reference unit system;

[0066] The preprocessing method for the motion response monitoring data includes filtering to remove abnormal values ​​and noise.

[0067] As a preferred solution of the offshore platform motion response prediction device based on CEEMDAN-Transformer, the monitoring data empirical mode decomposition module includes:

[0068] Operator definition submodule, used to define operator E m (·) Carry out m-order empirical mode decomposition on the signal; add multiple Gaussian white noises w that obey the standard normal distribution N(0,1) to the original motion response signal Y(t) i (t), and obtain the motion response signal Y i (t):

[0069] Y i (t) = Y (t) + ε i w i (t),i=1,2,…I

[0070] In the formula, ε i is the noise coefficient, which is used to determine the signal-to-noise ratio of each step; I is the number of Gaussian white noise added;

[0071] The first IMF definition submodule is used to obtain multiple motion response signals Y i (t) Carry out empirical mode decomposition to obtain multiple subsequences Define the first IMF as:

[0072]

[0073] The first residual signal R1(t) is:

[0074]

[0075] In the formula, is the first intrinsic mode function obtained by the CEEMDAN algorithm;

[0076] The second IMF definition submodule is used to decompose the result R1(t)+ε1E1(w i (t)) signal is subjected to empirical mode decomposition, and the second IMF is obtained as:

[0077]

[0078] In the formula, is the second intrinsic mode function obtained by CEEMDAN algorithm; is the intrinsic mode function of the second subsequence obtained by decomposing the residual R1(t) after adding multiple Gaussian white noises;

[0079] The k-th IMF definition submodule is used to obtain the k-th IMF. The obtained k-th IMF is:

[0080]

[0081] In the formula, is the kth intrinsic mode function obtained by CEEMDAN algorithm; is the residual R k-1 (t) After adding multiple Gaussian white noises, the second subsequence intrinsic mode function obtained by each decomposition; R k-1 (t) is the k-1th residual signal left after decomposition; ε k-1 is the noise coefficient of the k-1th step; E k-1 (w i (t)) is to carry out k-1 order empirical mode decomposition of the added Gaussian white noise;

[0082] The extreme value judgment analysis submodule is used to judge the k-th order residual signal R k (t) whether it contains at least two extreme values. If it does, continue to decompose and obtain the k+1th IMF; if it does not contain at least two extreme values, the decomposition process ends; finally, the time series signal Y(t) is decomposed into all subsequence IMFs that meet the conditions and the residual R representing the signal trend K (t) and.

[0083] As a preferred solution of the offshore platform motion response prediction device based on CEEMDAN-Transformer, in the subsequence division processing module:

[0084] The formula for standardizing the data in the training set and the test set is:

[0085]

[0086] In the formula, y i is the offshore floating platform motion monitoring value, y norm is the result after standardization, y max is the maximum value of the input data, y min is the minimum value of the input data.

[0087] As a preferred solution of the offshore platform motion response prediction device based on CEEMDAN-Transformer, in the Transformer neural network model construction module:

[0088] The neural network training parameter configuration objects for the Transformer neural network model include historical time step, prediction time step, data batch size and number of training rounds;

[0089] The neural network structure configuration objects of the Transformer neural network model include embedding layer, multi-head attention layer, normalization layer, feedforward layer, residual connection layer, and linearization layer; Softmax is selected to calculate the probability distribution of data, MSE is selected as the training loss function, and Adam optimizer is selected to update the neural network parameters.

[0090] As a preferred solution of the offshore platform motion response prediction device based on CEEMDAN-Transformer, the Transformer neural network model training module includes:

[0091] The vector generation submodule is used to input the data in the training set into the Transformer neural network model to generate the corresponding query vector Q, key vector K and value vector V:

[0092]

[0093] Where W Q , W K and W V are the weight matrices of query, key and value respectively, and X is the input data;

[0094] The attention score calculation submodule is used to calculate the attention score using the dot product of the query vector Q and the key vector K:

[0095]

[0096] Where Attention(Q,K,V) is the attention score of the input data; d k is the dimension of the key vector K;

[0097] Multi-head attention matrix calculation submodule, used to calculate the multi-head attention matrix:

[0098]

[0099] MultiHead(Q,K,V)=(head1,head2,…,head h )W O

[0100] In the formula, head i is the i-th attention weight; Corresponding to the weight matrices of the i-th Q, K, and V, W O is the final attention matrix after linear transformation; MultiHead(Q,K,V) is the multi-head attention score of the input data.

[0101] As a preferred solution for the offshore platform motion response prediction device based on CEEMDAN-Transformer, in the forecast result summary reconstruction module, multiple query vectors Q, key vectors K and value vectors V are generated, and the attention scores and weighted sums are calculated respectively; the results of all attention weights are concatenated and linearly transformed to obtain the final output, thereby obtaining the trained Transformer neural network.

[0102] As a preferred solution of the offshore platform motion response prediction device based on CEEMDAN-Transformer, in the prediction result quantitative evaluation module, the selected evaluation indicators are the root mean square error, mean absolute error and determination coefficient of the motion;

[0103] The formula for root mean square error RMSE is:

[0104]

[0105] The formula for mean absolute error MAE is:

[0106]

[0107] Determination coefficient R 2 The formula is:

[0108]

[0109] In the formula, y i is the platform motion monitoring value, is the platform motion prediction value, is the average value of the platform motion monitoring value; n is the number of motion data points.

[0110] The present invention has the following advantages: acquiring the motion response monitoring data of an offshore floating platform, preprocessing the data to remove outliers and noise; applying the CEEMDAN algorithm to decompose the preprocessed motion response data to obtain multiple subsequences including: multiple intrinsic mode functions IMF and a residual sequence. Each decomposed subsequence is divided into a training set and a test set, and standardized, and a Transformer neural network is built respectively, and the relevant structural parameters of the Transformer neural network are set, and the prediction work of the subsequence is carried out to obtain a trained neural network model; the test set of all subsequences is used to test the Transformer neural network; the prediction results of all subsequences are reconstructed to obtain the prediction results of the original motion response, and then a comprehensive quantitative evaluation is carried out on the prediction results of the motion response of the offshore floating platform to realize the prediction work of the motion response of the offshore floating platform. The present invention realizes the accurate prediction work of the motion response of the offshore floating platform, and can provide auxiliary reference for on-site decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0111] In order to more clearly illustrate the implementation methods of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the implementation methods or the description of the prior art. Obviously, the drawings in the following description are only exemplary, and for ordinary technicians in this field, other implementation drawings can be derived from the provided drawings without creative work.

[0112] Figure 1 A schematic flow chart of a method for predicting motion response of an offshore platform based on CEEMDAN-Transformer provided in an embodiment of the present invention;

[0113] Figure 2 A schematic diagram of the technical architecture of a method for predicting motion response of an offshore platform based on CEEMDAN-Transformer provided in an embodiment of the present invention;

[0114] Figure 3 A comparison diagram of the IMF5 monitoring value and the predicted value of the offshore floating platform motion response subsequence provided in an embodiment of the present invention;

[0115] Figure 4 A comparison diagram of the monitoring value and the predicted value of the offshore floating platform motion response subsequence IMF6 provided in an embodiment of the present invention;

[0116] Figure 5 A comparison diagram of the reconstructed offshore floating platform motion response monitoring value and the predicted value provided in an embodiment of the present invention;

[0117] Figure 6 A comparison chart of the reconstructed offshore floating platform motion response monitoring value provided in an embodiment of the present invention and the predicted value using the traditional LSTM method;

[0118] Figure 7 It is a schematic diagram of the architecture of an offshore platform motion response prediction device based on CEEMDAN-Transformer provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0119] The following is a description of the implementation of the present invention by specific embodiments. People familiar with the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0120] Example 1

[0121] See also Figure 1 and Figure 2 The embodiment of the present invention provides a method for predicting motion response of an offshore platform based on CEEMDAN-Transformer, comprising the following steps:

[0122] S1. Acquire motion response monitoring data of an offshore floating platform, and pre-process the motion response monitoring data;

[0123] S2, using CEEMDAN algorithm to perform empirical mode decomposition on the pre-processed motion response monitoring data to obtain a plurality of subsequences, wherein the subsequences include a plurality of intrinsic mode functions IMFs and a residual sequence;

[0124] S3, dividing the obtained subsequence into a training set and a test set, and performing standardization processing on the data in the training set and the test set;

[0125] S4, constructing a Transformer neural network model, configuring neural network training parameters and neural network structure for the Transformer neural network model; training the Transformer neural network model using the standardized training set to obtain the trained Transformer neural network model;

[0126] S5, testing the trained Transformer neural network model by using the standardized test set, summarizing and reconstructing the prediction results of all the test sets to obtain the prediction results of the motion response monitoring data;

[0127] S6. Conduct a comprehensive quantitative evaluation on the prediction results of the motion response of the offshore floating platform to achieve the prediction of the motion response of the offshore floating platform.

[0128] In this embodiment, in step S1, the motion response monitoring data of the offshore floating platform is obtained through the differential global positioning system and the motion reference unit system; the implementation process uses the sway motion response data of an offshore floating platform monitored for three hours from 5:35 to 8:35 on September 1, 2023, with a data sampling frequency of 1 Hz, and a total of 10,800 data points.

[0129] The preprocessing method of the motion response monitoring data includes filtering to remove abnormal values ​​and noise. There are related technologies for filtering and removing abnormal values ​​and noise, which will not be described in detail here.

[0130] In this embodiment, in step S2, the steps of using the CEEMDAN algorithm to perform empirical mode decomposition on the pre-processed motion response monitoring data to obtain a plurality of subsequences include:

[0131] a1) Define operator E m (·) Carry out m-order empirical mode decomposition on the signal; add multiple Gaussian white noises w that obey the standard normal distribution N(0,1) to the original motion response signal Y(t) i (t), and obtain the motion response signal Y i (t):

[0132] Y i (t) = Y (t) + ε i w i (t),i=1,2,…I

[0133] In the formula, ε i is the noise coefficient, which is used to determine the signal-to-noise ratio of each step; I is the number of Gaussian white noise added;

[0134] Among them, Gaussian white noise is a random process with zero mean, constant variance and uniform distribution of power spectrum density. The amplitude is an independent random variable at any time, and its probability density function P(a) can be expressed as:

[0135]

[0136] Where: a is a random variable, indicating the amplitude of the noise; σ 2 is the variance of the noise, which is the power of the noise.

[0137] In order to generate Gaussian white noise in practice, a normally distributed random number generator can be used. Specifically, the generation method is based on the Box-Muller transformation method, and the formula is as follows:

[0138]

[0139]

[0140] In the formula, u1 and u2 are uniformly distributed random numbers in the interval [0,1]. The Box-Muller transformation is used to generate standard normally distributed random numbers z1 and z2: z1 and z2 are random variables of standard normally distribution and are two independent samples of Gaussian white noise. Different Gaussian white noises can be obtained by changing the values ​​of u1 and u2.

[0141] a2) The obtained multiple motion response signals Y i (t) Carry out empirical mode decomposition to obtain multiple subsequences Define the first IMF as:

[0142]

[0143] The first residual signal R1(t) is:

[0144]

[0145] In the formula, is the first intrinsic mode function obtained by the CEEMDAN algorithm;

[0146] a3) For the decomposition result R1(t)+ε1E1(w i (t)) signal is subjected to empirical mode decomposition, and the second IMF is obtained as:

[0147]

[0148] In the formula, is the second intrinsic mode function obtained by CEEMDAN algorithm; is the intrinsic mode function of the second subsequence obtained by decomposing the residual R1(t) after adding multiple Gaussian white noises;

[0149] a4) Repeat step a3) to get the kth IMF:

[0150]

[0151] In the formula, is the kth intrinsic mode function obtained by CEEMDAN algorithm; is the residual R k-1 (t) After adding multiple Gaussian white noises, the second subsequence intrinsic mode function obtained by each decomposition; R k-1 (t) is the k-1th residual signal left after decomposition; ε k-1 is the noise coefficient of the k-1th step; E k-1 (w i (t)) is to carry out k-1 order empirical mode decomposition of the added Gaussian white noise;

[0152] a5) Determine the k-th order residual signal R k (t) whether it contains at least two extreme values. If it does, continue to decompose and repeat step a3) to obtain the k+1th IMF; if it does not contain at least two extreme values, the decomposition process ends; finally, the time series signal Y(t) is decomposed into all subsequence IMFs that meet the conditions and the residual R representing the signal trend K (t) Sum:

[0153]

[0154] Finally, the motion response monitoring data is decomposed into 9 IMF subsequences and 1 residual subsequence.

[0155] In this embodiment, in step S3, the formula for standardizing the data in the training set and the test set is:

[0156]

[0157] In the formula, y i is the offshore floating platform motion monitoring value, y norm is the result after standardization, y max is the maximum value of the input data, y min Is the minimum value of the input data. The first 80% of the data is divided into a training set, and the last 20% of the data is divided into a test set for training the Transformer neural network model.

[0158] In this embodiment, in step S4, the neural network training parameter configuration objects for the Transformer neural network model include the historical time step, the prediction time step, the data batch size and the number of training rounds;

[0159] Among them, the neural network training parameters are set as shown in Table 1:

[0160] Table 1 Neural network training parameters set

[0161]

[0162] In step S4, the neural network structure configuration objects of the Transformer neural network model include an embedding layer, a multi-head attention layer, a normalization layer, a feedforward layer, a residual connection layer, and a linearization layer; Softmax is used to calculate the probability distribution of the data, MSE is selected as the training loss function, and Adam optimizer is selected to update the neural network parameters.

[0163] Among them, the neural network structure parameters are set as shown in Table 2:

[0164] Table 2 Neural network structure parameters set

[0165]

[0166] Among them, the selected loss function MSE formula is as follows:

[0167]

[0168] Where: y i is the platform motion monitoring value, is the predicted value of platform motion.

[0169] In this embodiment, in step S5, the Transformer neural network model is trained by the standardized training set, and the step of obtaining the trained Transformer neural network model includes:

[0170] b1) Input the data in the training set into the Transformer neural network model to generate the corresponding query vector Q, key vector K and value vector V:

[0171]

[0172] Where W Q , W K and W V are the weight matrices of query, key and value respectively, and X is the input data;

[0173] b2) Calculate the attention score using the dot product of the query vector Q and the key vector K:

[0174]

[0175] Where Attention(Q,K,V) is the attention score of the input data; d k is the dimension of the key vector K;

[0176] b3) Calculate the multi-head attention matrix:

[0177]

[0178] MultiHead(Q,K,V)=(head1,head2,…,head h )W O

[0179] In the formula, head i is the i-th attention weight; Corresponding to the weight matrices of the i-th Q, K, and V, W O is the final attention matrix after linear transformation; MultiHead(Q,K,V) is the multi-head attention score of the input data;

[0180] b4) Repeat steps b1), b2), and b3) to generate multiple query vectors Q, key vectors K, and value vectors V, and calculate the attention scores and weighted sums respectively; concatenate the results of all attention weights and obtain the final output through linear transformation to obtain the trained Transformer neural network.

[0181] In this embodiment, in step S5, the test set obtained in S3 is used to test the Transformer neural network obtained in S4, and the prediction results of all subsequences are reconstructed to obtain the prediction results of the original motion response.

[0182] Specifically, the standardized subsequence test set obtained in step S3 is used to test the neural network obtained in step S4. The prediction results of each subsequence test set are obtained, such as Figure 3 , Figure 4 The following are comparisons of the monitoring and prediction values ​​of the offshore floating platform motion response subsequences IMF5 and IMF6, where the black curve is the subsequence value obtained by decomposing the offshore floating platform motion response, and the red curve is the subsequence prediction value of the method provided by the present invention. The prediction results of all subsequence test sets are summarized, the prediction results of the original motion response signal are reconstructed, and the relevant evaluation indicators are calculated to achieve the prediction of the offshore floating platform motion response. Figure 5 As shown, the black curve is the monitoring value of the motion response of the offshore floating platform, and the red curve is the predicted value of the method provided by the present invention.

[0183] In this embodiment, in step S6, in the process of realizing the prediction of the motion response of the offshore floating platform, the selected evaluation indicators are the root mean square error, mean absolute error and determination coefficient of the motion;

[0184] The formula for root mean square error RMSE is:

[0185]

[0186] The formula for mean absolute error MAE is:

[0187]

[0188] Determination coefficient R 2 The formula is:

[0189]

[0190] In the formula, y i is the platform motion monitoring value, is the platform motion prediction value, is the average value of the platform motion monitoring value; n is the number of motion data points.

[0191] Specifically, RMSE and MAE can better evaluate the error level of the prediction results. The smaller the value, the better the model prediction performance. 2 The fitting effect between the prediction result and the original data can be better evaluated. The value is distributed between [0,1]. The larger the value, the better the model performance. For comparison, the LSTM neural network is selected to carry out the prediction work on the same offshore floating platform motion response monitoring value. The results are as follows Figure 6 As shown, the black curve is the monitoring value of the motion response of the offshore floating platform, and the blue curve is the prediction value of the LSTM neural network. The prediction performance of the two methods is shown in Table 3:

[0192] Table 3 Prediction performance of the two methods

[0193]

[0194] The results obtained from the examples show that the prediction error of the method of the present invention is significantly smaller than that of the LSTM neural network, and the determination coefficient R 2 The larger the value, the better the fitting effect on the motion response of the floating platform, and the better the nonlinear characteristics of the monitoring data can be captured, which verifies that the method of the present invention has a significant improvement effect on the prediction accuracy of the motion response of the offshore floating platform.

[0195] In summary, the present invention obtains the motion response monitoring data of an offshore floating platform, pre-processes the motion response monitoring data; uses the CEEMDAN algorithm to perform empirical mode decomposition on the pre-processed motion response monitoring data to obtain several subsequences, wherein the subsequences include multiple intrinsic mode functions IMF and a residual sequence; divides the obtained subsequences into a training set and a test set, and performs standardization on the data in the training set and the test set; constructs a Transformer neural network model, and performs neural network training parameter and neural network structure configuration on the Transformer neural network model; trains the Transformer neural network model with the standardized training set to obtain the trained Transformer neural network model; tests the trained Transformer neural network model with the standardized test set, summarizes and reconstructs the prediction results of all the test sets, and obtains the prediction results of the motion response monitoring data; performs a comprehensive quantitative evaluation on the prediction results of the motion response of the offshore floating platform, and realizes the prediction of the motion response of the offshore floating platform. The present invention realizes the accurate prediction of the motion response of the offshore floating platform, and can provide auxiliary reference for on-site decision-making.

[0196] Example 2

[0197] See also Figure 7 Embodiment 2 of the present invention further provides a CEEMDAN-Transformer-based offshore platform motion response prediction device, comprising:

[0198] The motion response monitoring data processing module 100 is used to obtain the motion response monitoring data of the offshore floating platform and pre-process the motion response monitoring data;

[0199] The monitoring data empirical mode decomposition module 200 is used to perform empirical mode decomposition on the pre-processed motion response monitoring data using the CEEMDAN algorithm to obtain a plurality of subsequences, wherein the subsequences include a plurality of intrinsic mode functions IMFs and a residual sequence;

[0200] A subsequence division processing module 300 is used to divide the obtained subsequence into a training set and a test set, and perform standardization processing on the data in the training set and the test set;

[0201] Transformer neural network model construction module 400, used to construct a Transformer neural network model, and perform neural network training parameter and neural network structure configuration on the Transformer neural network model;

[0202] A Transformer neural network model training module 500 is used to train the Transformer neural network model using the standardized training set to obtain the trained Transformer neural network model;

[0203] A Transformer neural network model testing module 600 is used to test the trained Transformer neural network model using the standardized test set;

[0204] A forecast result aggregation and reconstruction module 700 is used to aggregate and reconstruct the forecast results of all the test sets to obtain the forecast results of the motion response monitoring data;

[0205] The forecast result quantitative evaluation module 800 is used to perform comprehensive quantitative evaluation on the forecast results of the motion response of the offshore floating platform, so as to realize the forecast of the motion response of the offshore floating platform.

[0206] In this embodiment, in the motion response monitoring data processing module 100:

[0207] Acquire the motion response monitoring data of offshore floating platforms through differential global positioning system and motion reference unit system;

[0208] The preprocessing method for the motion response monitoring data includes filtering to remove abnormal values ​​and noise.

[0209] In this embodiment, the monitoring data empirical mode decomposition module 200 includes:

[0210] Operator definition submodule 201, used to define operator E m(·) Carry out m-order empirical mode decomposition on the signal; add multiple Gaussian white noises w that obey the standard normal distribution N(0,1) to the original motion response signal Y(t) i (t), and obtain the motion response signal Y i (t):

[0211] Y i (t) = Y (t) + ε i w i (t),i=1,2,…I

[0212] In the formula, ε i is the noise coefficient, which is used to determine the signal-to-noise ratio of each step; I is the number of Gaussian white noise added;

[0213] The first IMF definition submodule 202 is used to obtain multiple motion response signals Y i (t) Carry out empirical mode decomposition to obtain multiple subsequences Define the first IMF as:

[0214]

[0215] The first residual signal R1(t) is:

[0216]

[0217] In the formula, is the first intrinsic mode function obtained by the CEEMDAN algorithm;

[0218] The second IMF definition submodule 203 is used to define the decomposition result R1(t)+ε1E1(w i (t)) signal is subjected to empirical mode decomposition, and the second IMF is obtained as:

[0219]

[0220] In the formula, is the second intrinsic mode function obtained by CEEMDAN algorithm; is the intrinsic mode function of the second subsequence obtained by decomposing the residual R1(t) after adding multiple Gaussian white noises;

[0221] The k-th IMF definition submodule 204 is used to obtain the k-th IMF. The obtained k-th IMF is:

[0222]

[0223] In the formula, is the kth intrinsic mode function obtained by CEEMDAN algorithm; is the residual Rk-1 (t) After adding multiple Gaussian white noises, the second subsequence intrinsic mode function obtained by each decomposition; R k-1 (t) is the k-1th residual signal left after decomposition; ε k-1 is the noise coefficient of the k-1th step; E k-1 (w i (t)) is to carry out k-1 order empirical mode decomposition of the added Gaussian white noise;

[0224] The extreme value judgment and analysis submodule 205 is used to judge the k-th order residual signal R k (t) whether it contains at least two extreme values. If it does, continue to decompose and obtain the k+1th IMF; if it does not contain at least two extreme values, the decomposition process ends; finally, the time series signal Y(t) is decomposed into all subsequence IMFs that meet the conditions and the residual R representing the signal trend K (t) and.

[0225] In this embodiment, in the subsequence division processing module 300:

[0226] The formula for standardizing the data in the training set and the test set is:

[0227]

[0228] In the formula, y i is the offshore floating platform motion monitoring value, y norm is the result after standardization, y max is the maximum value of the input data, y min is the minimum value of the input data.

[0229] In this embodiment, in the Transformer neural network model construction module 400:

[0230] The neural network training parameter configuration objects for the Transformer neural network model include historical time step, prediction time step, data batch size and number of training rounds;

[0231] The neural network structure configuration objects of the Transformer neural network model include embedding layer, multi-head attention layer, normalization layer, feedforward layer, residual connection layer, and linearization layer; Softmax is selected to calculate the probability distribution of data, MSE is selected as the training loss function, and Adam optimizer is selected to update the neural network parameters.

[0232] In this embodiment, the Transformer neural network model training module 500 includes:

[0233] The vector generation submodule 501 is used to input the data in the training set into the Transformer neural network model to generate the corresponding query vector Q, key vector K and value vector V:

[0234]

[0235] Where W Q , W K and W V are the weight matrices of query, key and value respectively, and X is the input data;

[0236] The attention score calculation submodule 502 is used to calculate the attention score using the dot product of the query vector Q and the key vector K:

[0237]

[0238] Where Attention(Q,K,V) is the attention score of the input data; d k is the dimension of the key vector K;

[0239] The multi-head attention matrix calculation submodule 503 is used to calculate the multi-head attention matrix:

[0240]

[0241] MultiHead(Q,K,V)=(head1,head2,…,head h )W O

[0242] In the formula, head i is the i-th attention weight; Corresponding to the weight matrices of the i-th Q, K, and V, W O is the final attention matrix after linear transformation; MultiHead(Q,K,V) is the multi-head attention score of the input data.

[0243] In this embodiment, in the forecast result summary and reconstruction module 700, multiple query vectors Q, key vectors K and value vectors V are generated, and the attention scores and weighted sums are calculated respectively; the results of all attention weights are concatenated and linearly transformed to obtain the final output, thereby obtaining the trained Transformer neural network.

[0244] In this embodiment, in the forecast result quantitative evaluation module 800, the evaluation indicators selected are the root mean square error, mean absolute error and determination coefficient of the motion;

[0245] The formula for root mean square error RMSE is:

[0246]

[0247] The formula for mean absolute error MAE is:

[0248]

[0249] Determination coefficient R 2 The formula is:

[0250]

[0251] In the formula, y i is the platform motion monitoring value, is the platform motion prediction value, is the average value of the platform motion monitoring value; n is the number of motion data points.

[0252] It should be noted that the information interaction, execution process and other contents between the modules of the above-mentioned device are based on the same concept as the method embodiment in Example 1 of the present application, and the technical effects they bring are the same as those of the method embodiment of the present application. For specific contents, please refer to the description in the method embodiment shown in the previous part of the present application, and will not be repeated here.

[0253] Example 3

[0254] Embodiment 3 of the present invention provides a non-transitory computer-readable storage medium, in which the program code of the offshore platform motion response prediction method based on CEEMDAN-Transformer is stored, and the program code includes instructions for executing the offshore platform motion response prediction method based on CEEMDAN-Transformer of embodiment 1 or any possible implementation thereof.

[0255] The computer-readable storage medium may be any available medium that can be accessed by a computer or a data storage device such as a server or a data center that includes one or more available media. The available medium may be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)).

[0256] Example 4

[0257] Embodiment 4 of the present invention provides an electronic device, including: a memory and a processor;

[0258] The processor and the memory communicate with each other via a bus; the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the offshore platform motion response prediction method based on CEEMDAN-Transformer of Example 1 or any possible implementation thereof.

[0259] Specifically, the processor can be implemented by hardware or by software. When implemented by hardware, the processor can be a logic circuit, an integrated circuit, etc.; when implemented by software, the processor can be a general-purpose processor implemented by reading software codes stored in a memory. The memory can be integrated in the processor or can be located outside the processor and exist independently.

[0260] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present invention is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions can be transmitted from a website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website site, computer, server or data center.

[0261] Obviously, those skilled in the art should understand that the above modules or steps of the present invention can be implemented by a general computing device, they can be concentrated on a single computing device, or distributed on a network composed of multiple computing devices, and optionally, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, and in some cases, the steps shown or described can be executed in a different order than here, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. Thus, the present invention is not limited to any specific combination of hardware and software.

[0262] Although the present invention has been described in detail above by general description and specific embodiments, it is obvious to those skilled in the art that some modifications or improvements can be made to the present invention. Therefore, these modifications or improvements made without departing from the spirit of the present invention all belong to the scope of protection claimed by the present invention.

Claims

1. The offshore platform motion response prediction method based on CEEMDAN-Transformer is characterized by: include: Acquiring motion response monitoring data of an offshore floating platform, and preprocessing the motion response monitoring data; Using CEEMDAN algorithm to perform empirical mode decomposition on the preprocessed motion response monitoring data to obtain a plurality of subsequences, wherein the subsequences include a plurality of intrinsic mode functions (IMFs) and a residual sequence; Dividing the obtained subsequences into a training set and a test set, and performing standardization processing on the data in the training set and the test set; Constructing a Transformer neural network model, and configuring neural network training parameters and neural network structure for the Transformer neural network model; training the Transformer neural network model using the standardized training set to obtain the trained Transformer neural network model; The trained Transformer neural network model is tested by using the standardized test set, and the prediction results of all the test sets are summarized and reconstructed to obtain the prediction results of the motion response monitoring data; A comprehensive quantitative evaluation is conducted on the prediction results of the motion response of the offshore floating platform to achieve the prediction of the motion response of the offshore floating platform.

2. The offshore platform motion response prediction method based on CEEMDAN-Transformer according to claim 1 is characterized in that: Acquire the motion response monitoring data of offshore floating platforms through differential global positioning system and motion reference unit system; The preprocessing method for the motion response monitoring data includes filtering to remove abnormal values ​​and noise.

3. The offshore platform motion response prediction method based on CEEMDAN-Transformer according to claim 1, characterized in that: The steps of performing empirical mode decomposition on the preprocessed motion response monitoring data using the CEEMDAN algorithm to obtain a plurality of subsequences include: a1) Define operator E m (·) Carry out m-order empirical mode decomposition on the signal; add multiple Gaussian white noises w that obey the standard normal distribution N(0,1) to the original motion response signal Y(t) i (t), and obtain the motion response signal Y i (t): Y i (t)=Y(t)+ε i w i (t),i=1,2,…I In the formula, ε i is the noise coefficient, which is used to determine the signal-to-noise ratio of each step of Gaussian white noise; I is the number of Gaussian white noise added; a2) The obtained multiple motion response signals Y i (t) Carry out empirical mode decomposition to obtain multiple subsequences Define the first IMF as: The first residual signal R1(t) is: In the formula, is the first intrinsic mode function obtained by the CEEMDAN algorithm; a3) For the decomposition result R1(t)+ε1E1(w i (t)) signal is subjected to empirical mode decomposition, and the second IMF is obtained as: In the formula, is the second intrinsic mode function obtained by CEEMDAN algorithm; is the intrinsic mode function of the second subsequence obtained by decomposing the residual R1(t) after adding multiple Gaussian white noises; a4) Repeat step a3) to get the kth IMF: In the formula, is the kth intrinsic mode function obtained by CEEMDAN algorithm; is the residual R k-1 (t) After adding multiple Gaussian white noises, the second subsequence intrinsic mode function obtained by each decomposition; R k-1 (t) is the k-1th residual signal left after decomposition; ε k-1 is the noise coefficient of the k-1th step; E k-1 (w i (t)) is to carry out k-1 order empirical mode decomposition of the added Gaussian white noise; a5) Determine the k-th order residual signal R k (t) whether it contains at least two extreme values. If it does, continue to decompose and repeat step a3) to obtain the k+1th IMF; if it does not contain at least two extreme values, the decomposition process ends; finally, the time series signal Y(t) is decomposed into all subsequence IMFs that meet the conditions and the residual R representing the signal trend K (t) and.

4. The offshore platform motion response prediction method based on CEEMDAN-Transformer according to claim 1, characterized in that: The formula for standardizing the data in the training set and the test set is: In the formula, y i is the offshore floating platform motion monitoring value, y norm is the result after standardization, y max is the maximum value of the input data, y min is the minimum value of the input data.

5. The offshore platform motion response prediction method based on CEEMDAN-Transformer according to claim 1 is characterized in that: The neural network training parameter configuration objects for the Transformer neural network model include historical time step, prediction time step, data batch size and number of training rounds; The neural network structure configuration objects of the Transformer neural network model include embedding layer, multi-head attention layer, normalization layer, feedforward layer, residual connection layer, and linearization layer; Softmax is selected to calculate the probability distribution of data, MSE is selected as the training loss function, and Adam optimizer is selected to update the neural network parameters.

6. The offshore platform motion response prediction method based on CEEMDAN-Transformer according to claim 5 is characterized in that: The step of training the Transformer neural network model by using the standardized training set to obtain the trained Transformer neural network model comprises: b1) Input the data in the training set into the Transformer neural network model to generate the corresponding query vector Q, key vector K and value vector V: Where W Q , W K and W V are the weight matrices of query, key and value respectively, and X is the input data; b2) Calculate the attention score using the dot product of the query vector Q and the key vector K: Where Attention(Q,K,V) is the attention score of the input data; d k is the dimension of the key vector K; b3) Calculate the multi-head attention matrix: MultiHead(Q,K,V)=(head1,head2,…,head h )W O In the formula, head i is the i-th attention weight; They correspond to the weight matrices of the i-th Q, K, and V respectively, and WO is the final attention matrix after linear transformation; MultiHead(Q,K,V) is the multi-head attention score of the input data; b4) Repeat steps b1), b2), and b3) to generate multiple query vectors Q, key vectors K, and value vectors V, and calculate the attention scores and weighted sums respectively; concatenate the results of all attention weights and obtain the final output through linear transformation to obtain the trained Transformer neural network.

7. The offshore platform motion response prediction method based on CEEMDAN-Transformer according to claim 6 is characterized in that: The prediction results of the motion response of the offshore floating platform are comprehensively and quantitatively evaluated. In the process of predicting the motion response of the offshore floating platform, the evaluation indicators selected are the root mean square error, mean absolute error and determination coefficient of the motion; The formula for root mean square error RMSE is: The formula for mean absolute error MAE is: Determination coefficient R 2 The formula is: In the formula, y i is the platform motion monitoring value, is the platform motion prediction value, is the average value of the platform motion monitoring value; n is the number of motion data points.

8. The offshore platform motion response prediction device based on CEEMDAN-Transformer is characterized by: include: A motion response monitoring data processing module is used to obtain the motion response monitoring data of the offshore floating platform and pre-process the motion response monitoring data; A monitoring data empirical mode decomposition module, used for performing empirical mode decomposition on the pre-processed motion response monitoring data using a CEEMDAN algorithm to obtain a plurality of subsequences, wherein the subsequences include a plurality of intrinsic mode functions IMFs and a residual sequence; A subsequence division processing module, used to divide the obtained subsequence into a training set and a test set, and perform standardization processing on the data in the training set and the test set; A Transformer neural network model construction module, used to construct a Transformer neural network model, and perform neural network training parameter and neural network structure configuration on the Transformer neural network model; A Transformer neural network model training module, used to train the Transformer neural network model using the standardized training set to obtain the trained Transformer neural network model; A Transformer neural network model testing module, used to test the trained Transformer neural network model using the standardized test set; A forecast result aggregation and reconstruction module, used for aggregating and reconstructing the forecast results of all the test sets to obtain the forecast results of the motion response monitoring data; The forecast result quantitative evaluation module is used to conduct a comprehensive quantitative evaluation of the forecast results of the motion response of the offshore floating platform, so as to realize the forecast of the motion response of the offshore floating platform.

9. The offshore platform motion response prediction device based on CEEMDAN-Transformer according to claim 8, characterized in that: In the motion response monitoring data processing module: Acquire the motion response monitoring data of offshore floating platforms through differential global positioning system and motion reference unit system; The preprocessing method of the motion response monitoring data includes filtering to remove abnormal values ​​and noise; The monitoring data empirical mode decomposition module includes: Operator definition submodule, used to define operator E m (·) Carry out m-order empirical mode decomposition on the signal; add multiple Gaussian white noises w that obey the standard normal distribution N(0,1) to the original motion response signal Y(t) i (t), and obtain the motion response signal Y i (t): Y i (t)=Y(t)+ε i w i (t),i=1,2,…I In the formula, ε i is the noise coefficient, which is used to determine the signal-to-noise ratio of each step; I is the number of Gaussian white noise added; The first IMF definition submodule is used to obtain multiple motion response signals Y i (t) Carry out empirical mode decomposition to obtain multiple subsequences Define the first IMF as: The first residual signal R1(t) is: In the formula, is the first intrinsic mode function obtained by the CEEMDAN algorithm; The second IMF definition submodule is used to decompose the result R1(t)+ε1E1(w i (t)) signal is subjected to empirical mode decomposition, and the second IMF is obtained as: In the formula, is the second intrinsic mode function obtained by CEEMDAN algorithm; is the intrinsic mode function of the second subsequence obtained by decomposing the residual R1(t) after adding multiple Gaussian white noises; The k-th IMF definition submodule is used to obtain the k-th IMF. The obtained k-th IMF is: In the formula, is the kth intrinsic mode function obtained by CEEMDAN algorithm; is the residual R k-1 (t) After adding multiple Gaussian white noises, the second subsequence intrinsic mode function obtained by each decomposition; R k-1 (t) is the k-1th residual signal left after decomposition; ε k-1 is the noise coefficient of the k-1th step; E k-1 (w i (t)) is to carry out k-1 order empirical mode decomposition of the added Gaussian white noise; The extreme value judgment analysis submodule is used to judge the k-th order residual signal R k (t) whether it contains at least two extreme values. If it does, continue to decompose and obtain the k+1th IMF; if it does not contain at least two extreme values, the decomposition process ends; finally, the time series signal Y(t) is decomposed into all subsequence IMFs that meet the conditions and the residual R representing the signal trend K (t) the sum; In the subsequence division processing module: The formula for standardizing the data in the training set and the test set is: In the formula, y i is the offshore floating platform motion monitoring value, y norm is the result after standardization, y max is the maximum value of the input data, y min is the minimum value of the input data; In the Transformer neural network model building module: The neural network training parameter configuration objects for the Transformer neural network model include historical time step, prediction time step, data batch size and number of training rounds; The neural network structure configuration objects of the Transformer neural network model include an embedding layer, a multi-head attention layer, a normalization layer, a feedforward layer, a residual connection layer, and a linearization layer; Softmax is used to calculate the probability distribution of the data, MSE is selected as the training loss function, and Adam optimizer is selected to update the neural network parameters; The Transformer neural network model training module includes: The vector generation submodule is used to input the data in the training set into the Transformer neural network model to generate the corresponding query vector Q, key vector K and value vector V: Where W Q , W K and W V are the weight matrices of query, key and value respectively, and X is the input data; The attention score calculation submodule is used to calculate the attention score using the dot product of the query vector Q and the key vector K: Where Attention(Q,K,V) is the attention score of the input data; d k is the dimension of the key vector K; Multi-head attention matrix calculation submodule, used to calculate the multi-head attention matrix: MultiHead(Q,K,V)=(head1,head2,…,head h )W O In the formula, head i is the i-th attention weight; Corresponding to the weight matrices of the i-th Q, K, and V, W O is the final attention matrix after linear transformation; MultiHead(Q,K,V) is the multi-head attention score of the input data; In the forecast result summary and reconstruction module, multiple query vectors Q, key vectors K and value vectors V are generated, and the attention scores and weighted sums are calculated respectively; the results of all attention weights are concatenated and linearly transformed to obtain the final output, thereby obtaining the trained Transformer neural network.

10. The offshore platform motion response prediction device based on CEEMDAN-Transformer according to claim 8, characterized in that: In the forecast result quantitative evaluation module, the selected evaluation indicators are the root mean square error, mean absolute error and determination coefficient of the motion; The formula for root mean square error RMSE is: The formula for mean absolute error MAE is: Determination coefficient R 2 The formula is: In the formula, y i is the platform motion monitoring value, is the platform motion prediction value, is the average value of the platform motion monitoring value; n is the number of motion data points.

Images