EWT-Transformer prediction method applied to vortex-induced vibration response of heat absorption tower structure
By combining empirical wavelet decomposition (EWT) and Transformer models, the EWT-Transformer combination prediction model is formed, which solves the calculation complexity and prediction error problems of traditional methods when predicting the eddy vibration response of the heat-absorbing tower structure, and realizes accurate prediction of the eddy vibration of the heat-absorbing tower structure.
Patent Information
- Application Number
- CN202411969512.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-30
- Publication Date
- 2025-05-02
AI Technical Summary
Traditional numerical simulation methods and machine learning models have problems of computational complexity and prediction error when predicting the eddy vibration response of the heat sink tower structure, especially when it is difficult to accurately capture the peaks in the face of complex nonlinear responses.
Empirical wavelet decomposition (EWT) technology is introduced and the Transformer model is combined to form an EWT-Transformer combination prediction model. The complex signal is decomposed into simple harmonic sub-signals through signal decomposition, and prediction is performed using the Transformer model, and the sub-signal results are finally superimposed to restore the prediction results of the original signal.
The prediction accuracy of the eddy vibration response of the heat-absorbing tower structure is improved, the prediction error of a single model is reduced, the processing capability of complex nonlinear responses is enhanced, and the accurate prediction of the eddy vibration of the heat-absorbing tower structure is achieved.
Smart Images
Figure CN119918134A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of vortex-induced vibration prediction, and specifically relates to an EWT-Transformer prediction method applied to vortex-induced vibration response of a heat absorption tower structure. Background Art
[0002] As a typical high-flexible building, the heat absorption tower is sensitive to the long-term wind load and is prone to vortex-induced vibration. Therefore, the study of vortex-induced vibration of the heat absorption tower structure is a key and hot research topic in the field of structural wind engineering. Traditional numerical simulation methods, such as constructing mathematical equations of motion and complex finite element analysis (CFD), are often complex and require high computational costs. The parameters required to construct mathematical equations of motion are often dynamic in the field of structural wind engineering, which makes it difficult to have fixed parameters to choose from, which indirectly leads to the difficulty in converging the calculation results of the model. Complex finite element analysis, such as CFD, has extremely high requirements for the precise division of the grid and the definition of boundary conditions. At the same time, CFD has extremely high requirements for computing resources, and it often takes several days or even more time to obtain relatively good numerical simulation results.
[0003] The vigorous development of machine learning in the field of structural seismic resistance also provides a good reference for the introduction of machine learning in the field of structural wind engineering. The Long Short-Term Memory Networks (LSTM) model commonly used in the field of seismic resistance is considered to be an extremely excellent structural response prediction model. Relying on its powerful time series information memory ability, it is widely used in the field of structural seismic resistance. In recent years, LSTM has also begun to appear in the field of structural wind engineering, completing the prediction of traditional buildings. However, the prediction performance of LSTM is not very ideal when facing complex nonlinear responses. It can capture the changing trend of structural response, but it is difficult to accurately capture the peak value. Summary of the invention
[0004] In order to overcome the above technical problems, the present invention provides an EWT-Transformer prediction method for the vortex vibration response of a heat absorption tower structure, introduces the Transformer model that has just appeared in recent years into the vortex vibration prediction of the heat absorption tower structure, and adds empirical wavelet decomposition (EWT), and proposes a combined prediction model of EWT-Transformer. This model fills the research gap of heat absorption tower structure in the field of machine learning, while further improving the prediction performance of traditional machine learning models, and more accurately predicts the vortex vibration response of the heat absorption tower structure.
[0005] The present invention adopts the following technical solutions: The EWT-Transformer prediction method applied to the vortex-induced vibration response of the heat absorption tower structure is characterized by comprising: S1. Determine the reasonable input features, output features, and division of training and test sets used by the machine learning model; S2. Introducing signal decomposition technology, the complex nonlinear response signal is decomposed into N sub-signals similar to simple harmonic waves through empirical wavelet decomposition (EWT); S3. Introduce the Transformer model and simplify the Transformer model; S4. Divide all sub-signals into training sets and test sets, match the corresponding input features, make predictions through the simplified Transformer, and then superimpose the prediction results of all sub-signals to restore the prediction results of the original signal.
[0006] Preferably, the input characteristics are wind speed, wind direction and turbulence, and the output characteristics are the displacement response of the heat absorption tower structure.
[0007] Preferably, the first 70% of the data set is divided into a training set, and the remaining 30% is divided into a test set. The training set and the test set are independent of each other, and the prediction performance of the model is evaluated on the test set.
[0008] Preferably, the specific steps of S2 are: Perform Fourier transform on the signal to obtain the spectrum of the signal in the frequency domain; According to the local extreme points in the signal spectrum, the spectrum is divided into multiple frequency bands; Construct an adaptive filter for each frequency band, and then extract the frequency components of the signal in each frequency band; The signal is decomposed in the frequency domain through the constructed filter.
[0009] Preferably, the filter: Low frequency filter (filter for the first band): The low-frequency filter is used to extract the signal component in the lowest frequency band, and the corresponding interval is : Bandpass filter (filter in the middle frequency band): For each intermediate frequency band, construct a bandpass filter to extract the signal component of this frequency band: High frequency filter (filter of the last band): The high-frequency filter is used to extract the signal components in the highest frequency band, and the corresponding interval is : .
[0010] Preferably, the signal is decomposed in the frequency domain by constructing a filter, and the sub-signals of each frequency band are It can be obtained by: in, is the Fourier transform of the signal, is the filter corresponding to the frequency band, is the inverse Fourier transform.
[0011] Preferably, the simplified Transformer model consists of a position encoding layer, a multi-head attention mechanism layer, a normalization layer and a feedforward neural network.
[0012] Compared with the prior art, the present invention has the following beneficial effects: The present invention fills the gap in the research of heat absorption tower structures in the field of machine learning; it introduces the empirical wavelet decomposition (EWT) technology and fully combines it with the novel Transformer machine learning model in recent years, reduces the prediction error of a single model, improves the performance of the model in the face of complex nonlinear responses, and realizes the accurate prediction of vortex vibration of the heat absorption tower structure. BRIEF DESCRIPTION OF THE DRAWINGS
[0013] Figure 1 It is a flow chart of the EWT-Transformer prediction method of the vortex-oscillation response of the heat absorption tower structure of the present invention.
[0014] Figure 2 This is the internal structure diagram of Transformer.
[0015] Figure 3 This is the EWT signal decomposition diagram.
[0016] Figure 4 Transformer prediction results with a damping ratio of 0.7%.
[0017] Figure 5 EWT-Transformer prediction results at a damping ratio of 0.7%.
[0018] Figure 6 The root mean square error and mean absolute error histograms of Transformer and EWT-Transformer. DETAILED DESCRIPTION
[0019] The embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. Unless otherwise specified, the raw materials and equipment used can be purchased from the market or are commonly used in the art. The methods in the embodiments are conventional methods in the art unless otherwise specified. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.
[0020] like Figure 1-3 As shown, an EWT-Transformer prediction method for vortex-induced vibration response of a heat absorption tower structure includes the following steps: S1. Determine the reasonable input features, output features, and division of training and test sets used by the machine learning model The selection of input features in machine learning models is crucial, and the selection of input features should ensure that they are closely related to the output features. Therefore, the input features of this method are wind speed, wind direction and turbulence, and the output features are the displacement response of the heat absorption tower structure. At the same time, the division of training sets and test sets plays a key role in the prediction effect of machine learning. After weighing the accuracy of model prediction and the rationality of training time, the first 70% of the data set is divided into training sets, and the remaining 30% is divided into test sets to ensure that the training sets and test sets are independent of each other, and the prediction performance of the model is evaluated on the test set.
[0021] S2. Introduce signal decomposition technology to decompose the complex nonlinear response signal into N sub-signals similar to simple harmonic waves through empirical wavelet decomposition (EWT) First, perform Fourier transform on the signal to obtain the spectrum of the signal in the frequency domain. The spectrum shows the energy distribution of the signal at different frequencies.
[0022] (1) in, is the original signal, is its Fourier transform in the frequency domain, and f is the frequency.
[0023] According to the local extreme points in the signal spectrum, the spectrum is divided into multiple frequency bands. Assume that N boundary frequencies are found by analyzing the spectrum. , the frequency band is divided into , corresponding to the frequency range of each sub-signal.
[0024] In order to extract the frequency components of the signal in each frequency band, an adaptive filter is constructed for each frequency band. The design of empirical wavelet filters usually uses a set of window functions with compact support. Assume that the filter bank we construct is , these filters are constructed as follows: Low frequency filter (filter for the first band): The low-frequency filter is used to extract the signal component in the lowest frequency band, and the corresponding interval is : (2) Bandpass filter (filter in the middle frequency band): For each intermediate frequency band, construct a bandpass filter to extract the signal component of this frequency band: (3) High frequency filter (filter of the last band): The high-frequency filter is used to extract the signal components in the highest frequency band, and the corresponding interval is : (4) The signal is decomposed in the frequency domain through the constructed filter. The sub-signals in each frequency band It can be obtained by: (5) in, is the Fourier transform of the signal, is the filter corresponding to the frequency band, is the inverse Fourier transform.
[0025] After this decomposition, the original signal can be represented as the sum of all sub-signals: (6) S3. Introduce the Transformer model and simplify it to predict the sub-signals after the empirical wavelet decomposition (EWT) one by one The traditional Transformer model encoder-decoder architecture is simplified, and the core components in its sublayers are retained, so that the simplified Transformer model consists of a position encoding layer, a multi-head attention mechanism layer, a normalization layer, and a feedforward neural network.
[0026] Positional Encoding is to make the model understand the order of elements in the input sequence, because the standard Transformer is based on the self-attention mechanism, which is inherently disordered, that is, the position information in the input sequence is not directly captured. Positional encoding enables the model to distinguish different positions in the sequence through a fixed position encoding method based on sine and cosine functions.
[0027] (7) (8) Where: It is the location; It is a dimension; is the model dimension.
[0028] The attention mechanism is the core component of the multi-head attention mechanism in Transformer. The core idea is to allow the model to assign different attention weights according to the importance when processing each input, and dynamically focus on the part most relevant to the current processing task, rather than treating all inputs equally. The multi-head attention mechanism improves the defects of the single attention mechanism. By processing the input in parallel with multiple independent attention heads, the model can capture information in the input sequence from different angles. (9) (10) Where: Q is the query vector; K is the key vector; V is the value vector; Represents the dimension of the key vector. Specifically, the dot product of the query vector Q and the key vector K divided by Scaling is performed to prevent the value of the dot product result from being too large, causing the gradient of the softmax function to disappear or explode; , , , and is the weight matrix.
[0029] The feedforward layer further processes the data through linear changes and nonlinear activation functions to improve the generalization ability of the model and enable the model to better handle complex nonlinear relationships.
[0030] (11) Where: and is the weight matrix; and is the bias vector; ReLU is the nonlinear activation function.
[0031] S4. Divide all sub-signals into training sets and test sets, match the corresponding input features, make predictions through Transformer, then superimpose the prediction results of all sub-signals, restore the prediction results of the original signal, and evaluate the prediction results on the test set The original signal processed by EWT will be decomposed into N sub-signals similar to simple harmonic waves. These sub-signals will be divided into training sets and test sets, and the corresponding input features, such as wind speed, wind direction and turbulence, will be matched and input into Transformer. The association between input features and output features is learned on the training set, and predictions are made on the test set of unknown data. Finally, the prediction results of all sub-signals on the test set are superimposed to evaluate the final prediction performance of the model.
[0032] This embodiment takes the wind tunnel test of the aeroelastic model as an example and takes the displacement response as the target to demonstrate the accuracy of the EWT-Transformer model of the present invention.
[0033] The detailed steps of this embodiment are as follows: 1. Using wind tunnel tests, the wind speed, wind direction and turbulence in the wind tunnel are measured by the Cobra detector as input features, and the displacement response monitored by the displacement meter is used as the output feature. Match the relationship between the input features and the output features, and divide 70% into the training set and 30% into the test set.
[0034] 2. Introducing signal decomposition technology, the complex nonlinear response signal obtained from the wind tunnel test is decomposed into N sub-signals similar to simple harmonic waves through empirical wavelet decomposition (EWT); 3. Pass the divided sub-signals to the simplified Transformer model, learn on the training set, and demonstrate the prediction performance on unknown data on the test set.
[0035] 4. Superimpose the prediction results of each sub-signal to restore the prediction results of the original signal.
[0036] 5. Compare the restored prediction results on the test set with the response results detected by the original wind tunnel test, and use the root mean square error (RMSE) and mean absolute error (MAE) as evaluation indicators.
[0037] like Figure 4-Figure 6 As shown, the EWT-Transformer model proposed in the present invention can obtain high-precision prediction results in the face of complex nonlinear responses, and compared with the single Transformer model, both the root mean square error (RMSE) and the mean absolute error (MAE) have been significantly improved. Therefore, the wind tunnel test example verifies the high accuracy and versatility of the EWT-Transformer. This reduces the problem of insufficient peak capture of the heat absorption tower structure caused by traditional models in actual engineering.
[0038] Although the embodiments of the present invention have been shown and described, those skilled in the art will appreciate that various changes, modifications, substitutions and variations may be made to the above embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.
Claims
1. The EWT-Transformer prediction method applied to the vortex-induced vibration response of the heat absorption tower structure is characterized by: include: S1. Determine the reasonable input features, output features, and division of training and test sets used by the machine learning model; S2. Introducing signal decomposition technology, the complex nonlinear response signal is decomposed into N sub-signals similar to simple harmonic waves through empirical wavelet decomposition (EWT); S3. Introduce the Transformer model and simplify the Transformer model; S4. Divide all sub-signals into training sets and test sets, match the corresponding input features, make predictions through the simplified Transformer, and then superimpose the prediction results of all sub-signals to restore the prediction results of the original signal.
2. The EWT-Transformer prediction method for vortex-induced vibration response of a heat absorbing tower structure according to claim 1 is characterized in that: The input features are wind speed, wind direction and turbulence, and the output features are the displacement response of the heat absorption tower structure.
3. The EWT-Transformer prediction method for vortex-induced vibration response of a heat absorbing tower structure according to claim 1 is characterized in that: The first 70% of the data set is divided into a training set, and the remaining 30% is divided into a test set. The training set and the test set are independent of each other, and the prediction performance of the model is evaluated on the test set.
4. The EWT-Transformer prediction method for vortex-induced vibration response of a heat absorbing tower structure according to claim 1 is characterized in that: The specific steps of S2 are: Perform Fourier transform on the signal to obtain the spectrum of the signal in the frequency domain; According to the local extreme points in the signal spectrum, the spectrum is divided into multiple frequency bands; Construct an adaptive filter for each frequency band, and then extract the frequency components of the signal in each frequency band; The signal is decomposed in the frequency domain through the constructed filter.
5. The EWT-Transformer prediction method for vortex-induced vibration response of a heat absorbing tower structure according to claim 4 is characterized in that: filter: Low frequency filter (filter for the first band): The low-frequency filter is used to extract the signal component in the lowest frequency band, and the corresponding interval is :
6. Bandpass filter (filter in the middle frequency band): For each intermediate frequency band, construct a bandpass filter to extract the signal component of this frequency band:
7. High frequency filter (filter of the last frequency band): The high-frequency filter is used to extract the signal components in the highest frequency band, and the corresponding interval is : 。 8. The EWT-Transformer prediction method for vortex-induced vibration response of a heat absorbing tower structure according to claim 5, characterized in that: Through the constructed filter, the signal is decomposed in the frequency domain, and the sub-signals in each frequency band It can be obtained by:
9. Among them, is the Fourier transform of the signal, is the filter corresponding to the frequency band, is the inverse Fourier transform.
10. The EWT-Transformer prediction method for vortex-induced vibration response of a heat absorbing tower structure according to claim 1, characterized in that: The simplified Transformer model consists of a position encoding layer, a multi-head attention mechanism layer, a normalization layer, and a feedforward neural network.