Ultra-short-term wind speed prediction method based on attention mechanism and orthogonal wind meter used for method
By using an attention-based ultra-short-term wind speed prediction method and a multi-rotor UAV orthogonal anemometer, the problems of low flexibility and prediction efficiency of existing equipment are solved, and efficient and accurate wind speed prediction is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-12
- Publication Date
- 2026-03-17
AI Technical Summary
Existing wind speed measurement equipment suffers from problems such as low flexibility, cost, and low spatial resolution. Furthermore, the decomposition process of commonly used prediction methods is cumbersome, resulting in low prediction efficiency.
An ultra-short-term wind speed prediction method based on an attention mechanism is adopted. The wind speed time series is decomposed using the Time2Vec time embedding algorithm, and the WDCNN-BiGRU-Attention model is used for training and prediction. Wind speed data is collected by using a multi-rotor UAV equipped with an orthogonal anemometer.
It improves the accuracy and efficiency of wind speed forecasting, reduces the impact of noise, and enhances the reliability and accuracy of forecasts.
Smart Images
Figure CN121679752A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to a wind speed prediction method, in particular to an ultra-short-term wind speed prediction method based on an attention mechanism and an orthogonal anemometer used for the method. BACKGROUND
[0002] Traditional wind speed observation equipment such as wind measuring masts, wind measuring towers and wind profile radars have the disadvantages of low flexibility, high cost and low spatial resolution. Multi-rotor unmanned aerial vehicles have the advantages of vertical take-off and landing, fixed-point hovering and convenient operation and are applied to wind field observation to provide effective data sources for wind speed prediction.
[0003] At present, commonly used anemometers mainly include mechanical anemometers, hot-wire anemometers and ultrasonic anemometers. However, each of them has some technical deficiencies. The mechanical anemometer adopts a rotating mechanical structure, and the rotating shaft is prone to wear during long-term use, thereby affecting the measurement accuracy. In addition, the mechanical anemometer is affected by the static friction force of the rotating shaft contact surface, and has a start-up wind speed problem, resulting in poor measurement accuracy in low wind speed environments. The heating wire of the hot-wire anemometer is relatively fragile and has poor reliability, so it is difficult to be widely used in field conditions. Although the ultrasonic anemometer has high accuracy, it has high cost, and the surface of the ultrasonic transducer is easily damaged, which can cause sound scattering due to scratches or bumps, thereby causing measurement errors. At the same time, the ultrasonic anemometer also has a shadow effect, and the measurement error is large at a specific angle.
[0004] At present, artificial intelligence methods are commonly used to predict wind speed. Empirical mode decomposition (EMD), variational mode decomposition (VMD) and singular spectrum analysis (SSA) are used to decompose wind speed measurement values, and then different subsequences are extracted or directly input into a deep learning model. However, the complicated decomposition process reduces the prediction efficiency. SUMMARY
[0005] The technical problem to be solved by the present application is to provide an efficient and accurate ultra-short-term wind speed prediction method based on an attention mechanism.
[0006] To solve the above technical problems, the technical solution adopted by the present application is: an ultra-short-term wind speed prediction method based on an attention mechanism, comprising the following specific steps:
[0007] Step S1: sampling wind speed data at a specified height of a to-be-measured location according to a predetermined sampling frequency;
[0008] Step S2: constructing an original wind speed data set including a wind speed time sequence according to the wind speed sampling data, the wind speed time sequence being used to represent wind speed-time information;
[0009] Step S3: Use the Time2Vec time embedding algorithm to decompose the wind speed time series in the original wind speed dataset to obtain the representations of non-periodic and periodic patterns.
[0010] The decomposed aperiodic and periodic components are concatenated with the original wind speed time series to obtain a new wind speed dataset.
[0011] Step S4: Divide the new wind speed dataset into a training set and a test set;
[0012] Step S5: Input the training sets of the two components into the WDCNN-BiGRU-Attention model for training, and use the grid search method to optimize the parameters and determine the hyperparameters of the model.
[0013] Step S6: After training, save the WDCNN-BiGRU-Attention model and its optimal hyperparameters.
[0014] Step S7: Input the test set into the WDCNN-BiGRU-Attention model for testing and performance evaluation, verify the superiority of the hyperparameters, and obtain the optimal prediction model;
[0015] The WDCNN-BiGRU-Attention model with the highest hyperparameters is used as the trained model.
[0016] Step S8: The collected real-time wind speed data is decomposed into non-periodic and periodic pattern representations using the Time2Vec time embedding algorithm; then, it is input into the trained WDCNN-BiGRU-Attention model to obtain the prediction results.
[0017] As a preferred embodiment, the expression for the time embedding algorithm in step S3 is:
[0018]
[0019] For any time series τ, its corresponding Time2Vec is represented as a vector of length k+1; in this vector, represents the i-th feature; the sum of the trainable parameters in the model constitutes a linear function used to extract the non-periodic features of the time series; and the function is used to capture the periodic features.
[0020] As a preferred embodiment, the WDCNN-BiGRU-Attention model includes: a wide-depth convolutional neural network WDCNN module, a multi-layer BiGRU module, and an Attention module.
[0021] The WDCNN module mainly includes:
[0022] The first layer is a convolutional layer, which uses a 64×1 convolutional kernel and the activation function is ReLU;
[0023] The second layer is a pooling layer, which uses 2×2 max pooling;
[0024] The third layer is a convolutional layer, which uses a 3×1 convolutional kernel and the activation function is ReLU;
[0025] The fourth layer is the pooling layer, which uses 2×2 max pooling;
[0026] The structure then consists of three convolutional layers and three pooling layers arranged alternately, with the convolutional kernels of the convolutional layers all being 3×1 and the pooling kernels of the pooling layers being 2×2.
[0027] The last layer is the output layer, and the softmax function is used as the activation function for the output layer.
[0028] As a preferred embodiment, the multi-layer BiGRU module is a 3-layer BiGRU module, including 3 layers of BiGRU and 3 layers of Dropout. The spatiotemporal correlation features within the time window are extracted through the operation of the 3-layer BiGRU module, and the dimensionality of subsequent processing is reduced.
[0029] BiGRU adopts a bidirectional structure, using two gated recurrent network models GRU to extract the temporal relevance features of historical information from the historical end and the future end, respectively.
[0030] The formula for resetting the door is expressed as follows:
[0031] R t =σ(X) t W XR +H t-1 W HR +b R (2)
[0032] In the formula, X t W represents the input data at time step t. XR X represents t Reset gate weight parameters, H t-1 W represents the hidden state at time step t-1, which indicates the spatiotemporal correlation characteristics of the input data. HR H represents t-1 Reset gate weight parameters, b R This represents the bias parameter of the reset gate, and σ represents the sigmoid activation function;
[0033] The update gate is represented as follows:
[0034] Z t =σ(X) t W XZ +H t-1 W HZ +b Z (3)
[0035] In the formula, W XZ X represents t Update the gate weight parameters, W HZ H represents t-1 Update the gate weight parameters, b Z This indicates that the gate's bias parameters are being updated;
[0036]
[0037] In the formula, W XH W HH They represent X respectively t The weight parameters and H t-1 The weight parameter, b H The bias parameter for the candidate hidden state is tanh. The value of the candidate hidden state is guaranteed to be between -1 and 1 by using the nonlinear activation function tanh. ⊙ represents the Hadamard product.
[0038]
[0039] In the formula, Z t Represents the update gate at time step t;
[0040] The expression for the BiGRU model is:
[0041]
[0042] In the formula, X t Let H be the wind speed input at time t. t Let be the hidden layer state at time t; This represents the output of the forward hidden layer at time t. This represents the output of the reverse hidden layer at time t. and Let b represent the weights corresponding to the forward and backward hidden layer states of the BiGRU at time t, respectively; t This represents the bias corresponding to the hidden layer state at time t.
[0043] As a preferred embodiment, the expression for the attention mechanism module is:
[0044] e t =V e tanh(W W H t +bW (9)
[0045]
[0046] H t V represents the output of the BiGRU hidden layer, and V represents the attention probability distribution from the attention mechanism output to the BiGRU hidden layer. e and W W This represents the weighting coefficient.
[0047] As a preferred approach, the WDCNN-BiGRU-Attention model is trained using the Adam optimizer with a learning rate of 0.001, and the parameters are tuned. The loss function is the root mean square error (RMSE), and the formula is:
[0048]
[0049] In the formula, n represents the number of samples, i represents the sample number, Y represents the true label of the sample, Y' represents the predicted value, and the smaller the RMSE value, the better the prediction result.
[0050] Another technical problem to be solved by the present invention is to provide an orthogonal anemometer based on a multi-rotor UAV for an ultra-short-term wind speed prediction method based on an attention mechanism.
[0051] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: an orthogonal anemometer based on a multi-rotor drone, comprising a multi-rotor drone and an anemometer body mounted on the multi-rotor drone. The anemometer body is provided with a vertically erected connecting rod mechanism of adjustable length. The connecting rod mechanism is provided with four orthogonally arranged Pitot tubes serving as air guides. The length of the connecting rod mechanism is such that the Pitot tubes are located outside the wing turbulence area of the multi-rotor drone. The ends of the four Pitot tubes are respectively connected to four micro-pressure sensors. The main control circuit is electrically connected to each micro-pressure sensor. The main control circuit calculates the wind speed data through the wind pressure data measured by each micro-pressure sensor and the pitch and roll angle data of the multi-rotor drone.
[0052] The beneficial effects of this invention are:
[0053] This wind speed prediction method uses Time2Vec to decompose the wind speed time series to reduce the complexity and variability of the original wind speed time series. The linear and periodic representations obtained after decomposition are concatenated with the original wind speed data to obtain a new wind speed dataset. The new wind speed dataset is then input into a trained WDCNN-BiGRU-Attention model for prediction. WDCNN is used to obtain a wider receptive field while reducing the influence of high-frequency noise. After passing through BiGRU, the temporal features of each operating condition are discarded according to the importance of the information, followed by feature extraction. These time series features are influenced by historical and future data, and the extracted operating condition features are reprocessed through an attention mechanism. The attention mechanism module assigns different weights to hidden layer units according to the importance of feature information. The combination of BiGRU and Attention effectively extracts the dependence of historical information, improving the accuracy, efficiency, and reliability of wind speed prediction. Attached Figure Description
[0054] Figure 1 A flowchart illustrating the wind speed prediction method based on a combined model provided in an embodiment of the present invention;
[0055] Figure 2 A schematic diagram of the Time2Vec time embedding algorithm provided for a real-time example of the present invention;
[0056] Figure 3 This is a schematic diagram of a bidirectional gated loop unit.
[0057] Figure 4 A flowchart for wind speed prediction;
[0058] Figure 5 A schematic diagram showing the comparison between the predicted wind speed values provided in the embodiments of the present invention and the actual values measured in the embodiments;
[0059] Figure 6 A schematic diagram of the pitot tube arrangement of a wind gauge based on a multi-rotor UAV provided in an embodiment of the present invention;
[0060] Figure 7 This invention relates to the XZ plane velocity contour map in Fluent software for the DJI M300RTK facing the wind. Detailed Implementation
[0061] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0062] like Figures 1-5 As shown, an ultra-short-term wind speed prediction method based on an attention mechanism includes the following steps:
[0063] Step S1: Sample the wind speed data at a specified height at the location to be measured according to the predetermined sampling frequency;
[0064] Step S2: Based on the wind speed sampling data, construct an original wind speed dataset including a wind speed time series. The wind speed time series is used to characterize wind speed-time information.
[0065] Step S3: Use the Time2Vec time embedding algorithm to decompose the wind speed time series in the original wind speed dataset to obtain the representations of non-periodic and periodic patterns.
[0066] The expression for the time embedding algorithm is:
[0067]
[0068] For any time series τ, its corresponding Time2Vec is represented as a vector of length k+1; in this vector, represents the i-th feature; the sum of the trainable parameters in the model constitutes a linear function used to extract the non-periodic features of the time series; and the function is used to capture the periodic features.
[0069] The decomposed aperiodic and periodic components are concatenated with the original wind speed time series to obtain a new wind speed dataset.
[0070] Step S4: Divide the new wind speed dataset into a training set and a test set;
[0071] Step S5: Input the training set of each modal component into the WDCNN-BiGRU-Attention model for training, and use the grid search method to optimize the parameters and determine the hyperparameters of the model.
[0072] The WDCNN-BiGRU-Attention model includes: a wide-depth convolutional neural network (WDCNN) module, a multi-layer BiGRU module, and an Attention module.
[0073] The WDCNN module mainly includes:
[0074] The first layer is a convolutional layer, which uses a 64×1 convolutional kernel and the activation function is ReLU;
[0075] The second layer is a pooling layer, which uses 2×2 max pooling;
[0076] The third layer is a convolutional layer, which uses a 3×1 convolutional kernel and the activation function is ReLU;
[0077] The fourth layer is the pooling layer, which uses 2×2 max pooling;
[0078] The structure then consists of three convolutional layers and three pooling layers arranged alternately, with the convolutional kernels of the convolutional layers all being 3×1 and the pooling kernels of the pooling layers being 2×2.
[0079] The last layer is the output layer, and the softmax function is used as the activation function for the output layer.
[0080] The multi-layer BiGRU module is a 3-layer BiGRU module, including 3 layers of BiGRU and 3 layers of Dropout. The spatiotemporal correlation features within the time window are extracted through the operation of the 3-layer BiGRU module, and the dimensionality of subsequent processing is reduced.
[0081] BiGRU adopts a bidirectional structure, using two gated recurrent network models GRU to extract the temporal relevance features of historical information from the historical end and the future end, respectively.
[0082] Gated Recurrent Network (GRU) is a gated recurrent network model. Each GRU model contains only two gates: a reset gate and an update gate. The reset gate determines how new inputs are combined with previous memories, which helps capture short-term dependencies in the sequence. The update gate determines how much of the previous memory is effective, which helps capture long-term dependencies in the sequence.
[0083] The formula for resetting the door is expressed as follows:
[0084] R t =σ(X) t W XR +H t-1 W HR +b R (2)
[0085] In the formula, X t W represents the input data at time step t. XR X represents t Reset gate weight parameters, H t-1 W represents the hidden state at time step t-1, which indicates the spatiotemporal correlation characteristics of the input data. HR H represents t-1 Reset gate weight parameters, b R This represents the bias parameter of the reset gate, and σ represents the sigmoid activation function;
[0086] The update gate is represented as follows:
[0087] Z t =σ(X) t W XZ +H t-1 W HZ +b Z (3)
[0088] In the formula, W XZ X represents t Update the gate weight parameters, W HZ H represents t-1 Update the gate weight parameters, b Z This indicates that the gate's bias parameters are being updated;
[0089]
[0090] In the formula, W XH W HH They represent X respectively t The weight parameters and H t-1 The weight parameter, b H The bias parameter for the candidate hidden state is tanh. The value of the candidate hidden state is guaranteed to be between -1 and 1 by using the nonlinear activation function tanh. ⊙ represents the Hadamard product.
[0091]
[0092] In the formula, Z t Represents the update gate at time step t;
[0093] The expression for the BiGRU model is:
[0094]
[0095] In the formula, X t Let H be the wind speed input at time t. t Let be the hidden layer state at time t; This represents the output of the forward hidden layer at time t. This represents the output of the reverse hidden layer at time t. and Let b represent the weights corresponding to the forward and backward hidden layer states of the BiGRU at time t, respectively; t This represents the bias corresponding to the hidden layer state at time t.
[0096] After passing through BiGRU, the time-series features of each operating condition are discarded based on the importance of the information, and then feature extraction is performed. These time-series features are influenced by historical and future data, and the extracted operating condition features are reprocessed through an attention mechanism. The attention mechanism module assigns different weights to the hidden layer units according to the importance of the feature information.
[0097] The expression I for the Attention module of the aforementioned attention mechanism is:
[0098] e t =V e tanh(W W Ht +b W (9)
[0099]
[0100] H t V represents the output of the BiGRU hidden layer, and V represents the attention probability distribution from the attention mechanism output to the BiGRU hidden layer. e and W W This represents the weighting coefficient.
[0101] The constructed WDCNN-BiGRU-Attention model was trained using the training set and used to predict wind speeds for the next 2 minutes. The Adam optimizer was used to train the WDCNN-BiGRU-Attention model with a learning rate of 0.001, and parameter tuning was performed. The loss function was the root mean square error (RMSE), and the formula is as follows:
[0102]
[0103] In the formula, n represents the number of samples, i represents the sample number, Y represents the true label of the sample, Y' represents the predicted value, and the smaller the RMSE value, the better the prediction result.
[0104] Step S6: After training, save the WDCNN-BiGRU-Attention model and its optimal hyperparameters.
[0105] Step S7: Input the test set into the WDCNN-BiGRU-Attention model for testing and performance evaluation, verify the superiority of the hyperparameters, and obtain the optimal prediction model;
[0106] The WDCNN-BiGRU-Attention model with the highest hyperparameters is used as the trained model.
[0107] Step S8: The collected real-time wind speed data is decomposed into non-periodic and periodic pattern representations using the Time2Vec time embedding algorithm; then, it is input into the trained WDCNN-BiGRU-Attention model to obtain the prediction results.
[0108] An orthogonal anemometer based on a multi-rotor UAV for an attention-based ultra-short-term wind speed prediction method includes a multi-rotor UAV, an anemometer body mounted on the multi-rotor UAV, and a vertically erected, adjustable-length connecting rod mechanism on the anemometer body. The connecting rod mechanism has four orthogonally arranged Pitot tubes serving as wind guides. Figure 6As shown, the length of the connecting rod mechanism is such that the pitot tube is located outside the wing turbulence area of the multi-rotor UAV. The ends of the four pitot tubes are connected to four micro-pressure sensors respectively. The main control circuit is electrically connected to each micro-pressure sensor. The main control circuit calculates the wind speed data through the wind pressure data measured by each micro-pressure sensor and the pitch and roll angle data of the multi-rotor UAV.
[0109] In this embodiment, the pitot tube has an inner diameter of 2mm, and the micro-pressure sensor is an XGZP6847A sensor. The drone model is DJIM300RTK (DJI brand).
[0110] Depend on Figure 7 Simulation results show that the farther away from the rotor plane, the less the impact of the drone's rotation; the farther away from the quadcopter drone's body, the less the impact of the rotor rotation. The suitable location for installing the wind measuring device on the quadcopter drone is 300-450mm above the drone. To avoid affecting the drone's center of gravity, the connecting rod mechanism in this embodiment is set to a length of 350mm to prevent rotor turbulence from affecting the air duct.
[0111] Wind pressure measurement principle: Based on the assumption that air is an incompressible gas, in an ideal flow field, at any two points on the same streamline, the sum of the fluid's kinetic energy, gravitational potential energy, and pressure potential energy remains constant, as shown in equation (13):
[0112]
[0113] Where ρ is the fluid density, v is the fluid velocity, g is the gravitational acceleration, h is the height, P is the pressure potential energy, and C is a constant;
[0114] Pitot tubes 1 and 3, which are positioned opposite each other, satisfy equation (14):
[0115]
[0116] In the formula, v1 and v3 are the fluid velocities in Pitot tube 1 and Pitot tube 3, respectively; h1 and h3 are the heights of Pitot tube 1 and Pitot tube 3, respectively; and P1 and P3 are the pressure potential energies of Pitot tube 1 and Pitot tube 3, respectively.
[0117] Since the opening heights of Pitot tube 1 and Pitot tube 3 are the same, i.e., h1 = h3, the north-south wind speed v′ can be obtained from equation (14):
[0118]
[0119] The east-west wind speed is v″:
[0120]
[0121] According to equation (17), the pitch angle θ and roll angle of the UAV are collected by the data acquisition device. The data is used to establish a rotation matrix R(Θ), and after correction, the east-west wind speed and north-south wind speed in the geographic coordinate system are obtained. The correction calculation is as shown in equation (18):
[0122]
[0123] v 24 =R(Θ) T v″ (18)
[0124] v 13 =R(Θ) T v′ (19)
[0125] Where, R(Θ) T The final wind speed value v is calculated using the transpose of the rotation matrix R(Θ) according to equation (19):
[0126]
[0127] This example uses wind speed data collected at different flight altitudes by a drone wind measurement platform. The sampling time between wind speed data is 0.5 seconds. Then, the measured wind speed data is corrected to the real wind speed data based on the three-axis attitude data of the drone to construct the original wind speed dataset and wind speed time series.
[0128] The above embodiments are merely illustrative of the principles and effects of the present invention, as well as some examples of its application, and are not intended to limit the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these modifications and improvements are all within the scope of protection of the present invention.
Claims
1. An attention mechanism-based ultra-short-term wind speed prediction method, comprising the following specific steps: Step S1, sampling the wind speed data of a specified height at a test site according to a predetermined sampling frequency; Step S2, constructing an original wind speed data set including a wind speed time series according to the wind speed sampling data, the wind speed time series being used to represent wind speed-time information; Step S3, decomposing the wind speed time series in the original wind speed data set by using a Time2Vec time embedding algorithm to obtain representations of non-periodic patterns and periodic patterns; splicing the decomposed non-periodic components and periodic components with the original wind speed time series to obtain a new wind speed data set; Step S4, dividing the new wind speed data set into a training set and a test set; Step S5, inputting the training set of each modal component into a WDCNN-BiGRU-Attention model for training, and performing parameter optimization by using a grid search method to determine the hyperparameters of the model; Step S6, saving the WDCNN-BiGRU-Attention model and its optimal hyperparameters after the training is completed; Step S7, inputting the test set into the WDCNN-BiGRU-Attention model for testing and performance evaluation, verifying the superiority of the hyperparameters, and obtaining an optimal prediction model; using the WDCNN-BiGRU-Attention model with the optimal hyperparameters as the trained model; Step S8, decomposing the wind speed time series of real-time wind speed data collected by using the Time2Vec time embedding algorithm into representations of non-periodic patterns and periodic patterns, and then inputting the wind speed time series into the trained WDCNN-BiGRU-Attention model to obtain a prediction result.
2. The ultra-short-term wind speed prediction method based on attention mechanism according to claim 1, characterized in that: The expression of the time embedding algorithm in step S3 is as follows: wherein for any time series τ, its corresponding Time2Vec is represented as a vector with a length of k+1; in the vector, represents the i-th feature; the sum of in the model is a trainable parameter, constituting a linear function for extracting non-periodic features of the time series; the function is used to capture periodic features. 3.The ultra-short-term wind speed prediction method based on attention mechanism according to claim 1, wherein: The WDCNN-BiGRU-Attention model comprises a wide and deep convolutional neural network (WDCNN) module, a multi-layer BiGRU module, and an Attention module. The WDCNN module mainly comprises: The first layer is a convolutional layer, which uses a convolutional kernel of 64x1 and a ReLU activation function; The second layer is a pooling layer, which uses a maximum pooling of 2x2; The third layer is a convolutional layer, which uses a convolutional kernel of 3x1 and a ReLU activation function; The fourth layer is a pooling layer, which uses a maximum pooling of 2x2; Then, there is an alternating arrangement structure of three convolutional layers and three pooling layers, wherein the convolutional kernel size of the convolutional layers is 3x1, and the pooling kernel of the pooling layers is 2x2; The last layer is an output layer, which uses a softmax function as the activation function.
4. The ultra-short-term wind speed prediction method based on attention mechanism according to claim 3, characterized in that: The multi-layer BiGRU module is a 3-layer BiGRU module, including 3-layer BiGRU and 3-layer Dropout, and the time window space correlation feature is extracted through the operation of the 3-layer BiGRU module, and the dimension size of subsequent processing is reduced; The BiGRU adopts a bidirectional structure, and two GRU models are used to extract the time correlation features of historical information from the history end and the future end respectively; The calculation formula of the reset gate is represented as: R t = σ(X t W XR + H t-1 W HR + b R )(2) In the formula, X t represents the input data at the t-th time step, W XR represents the reset gate weight parameter of X t , H t-1 represents the hidden state at the t-1-th time step, which represents the spatio-temporal correlation feature of the input data, W HR represents the reset gate weight parameter of H t-1 , b R represents the bias parameter of the reset gate, and σ represents the sigmoid activation function. The update gate is represented as: Z t = σ(X t W XZ +H t-1 W HZ +b Z )(3) where W XZ represents the update gate weight parameter of X t HZ represents the update gate weight parameter of H t-1 Z represents the bias parameter of the update gate wherein W XH , W HH represent the weight parameters of X t and H t-1 , respectively, b H is a bias parameter of the candidate hidden state, and is guaranteed to be between -1 and 1 by using a tanh non-linear activation function, and represents the Hadamard product; In the formula, Z t denotes the update gate for the t-th time step; The expression of the BiGRU model is: wherein X t is the wind speed input at time t, H t is the hidden layer state at time t; is the output of the forward hidden layer at time t; is the output of the backward hidden layer at time t; and respectively represent the weights corresponding to the forward hidden layer state and the backward hidden layer state of the BiGRU at time t; b t represents the bias corresponding to the hidden layer state at time t.
5. The ultra-short-term wind speed prediction method based on attention mechanism according to claim 1, characterized in that: The expression I of the Attention module is: e t = V e tanh(W W H t + b W )(9) H t represents the output of the BiGRU hidden layer, and represents the attention probability distribution from the attention mechanism output to the BiGRU hidden layer; V e and W W represent the weighting coefficients.
6. The ultra-short-term wind speed prediction method based on attention mechanism according to claim 4, characterized in that: The WDCNN-BiGRU-Attention model is trained by using an Adam optimizer, the learning rate is 0.001, and parameter optimization is performed; the loss function is a root mean square error (RMSE), and the formula is: In the formula, n represents the sample quantity, i represents the sample serial number, Y represents the real label of the sample, Y' represents the predicted value, and the smaller the RMSE value, the better the prediction result.
7. A multi-copter drone based orthogonally mounted anemometer for the attention mechanism based ultra-short term wind speed prediction method according to any one of claims 1-6, characterized in that: The wind speed sensor includes a multi-rotor unmanned aerial vehicle, a wind speed sensor body arranged on the multi-rotor unmanned aerial vehicle, a vertically arranged connecting rod mechanism with adjustable length arranged on the wind speed sensor body, four Pitot tubes arranged orthogonally and serving as wind guide pipes arranged on the connecting rod mechanism, and the length of the connecting rod mechanism satisfies that the Pitot tubes are located outside a wing disturbance area of the multi-rotor unmanned aerial vehicle, four micro pressure sensors are connected to the four Pitot tubes, a main control circuit is electrically connected to each micro pressure sensor, and the main control circuit calculates wind speed data by using wind pressure data measured by each micro pressure sensor and pitch angle and roll angle data of the multi-rotor unmanned aerial vehicle.