Buoy measurement wind speed correction method and device based on deep learning, and electronic equipment

Abstract

The invention discloses a buoy measurement wind speed correction method and device based on deep learning, and electronic equipment. The method comprises the steps of obtaining buoy measurement wind speed data, buoy attitude angle data and buoy angular velocity data; performing data enhancement on the wind speed data; preprocessing the wind speed data, the attitude angle data, the angular speed data and the enhanced wind speed data; the preprocessed data are input into a deep learning model, a wind speed correction value is output, the deep learning model comprises a sub-model 1, a sub-model 2 and an output layer, and particularly, the preprocessed wind speed data, the enhanced wind speed data and yaw angle data in the attitude angle data are input into the sub-model 1 for feature extraction; inputting other preprocessed data into the sub-model 2 for feature extraction; and finally, combining the outputs of the two sub-models at an output layer to obtain a final correction result. The method can effectively correct the wind speed measured by the buoy, thereby improving the precision of wind speed measurement.

Description

Technical Field

[0001] The present application relates to the technical field of buoy-measured wind speed correction, and particularly relates to a buoy-measured wind speed correction method and device based on deep learning, and an electronic device. Background Art

[0002] Buoy plays an important role in aspects such as marine environmental monitoring, early warning and forecasting, resource development, maritime traffic, and military activity support, and is one of the important means to obtain offshore marine information. In the observation of marine environmental information such as hydrology, meteorology, and ecology, wind is an important observation element. Wind speed measurement is of great significance for navigation safety, meteorological observation, wind resource assessment, etc. However, most of the currently measured wind speed data are average values over a period of time, which are difficult to meet the requirements for accurate wind data in applications such as sea-air flux calculation and offshore wind turbine design. Therefore, how to obtain high-precision instantaneous wind data has become an urgent problem to be solved.

[0003] When measuring wind speed with a buoy, the measurement process will be affected by the movement of the buoy, resulting in measurement errors. Therefore, it is necessary to establish a model to correct the wind speed measured by the buoy. The current wind speed correction methods mainly target ultrasonic wind sensors and correct the wind speed through formulas, lacking a high-precision correction method for mechanical wind sensors. Summary of the Invention

[0004] The purpose of the embodiments of the present application is to provide a buoy-measured wind speed correction method and device based on deep learning, and an electronic device, so as to solve the problems of poor correction effect and lack of a high-precision correction method for mechanical wind sensors in the related art.

[0005] According to the first aspect of the embodiments of the present application, a buoy-measured wind speed correction method based on deep learning is provided, including: Obtaining the wind speed data measured by the buoy, the attitude angle data of the buoy, and the angular velocity data of the buoy; Performing data augmentation on the wind speed data; Preprocessing the wind speed data, the attitude angle data, the angular velocity data, and the wind speed data after data augmentation; Inputting the preprocessed data into a deep learning model to output a wind speed correction value, where the deep learning model includes sub-model 1, sub-model 2, and an output layer, and: Inputting the yaw angle data in the preprocessed wind speed data, the wind speed data after data augmentation, and the attitude angle data into sub-model 1 for feature extraction. Sub-model 1 first processes the input data with 1DCNN to obtain high-dimensional features; then, an improved LSTM network fused with wavelet decomposition and an attention mechanism are used to extract the relationships between variables and the relationships in the time dimension; Input the pre - processed data into Sub - model 2 for feature extraction. The Sub - model 2 adopts a frequency attention mechanism to extract periodic features; Finally, merge the outputs of the two sub - models in the output layer to obtain the final corrected result.

[0006] According to the second aspect of the embodiments of the present application, a buoy - based wind speed correction device for deep learning is provided, including: An acquisition module, configured to acquire the wind speed data measured by the buoy, the attitude angle data of the buoy, and the angular velocity data of the buoy; A data augmentation module, configured to perform data augmentation on the wind speed data; A pre - processing module, configured to pre - process the wind speed data, the attitude angle data, the angular velocity data, and the wind speed data after data augmentation; A correction module, configured to input the pre - processed data into a deep learning model and output a wind speed correction value. The deep learning model includes Sub - model 1, Sub - model 2, and an output layer, where: Input the pre - processed wind speed data, the wind speed data after data augmentation, and the yaw angle data in the attitude angle data into Sub - model 1 for feature extraction. The Sub - model 1 first processes the input data with 1DCNN to obtain high - dimensional features; then, uses an improved LSTM network fused with wavelet decomposition and an attention mechanism to extract the relationships between variables and the relationships in the time dimension; Input the roll angle and pitch angle data in the pre - processed attitude angle data and the angular velocity data into Sub - model 2 for feature extraction. The Sub - model 2 adopts a frequency attention mechanism to extract periodic features; Finally, merge the outputs of the sub - models in the output layer to obtain the final corrected result.

[0007] According to the third aspect of the embodiments of the present application, an electronic device is provided, including: One or more processors; A memory, configured to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in the first aspect.

[0008] According to the fourth aspect of the embodiments of the present application, a computer - readable storage medium is provided, on which computer instructions are stored, and when the instructions are executed by a processor, the steps of the method as described in the first aspect are implemented.

[0009] The technical solutions provided by the embodiments of the present application may include the following beneficial effects: As can be seen from the above embodiments, the embodiments of the present application propose a method for correcting the measured wind speed of a buoy based on deep learning, effectively overcoming the problems of poor correction effect and lack of a high-precision correction method for mechanical wind sensors in the existing correction methods.

[0010] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] The accompanying drawings herein are incorporated into the specification and form a part of the specification, showing embodiments consistent with the present application and, together with the specification, are used to explain the principles of the present application.

[0012] Figure 1 is a flowchart of a method for correcting the measured wind speed of a buoy based on deep learning shown according to an exemplary embodiment.

[0013] Figure 2 is a structural diagram of an LSTM network integrated with wavelet decomposition shown according to an exemplary embodiment.

[0014] Figure 3 is a structural diagram of frequency attention shown according to an exemplary embodiment.

[0015] Figure 4 is a block diagram of a device for correcting the measured wind speed of a buoy based on deep learning shown according to an exemplary embodiment. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0016] Here, the exemplary embodiments will be described in detail, and the examples are shown in the accompanying drawings. When the following description refers to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present application. On the contrary, they are only examples of devices and methods consistent with some aspects of the present application as detailed in the appended claims.

[0017] The terms used in the present application are only for the purpose of describing specific embodiments and are not intended to limit the present application. The singular forms "a", "the", and "said" used in the present application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0018] Figure 1 is a flowchart of a method for correcting the measured wind speed of a buoy based on deep learning shown according to an exemplary embodiment. As Figure 1 shown, the method may include the following steps: S1: Obtain the wind speed data measured by the buoy, the attitude angle data of the buoy, and the angular velocity data of the buoy; Specifically, use the anemometer on the buoy to obtain the wind speed data , use the combined attitude measurement system to obtain the roll angle , pitch angle , yaw angle , calculate the roll angular velocity according to the forward difference , pitch angular velocity , yaw angular velocity , and construct a data set. This example uses the data collected on a certain day in summer in a certain sea area, selects 7,000 of them, and the sampling interval is uniformly 1 second.

[0019] S2: Perform data augmentation on the wind speed data; it may include the following sub-steps: S21: Perform difference calculation on the wind speed data to obtain the wind speed data after difference; Specifically, adopt the backward difference method, and the calculation formula is as follows: ; In the formula, is the wind speed data after difference at time , is the wind speed at time , is the wind speed at time .

[0020] S22: Perform moving average calculation on the wind speed data to obtain the wind speed data after moving average; Specifically, adopt a moving average with a window size of 10, and the calculation formula is as follows: ; In the formula, is the wind speed data after moving average at time , is the wind speed at time .

[0021] S23: Perform a fast Fourier transform on the wind speed data, retain the part less than the frequency threshold, and finally perform an inverse fast Fourier transform to obtain the denoised wind speed data ; Specifically, first perform a sliding window with a step size of 1 on the wind speed data, then within each window, perform a fast Fourier transform on the wind speed data, retain the part less than the frequency threshold, and finally perform an inverse fast Fourier transform to obtain the denoised result .

[0022] Through the above feature enhancement operation, the original wind speed data is transformed and augmented, which can improve the accuracy of the model's correction results and enable the model to capture more potential patterns and relationships in the data.

[0023] S3: Preprocess the wind speed data, attitude angle data, angular velocity data, and the augmented wind speed data; it may include the following sub-steps: S31: Normalize the data to be measured and the augmented wind speed data; Specifically, perform Z-Score normalization, and the calculation formula is as follows: ; In the formula, is the normalized data, is the sample data, is the mean of the corresponding sample data, is the standard deviation of the corresponding sample data.

[0024] S32: Perform a sliding window operation on the normalized data; Specifically, in the embodiment of the present invention, the window size is selected as 64, and after performing the sliding window operation, three-dimensional data is obtained as the model input.

[0025] S4: Input the preprocessed data into the deep learning model and output the wind speed correction value; it may include the following sub-steps: S41: Input the preprocessed wind speed data, the augmented wind speed data, and the yaw angle data ( 、 、 、 、 ) in the attitude angle data into sub-model 1 for feature extraction; Specifically, the sub-model first processes the input data with 1DCNN to obtain high-dimensional features; then, uses an LSTM network fused with wavelet decomposition and an attention mechanism to extract the relationships between variables and the relationships in the time dimension.

[0026] The structure of the LSTM network fused with wavelet decomposition is as Figure 2 shown. Compared with the ordinary LSTM network, the LSTM network fused with wavelet decomposition performs wavelet decomposition on the cell state at time t-1, and adds a fully connected layer to extract time-frequency domain features. The calculation formula is as follows: ; ; ; ; ; ; ; ; wherein 、 、 、 are respectively the forget gate, input gate, output gate, cell state, and input state of the memory unit at the 、 、 、 the weight matrices of the forget gate, input gate, output gate, and cell state respectively, 、 、 、 the bias terms of the forget gate, input gate, output gate, and cell state respectively, 、 are the weight matrix and bias term of the first fully connected layer, 、 are the weight matrix and bias term of the second fully connected layer, is the wavelet transform matrix, is the input at the is the cell state at time t-1 after wavelet decomposition and the fully connected layer, is the output of the hidden layer at time t, is the output of the hidden layer at time t-1, is the sigmoid function, and tanh is the hyperbolic tangent function.

[0027] The attention mechanism includes feature attention and temporal attention. The feature attention calculates the similarity between the input at each time step and the previous hidden state to obtain the weighted feature. The calculation formula is as follows: ; ; ; In the formula, , , and are the parameters updated during backpropagation, tanh is the hyperbolic tangent function, is the output of the hidden layer at time -1, is the similarity between the feature and , is the weight, is the input at time is the weighted feature, is the Hadamard product. Combine the feature attention with the improved LSTM network, and use as the input of the improved LSTM network.

[0028] The time attention is applied after obtaining the set of hidden states at all time steps, and the hidden states are weighted and summed. The calculation formula is as follows: ; ; ; In the formula , are the weights and bias terms updated during backpropagation, is the output of the hidden layer at time is the attention score, is the weight, is the output result of the weighted hidden layer.

[0029] The sub - model 1 uses 1DCNN to obtain high - dimensional features, combines the LSTM network to process temporal dependencies, uses wavelet decomposition to enhance the frequency - domain feature expression, and then fuses the feature attention and time attention mechanisms, thus realizing the effective extraction of multi - scale features and highlighting key information.

[0030] S42: Input the roll angle, pitch angle data, and angular velocity data ( , , , , ) in the attitude angle data after data pre - processing into the sub - model 2 for feature extraction; Specifically, considering that these features have obvious periodicity and change rapidly, a frequency attention mechanism is used to extract features, such as Figure 3As shown below. First, each feature is mapped to a high-dimensional space using linear embedding, and the calculation formula is as follows: ; where is the input feature, , are the weights and bias terms updated during backpropagation, is the result after embedding.

[0031] Then, a linear transformation is used to obtain matrices , , , and the calculation formula is as follows: ; where , are the weight matrices of , , , and , are the bias terms.

[0032] Perform a fast Fourier transform on , , to obtain the real and imaginary parts of the result, and the calculation formula is as follows: ; Then, perform multi-head attention on the real part results , , , and the imaginary part results , , , respectively, and the calculation formula is as follows: ; ; ; ; where , are the attention outputs of the real and imaginary parts respectively, is the scaling factor, softmax is the normalization exponential function, , are the linear transformation matrices of the real and imaginary parts respectively, concat represents concatenation, , are the final outputs of the real and imaginary parts.

[0033] Finally, merge the real and imaginary part results to obtain the final output, and the calculation formula is as follows: ​​ ; The sub-model 2 utilizes the frequency-domain attention mechanism to achieve the ability to capture long-range dependencies and periodic patterns in the frequency domain space, thereby enhancing the model's perception and expression ability for complex frequency features.

[0034] S43: Combine the outputs of the two sub-models at the output layer to obtain the corrected wind speed value; Specifically, for the output of sub-model 1, further process it with a fully connected layer; for the output of sub-model 2, reduce the dimension with average pooling. Pass the final results of the two parts through a fully connected layer with 1 neuron to obtain the final corrected result.

[0035] After the above deep learning model is trained, wind speed correction can be performed.

[0036] In this embodiment, two statistical indicators, the root mean square error RMSE (Root Mean Square Error) and the mean absolute error MAE (Mean Absolute Error), are used to evaluate the correction effect of the model. Their calculation formulas are as follows: ; ; In the formula, and are the true value and the corrected value of the data respectively.

[0037] Compare the LSTM network with the effect of the buoy-measured wind speed correction method based on deep learning in this embodiment of the present invention. The results are shown in the following table: It can be seen that the buoy-measured wind speed correction method based on deep learning in this embodiment of the present invention is superior to the LSTM network in terms of accuracy, and the error after correction is significantly reduced.

[0038] As can be seen from the above embodiments, the deep learning model proposed in this application effectively overcomes the problems of poor correction effect and lack of a high-precision correction method for mechanical wind sensors in the existing correction methods.

[0039] Corresponding to the foregoing embodiment of the buoy-measured wind speed correction method based on deep learning, this application also provides an embodiment of a buoy-measured wind speed correction device based on deep learning.

[0040] Figure 4 is a block diagram of a buoy-measured wind speed correction device based on deep learning shown according to an exemplary embodiment. Referring to Figure 4 , the device includes: Acquisition module 1, configured to acquire wind speed data measured by a buoy, attitude angle data of the buoy, and angular velocity data of the buoy; Data enhancement module 2, configured to perform data enhancement on the wind speed data; Preprocessing module 3, configured to preprocess the wind speed data, attitude angle data, angular velocity data, and the data-enhanced wind speed data; Correction module 4, configured to input the preprocessed data into a deep learning model, and output a wind speed correction value. The deep learning model includes sub-model 1, sub-model 2, and an output layer, where: Input the preprocessed wind speed data, the data-enhanced wind speed data, and the yaw angle data in the attitude angle data into sub-model 1 for feature extraction. Sub-model 1 first processes the input data using 1DCNN to obtain high-dimensional features; then, an improved LSTM network integrated with wavelet decomposition and an attention mechanism is used to extract the relationships between variables and the relationships in the time dimension; Input the roll angle and pitch angle data in the preprocessed attitude angle data and the angular velocity data into sub-model 2 for feature extraction. Sub-model 2 adopts a frequency attention mechanism to extract periodic features; Finally, merge the outputs of the two sub-models in the output layer to obtain the final correction result.

[0041] Regarding the device in the above embodiments, the specific manners in which each module performs operations have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0042] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can refer to the partial descriptions of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place, or may be distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of the present application. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0043] Correspondingly, the present application further provides an electronic device, including: one or more processors; a memory, configured to store one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the method for correcting the wind speed measured by a buoy based on deep learning as described above.

[0044] Correspondingly, the present application also provides a computer-readable storage medium, on which computer instructions are stored, and when the instructions are executed by a processor, the above-mentioned method for correcting the wind speed measured by the buoy based on deep learning is implemented.

[0045] After considering the specification and practicing the content disclosed herein, those skilled in the art will readily conceive of other embodiments of the present application. The present application is intended to cover any variations, uses, or adaptations of the present application, which follow the general principles of the present application and include well-known common general knowledge or conventional technical means in the technical field not disclosed in the present application. The specification and examples are only regarded as exemplary, and the true scope and spirit of the present application are pointed out by the claims.

[0046] It should be understood that the present application is not limited to the exact structures described above and shown in the drawings, and various modifications and changes can be made without departing from its scope. The scope of the present application is only limited by the appended claims.

Claims

1. A buoy measurement wind speed correction method based on deep learning, characterized in that: include: Obtain wind speed data measured by the buoy, buoy attitude angle data, and buoy angular velocity data; Performing data enhancement on the wind speed data; Preprocessing the wind speed data, attitude angle data, angular velocity data, and data-enhanced wind speed data; The preprocessed data is input into a deep learning model to output a wind speed correction value. The deep learning model includes sub-model 1, sub-model 2 and an output layer, wherein: The preprocessed wind speed data, the data-enhanced wind speed data, and the yaw angle data in the attitude angle data are input into the sub-model 1 for feature extraction. The sub-model 1 first processes the input data with 1DCNN to obtain high-dimensional features; then, the LSTM network integrated with wavelet decomposition and the attention mechanism are used to extract the relationship between variables and the relationship in the time dimension; The roll angle, pitch angle data and angular velocity data in the preprocessed attitude angle data are input into sub-model 2 for feature extraction, and the sub-model 2 adopts a frequency attention mechanism to extract periodic features; Finally, the outputs of the two sub-models are merged at the output layer to obtain the final correction result.

2. The method according to claim 1, characterized in that: Performing data enhancement on the wind speed data includes: (1) Perform differential calculation on the wind speed data to obtain differential wind speed data. The calculation formula is as follows: ； In the formula, for Wind speed data after time difference, for The wind speed at the moment, for Wind speed at the moment; (2) Perform moving average calculation on the wind speed data to obtain the moving average wind speed data. The calculation formula is as follows: ； In the formula, for The wind speed data after moving average at all times, for The wind speed at the moment, is the window size; (3) Perform fast Fourier transform on the wind speed data, retain the part less than the frequency threshold, and finally perform inverse fast Fourier transform to obtain the denoised wind speed data .

3. The method according to claim 1, characterized in that: The LSTM network integrated with wavelet decomposition is -1 cell status at time Perform wavelet decomposition and add a fully connected layer to extract time-frequency domain features. The calculation formula is as follows: ； ； ； ； ； ； ； In the formula , , , They are The forget gate, input gate, output gate, cell state and input state of the memory unit at each moment. 、 、 、 are the weight matrices of the forget gate, input gate, output gate and cell state, respectively. 、 、 、 are the bias items of forget gate, input gate, output gate and cell state respectively, , is the weight matrix and bias term of the first fully connected layer, , is the weight matrix and bias term of the second fully connected layer, is the wavelet transform matrix, for Input at the moment, for -1 moment cell status The result after wavelet decomposition and full connection layer, for The hidden layer output at the moment, for -1 hidden layer output, is the sigmoid function, and tanh is the hyperbolic tangent function.

4. The method according to claim 1, characterized in that: The attention mechanism includes feature attention and time attention, where: Through the feature attention, the weighted features are obtained, and the calculation formula is as follows: ； ； ； In the formula, , , and is the parameter updated during back propagation, tanh is the hyperbolic tangent function, for -1 hidden layer output, Characterized by The similarity of is the weight, for Input at the moment, is the weighted feature, It is Hadamard; The temporal attention is weighted summed over the hidden states, and the calculation formula is as follows: ； ； ； In the formula , are the weights and biases updated during back propagation, for The hidden layer output at the moment, Score for attention, is the weight, Output result of the hidden layer after weighting.

5. The method according to claim 1, characterized in that: The sub-model 2 adopts the frequency attention mechanism to extract periodic features, which specifically includes: first, using linear embedding to map each feature to a high dimension, and then using linear transformation to obtain the matrix , , ,right , , Perform fast Fourier transform to get the real and imaginary results, then perform multi-head attention separately, and finally merge the real and imaginary results to get the final output.

6. The method according to claim 1, characterized in that The outputs of the two sub-models are merged at the output layer to obtain the final correction result, including: For the output of sub-model 1, a fully connected layer is used for further processing; For the output of sub-model 2, average pooling is used to reduce the dimension; Finally, after passing through the fully connected layer, the final correction result is obtained.

7. A buoy wind speed correction device based on deep learning, characterized in that: include: An acquisition module is used to acquire wind speed data measured by the buoy, attitude angle data of the buoy, and angular velocity data of the buoy; A data enhancement module, used for performing data enhancement on the wind speed data; A preprocessing module, used for preprocessing the wind speed data, attitude angle data, angular velocity data and data-enhanced wind speed data; A correction module is used to input the preprocessed data into a deep learning model and output a wind speed correction value. The deep learning model includes a sub-model 1, a sub-model 2 and an output layer, wherein: The preprocessed wind speed data, the data-enhanced wind speed data, and the yaw angle data in the attitude angle data are input into the sub-model 1 for feature extraction. The sub-model 1 first processes the input data with 1DCNN to obtain high-dimensional features; then, the improved LSTM network integrated with wavelet decomposition and the attention mechanism are used to extract the relationship between variables and the relationship in the time dimension; The roll angle, pitch angle data and angular velocity data in the preprocessed attitude angle data are input into the sub-model 2 for feature extraction. The sub-model 2 adopts a frequency attention mechanism to extract periodic features; Finally, the outputs of the two sub-models are merged at the output layer to obtain the final correction result.

8. An electronic device, characterized in that: include: one or more processors; A memory for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer instructions stored thereon, characterized in that: When the instruction is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images