Wave prediction method based on MEEMD-ATTBiLSTM

Abstract

The invention discloses a wave prediction method based on MEEMD-ATTBiLSTM, and relates to the technical field of marine monitoring, and the method comprises the steps: carrying out the boundary extension of a training set through employing an LSTM algorithm, and recognizing adjacent first extreme points; extending the first extreme point into a symmetric closed sequence by using mirror image continuation, and forming a decomposition training set by using the symmetric closed sequence and the training set after boundary continuation; processing and decomposing the training set by adopting an empirical mode decomposition method to obtain a plurality of groups of imf components and a group of res residual data; frequencies of each group of imf components are different and are arranged from high to low; respectively inputting each group of imf component and res residual data into an Attention network, inputting the original sequence into the Attention network to establish a weight index, and obtaining a weight coefficient matrix of each group of imf component and res residual data relative to the training set; and inputting each group of weight coefficient matrixes into the Bi-LSTM model, predicting each group of imf components, and calculating and predicting the SWH according to the predicted imf components. According to the invention, short-term wave prediction can be realized, and the wave prediction effect is improved.

Description

Technical Field

[0001] The present invention relates to the field of ocean monitoring technology, and more specifically to a wave prediction method based on MEEMD-ATTBi LSTM. Background Art

[0002] At present, wave prediction methods are mainly divided into four categories: numerical models, classical time series models, machine learning models and hybrid models. Among the numerical models, the third-generation wave numerical WAM (wave modeling group) prediction method is widely used. WAM can predict SWH (Significant wave height) based on various terms in the energy balance equation. It is a prediction method widely used in shallow and deep waters around the world. Although the numerical prediction method takes into account various factors of wave formation and calculates waves according to the principles of fluid mechanics, it has high computational cost and complexity, and requires a large amount of data and powerful computers; time series models, such as autoregressive (AR) models and autoregressive moving average (ARMA) models, are basically linear models, and the data to be analyzed and predicted need to be linear and stable when used; in order to overcome the adverse effects of nonlinear factors in wave time series on model prediction accuracy, machine learning methods, such as support vector machines (SVM) and support vector regression (SVR), have also been tried. At the same time, hybrid models of machine learning are also widely used in ocean element prediction, such as: a hybrid algorithm (GGA-ELM) based on the Group Genetic Algorithm (GGA) and the Extreme Learning Machine (ELM) is used to predict future wave heights and energy flows. In addition, due to the strong nonlinearity in the generation and evolution of waves, it is difficult to make accurate predictions.

[0003] Therefore, additional data preprocessing methods are needed, such as the widely used empirical wavelet transform (EWT), empirical mode decomposition (EMD), seasonal-trend decomposition procedure based on loess (STL), etc., to reduce the non-stationarity of the wave height series. Among them, although the EMD algorithm can adaptively decompose the signal, it is necessary to obtain the upper and lower envelopes of the signal multiple times in the process of using the EMD method to decompose the signal IMF. However, the endpoints of the signal data sequence are not necessarily local extreme points, which may cause fitting errors at the endpoints, resulting in divergence at both ends of the decomposed IMF, and the error will be transmitted to each order IMF component step by step, so that the entire IMF set is distorted, resulting in poor signal processing effect based on the EMD method. To address this shortcoming, methods such as mirror extension method, proportional extension method, linear prediction method, etc. can be used to reduce its impact on internal data, but the traditional extension method does not have a good improvement effect on the endpoint effect, and cannot be adaptively extended with time series transformation. Moreover, most current studies focus on long-term SWH prediction, with sampling frequencies of 1 h and 2 h, while research in the field of short-term prediction is relatively lacking.

[0004] Therefore, how to achieve short-term wave prediction and improve the wave prediction effect is an urgent problem to be solved by those skilled in the art. Summary of the invention

[0005] In view of this, the present invention provides a wave prediction method based on MEEMD-ATTBiLSTM to achieve accurate prediction of waves, which is helpful for making decisions on sea surface operations and improving the utilization rate of ocean energy.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] A wave prediction method based on MEEMD-ATTBiLSTM, comprising the following steps:

[0008] Step 1: Collect historical wave data to construct a training set, and use the LSTM algorithm to extend the boundary of the training set to identify the adjacent first extreme value points;

[0009] Step 2: Use mirror extension to extend the first extreme point into a symmetric closed sequence. The symmetric closed sequence and the training set after boundary extension constitute the decomposition training set.

[0010] Step 3: Use the empirical mode decomposition method to process the decomposition training set to obtain several groups of IMF components and a group of res residual data; the frequency of each group of IMF components is different and arranged from high to low for easy visualization;

[0011] Step 4: Input each group of IMF components and res residual data into the Attention network respectively, and input the training set into the Attention network to establish the weight index, and obtain the weight coefficient matrix of each group of IMF components and res residual data relative to the training set;

[0012] Step 5: Input each group of weight coefficient matrices into the Bi-LSTM model, predict each group of IMF components, and calculate the predicted SWH based on the predicted IMF components.

[0013] The technical effect of the above technical solution is that, on the basis of the traditional mirror extension method, an LSTM network is added to more accurately judge adjacent mirror extreme extension points, and with the addition of the neural network, the original mirror module can adaptively adjust the adjacent extreme points found after the sequence changes, and use the found adjacent extreme points as mirror extreme extension points to extend them at the end of the signal to suppress the endpoint effect.

[0014] Preferably, in step 4, the Attention network uses a self-attention mechanism to calculate the weight coefficient matrix, and the specific process is:

[0015] Step 41: According to the two weight matrices W input q , W v Linearly map the imf component and res residual data respectively to obtain the query vector Q and the value vector V;

[0016] Step 42: According to the input weight matrix W k Perform linear mapping on the training set to obtain the key vector K;

[0017] Step 43: Use the attention score function to calculate the dot product of the query vector and the key vector to obtain the attention weight matrix A;

[0018] Step 44: Use the attention weight matrix A to weight and sum the value vector V to obtain the weight coefficient matrix H.

[0019] Preferably, in step 4, each group of IMF components is used as an index and the training set is used as a dictionary to establish a correlation weight matrix between each group of IMF components and the training set to obtain a weight coefficient matrix; the training set is transformed according to the dimensions of the query vector and the value vector, and the matrix of the training set, the query vector, and the value vector are aligned to obtain a key vector; and the index of the weight matrix of the query vector is established according to the key vector.

[0020] In the traditional attention mechanism, the query vector Q and the input matrix X are used to generate the attention distribution matrix A, and then the matrix is ​​multiplied by X to derive the attention weighted output matrix. In contrast, the self-attention mechanism uses the query key value to replace this process. It establishes the relationship between each position in the input sequence by calculating the interaction between each element in the sequence and other elements. The result of this calculation is to determine the weight assigned to each element, thereby learning the correlation between different positions in the input sequence. In the self-attention mechanism, the input sequence X is linearly mapped to three different matrices through three different weight matrices (Wq, Wk, Wv); the attention score function is obtained by calculating the dot product of the query vector Q and the key vector K, thereby generating the attention weight matrix A, and finally, the value vector V is weighted and summed using the weight matrix A to obtain the attention output matrix H. The attention mechanism used in the present invention is based on the self-attention mechanism. The original sequence data is converted according to the dimensions of Query and value to align it with the Query and value matrices, and then input into the network as a Key. The reconstructed Key can establish the index of the weight matrix of the original sequence Query and enhance the relationship between the decomposed components and the original sequence. This process captures the relationship between different positions in the input sequence and the original sequence, amplifies the important features of the sequence, and reduces the impact of irrelevant aspects.

[0021] Preferably, in step 5, each group of weight coefficient matrices is input into the Bi-LSTM model in a sliding window manner to obtain the weighted IMF components and weighted RES residual data after training weighting, and each group of weighted IMF components and weighted RES residual data are superimposed to obtain the predicted SWH.

[0022] It can be seen from the above technical solution that, compared with the prior art, the present invention discloses a wave prediction method based on MEEMD-ATTBiLSTM, which includes the following beneficial effects:

[0023] 1) The empirical mode decomposition method is introduced to decompose non-stationary data into intrinsic mode functions and residual terms of different frequencies, which is conducive to the mining of sequence features. In order to eliminate the endpoint effect brought by the EMD algorithm, the improved MEEMD algorithm is introduced;

[0024] 2) In the deep learning module, a new attention mechanism is proposed to associate the attention mechanism with the original sequence, and a correlation weight matrix between each imf component and the original sequence is established. Finally, Bi-LSTM is used to improve the utilization of historical data, mine deeper sequence information and make predictions on wave sequences;

[0025] 3) Compared with traditional machine learning models, the short-term wave prediction model of MEEMD-ATTBiLSTM proposed in this paper can effectively improve the accuracy of short-term wave data prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0027] Figure 1 A flow chart of the wave prediction method based on MEEMD-ATTBiLSTM provided by the present invention;

[0028] Figure 2 ATTBiLSTM model structure diagram provided by the present invention;

[0029] Figure 3 Schematic diagram of LSTM mirror extension provided by the present invention;

[0030] Figure 4 A schematic diagram of EMD decomposition components provided by the present invention;

[0031] Figure 5 A schematic diagram of the decomposition effect of important components of MEEMD and EMD decomposition provided by the present invention;

[0032] Figure 6 This is a schematic diagram comparing the prediction results of various models provided by the present invention. DETAILED DESCRIPTION

[0033] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only 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.

[0034] The embodiment of the present invention discloses a wave prediction method based on MEEMD-ATTBiLSTM, and proposes a hybrid prediction model that associates the Attention mechanism with the original sequence and uses Bi-LSTM. Historical wave data is collected as the original sequence, and in the traditional Attention mechanism, the original sequence is added as a dictionary, each imf component is used as an index, and a correlation weight matrix between each imf component and the original sequence is established. Because each imf component decomposed by EMD has a certain connection with the original sequence, the decomposed components can be divided into strongly correlated linear parts and unimportant noise interference according to the size of the correlation coefficient, and this association can be reflected in the correlation weight matrix in the above-mentioned Attention mechanism. After obtaining the correlation weight matrix, it is input into the Bi-LSTM network for component prediction. Compared with the traditional LSTM network, the Bi-LSTM network can more comprehensively understand the meaning of time series data, which is very useful when processing non-stationary and nonlinear time series tasks such as wave data. The Bi-LSTM network can also improve the performance and stability of the model. By stacking LSTM layers, Bi-LSTM can get rid of the limitation of relying only on current time information and better combine forward and backward waves for prediction, which makes Bi-LSTM perform well when processing complex sequence structures. When all the imf components are predicted by the Bi-LSTM layer and output through the fully connected layer, all the predicted values ​​are linearly superimposed to obtain the trend of the future movement of the wave. The specific process of the wave prediction method based on MEEMD-ATTBiLSTM is as follows:

[0035] S1: Use the LSTM algorithm to perform boundary extension on the training set data (original sequence) and identify the first adjacent extreme point, then use mirror extension to extend this extreme point into a symmetrical and closed sequence; use the LSTM algorithm to extend the first adjacent extreme point at the endpoint of the training data set, and then use the mirror symmetric extension method to extend from the boundary extreme point into a symmetrical and closed time series. The length of the boundary extension generally does not exceed half the average period of the time series, which can eliminate the endpoint effect and avoid the problem of increased error when using LSTM alone; the extension process is as follows Figure 3 As shown, the lines with diamond patterns represent the original sequence, the lines with upper triangle patterns represent the sequence extended from the boundary, and the lines with lower triangle patterns represent the sequence extended from the mirror image;

[0036] S2: Use the empirical mode decomposition method to process the training set data that has been extended to obtain a set of IMF components and res residual data with frequencies from high to low; use the adaptive sequence decomposition ability of the empirical mode decomposition method, combined with the LSTM algorithm of end-point mirror extension, to process the nonlinear and non-smooth wave motion data into a stable time series and minimize the impact of the end-point effect on the sequence prediction;

[0037] S3: Input each group of imf components and res residual data into the attention mechanism weight network, and input the original sequence into the attention network of each group to establish the weight index;

[0038] S4: The weight coefficient sequence reconstructed by the Attention network is input into the neural network by means of a sliding window; each group of sequences in the training set is input into the Bi-LSTM model, and the final predicted SWH is obtained by superimposing each group of imf and res components after training. For the prediction of short-term wave height time series, the improved MEEMD algorithm and Bi-LSTM model in S1 are combined, and the ATTBiLSTM neural network is used for prediction; the ATTBiLSTM neural network uses the attention mechanism layer combined with the index enhancement component features of the original sequence to show the movement trend of the original sequence, and then uses the Bi-LSTM model to improve the utilization of historical data, dig deeper sequence information and make predictions.

[0039] Example 1

[0040] In this example, a training set consisting of the first 1800 data points and a test set consisting of the last 200 data points are used. Both the training samples and the test samples are generated by the sliding window method, and each sample is in a sequence-to-sequence format. Specifically, the input vector is defined as {x1, x2, ..., x I}, the target vector is defined as {x I+1 , x I+2 , ..., x I+O}, where I represents the historical window size and O represents the forecasting horizon. The parameter update of the model is realized by the difference between the target vector and the true value. In this embodiment, the random seed is fixed, which can reduce the influence of random variables and ensure the repeatability of the experimental results.

[0041] This embodiment is an improved experiment based on EMD decomposition (empirical mode decomposition). Figure 4The 7 intrinsic mode functions (imf) and 1 residual component (Res) in the figure are generated by EMD decomposition of the input original wave sequence. It can be observed in the figure that the eigenmode functions of different frequencies reflect the characteristic information of the original sequence on different time scales, and the residual component reflects the long-term change trend of the sequence. This operation can transform the original non-stationary sequence into multiple stationary sequences; and among the multiple intrinsic mode functions generated, different intrinsic mode functions have huge differences in the performance of the original sequence characteristics. Figure 4 As shown in the figure, the characteristics of each imf and RES are different. From imf1 to imf7 and then to residual, the complexity and frequency of the sequence gradually decrease, and the period gradually increases; the trends of the first four imfs are the most complex and changeable; the Res value fluctuates the least, which reflects the overall trend of wave movement.

[0042] The trend of the decomposed sequence is more obvious than that of the original sequence. Therefore, using the decomposition algorithm to predict and process the original data can make it easier for the deep learning module to learn the potential mapping between input and output. In this set of decomposed sequences, the imf1 and imf2 functions play a very important role in the composition of the original sequence. Figure 5 It shows the decomposition effects of MEEMD decomposition and EMD decomposition for important imf components. The straight line represents the Original Series EMD decomposition, that is, the real wave trajectory EMD decomposition, the line with a dot pattern represents the Truncated Series EMD decomposition, and the line with a triangle pattern represents the Truncated Series MEEMD decomposition. The left figure shows the decomposition effect of imf1, and the right figure shows the decomposition effect of imf2.

[0043] Example 2

[0044] Based on the above embodiment, in order to evaluate the prediction effect of the proposed model, three commonly used statistical indicators are used, namely root mean square error (RMSE), mean absolute error (MAE) and NSE to evaluate the performance of the prediction model. The definitions of the statistical indicators are as follows:

[0045]

[0046] Among them, x i and x are the observed and average significant wave heights, respectively; y i and y are the predicted significant wave height and the average predicted wave height respectively; the length of the time series is n.

[0047] NSE is a commonly used model evaluation indicator in hydrology and environmental science to evaluate the degree of fit of the model to the observed data. The range of NSE values ​​is usually between -∞ and 1, which can be used to quantify the consistency between the model prediction value and the actual observation value. For wave time series, a high NSE value can indicate that the predicted sequence has a similar trend to the actual sequence, but it cannot guarantee whether the predicted value has real-time followability, especially in some non-stationary parts of the sequence. At this time, the supplement of RMSE and MAE can reflect the follow-up of the predicted sequence and the real sequence, which can increase the comprehensiveness of the evaluation.

[0048] In order to further establish the superior performance of the model and evaluate the prediction performance of MEEMD-ATTBiLSTM for SWH, the SWH prediction results of MEEMD-ATTBiLSTM were compared with those of LSTM, ATTBiLSTM and EMD-LSTM models using the optimized data after EMD decomposition.

[0049] The SWH prediction results of the four models are shown in Table 1. It can be clearly seen from the results that MEEMD-ATTBiLSTM is significantly better than other models. The RMSE, MAE and NSE of SWH prediction using MEEMD-ATTBiLSTM are 0.267m, 0.209m and 0.984 respectively.

[0050] Table 1 Model prediction results

[0051] Model RMSE MAE NSE MEEMD-ATTBiLSTM 0.267 0.209 0.984 LSTM 0.952 0.757 0.704 ATTBiLSTM 0.550 0.445 0.934 EMD-LSTM 1.625 1.30 0.432

[0052] The results of the comparative experiment are as follows Figure 6 As shown, the black line represents the real wave motion trajectory, the red line is the predicted value of MEEMD-ATTBiLSTM, the blue line is the predicted value of LSTM, the green line is the predicted value of ATTBiLSTM, and the orange line is the predicted value of EMD-LSTM. The top and bottom layers are both partial enlarged views of the middle layer. It can be seen from the figure that the MEEMD-ATTLSTM self-attention mechanism model has the highest degree of fit with the real load.

[0053] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0054] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A wave prediction method based on MEEMD-ATTBiLSTM, characterized in that: The following steps are involved: Step 1: Collect historical wave data to construct a training set, and use the LSTM algorithm to extend the boundary of the training set to identify the adjacent first extreme value points; Step 2: Use mirror extension to extend the first extreme point into a symmetric closed sequence. The symmetric closed sequence and the training set after boundary extension constitute the decomposition training set. Step 3: Use the empirical mode decomposition method to process and decompose the training set to obtain several groups of IMF components and a group of res residual data; Step 4: Input each group of IMF components and res residual data into the Attention network respectively, and input the training set into the Attention network to establish the weight index, and obtain the weight coefficient matrix of each group of IMF components and res residual data relative to the training set; Step 5: Input each group of weight coefficient matrices into the Bi-LSTM model, predict each group of IMF components, and calculate the predicted SWH based on the predicted IMF components.

2. A wave prediction method based on MEEMD-ATTBiLSTM according to claim 1, characterized in that: In step 3, the frequencies of each group of IMF components are different and are arranged from high to low.

3. A wave prediction method based on MEEMD-ATTBiLSTM according to claim 1, characterized in that: In step 4, the Attention network uses the self-attention mechanism to calculate the weight coefficient matrix. The specific process is: Step 41: linearly map the imf component and the res residual data according to the two input weight matrices to obtain a query vector and a value vector; Step 42: Perform linear mapping on the training set according to the input weight matrix to obtain the key vector; Step 43: Use the attention score function to calculate the dot product of the query vector and the key vector to obtain the attention weight matrix; Step 44: Use the attention weight matrix to weight and sum the value vector to obtain the weight coefficient matrix.

4. A wave prediction method based on MEEMD-ATTBiLSTM according to claim 3, characterized in that: In step 4, each group of IMF components is used as an index and the training set is used as a dictionary to establish a correlation weight matrix between each group of IMF components and the training set to obtain a weight coefficient matrix; the training set is transformed according to the dimensions of the query vector and the value vector, and the matrix of the training set, the query vector, and the value vector are aligned to obtain the key vector; the index of the weight matrix of the query vector is established based on the key vector.

5. The wave prediction method based on MEEMD-ATTBiLSTM according to claim 1, characterized in that: In step 5, each group of weight coefficient matrices is input into the ATTBiLSTM network in a sliding window manner to obtain the weighted IMF components and weighted RES residual data after training weighting, and each group of weighted IMF components and weighted RES residual data are superimposed to obtain the predicted SWH.

Images