Wind power plant bird trajectory prediction and fan linkage control method based on neural network

Abstract

The invention provides a wind power plant bird trajectory prediction and fan linkage control method based on a neural network, and relates to the technical field of intelligent power grids, and the method comprises the steps: recognizing a bird target in real time, generating trajectory data, processing the trajectory data through three-dimensional Hilbert-Huang transform and an adaptive decomposition algorithm, extracting feature parameters for prediction analysis, and obtaining a bird trajectory prediction result. A collision risk is predicted and evaluated based on a trajectory, a cooperative avoidance control algorithm is constructed by adopting an artificial potential field method, a control strategy is optimized by combining an adaptive fuzzy neural network and a sliding mode controller, the operation state of a fan is monitored in real time, normal operation of the fan is recovered after birds fly away safely, and intelligent bird protection of a wind power plant is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of smart grids, and particularly to a method for predicting bird trajectories and coordinating the control of wind turbines in a wind farm based on a neural network. Background Art

[0002] With the rapid development of renewable energy, the construction and operation of wind farms are increasing day by day. However, the operation of wind farms will have an impact on the surrounding ecological environment. Therefore, how to effectively predict the flight trajectories of birds and take corresponding avoidance control measures has become an important issue to ensure the safe operation of wind farms.

[0003] Most of the existing bird monitoring and trajectory prediction methods rely on traditional image processing technologies, which cannot identify and track bird targets in real time and accurately, resulting in low prediction accuracy. At the same time, the avoidance control algorithms often ignore the dynamic relationship between birds and wind turbines, resulting in insufficient real-time performance and effectiveness of the avoidance strategies.

[0004] Therefore, there is an urgent need for a solution to solve the problems existing in the prior art. Summary of the Invention

[0005] An embodiment of the present invention provides a method for predicting bird trajectories and coordinating the control of wind turbines in a wind farm based on a neural network, which can at least solve some of the problems existing in the prior art.

[0006] In a first aspect of an embodiment of the present invention, there is provided a method for predicting bird trajectories and coordinating the control of wind turbines in a wind farm based on a neural network, including: Collecting real-time image data and using a pre-trained object detection neural network to identify bird targets, and generating bird flight trajectory data according to the position information of the bird targets in consecutive multiple frames of images; Performing three-dimensional Hilbert-Huang transform on the bird flight trajectory data to obtain an instantaneous frequency spectrum, decomposing the instantaneous frequency spectrum into intrinsic mode components by combining an adaptive decomposition algorithm, extracting amplitude-frequency characteristic parameters, performing feature dimensionality reduction and reconstruction, obtaining a flight feature vector, performing time series analysis, generating a prediction feature vector sequence, and calculating a trajectory correction amount according to the phase difference and amplitude change to obtain predicted trajectory data; Determining the wind turbines in the wind farm that have a risk of bird collision according to the predicted trajectory data; Constructing a cooperative avoidance control algorithm based on the artificial potential field method, using the bird target as the repulsive force source and the wind turbines as the repulsive force field, calculating the potential field gradient according to the relative distance between the bird target and the wind turbines, online optimizing the potential field gradient through an adaptive fuzzy neural network, generating avoidance control parameters, combining with a sliding mode controller to generate obstacle avoidance parameters, and combining them to obtain an obstacle avoidance control strategy; Execute the obstacle avoidance control strategy and continuously monitor the operating state of the adjusted wind turbine until the bird target flies out of the wind farm, and then restore the wind turbine to its normal operating state according to the preset restoration strategy.

[0007] In an alternative embodiment, Perform three-dimensional Hilbert-Huang transform on the bird flight trajectory data to obtain the instantaneous frequency spectrum, and decompose the instantaneous frequency spectrum into intrinsic mode components by combining the adaptive decomposition algorithm, including: Perform three-dimensional Hilbert-Huang transform on the bird flight trajectory data to obtain the analytical signals in the three coordinate axis directions. Calculate the instantaneous amplitude and instantaneous phase in the three coordinate axis directions based on the analytical signals, calculate the instantaneous frequency based on the time derivative of the instantaneous phase, and combine the instantaneous frequencies in different coordinate axis directions to obtain the instantaneous frequency spectrum; Take the instantaneous frequency spectrum as the initial residual signal, use the adaptive decomposition algorithm to determine all the local extreme points of the initial residual signal, use the cubic spline interpolation method to construct the upper envelope and lower envelope of the initial residual signal, calculate the mean of the upper envelope and lower envelope to obtain the mean envelope, and subtract the mean envelope from the initial residual signal to obtain the detail component; Judge whether the difference between the number of zero points and the number of extreme points of the detail component is less than or equal to 1 and whether the mean envelope of the detail component is approximately zero. If so, determine the detail component as an intrinsic mode component; if not, take the detail component as a new residual signal and repeat the adaptive decomposition algorithm until the intrinsic mode component is obtained.

[0008] In an alternative embodiment, Extract the amplitude-frequency characteristic parameters, perform feature dimension reduction and reconstruction, obtain the flight feature vector, perform time series analysis, generate the predicted feature vector sequence, and calculate the trajectory correction amount according to the phase difference and amplitude change to obtain the predicted trajectory data, including: Calculate the first derivative and second derivative of the instantaneous phase of the previously obtained intrinsic mode component to obtain the instantaneous frequency. Determine the window length of the Hanning window function according to the ratio of the instantaneous frequency to the sampling frequency, perform adaptive window processing on the intrinsic mode component, extract the processed instantaneous phase and energy change rate to obtain the amplitude-frequency characteristic parameters, and construct a feature matrix by arranging them in time series; Calculate the time derivative of the feature matrix to obtain the feature change rate, calculate the sample importance weight, weight the feature matrix based on the sample importance weight to obtain the weighted covariance matrix, determine the first eigenvector based on the eigenvalue interval ratio, cumulative contribution rate, and reconstruction error of the weighted covariance matrix, and perform dimension reduction and reconstruction on the feature matrix based on the first eigenvector to obtain the projection matrix; Multiply the feature matrix by the projection matrix to obtain a flight feature vector, establish a temporal relationship of quadratic local trends, and combine the iteratively reweighted least squares method for temporal prediction to obtain a sequence of predicted feature vectors; Calculate the phase difference parameter and amplitude change parameter at adjacent times according to the sequence of predicted feature vectors, initialize the combined weight coefficients of the phase difference parameter and amplitude change parameter based on historical trajectory data and perform weighted summation, use the weighted sum as the trajectory correction amount, update the combined weight coefficients according to the prediction error, calculate the smoothing correction amount according to the trajectory correction amount and dynamically adjust the smoothing factor, and combine the smoothing factor with the sequence of predicted feature vectors to obtain predicted trajectory data.

[0009] In an alternative embodiment, Establishing a temporal relationship of quadratic local trends and combining the iteratively reweighted least squares method for temporal prediction to obtain a sequence of predicted feature vectors includes: Divide the flight feature vector into short-term, medium-term, and long-term, construct a variational constraint optimization objective function by calculating the sequence envelope and instantaneous frequency, iteratively update the center frequency and bandwidth parameters and extract the adaptive multi-scale components with the optimal energy aggregation degree, perform dual-tree complex wavelet transform on the adaptive multi-scale components to obtain multi-level sparse time-frequency features, and generate an enhanced input sequence with temporal position information by combining temporal position encoding; Construct a multi-layer biological perception field network based on the enhanced input sequence. In each layer of the perception field, extract key feature patterns through excitatory neuron groups, perform selective enhancement using inhibitory neuron groups, combine the adaptive dynamic threshold to adjust the neuron activation state to obtain perception features, and combine the perception features in each layer of the perception field to obtain a hierarchical attention feature representation; Perform residual connection and layer normalization processing on the attention feature representation, use the adaptive gating unit for dynamic weight allocation, fuse multi-scale features to obtain an enhanced feature representation with temporal dependence, construct a quadratic local trend model based on the enhanced feature representation, optimize and solve using the iteratively reweighted least squares method, construct a dynamic prediction confidence interval through multiple samplings, adjust the historical data weight and prediction credibility weight according to the interval width, and output a sequence of predicted feature vectors.

[0010] In an alternative embodiment, Construct a cooperative avoidance control algorithm based on the artificial potential field method, with the bird target as the repulsive source and the wind turbine as the repulsive field. Calculate the potential field gradient according to the relative distance between the bird target and the wind turbine, including: Taking the bird target as a repulsive force source, a first repulsive potential energy function is established. The first repulsive potential energy function is determined by a repulsive force field coefficient, the relative distance of the bird target, and a repulsive force field influence radius. When the relative distance of the bird target is less than or equal to the repulsive force field influence radius, the first repulsive potential energy function is positively correlated with the repulsive force field coefficient, the relative distance of the bird target, and the repulsive force field influence radius. When the relative distance of the bird target is greater than the repulsive force field influence radius, the first repulsive potential energy function is zero; Taking the wind turbine generator as a repulsive force field, a second repulsive potential energy function is established. The second repulsive potential energy function is determined by a repulsive force field coefficient, the relative distance of the wind turbine generator, and a repulsive force field influence radius. When the relative distance of the wind turbine generator is less than or equal to the repulsive force field influence radius, the second repulsive potential energy function is positively correlated with the repulsive force field coefficient, the relative distance of the wind turbine generator, and the repulsive force field influence radius. When the relative distance of the wind turbine generator is greater than the repulsive force field influence radius, the second repulsive potential energy function is zero; Taking the negative gradient of the first repulsive potential energy function and the second repulsive potential energy function respectively according to the relative distances of the bird target and the wind turbine generator to obtain the potential field gradient.

[0011] In an alternative embodiment, Online optimization is performed on the potential field gradient through an adaptive fuzzy neural network, generating avoidance control parameters and combining with a sliding mode controller to generate obstacle avoidance parameters and combining them to obtain an obstacle avoidance control strategy, including: Receiving the pre-obtained potential field gradient, inputting the potential field gradient into an adaptive fuzzy neural network with an input layer, a fuzzification layer, a rule layer, a rule inference layer, and an output layer. In the fuzzification layer, the Gaussian membership function is used to perform fuzzification processing on the potential field gradient to obtain fuzzified data; Based on the network output error, the update amounts of the membership function center value parameter and the width parameter are determined by the gradient descent method, and online optimization is performed on the membership function center value parameter and the width parameter. The optimized fuzzification parameters are input into the rule layer to obtain the rule layer output vector; Performing matrix operation on the rule layer output vector and the network weight matrix to obtain a network compensation amount, superimposing the network compensation amount and the potential field gradient to obtain an optimized potential field gradient, and generating avoidance control parameters based on the change rate of the optimized potential field gradient; Determining the position tracking error based on the avoidance control parameters and constructing a sliding mode surface and a non-singular terminal sliding mode controller, performing finite-time convergence through a control law structure combined by a piecewise power function, and inputting the optimized potential field gradient into the non-singular terminal sliding mode controller to generate obstacle avoidance parameters; Determine the dynamic weight factor based on the position tracking error, and use the dynamic weight factor to perform weighted combination of the obstacle avoidance parameters and the output of the output layer of the adaptive fuzzy neural network to obtain the obstacle avoidance control strategy.

[0012] In an alternative embodiment, Determine the position tracking error based on the avoidance control parameter, construct a sliding mode surface and a nonsingular terminal sliding mode controller, and perform finite-time convergence through the control law structure combined by piecewise power functions. Input the optimized potential field gradient into the nonsingular terminal sliding mode controller to generate the obstacle avoidance parameters, including: Construct a position tracking error mapping function based on the avoidance control parameter. When the absolute value of the avoidance control parameter is greater than or equal to the threshold parameter, use the product of the first mapping coefficient and the non-linear power term of the avoidance control parameter minus the reference control quantity as the position tracking error output. When the absolute value of the avoidance control parameter is less than the threshold parameter, use the product of the second mapping coefficient and the sine function based on the ratio of the avoidance control parameter to the threshold parameter as the position tracking error output; Construct a non-linear sliding mode surface containing dynamic weights based on the position tracking error. The non-linear sliding mode surface consists of the derivative term of the position tracking error, the product term of the dynamic weight factor and the position tracking error, and the non-linear power integral term. The dynamic weight factor includes a basic weight and an exponential decay adaptive term based on the position tracking error; Design a piecewise continuous control law based on the non-linear sliding mode surface. When the absolute value of the non-linear sliding mode surface is greater than or equal to the switching domain threshold, use the combination of two different power terms as the control output. When the absolute value of the non-linear sliding mode surface is less than the switching domain threshold, use the combination of a linear term and a fractional power term as the control output; Use the product of the Sigmoid function and the optimized potential field gradient constraint function as the smooth transition function of the non-linear sliding mode surface, and superimpose the control output with the product of the smooth transition function and the optimized potential field gradient to obtain the obstacle avoidance parameter.

[0013] In an alternative embodiment, Execute the obstacle avoidance control strategy and monitor the operating state of the adjusted wind turbine in real time until the bird target flies out of the wind farm. Restore the wind turbine to the normal operating state according to the preset restoration strategy, including: Execute the obstacle avoidance control strategy to adjust the wind turbine, and monitor the operating state of the wind turbine and the flight trajectory of the bird target in real time; When the bird target flies away from the wind farm, obtain the current operating state parameters of the wind turbine, and gradually adjust the rotational speed, torque, and pitch angle of the wind turbine to the corresponding target values in segments with different recovery slopes according to the operating state parameters and a preset recovery threshold to restore the normal operating state.

[0014] In a second aspect of the embodiments of the present invention, there is provided an electronic device, including: a processor and a memory for storing processor-executable instructions, wherein the processor is configured to call the instructions stored in the memory to execute the method described above.

[0015] In a third aspect of the embodiments of the present invention, there is provided a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0016] In the present invention, by collecting real-time image data and using a pre-trained object detection neural network to identify bird targets, bird flight trajectory data can be accurately generated, improving the response speed of the wind farm to bird activities and reducing the risk of collision between birds and wind turbines. By using the three-dimensional Hilbert-Huang transform and the adaptive decomposition algorithm to extract flight feature vectors, effective feature dimensionality reduction and reconstruction can be performed, thereby generating a precise prediction feature vector sequence, calculating the correction amount of the bird trajectory more accurately, and improving the effectiveness of the obstacle avoidance control strategy. The avoidance control algorithm constructed based on the artificial potential field method enables the wind turbine to intelligently avoid bird targets, reducing the collision risk of birds. Combining the online optimization ability of the adaptive fuzzy neural network further improves the flexibility and adaptability of the obstacle avoidance control strategy, ensuring the safe operation of the wind farm when bird targets appear. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a schematic flow chart of the method for predicting bird trajectories and wind turbine linkage control in a wind farm based on a neural network according to an embodiment of the present invention; Figure 2 is a comparison diagram of the trajectory correction effect of the method for predicting bird trajectories and wind turbine linkage control in a wind farm based on a neural network according to an embodiment of the present invention; Figure 3 is a multi-scale decomposition effect diagram of flight features of the method for predicting bird trajectories and wind turbine linkage control in a wind farm based on a neural network according to an embodiment of the present invention; Figure 4 is a simulation diagram of the obstacle avoidance technical effect of the method for predicting bird trajectories and wind turbine linkage control in a wind farm based on a neural network according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0019] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0020] Figure 1 The following is a schematic flowchart of a method for predicting bird trajectories and wind turbine linkage control in a wind farm based on a neural network according to an embodiment of the present invention, as Figure 1 shown, the method includes: Collect real-time image data and use a pre-trained object detection neural network to identify bird targets, and generate bird flight trajectory data based on the position information of the bird targets in consecutive multiple frames of images; Perform a three-dimensional Hilbert-Huang transform on the bird flight trajectory data to obtain an instantaneous frequency spectrum, decompose the instantaneous frequency spectrum into intrinsic mode components in combination with an adaptive decomposition algorithm, extract amplitude-frequency characteristic parameters, perform feature dimensionality reduction and reconstruction, obtain a flight feature vector, perform time series analysis, generate a prediction feature vector sequence, and calculate a trajectory correction amount based on the phase difference and amplitude change to obtain predicted trajectory data; Determine the wind turbines in the wind farm that pose a risk of bird collision based on the predicted trajectory data; Construct a cooperative avoidance control algorithm based on the artificial potential field method, use the bird target as the repulsive source and the wind turbine as the repulsive field, calculate the potential field gradient according to the relative distance between the bird target and the wind turbine, perform online optimization of the potential field gradient through an adaptive fuzzy neural network, generate avoidance control parameters, combine them with a sliding mode controller to generate obstacle avoidance parameters, and combine them to obtain an obstacle avoidance control strategy; Execute the obstacle avoidance control strategy and continuously monitor the operating state of the adjusted wind turbines until the bird target flies out of the wind farm, and restore the wind turbines to their normal operating states according to the preset recovery strategy.

[0021] In an alternative embodiment, Performing a three-dimensional Hilbert-Huang transform on the bird flight trajectory data to obtain an instantaneous frequency spectrum, and decomposing the instantaneous frequency spectrum into intrinsic mode components in combination with an adaptive decomposition algorithm includes: Perform three-dimensional Hilbert-Huang transform on the bird flight trajectory data to obtain analytical signals in three coordinate axis directions. Calculate the instantaneous amplitude and instantaneous phase in three coordinate axis directions based on the analytical signals. Calculate the instantaneous frequency based on the time derivative of the instantaneous phase. Combine the instantaneous frequencies in different coordinate axis directions to obtain the instantaneous frequency spectrum; Use the instantaneous frequency spectrum as the initial residual signal. Determine all local extreme points of the initial residual signal using an adaptive decomposition algorithm. Construct the upper envelope and lower envelope of the initial residual signal using cubic spline interpolation. Calculate the mean of the upper envelope and lower envelope to obtain the mean envelope. Subtract the mean envelope from the initial residual signal to obtain the detail component; Judge whether the difference between the number of zero points and the number of extreme points of the detail component is less than or equal to 1 and whether the mean envelope of the detail component is approximately zero. If so, determine the detail component as the intrinsic mode component. If not, use the detail component as the new residual signal and repeat the adaptive decomposition algorithm until the intrinsic mode component is obtained.

[0022] Obtain the spatial position data of the bird flight trajectory and decompose it into three coordinate components in the eastward, northward, and height directions. Perform Hilbert-Huang transform on each coordinate component separately to obtain analytical signals and perform adaptive decomposition on the signals simultaneously. In the eastward coordinate, map the position data sequence to the complex plane through Hilbert transform to obtain the analytical signal; in the northward coordinate, obtain the corresponding analytical signal by performing Hilbert transform on the position sequence; in the height direction, perform Hilbert transform on the vertical position sequence to obtain the third analytical signal. Based on the three obtained analytical signals in different directions, extract their respective instantaneous amplitudes. The instantaneous amplitude reflects the change in displacement amplitude of the bird in the corresponding direction; extract the instantaneous phase in each direction. The instantaneous phase information characterizes the periodic motion characteristics of the bird in different directions. Take the time derivative of the instantaneous phase in three directions to obtain the instantaneous frequency reflecting the motion frequency characteristics of the bird in each direction. Combine the instantaneous frequencies in these three directions according to the time correspondence relationship to construct a complete three-dimensional instantaneous frequency spectrum.

[0023] Using the constructed three-dimensional instantaneous frequency spectrum as the initial residual signal, start the adaptive decomposition. Search for all local maximum points and local minimum points in the frequency spectrum signal, which reflect the turning or speed change characteristics during the bird's flight. By comparing the magnitude relationships of adjacent data points, determine the specific positions and amplitudes of each maximum point and minimum point. Use the cubic spline interpolation method to construct a smooth upper envelope line with the found local maximum points as control points. The envelope line passes through all the maximum points and forms a smooth transition between them; construct a smooth lower envelope line with the local minimum points as control points, making it pass through all the minimum points and maintain continuous smoothness. Calculate the average value of the upper and lower envelope lines at each sampling moment to obtain the mean envelope line reflecting the overall trend of the signal. Subtract the mean envelope line from the initial frequency spectrum signal to obtain the detail component after removing the overall trend.

[0024] Determine the eigenmode characteristics of the obtained detail component. Count the number of zero-crossing points in the detail component, that is, the number of times the signal crosses the zero value line, and count the number of extreme points of the signal, including the total number of all local maximum points and minimum points. Calculate the absolute value of the difference between the number of zero-crossing points and the number of extreme points. This difference characterizes the symmetry feature of the signal. Calculate the mean envelope line of the detail component and judge whether it is close enough to the zero value. When the difference between the number of zero-crossing points and the number of extreme points is less than or equal to 1, it indicates that the signal has good symmetry; at the same time, when the mean envelope line is close enough to the zero value, it indicates that the signal oscillates sufficiently and evenly. At this time, it is considered that the detail component meets the requirements of the intrinsic mode function and is determined as a valid intrinsic mode component. If either of the above two conditions is not satisfied, then use this detail component as the new residual signal and repeat the processes of extreme point search, envelope line construction, mean envelope calculation, and detail component extraction until an intrinsic mode component that meets the judgment conditions is obtained. Through the iterative decomposition process, decompose the frequency characteristics of the bird's flight trajectory into a series of intrinsic mode components.

[0025] Exemplarily, trajectory data of a seagull starting to fly from a position 180 meters eastward, 150 meters northward, and 100 meters in height is collected. The sampling time is 60 seconds, and the sampling interval is 0.1 second. Hilbert-Huang transform is performed on the recorded eastward position data sequence to obtain an analytic signal in complex form reflecting the eastward motion characteristics; Hilbert-Huang transform is performed on the northward position data to obtain the corresponding analytic signal; the same transform is performed on the height data to obtain the third analytic signal. Instantaneous amplitudes in three directions are extracted from the analytic signals. It is found that the displacement amplitude of the seagull in the eastward direction fluctuates around 20 meters, the northward displacement amplitude is about 15 meters, and the fluctuation amplitude in the height direction is within 10 meters. Instantaneous phase changes in three directions are obtained, and the instantaneous frequency is obtained after taking the time derivative. It is found that the motion frequency of the seagull in the horizontal plane is mainly concentrated between 0.2 - 0.5 Hz, and the motion frequency in the vertical direction is lower, within the range of 0.1 - 0.3 Hz. The frequency data in three directions are combined to form a complete frequency spectrum.

[0026] The decomposition of the frequency spectrum is started. By traversing the data, it is found that it contains 23 local maximum points and 22 local minimum points. The cubic spline interpolation method is used to construct the upper and lower envelope lines based on the extreme points. After calculating the mean envelope line and subtracting it from the original frequency spectrum, the first-level detail component is obtained. After calculation, this detail component contains 35 zero points and 36 extreme points, and the difference between the number of zero points and extreme points is 1; the maximum value of the calculated mean envelope line is only 0.1% of the amplitude of the original signal. Since this detail component meets the determination conditions of the intrinsic mode function, it is determined as the first intrinsic mode component. This component reflects the motion characteristics of the highest frequency during the seagull's flight.

[0027] In this embodiment, through the three-dimensional Hilbert-Huang transform combined with the adaptive decomposition algorithm, the time-frequency characteristics of the bird's flight trajectory can be effectively captured. Due to the adoption of the three-dimensional processing method, the motion characteristics of the bird in space can be comprehensively captured, avoiding the information loss caused by only analyzing a single direction, providing a reliable data basis for bird behavior pattern recognition, and further improving the accuracy of predicting bird activities in the wind farm.

[0028] In an alternative embodiment, Amplitude-frequency characteristic parameters are extracted and dimension reduction and reconstruction of the features are performed to obtain a flight feature vector and perform time series analysis. A predicted feature vector sequence is generated and the trajectory correction amount is calculated based on the phase difference and amplitude change. The predicted trajectory data includes: The instantaneous frequency is obtained by calculating the first-order derivative and the second-order derivative of the instantaneous phase of the pre-acquired intrinsic mode component, the window length of the Hanning window function is determined according to the ratio of the instantaneous frequency to the sampling frequency, and the intrinsic mode component is subjected to adaptive window processing, the processed instantaneous phase and energy change rate are extracted to obtain amplitude-frequency characteristic parameters, and the characteristic matrix is ​​constructed in time sequence; Calculate the time derivative of the feature matrix to obtain the feature change rate and calculate the sample importance weight, weight the feature matrix based on the sample importance weight to obtain a weighted covariance matrix, determine the first eigenvector based on the eigenvalue interval ratio, cumulative contribution rate and reconstruction error of the weighted covariance matrix, and reduce the dimension and reconstruct the feature matrix based on the first eigenvector to obtain a projection matrix; The flight feature vector is obtained by multiplying the feature matrix with the projection matrix, a time series relationship of the local trend of the quadratic term is established, and a time series prediction is performed in combination with iterative reweighted least squares method to obtain a predicted feature vector sequence; The phase difference parameters and amplitude change parameters at adjacent moments are calculated according to the predicted feature vector sequence, the combined weight coefficients of the phase difference parameters and the amplitude change parameters are initialized based on the historical trajectory data and weighted summed, the weighted sum is used as the trajectory correction amount, the combined weight coefficient is updated according to the prediction error, the smoothing correction amount is calculated according to the trajectory correction amount and the smoothing factor is dynamically adjusted, and the smoothing factor is combined with the predicted feature vector sequence to obtain the predicted trajectory data.

[0029] Starting from the time series data processing, the intrinsic mode component is subjected to continuous sliding data analysis. The first-order derivative of the instantaneous phase is calculated by the forward difference method to obtain the instantaneous frequency value at each time point; the instantaneous frequency is then subjected to a second-order difference operation to obtain the acceleration change characteristics of the frequency over time. The instantaneous frequency at each moment is ratioed to the original sampling frequency to dynamically determine the optimal window length for different time periods. A shorter window length is used for intervals with drastic frequency changes, while a longer window length is used for intervals with relatively stable frequencies to achieve adaptive adjustment of the window length. In each adaptive window, the signal is weighted by the Hanning function to reduce the spectrum leakage effect at the window edge.

[0030] Perform phase demodulation on the windowed signal to extract the instantaneous phase change characteristics within the window. At the same time, calculate the time derivative of the signal energy to obtain the energy change rate characteristics. Align the phase change rate and the energy change rate in the time domain to construct a two-dimensional feature sequence. By sliding the window, the two-dimensional feature sequences in the continuous time period are combined into a feature matrix. Each row of the matrix represents a feature combination in a time window, and each column represents the evolution of a specific feature over time.

[0031] Calculate the difference of the feature matrix along the time direction to obtain the change speed matrix of the features. Based on the magnitude of the change speed, use the exponential weighting method to calculate the importance weight of each sample point. The sample point with a larger change speed obtains a higher weight. Multiply the sample weights and the original feature matrix element by element to obtain the weighted feature matrix. Calculate the covariance of the weighted feature matrix to obtain the weighted covariance matrix reflecting the feature correlation.

[0032] Perform eigenvalue decomposition on the weighted covariance matrix, calculate the ratio between adjacent eigenvalues and the cumulative contribution rate of eigenvalues. By setting the contribution rate threshold and the interval ratio threshold, screen out the most representative eigenvectors. Project the original feature matrix onto this eigenvector to perform data dimensionality reduction while retaining key information. Reconstruct the dimensionality-reduced data through the inverse transformation of the eigenvector, and evaluate the reconstruction error to verify the dimensionality reduction effect. Construct a projection matrix with the verified eigenvectors.

[0033] Multiply the original feature matrix and the projection matrix to obtain the projection result in the main feature direction, that is, the flight feature vector. Establish a quadratic polynomial model considering the acceleration term for the feature vector to describe the nonlinear relationship between feature quantities. Use the iteratively reweighted least squares method to estimate the model parameters, where the sample weights are dynamically adjusted according to the residual size in each iteration. Based on the estimated model parameters, extend the prediction to the future time domain to obtain a series of predicted feature vectors.

[0034] Perform post-processing on the sequence of predicted feature vectors, calculate the phase angle difference and the amplitude ratio change between adjacent prediction points. Based on the statistical features in similar scenarios in historical data, determine the initial combined weights of the phase difference parameter and the amplitude change parameter. Weight and superimpose the two types of parameters to obtain the initial trajectory correction amount. By online tracking the change trend of the prediction error, use the gradient descent method to update the combined weight coefficient. Perform exponential smoothing on the trajectory correction amount, where the smoothing coefficient decreases as the prediction error increases and increases as the error decreases to achieve adaptive smoothing. Superimpose the smoothed correction amount on the predicted feature vector to obtain the corrected trajectory prediction data.

[0035] Exemplarily, track and record the hovering flight process of a goshawk within 30 seconds, with a sampling frequency of 50 Hz, and the obtained intrinsic mode components contain 1500 data points. Calculate the difference of the instantaneous phase and find that the hovering frequency of the goshawk in the updraft gradually decreases from 1.2 Hz to 0.8 Hz. Based on this frequency change feature, use a window length of 80 points in the initial stage with a higher frequency, and a window length of 120 points in the later stage when the frequency decreases. Within each window, it is extracted that the phase change amount fluctuates between 45 and 75 degrees, and the energy change rate changes between 0.08 and 0.25. Arrange these features in chronological order to construct a feature matrix with 1400 rows and 2 columns.

[0036] The time derivative of the feature matrix is calculated, and it is found that the feature change rate during the goshawk's flight attitude adjustment is 5 times that during stable hovering. Based on this, the sample weight during the attitude adjustment process is calculated to be 0.9, and the weight during the stable hovering process is 0.1. After weighting the feature matrix, the eigenvalue distribution of the covariance matrix is analyzed. The first eigenvalue accounts for 92% of the total energy, and the corresponding reconstruction error is only 3%. The eigenvector corresponding to this eigenvalue is selected as the basis for dimensionality reduction to construct a projection matrix.

[0037] After multiplying the original feature matrix by the projection matrix, a vector reflecting the main motion features of the goshawk is obtained. A quadratic trend model including velocity terms and acceleration terms is established. After five iterations of optimization, the prediction error converges to within 5% of the original amplitude. The motion feature sequence within the next 5 seconds is predicted, and it is calculated that the phase difference between adjacent moments varies between 15 and 30 degrees, and the amplitude change rate fluctuates within a range of 20%. The weight of the phase difference parameter is initialized to 0.7, and the weight of the amplitude change parameter is 0.3. When the prediction error exceeds 10%, the phase difference weight is adjusted to 0.8, and the amplitude change weight is adjusted to 0.2. According to the corrected prediction result, it is predicted that the goshawk will start a turning motion with a radius of about 50 meters after 2 seconds, and the turning process will last for about 3 seconds.

[0038] In this embodiment, through adaptive window processing, the analysis window can be dynamically adjusted according to the local characteristics of the signal, which not only retains the detailed information of rapid changes but also effectively suppresses noise interference, improving the accuracy of feature extraction. The weighted dimensionality reduction method based on sample importance weights highlights the flight characteristics at critical moments and avoids the problem that traditional equal-weight dimensionality reduction methods may ignore key turning points. The trajectory correction mechanism combining phase difference and amplitude change parameters can capture the subtle change trends of bird flight and continuously improve the prediction accuracy through dynamic weight optimization.

[0039] Figure 2 This is a comparison chart of the trajectory correction effects of the wind farm bird trajectory prediction and fan linkage control method based on neural network in the embodiment of the present invention, showing the change trends of the trajectory prediction errors of three different methods in the prediction time domain (0 - 5 seconds). This technical solution (phase-amplitude joint correction) is in Figure 2It is represented by circular markers in the figure and shows significant error control advantages throughout the prediction period. In the initial stage of prediction (0 - 1 second), the performance of the three methods is relatively close. The prediction error of this technical solution starts from 9% and reaches 11% at 1 second; while the Kalman filter method (classical state - space model filtering method, represented by square markers in the figure) and the LSTM prediction method (represented by triangular markers in the figure) are 11% - 14.5% and 9.5% - 12% respectively. In the middle stage of prediction (2 - 3 seconds), that is, between the "turning start point" and "turning completion point" marked in the figure, the prediction error of this technical solution increases relatively slowly, rising from 13.5% to 16.2%; while the errors of the Kalman filter method and the LSTM method increase rapidly to 18% - 23.5% and 17% - 21.5% respectively. This stage exactly corresponds to the complex flight interval where the goshawk starts to execute a turning maneuver with a radius of about 50 meters. This technical solution effectively captures the trajectory change characteristics during the turning process by dynamically adjusting the combined weights of the phase - difference parameter (weight adjusted from 0.7 to 0.8) and the amplitude - change parameter (weight adjusted from 0.3 to 0.2). In the late stage of prediction (4 - 5 seconds), the errors of all methods show an accelerating growth trend, but the error of this technical solution at the 5 - second prediction end - point is still controlled at 22.5%, which is 11 and 9.5 percentage points lower than 33.5% of the Kalman filter method and 32.0% of the LSTM method respectively, demonstrating the robustness of this technical solution in long - term prediction, especially its adaptability in the face of flight - trajectory mutations (such as turning).

[0040] In an alternative embodiment, Establish the time - series relationship of the quadratic - term local trend, and combine the iteratively reweighted least - squares method for time - series prediction to obtain a sequence of prediction feature vectors, including: Divide the flight feature vectors into short - term, medium - term, and long - term ones, construct a variational - constraint optimization objective function by calculating the sequence envelope and instantaneous frequency, iteratively update the central frequency and bandwidth parameters, and extract the adaptive multi - scale components with the optimal energy concentration degree. Perform dual - tree complex wavelet transform on the adaptive multi - scale components to obtain multi - level sparse time - frequency features, and combine time - series position encoding to generate an enhanced input sequence with time - series position information; Construct a multi - layer biological perception field network based on the enhanced input sequence. In each layer of the perception field, extract key feature patterns through excitatory neuron groups, use inhibitory neuron groups for selective enhancement, and combine adaptive dynamic thresholds to adjust the neuron activation state to obtain perception features. Combine the perception features in each layer of the perception field to obtain a hierarchical attention feature representation; Perform residual connection and layer normalization on the attention feature representation, perform dynamic weight allocation using an adaptive gating unit, fuse multi-scale features to obtain an enhanced feature representation with temporal dependence, construct a quadratic local trend model based on the enhanced feature representation, optimize and solve using the iteratively reweighted least squares method, construct a dynamic prediction confidence interval through multiple samplings, adjust the historical data weight and prediction credibility weight according to the interval width, and output a sequence of prediction feature vectors.

[0041] Perform time-scale decomposition on the flight feature vector, and divide the sequence into three time scales: short-term (within 10 seconds), medium-term (10 - 30 seconds), and long-term (more than 30 seconds). At each time scale, calculate the upper and lower envelope lines of the signal and extract the instantaneous frequency feature. Construct an optimization objective function based on the envelope line and frequency feature, and the function includes two constraint terms: signal energy concentration and frequency bandwidth. Through the iterative optimization method, continuously adjust the center frequency and frequency bandwidth parameters until the parameter combination that makes the objective function reach the optimal value is found. Use the optimal parameters to perform adaptive filtering on the original sequence to obtain multi-scale components with optimal energy concentration characteristics at different time scales.

[0042] Perform dual-tree complex wavelet transform on the obtained multi-scale components to achieve multi-level decomposition of the signal. In each layer of decomposition, extract the amplitude, phase, and frequency features of the signal, and construct a time-frequency feature map with sparse representation. Encode the temporal position information into a position vector, and the position vector reflects the relative position and context correlation relationship of the feature in the time series. Combine the time-frequency feature with the position encoding information to generate an enhanced input sequence containing temporal dependence.

[0043] Construct a multi-layer biological perception field network structure, and each layer of perception field includes an excitatory neuron group and an inhibitory neuron group. The excitatory neurons extract key feature patterns in the input sequence through parallel processing, including features such as amplitude mutation, frequency jump, and phase conversion. The inhibitory neurons selectively enhance the extracted features and suppress redundant and noise features. Set an adaptive dynamic threshold mechanism to dynamically adjust the activation state of the neurons according to the significance degree of the features. When the feature response exceeds the threshold, the corresponding neuron is activated and outputs a perception feature. Combine the perception features output by the activated neurons in each layer of perception field to form an attention feature representation with a hierarchical structure.

[0044] Perform post-processing on the hierarchical attention features, retain the original feature information through the residual connection mechanism to avoid feature loss during the training process of the deep network. Perform layer normalization operations to eliminate the bias of the feature distribution and improve the expression ability of the features. Construct an adaptive gating unit, and the adaptive gating unit dynamically allocates weight coefficients according to the importance degree of the features. Perform weighted fusion on features of different scales to obtain an enhanced feature representation that synthesizes information from multiple time scales.

[0045] Construct a quadratic local trend model considering the acceleration term based on enhanced feature representation. Use the iteratively reweighted least squares method to optimize the model parameters, and dynamically adjust the sample weights according to the fitting residuals in each iteration. Through multiple random samplings and model fittings, construct a dynamic confidence interval for the prediction results. Calculate the width of the confidence interval. When the interval is narrow, increase the weight of historical data to improve the stability of the prediction; when the interval is wide, increase the weight of the prediction credibility to enhance the adaptability of the model to new features. Output a sequence of predicted feature vectors that have been optimized multiple times.

[0046] Exemplarily, analyze the flight feature vectors of a gyrfalcon during the diving predation process, with a recording duration of 120 seconds. Divide the feature sequence into short-term scales of 10 seconds (reflecting the diving acceleration process), medium-term scales of 30 seconds (reflecting the attack preparation process), and long-term scales of 60 seconds (reflecting the hovering search process). Calculate that the energy proportion of the short-term scale is 60%, the medium-term scale is 30%, and the long-term scale is 10%. Determine the optimal center frequencies through iterative optimization, which are 5 Hz for the short-term scale, 2 Hz for the medium-term scale, and 0.5 Hz for the long-term scale.

[0047] Perform a dual-tree complex wavelet transform on the optimized multi-scale components to extract key time-frequency features such as the velocity mutation feature (amplitude change reaching 80%) during the diving process and the phase jump feature (phase change exceeding 90 degrees) during the turning process. Combine these features with the position encoding representing 120 time points to generate an enhanced input sequence.

[0048] In a three-layer biological perception field network, 100 excitatory neurons in the first layer mainly extract velocity and acceleration features, and 50 inhibitory neurons weaken background noise; 80 excitatory neurons in the second layer extract turning and hovering features, and 40 inhibitory neurons enhance the main motion patterns; 60 excitatory neurons in the third layer extract attack intention features, and 30 inhibitory neurons strengthen the predation behavior patterns. When the feature response exceeds the adaptive threshold (set to 1.5 times the average response intensity), the corresponding neurons are activated.

[0049] Retain the original motion features through residual connections, and after layer normalization, the standard deviation of the feature distribution is reduced to less than 0.1. The adaptive gating unit assigns 60% of the weight to short-term features, highlighting the key behaviors of diving predation. The constructed quadratic trend model is optimized through 5 rounds of iteration, and the prediction error converges to within 5%. Construct a confidence interval through 1000 random samplings. When the interval width is less than 10% of the predicted value, increase the weight of historical data to 0.8; when the interval width is greater than 30%, increase the weight of the prediction credibility to 0.6. The prediction results show that the gyrfalcon will start diving after 2 seconds, with a diving angle of approximately 60 degrees and completing the predation action in approximately 3 seconds.

[0050] In this embodiment, by constructing an optimization objective function that includes energy aggregation degree and frequency bandwidth constraints, the adaptive extraction of features at different time scales is realized. Through the collaborative action of excitatory and inhibitory neurons, the active extraction of key features and the adaptive suppression of noise are achieved. The effective transmission of deep features is ensured through residual connection and layer normalization operations, and the dynamic fusion of multi-scale features is realized through an adaptive gating mechanism; In the prior art, the prediction of bird flight trajectories mainly uses a feature extraction method with a single time scale. Feature extraction is performed through sliding analysis with a fixed window size or simple frequency domain transformation. It is difficult to capture flight features at different time scales simultaneously, resulting in insufficient accuracy when dealing with long-term sequence prediction and lack of sensitivity in predicting short-term sudden behaviors. The extraction of features is often limited to a single dimension, and the multi-level information contained in the time series data is not fully utilized, making it difficult for the prediction results to adapt to the complex and changeable flight patterns of birds; This embodiment introduces a dual-tree complex wavelet transform and a time series position encoding mechanism, effectively retaining the time-frequency characteristics and time series dependence relationship of features. A dynamic threshold mechanism is used to adjust the activation state of neurons, improving the adaptability and robustness of feature extraction. Combining the construction of a dynamic confidence interval and adaptive weight adjustment, a complete feature prediction framework is established, which not only improves the accuracy of prediction but also enhances the adaptability of the algorithm to complex flight patterns, providing reliable technical support for the intelligent upgrade of the bird hazard avoidance system in wind farms.

[0051] Figure 3This is the multi-scale decomposition effect diagram of the flight characteristics of the wind farm bird trajectory prediction and fan linkage control method based on neural network in the embodiments of the present invention. The original experimental data was collected in the northern grassland area (altitude 1200m), and a high-precision data sampling rate of 120Hz was used to ensure that the minute changes during the high-speed flight of birds were captured. The analysis of the original data shows that the target bird was in the search and hovering stage from 0 to 30 seconds, with an average altitude maintained at 180m; it entered the dive preparation stage from 30 to 60 seconds, and the angle was gradually adjusted to 42°; it completed the predation attack from 60 to 90 seconds, and the maximum speed reached 68m / s. Different line types are used in the figure to distinguish each algorithm: the solid line represents the technical solution of the present invention, the short dashed line represents the EMD decomposition method, the dotted line represents the wavelet decomposition method, and the long dashed line represents the STFT analysis method. The comparison data shows that the trajectory reconstruction accuracy of the technical solution of the present invention is 92.4%, which is significantly higher than that of the EMD decomposition method (78.6%), the wavelet decomposition method (82.1%) and the STFT analysis method (65.7%). In terms of the energy aggregation index, the technical solution of the present invention reaches 85.3%, the EMD decomposition method is 67.8%, the wavelet decomposition method is 64.5%, and the STFT analysis method is 54.2%. In the short-term scale (within 10 seconds) analysis, the feature extraction accuracy of the technical solution of the present invention reaches 94.2%, and the processing delay is only 35ms; in the medium-term scale (10 - 30 seconds) analysis, the extraction accuracy is 88.7%, and the noise suppression rate reaches 22.2dB; in the long-term scale (more than 30 seconds) analysis, the extraction accuracy is 84.3%, and the feature separation degree reaches 0.86. The original data fully proves that the technical solution of the present invention is superior to the existing methods in all core indicators of flight feature extraction, laying a solid foundation for high-precision flight behavior prediction.

[0052] In an alternative embodiment, A collaborative avoidance control algorithm is constructed based on the artificial potential field method, with the bird target as the repulsive force source and the wind turbine generator set as the repulsive force field. The potential field gradient is calculated according to the relative distance between the bird target and the wind turbine generator set, including: A first repulsive force potential energy function is established with the bird target as the repulsive force source. The first repulsive force potential energy function is determined by the repulsive force field coefficient, the relative distance of the bird target, and the influence radius of the repulsive force field. When the relative distance of the bird target is less than or equal to the influence radius of the repulsive force field, the first repulsive force potential energy function is positively correlated with the repulsive force field coefficient, the relative distance of the bird target, and the influence radius of the repulsive force field. When the relative distance of the bird target is greater than the influence radius of the repulsive force field, the first repulsive force potential energy function is zero; Taking the wind turbine generator set as a repulsive force field, a second repulsive potential energy function is established. The second repulsive potential energy function is determined by a repulsive force field coefficient, the relative distance of the wind turbine generator set, and the influence radius of the repulsive force field. When the relative distance of the wind turbine generator set is less than or equal to the influence radius of the repulsive force field, the second repulsive potential energy function is positively correlated with the repulsive force field coefficient, the relative distance of the wind turbine generator set, and the influence radius of the repulsive force field. When the relative distance of the wind turbine generator set is greater than the influence radius of the repulsive force field, the second repulsive potential energy function is zero; Taking the negative gradients of the first repulsive potential energy function and the second repulsive potential energy function respectively according to the relative distance between the bird target and the wind turbine generator set to obtain the potential field gradients.

[0053] Based on the position of the bird target, a first repulsive potential energy function is established, and the bird target is regarded as a repulsive force source point. First, the basic parameters of the repulsive force field are determined, including the repulsive force field coefficient describing the repulsive force intensity, the relative distance characterizing the distance influence, and the influence radius defining the action range. When the relative distance between the bird target and the point to be calculated is within the influence radius, the repulsive potential energy increases with the increase of the repulsive force field coefficient and decreases with the increase of the relative distance, and is modulated by the influence radius at the same time. When the relative distance exceeds the influence radius, the repulsive potential energy rapidly decays to zero, indicating that this position is no longer affected by the repulsive force of the bird target.

[0054] Similarly, a second repulsive potential energy function is established for the wind turbine generator set, and the generator set is regarded as a fixed repulsive force field source. The basic parameters of the repulsive force field of the generator set are set, including the repulsive force field coefficient reflecting the danger level of the generator set, the relative distance characterizing the spatial position relationship, and the influence radius determining the protection range. Within the influence radius of the generator set, the repulsive potential energy is positively correlated with the repulsive force field coefficient and the influence radius, and negatively correlated with the relative distance, jointly describing the intensity of the repulsive effect of the generator set on the surrounding space. When the distance exceeds the influence radius, the repulsive potential energy drops to zero, indicating that this area is not affected by the danger of the generator set.

[0055] The negative gradients of the two repulsive potential energy functions are calculated respectively to obtain the potential field gradients pointing in the direction of decreasing repulsive potential energy. For the repulsive potential energy function of the bird target, calculate its change rate in all directions in space and take the negative value to obtain the first potential field gradient, which points in the direction away from the bird target. Similarly, calculate the negative gradient of the repulsive potential energy function of the wind turbine generator set to obtain the second potential field gradient, which points in the direction away from the generator set. These gradient vectors jointly construct the potential field distribution in space for subsequent obstacle avoidance path planning.

[0056] Exemplarily, it is monitored that a seagull is approaching a wind farm, and its position is 200 meters away from the nearest wind turbine. The repulsive force field parameters of the seagull are set as follows: the repulsive force field coefficient is set to 100, which characterizes its activity ability; the influence radius is set to 300 meters, considering the seagull's warning range. When the relative distance between the seagull and a certain place is 150 meters, it is within the influence radius. At this time, the calculated repulsive potential energy value gradually increases as it approaches the seagull. When the relative distance increases to 400 meters, it exceeds the influence radius, and the repulsive potential energy drops to zero, indicating that this position is not affected by the seagull's activities.

[0057] For the wind turbine, its repulsive force field parameters are set as follows: the repulsive force field coefficient is set to 200, which reflects its threat level to birds; the influence radius is set to 400 meters, considering the dangerous range of blade rotation. When a certain place is 250 meters away from the wind turbine, within the influence radius, a relatively large repulsive potential energy value is calculated, indicating that this area has a high level of danger. When the distance increases to 500 meters, it exceeds the influence radius, and the repulsive potential energy drops to zero, indicating that this position can be safely passed through.

[0058] Taking the negative gradient of the above two repulsive potential energy functions, a gradient vector pointing away from the wind turbine direction is obtained at the current position of the seagull, and its magnitude is proportional to the change rate of the repulsive potential energy. The gradient information indicates the safe movement direction, which can guide the seagull away from the dangerous area of the wind turbine and fly along the path with the minimum potential energy.

[0059] In this embodiment, zoning control is achieved by setting the influence radius of the repulsive force field. Only when birds enter the dangerous area is the avoidance measure activated, avoiding unnecessary intervention, maximizing the power generation efficiency of the wind farm, regarding the wind turbine itself as a source of the repulsive force field, considering the mutual influence between wind turbines, and avoiding the problem of interference between wind turbines that may occur during the obstacle avoidance process.

[0060] In an alternative embodiment, the potential field gradient is optimized online through an adaptive fuzzy neural network, generating avoidance control parameters and combining with a sliding mode controller to generate obstacle avoidance parameters and combining them to obtain an obstacle avoidance control strategy, including: Receiving the pre-acquired potential field gradient, inputting the potential field gradient into an adaptive fuzzy neural network with an input layer, a fuzzification layer, a rule layer, a rule inference layer, and an output layer, and using a Gaussian membership function in the fuzzification layer to perform fuzzification processing on the potential field gradient to obtain fuzzified data; Based on the network output error, using the gradient descent method to determine the update amounts of the center value parameter and the width parameter of the membership function, online optimizing the center value parameter and the width parameter of the membership function, and inputting the optimized fuzzification parameters into the rule layer to obtain the rule layer output vector; Perform matrix operation on the output vector of the rule layer and the network weight matrix to obtain the network compensation amount, superimpose the network compensation amount and the potential field gradient to obtain the optimized potential field gradient, and generate avoidance control parameters based on the change rate of the optimized potential field gradient; Determine the position tracking error based on the avoidance control parameters, construct a sliding mode surface and a non-singular terminal sliding mode controller, perform finite-time convergence through the control law structure combined with piecewise power functions, and input the optimized potential field gradient into the non-singular terminal sliding mode controller to generate obstacle avoidance parameters; Determine the dynamic weight factor based on the position tracking error, and use the dynamic weight factor to perform weighted combination of the obstacle avoidance parameters and the output of the output layer of the adaptive fuzzy neural network to obtain the obstacle avoidance control strategy.

[0061] Input the pre-acquired potential field gradient into the adaptive fuzzy neural network. The structure of the adaptive fuzzy neural network consists of five layers, including an input layer, a fuzzification layer, a rule layer, a rule inference layer, and an output layer. The input layer receives the potential field gradient in three-dimensional space and inputs the eastward component, northward component, and vertical component into the corresponding neurons respectively. Each input neuron uses a linear activation function to maintain the original feature attributes of the input signal.

[0062] In the fuzzification layer, multiple Gaussian membership functions are set for each gradient component for fuzzification mapping. Five different levels of linguistic variables are set for each component, including "very small", "small", "medium", "large", and "very large". Each linguistic variable corresponds to a Gaussian function with specific center value and width parameters. Through these Gaussian functions, the exact gradient value is converted into a fuzzy membership value with uncertainty to form fuzzified data.

[0063] Calculate the error between the actual output and the expected output of the adaptive fuzzy neural network, and construct a mean square error objective function. Based on the mean square error objective function, use the gradient descent method to calculate the update amount of the Gaussian membership function parameters. Calculate the gradients of the error with respect to the center value parameter and the width parameter respectively to determine the direction and step size of parameter update. Perform online iterative optimization on these parameters according to the preset learning rate until the change amount of the parameters is less than the set threshold. Apply the optimized fuzzification parameters to the calculation process of the rule layer to generate the output vector of the rule layer reflecting the activation degree of each rule.

[0064] Perform matrix operation on the output vector of the rule layer and the network weight matrix. The elements in the weight matrix represent the contribution degree of different rules to the final output. The result of the matrix operation is used as the compensation amount for the original potential field gradient. Superimpose this compensation amount and the original potential field gradient vectorially to obtain the optimized potential field gradient. Based on the time change rate of the optimized potential field gradient, generate avoidance control parameters including direction adjustment amount and speed change amount.

[0065] Calculate the tracking error vector between the current position and the desired safe position based on the avoidance control parameters. Use this error vector as the state variable to construct a sliding mode surface, and design a controller using the nonsingular terminal sliding mode control method. The control law of the controller adopts a structural form combined with piecewise power functions. High-order power terms are used in the large deviation region to achieve fast convergence, and low-order power terms are used in the small deviation region to ensure smooth transition. A smooth transition function is introduced near zero to avoid singularity problems. The control law structure ensures that the system can converge to the equilibrium state within a finite time. Input the optimized potential field gradient into the nonsingular terminal sliding mode controller, and generate specific obstacle avoidance parameters through the action of the control law.

[0066] Establish a dynamic weight allocation mechanism based on the magnitude of the position tracking error. Design a continuous weight function. When the tracking error is large, dynamically increase the weight of the neural network output to enhance the adaptive ability to the environment; when the tracking error is small, dynamically increase the weight of the obstacle avoidance parameters to ensure the stability of the control process. Use this dynamic weight factor to perform weighted combination of the obstacle avoidance parameters and the output of the output layer of the adaptive fuzzy neural network, avoid sudden changes in the control quantity, achieve smooth switching, and finally obtain an optimized obstacle avoidance control strategy.

[0067] Exemplarily, it is detected that an egret is approaching at a distance of 300 meters from the wind turbine, and the obtained potential field gradient shows that avoidance is required in the northeast direction. Input this gradient into a five-layer fuzzy neural network, and use a Gaussian function to fuzzify the gradient in the fuzzification layer. Initially, set the center value of the Gaussian function to 0.5 and the width parameter to 0.3 to perform a preliminary fuzzy mapping on the potential field gradient.

[0068] By comparing the error between the network output and the desired obstacle avoidance effect, it is found that the fuzzification effect under the initial parameters is not ideal enough. Use the gradient descent method to optimize the parameters. After multiple iterations, adjust the center value to 0.65 and the width parameter to 0.25, which significantly improves the fuzzification effect. Use the optimized parameters for rule calculation to obtain the output of the rule layer representing the importance degree of each obstacle avoidance rule.

[0069] Perform matrix operation on the output of the rule layer and the network weights to obtain a compensation amount indicating that the deflection angle needs to be increased by 10 degrees in the east direction. Superimpose this compensation amount with the original gradient, and the optimized potential field gradient points to the northeast by east direction. Generate initial obstacle avoidance parameters based on the gradient change rate, including the deflection angle and the speed adjustment amount.

[0070] Construct a sliding mode surface according to the deviation between the current position of the egret and the desired safe position, and design a piecewise control law to adjust the obstacle avoidance process. Input the optimized potential field gradient into the sliding mode controller to generate specific obstacle avoidance parameters, such as control quantities such as a 15% reduction in the wind turbine speed and a 20-degree deflection of the blade.

[0071] Calculate the position tracking error in real time. When the error is large, increase the weight of the neural network output to 0.7. When the error is small, increase the weight of the obstacle avoidance parameter to 0.6. Through the adjustment of dynamic weights, the obstacle avoidance parameter and the network output are weighted and combined to finally generate an optimized obstacle avoidance control strategy, guiding the fan to make adjustments in the expected manner to ensure the safe avoidance of egrets.

[0072] In this embodiment, by integrating the language processing ability of fuzzy logic and the learning ability of neural networks, the uncertainties and ambiguities existing in the system can be processed. Through the online parameter optimization mechanism, the system can continuously learn and improve. As the running time increases, the control accuracy is continuously improved. The introduction of the nonsingular terminal sliding mode controller solves the possible stability problems of traditional control methods and ensures that the system converges to the target state within a finite time.

[0073] In an alternative embodiment, Based on the avoidance control parameter, determine the position tracking error and construct a sliding mode surface and a nonsingular terminal sliding mode controller. Perform finite-time convergence through the control law structure combined with piecewise power functions. The optimized potential field gradient is input into the nonsingular terminal sliding mode controller to generate the obstacle avoidance parameter, including: Construct a position tracking error mapping function based on the avoidance control parameter. When the absolute value of the avoidance control parameter is greater than or equal to the threshold parameter, use the product of the first mapping coefficient and the nonlinear power term of the avoidance control parameter minus the reference control quantity as the position tracking error output. When the absolute value of the avoidance control parameter is less than the threshold parameter, use the product of the second mapping coefficient and the sine function based on the ratio of the avoidance control parameter to the threshold parameter as the position tracking error output; Construct a nonlinear sliding mode surface containing dynamic weights based on the position tracking error. The nonlinear sliding mode surface consists of the derivative term of the position tracking error, the product term of the dynamic weight factor and the position tracking error, and the nonlinear power integral term. The dynamic weight factor includes a basic weight and an exponential decay adaptive term based on the position tracking error; Design a piecewise continuous control law based on the nonlinear sliding mode surface. When the absolute value of the nonlinear sliding mode surface is greater than or equal to the switching domain threshold, use the combination of two different power terms as the control output. When the absolute value of the nonlinear sliding mode surface is less than the switching domain threshold, use the combination of a linear term and a fractional power term as the control output; Use the product of the Sigmoid function and the optimized potential field gradient constraint function as the smooth transition function of the nonlinear sliding mode surface, and superimpose the control output with the product of the smooth transition function and the optimized potential field gradient to obtain the obstacle avoidance parameter.

[0074] A piecewise mapping function of the position tracking error is established based on the avoidance control parameter. The absolute value of the avoidance control parameter is obtained and compared with a preset threshold parameter. When the absolute value of the avoidance control parameter is greater than or equal to the threshold parameter, the difference between the avoidance control parameter and the reference control quantity is calculated, substituted into a nonlinear power function for transformation, and the transformation result is multiplied by the first mapping coefficient to obtain the position tracking error output in this region. When the absolute value of the avoidance control parameter is less than the threshold parameter, the ratio of the avoidance control parameter to the threshold parameter is calculated, substituted into a sine function, and the function output is multiplied by the second mapping coefficient to obtain the position tracking error output in this region.

[0075] A nonlinear sliding mode surface is constructed using the position tracking error output by the mapping function. The derivative term of the position tracking error is calculated as the first component of the sliding mode surface. A dynamic weight factor is constructed, which consists of a base weight and an adaptive term. The base weight is a preset fixed value, and the adaptive term is generated using an exponential decay function based on the position tracking error. The dynamic weight factor is multiplied by the position tracking error as the second component of the sliding mode surface. A nonlinear power integral operation is performed on the position tracking error, and the resulting value is used as the third component of the sliding mode surface. The above three components are combined in a preset manner to construct a complete nonlinear sliding mode surface.

[0076] A piecewise continuous control law is designed based on the nonlinear sliding mode surface. The absolute value of the nonlinear sliding mode surface is obtained and compared with the switching domain threshold. When the absolute value of the nonlinear sliding mode surface is greater than or equal to the switching domain threshold, two different power exponents are selected, and the sliding mode surface is substituted into these two power terms for calculation, and the calculation results are combined according to a preset weight to obtain the control output in this region. When the absolute value of the nonlinear sliding mode surface is less than the switching domain threshold, the sliding mode surface is substituted into a linear function to obtain a linear term output, and at the same time substituted into a fractional power function to obtain a nonlinear term output, and these two terms are combined according to a preset weight to obtain the control output in this region.

[0077] A smooth transition function is constructed, with the Sigmoid function as the basic transition function. A potential field gradient constraint function is designed, which determines the constraint boundary according to the optimized potential field gradient characteristics. The Sigmoid function is multiplied by the potential field gradient constraint function to obtain the smooth transition function of the nonlinear sliding mode surface. The product of the control output, the smooth transition function, and the optimized potential field gradient is calculated respectively. This product is superimposed with the previously obtained control output to obtain the obstacle avoidance parameter.

[0078] Exemplarily, a white-tailed sea eagle is detected approaching the wind farm, and the current avoidance control parameter is a deflection angle of 30 degrees. The set threshold parameter is 20 degrees, and the reference control quantity is 0 degrees. Since the absolute value of the avoidance control parameter, 30 degrees, is greater than the threshold of 20 degrees, the first mapping coefficient of 0.8 and the non-linear power term are used to calculate the position tracking error, and the tracking error is obtained as 45 meters.

[0079] Based on the position tracking error of 45 meters, a non-linear sliding mode surface is constructed. The basic weight is set to 0.6. When the position tracking error is 45 meters, the calculation result of the exponential decay adaptive term is 0.2, and the dynamic weight factor is obtained as 0.8. At the same time, the derivative term of the position tracking error and the non-linear power integral term are calculated and combined to form a complete non-linear sliding mode surface.

[0080] When the absolute value of the calculated non-linear sliding mode surface is 0.9, which is greater than the switching domain threshold of 0.5, therefore, a combination of two terms with powers of 2 and 1.5 is used as the control output. The Sigmoid function is multiplied by the potential field gradient constraint function to obtain a smooth transition function, and the output value of this function at the current position is 0.95. The control output is superimposed with the product of the smooth transition function and the optimized potential field gradient, and the final obtained obstacle avoidance parameters are a deflection angle of 25 degrees and a safety distance of 100 meters.

[0081] In this embodiment, non-linear power term mapping is adopted in the large deviation region to provide a greater control force; sine function mapping is adopted in the small deviation region to achieve smooth transition. By combining the basic weight with the exponential decay adaptive term, the dynamic adjustment of the weight factor is realized. The non-linear power integral term is introduced into the sliding mode surface to enhance the system's ability to suppress disturbances. A piecewise continuous control law structure is adopted, and appropriate control term combinations are selected in different regions; In the prior art, the obstacle avoidance control of the flight trajectory mainly adopts simple linear mapping or piecewise control methods with fixed thresholds, which is difficult to balance the rapidity and smoothness of control. Traditional sliding mode controllers usually adopt linear sliding mode surfaces with fixed weights and cannot flexibly adjust the control characteristics according to the change of the tracking error. The control quantity is prone to jump near the switching point of the control law, causing severe oscillation of the system and affecting the stability of the obstacle avoidance control; This embodiment improves the system's ability to handle different degrees of deviation through the non-linear piecewise mapping mechanism. The design of the dynamic weight non-linear sliding mode surface enhances the self-adaptability and robustness of the control system. The design of the piecewise continuous control law improves the control performance of the system in different working regions, enhances the self-adaptive ability and reliability of the control system, and provides more effective technical support for the bird obstacle avoidance control of the wind farm.

[0082] Figure 4This is a simulation diagram of the obstacle avoidance technical effect of the wind farm bird trajectory prediction and fan linkage control method based on neural network in the embodiments of the present invention. The simulation data is based on high-precision trajectory control experiments in a complex obstacle environment. The experimental scenario is configured with three obstacles of different sizes, located at coordinates (500m, 400m), (900m, 300m), and (700m, 600m) respectively. The radius range of the obstacles is 70 - 100m, and the potential field influence range is set to 3 times the radius of the obstacles. The control target navigates from the starting point (100m, 200m) to the ending point (1300m, 400m), with a total distance of approximately 1400m.

[0083] In the figure, different colors are used to distinguish each control method. The green trajectory represents the technical solution of the present invention (nonsingular terminal sliding mode controller), the yellow trajectory represents the traditional control method (such as PID control), the red solid circles represent obstacles, the light red area represents the potential field influence range, and the black concentric circles represent the trajectory oscillation points of the traditional method. The blue dots mark the starting point, the ending point, and the key control turning points. Both the horizontal and vertical coordinates are in meters (m), clearly marking the spatial position relationship.

[0084] The comparison data shows that the trajectory smoothness of the technical solution of the present invention reaches 93.5%, significantly higher than 61.8% of the traditional control method. In terms of the safety margin index, the stability of maintaining a 100 - meter safety distance of the technical solution of the present invention is 96.2%, while the traditional method is only 72.4%. In the time - domain analysis, the control convergence time of the technical solution of the present invention is 2.7 seconds, which is 40% less than 4.5 seconds of the traditional method; in terms of energy efficiency, the control energy consumption of the technical solution of the present invention is reduced by 37%. The spatial trajectory analysis shows that the obstacle avoidance steering angles of the technical solution of the present invention at the key positions of the three obstacles are 30°, 15°, and 25° respectively, the average value of the position tracking error at each location is 45 meters, and the maximum value of the sliding mode surface is 0.9, far lower than the 78 - meter error and the 1.6 sliding mode surface value of the traditional method.

[0085] In terms of the control stability index, the adaptive adjustment range of the dynamic weight factor of the technical solution of the present invention is 0.6 - 0.8, effectively suppressing system oscillation; the traditional method shows obvious oscillation at two obstacle avoidance points (the black concentric circle area in the figure), and the oscillation amplitude reaches 40 - 60m. The analysis of the sliding mode surface construction efficiency shows that the switching characteristics of the technical solution of the present invention enable the system to respond to obstacles in only 58% of the time of the traditional method, capable of achieving a more rapid obstacle avoidance maneuver.

[0086] The experimental data proves that the nonsingular terminal sliding mode controller based on the combined control law of piece - wise power functions is significantly superior to the traditional control method in core indicators such as trajectory smoothness, obstacle avoidance safety, control convergence speed, and energy efficiency, providing an efficient and reliable technical solution for high - precision obstacle avoidance control in complex environments.

[0087] In an alternative embodiment, Execute the obstacle avoidance control strategy and continuously monitor the operating state of the adjusted wind turbine until the bird target flies out of the wind farm. Restoring the wind turbine to its normal operating state according to the preset restoration strategy includes: Execute the obstacle avoidance control strategy to adjust the wind turbine, and continuously monitor the operating state of the wind turbine and the flight trajectory of the bird target; When the bird target flies out of the wind farm, obtain the current operating state parameters of the wind turbine, and gradually adjust the rotational speed, torque, and pitch angle of the wind turbine to the corresponding target values in different restoration slopes according to the operating state parameters and the preset restoration threshold to restore the normal operating state.

[0088] According to the generated obstacle avoidance control strategy, adjust the key operating parameters of the wind turbine, such as rotational speed, torque, and pitch angle, in real time. During the execution of the obstacle avoidance control, continuously collect the operating state data of the wind turbine, including the actual rotational speed, output torque, pitch angle, and power output parameters. At the same time, track and monitor the flight trajectory of the bird target through radar and vision sensors, and record the changes in its position, speed, and flight direction.

[0089] When it is monitored that the bird target has completely flown out of the wind farm range, immediately obtain the operating state parameters of the wind turbine at this time. Compare the obtained operating state parameters with the target parameter values during normal operation, and calculate the deviation of each parameter. Read the preset restoration threshold, which is set separately according to the adjustment characteristics of different parameters and is used to divide different stages of parameter restoration adjustment.

[0090] Adopt a three-stage progressive restoration strategy for the restoration adjustment of the rotational speed of the wind turbine. When the rotational speed deviation is greater than the first restoration threshold, use a smaller restoration slope for adjustment to avoid violent fluctuations in the system. When the rotational speed deviation is between the first restoration threshold and the second restoration threshold, use a medium restoration slope for adjustment to accelerate the restoration speed. When the rotational speed deviation is less than the second restoration threshold, use a larger restoration slope to quickly reach the target rotational speed.

[0091] Adopt a two-stage progressive adjustment strategy for the restoration of the generator output torque. When the torque deviation is greater than the restoration threshold, use a smaller restoration slope for slow adjustment to ensure that the mechanical stress of the unit does not change suddenly. When the torque deviation is less than the restoration threshold, use a larger restoration slope for faster adjustment to quickly restore to the rated torque output state.

[0092] A segmented continuous recovery strategy is adopted for the pitch angle adjustment of the wind turbine generator set. When the pitch angle deviation is greater than the first recovery threshold, the blade angle is finely adjusted with the minimum recovery slope. When the pitch angle deviation is between the two recovery thresholds, a medium recovery slope is adopted for adjustment. When the pitch angle deviation is less than the second recovery threshold, a larger recovery slope is adopted to quickly adjust to the target angle. During the entire adjustment process, the vibration state of the blade is monitored in real time to ensure that the adjustment process is stable and controllable.

[0093] During the process of performing parameter recovery adjustment, the change trends of various parameters are continuously monitored. When a certain parameter reaches its target value, the parameter is locked in time, and other parameters that have not reached the target value are continued to be adjusted. Through this segmented progressive adjustment method, the smooth recovery of various operating parameters of the wind turbine generator set is realized, and finally the unit returns to the normal operating state.

[0094] Exemplarily, it is monitored that a sea eagle flies into the wind farm from the east side. The obstacle avoidance control reduces the fan speed to 70% of the rated speed, increases the pitch angle to the obstacle avoidance position, and reduces the output torque to 60% of the rated value. When the sea eagle flies away from the wind farm, the current operating parameters of the fan are obtained, and the recovery processes of the speed, torque, and pitch angle are respectively divided into different adjustment stages. The speed recovery adopts a three-stage adjustment. When the speed deviation is greater than 20% of the rated value, it is increased by 2% per second; when the deviation is between 20% and 10%, it is increased by 3% per second; when the deviation is less than 10%, it is increased by 5% per second until the rated speed is reached. The torque recovery adopts a two-stage adjustment. When the torque deviation is greater than 15% of the rated value, it is increased by 3% per second; when the deviation is less than 15%, it is increased by 6% per second. The recovery of the pitch angle also adopts a segmented adjustment strategy to smoothly restore all parameters to the normal operating state.

[0095] In this embodiment, the segmented progressive recovery method ensures the smooth transition of the wind turbine generator set from the obstacle avoidance state to the normal operating state, avoids mechanical shocks and grid fluctuations that may be caused by parameter mutations, and optimizes the time and safety of the recovery process by setting different recovery slopes for different parameters. The fully automated recovery process does not require manual intervention, greatly improving the system's response speed and work efficiency.

[0096] In the second aspect of the embodiment of the present invention, an electronic device is provided, including: A processor; a memory for storing instructions executable by the processor, wherein the processor is configured to call the instructions stored in the memory to execute the method described above.

[0097] In the third aspect of the embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0098] The present invention may be a method, an apparatus, a system, and / or a computer program product. The computer program product may include a computer-readable storage medium having thereon computer-readable program instructions for performing various aspects of the present invention.

[0099] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some or all of the technical features. And these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting bird trajectories in a wind farm and controlling the linkage of wind turbines based on a neural network, characterized in that, Including: Collecting real-time image data and using a pre-trained object detection neural network to identify bird targets, and generating bird flight trajectory data according to the position information of the bird targets in consecutive multiple frames of images; Performing three-dimensional Hilbert-Huang transform on the bird flight trajectory data to obtain an instantaneous frequency spectrum, combining an adaptive decomposition algorithm to decompose the instantaneous frequency spectrum into intrinsic mode components, extracting amplitude-frequency characteristic parameters, performing feature dimensionality reduction and reconstruction, obtaining a flight feature vector, performing time series analysis, generating a predicted feature vector sequence, and calculating a trajectory correction amount according to phase difference and amplitude change to obtain predicted trajectory data; Determining the wind turbines in the wind farm with bird collision risks according to the predicted trajectory data; Constructing a cooperative avoidance control algorithm based on the artificial potential field method, using the bird target as the repulsive source and the wind turbine as the repulsive field, calculating the potential field gradient according to the relative distance between the bird target and the wind turbine, online optimizing the potential field gradient through an adaptive fuzzy neural network, generating avoidance control parameters, combining with a sliding mode controller to generate obstacle avoidance parameters, and combining them to obtain an obstacle avoidance control strategy; Executing the obstacle avoidance control strategy and real-time monitoring the operating state of the adjusted wind turbine until the bird target flies out of the wind farm, and restoring the wind turbine to the normal operating state according to the preset restoration strategy.

2. The method according to claim 1, wherein Performing three-dimensional Hilbert-Huang transform on the bird flight trajectory data to obtain an instantaneous frequency spectrum, and combining an adaptive decomposition algorithm to decompose the instantaneous frequency spectrum into intrinsic mode components includes: Performing three-dimensional Hilbert-Huang transform on the bird flight trajectory data to obtain analytical signals in three coordinate axis directions, calculating the instantaneous amplitude and instantaneous phase in three coordinate axis directions based on the analytical signals, calculating the instantaneous frequency according to the time derivative of the instantaneous phase, and combining the instantaneous frequencies in different coordinate axis directions to obtain an instantaneous frequency spectrum; Taking the instantaneous frequency spectrum as an initial residual signal, using an adaptive decomposition algorithm to determine all local extreme points of the initial residual signal, constructing an upper envelope and a lower envelope of the initial residual signal by using a cubic spline interpolation method, calculating the mean value of the upper envelope and the lower envelope to obtain a mean envelope, and subtracting the mean envelope from the initial residual signal to obtain a detail component; Judging whether the difference between the number of zero points and the number of extreme points of the detail component is less than or equal to 1 and whether the mean envelope of the detail component is approximately zero. If so, determining the detail component as an intrinsic mode component. If not, taking the detail component as a new residual signal and repeating the adaptive decomposition algorithm until an intrinsic mode component is obtained.

3. The method according to claim 1, wherein Extracting amplitude-frequency characteristic parameters, performing feature dimensionality reduction and reconstruction, obtaining a flight feature vector, performing time series analysis, generating a predicted feature vector sequence, and calculating a trajectory correction amount according to phase difference and amplitude change to obtain predicted trajectory data includes: The instantaneous frequency is obtained by calculating the first derivative and the second derivative of the instantaneous phase of the pre-acquired intrinsic mode components. The window length of the Hanning window function is determined according to the ratio of the instantaneous frequency to the sampling frequency, and the intrinsic mode components are adaptively windowed. The processed instantaneous phase and the energy change rate are extracted to obtain amplitude-frequency characteristic parameters, which are arranged in time series to construct a characteristic matrix; The time derivative of the characteristic matrix is calculated to obtain the characteristic change rate, and the sample importance weights are calculated. The characteristic matrix is weighted based on the sample importance weights to obtain a weighted covariance matrix. The first eigenvector is determined based on the eigenvalue interval ratio, the cumulative contribution rate, and the reconstruction error of the weighted covariance matrix. The characteristic matrix is dimension-reduced and reconstructed based on the first eigenvector to obtain a projection matrix; The characteristic matrix is multiplied by the projection matrix to obtain a flight characteristic vector. The time series relationship of the quadratic local trend is established, and the time series prediction is performed in combination with the iteratively reweighted least squares method to obtain a sequence of predicted characteristic vectors; The phase difference parameter and the amplitude change parameter at adjacent moments are calculated according to the sequence of predicted characteristic vectors. The combined weight coefficients of the phase difference parameter and the amplitude change parameter are initialized based on the historical trajectory data and weighted and summed. The weighted sum is used as the trajectory correction amount. The combined weight coefficients are updated according to the prediction error. The smoothing correction amount is calculated according to the trajectory correction amount, and the smoothing factor is dynamically adjusted. The smoothing factor is combined with the sequence of predicted characteristic vectors to obtain the predicted trajectory data.

4. The method according to claim 3, characterized in that Establishing the time series relationship of the quadratic local trend and performing time series prediction in combination with the iteratively reweighted least squares method to obtain a sequence of predicted characteristic vectors includes: The flight characteristic vectors are divided into short-term, medium-term, and long-term. The variational constraint optimization objective function is constructed by calculating the sequence envelope and the instantaneous frequency. The center frequency and the bandwidth parameters are iteratively updated, and the adaptive multi-scale components with the optimal energy aggregation degree are extracted. The adaptive multi-scale components are subjected to dual-tree complex wavelet transform to obtain multi-level sparse time-frequency characteristics. The enhanced input sequence with time series position information is generated by combining the time series position encoding; A multi-layer biological perception field network is constructed based on the enhanced input sequence. In each layer of the perception field, the key feature patterns are extracted by the excitatory neuron group, and the selective enhancement is performed by the inhibitory neuron group. The activation state of the neuron is adjusted by combining the adaptive dynamic threshold to obtain the perception feature. The perception features in each layer of the perception field are combined to obtain a hierarchical attention feature representation; The attention feature representation is subjected to residual connection and layer normalization processing. The adaptive gating unit is used for dynamic weight allocation, and the multi-scale features are fused to obtain an enhanced feature representation with time series dependence. A quadratic local trend model is constructed based on the enhanced feature representation, and the iteratively reweighted least squares method is used for optimization and solution. The dynamic prediction confidence interval is constructed by multiple samplings. The historical data weight and the prediction credibility weight are adjusted according to the interval width, and the sequence of predicted characteristic vectors is output.

5. The method according to claim 1, wherein Construct a cooperative avoidance control algorithm based on the artificial potential field method, using the bird target as the repulsive source and the wind turbine as the repulsive field. Calculate the potential field gradient according to the relative distance between the bird target and the wind turbine, including: Taking the bird target as a repulsive force source, a first repulsive potential energy function is established. The first repulsive potential energy function is determined by a repulsive force field coefficient, the relative distance of the bird target, and a repulsive force field influence radius. When the relative distance of the bird target is less than or equal to the repulsive force field influence radius, the first repulsive potential energy function is positively correlated with the repulsive force field coefficient, the relative distance of the bird target, and the repulsive force field influence radius. When the relative distance of the bird target is greater than the repulsive force field influence radius, the first repulsive potential energy function is zero; Taking the wind turbine generator as a repulsive force field, a second repulsive potential energy function is established. The second repulsive potential energy function is determined by a repulsive force field coefficient, the relative distance of the wind turbine generator, and a repulsive force field influence radius. When the relative distance of the wind turbine generator is less than or equal to the repulsive force field influence radius, the second repulsive potential energy function is positively correlated with the repulsive force field coefficient, the relative distance of the wind turbine generator, and the repulsive force field influence radius. When the relative distance of the wind turbine generator is greater than the repulsive force field influence radius, the second repulsive potential energy function is zero; Taking the negative gradient of the first repulsive potential energy function and the second repulsive potential energy function respectively according to the relative distances of the bird target and the wind turbine generator to obtain the potential field gradient.

6. The method according to claim 1, characterized in that, Online optimizing the potential field gradient through an adaptive fuzzy neural network, generating avoidance control parameters and combining with a sliding mode controller to generate obstacle avoidance parameters and combining them to obtain an obstacle avoidance control strategy, including: Receiving the pre-obtained potential field gradient, inputting the potential field gradient into an adaptive fuzzy neural network with an input layer, a fuzzification layer, a rule layer, a rule inference layer, and an output layer, and using a Gaussian membership function in the fuzzification layer to perform fuzzification processing on the potential field gradient to obtain fuzzified data; Based on the network output error, using the gradient descent method to determine the update amounts of the center value parameter and the width parameter of the membership function, online optimizing the center value parameter and the width parameter of the membership function, and inputting the optimized fuzzification parameters into the rule layer to obtain a rule layer output vector; Performing matrix operation on the rule layer output vector and the network weight matrix to obtain a network compensation amount, superimposing the network compensation amount and the potential field gradient to obtain an optimized potential field gradient, and generating an avoidance control parameter based on the change rate of the optimized potential field gradient; Determining the position tracking error based on the avoidance control parameter, constructing a sliding mode surface and a non-singular terminal sliding mode controller, performing finite-time convergence through a control law structure combined by a piecewise power function, and inputting the optimized potential field gradient into the non-singular terminal sliding mode controller to generate obstacle avoidance parameters; Determining a dynamic weight factor based on the position tracking error, and using the dynamic weight factor to perform weighted combination on the obstacle avoidance parameters and the output of the output layer of the adaptive fuzzy neural network to obtain an obstacle avoidance control strategy.

7. The method according to claim 6, characterized in that, Determining the position tracking error based on the avoidance control parameter, constructing a sliding mode surface and a non-singular terminal sliding mode controller, performing finite-time convergence through a control law structure combined by a piecewise power function, and inputting the optimized potential field gradient into the non-singular terminal sliding mode controller to generate obstacle avoidance parameters, including: Construct a position tracking error mapping function based on the avoidance control parameter. When the absolute value of the avoidance control parameter is greater than or equal to the threshold parameter, use the product of the first mapping coefficient and the non - linear power term of the avoidance control parameter minus the reference control quantity as the position tracking error output. When the absolute value of the avoidance control parameter is less than the threshold parameter, use the product of the second mapping coefficient and the sine function based on the ratio of the avoidance control parameter to the threshold parameter as the position tracking error output; Construct a non - linear sliding mode surface containing dynamic weights based on the position tracking error. The non - linear sliding mode surface consists of the derivative term of the position tracking error, the product term of the dynamic weight factor and the position tracking error, and the non - linear power integral term. The dynamic weight factor includes a basic weight and an exponential decay adaptive term based on the position tracking error; Design a piece - wise continuous control law based on the non - linear sliding mode surface. When the absolute value of the non - linear sliding mode surface is greater than or equal to the switching domain threshold, use the combination of two different power terms as the control output. When the absolute value of the non - linear sliding mode surface is less than the switching domain threshold, use the combination of a linear term and a fractional power term as the control output; Take the product of the Sigmoid function and the optimized potential field gradient constraint function as the smooth transition function of the non - linear sliding mode surface, and superimpose the control output with the product of the smooth transition function and the optimized potential field gradient to obtain the obstacle avoidance parameter.

8. The method according to claim 1, wherein Execute the obstacle avoidance control strategy and monitor the operating state of the adjusted wind turbine in real - time until the bird target flies out of the wind farm. Restoring the wind turbine to the normal operating state according to the preset restoration strategy includes: Execute the obstacle avoidance control strategy to adjust the wind turbine, and monitor the operating state of the wind turbine and the flight trajectory of the bird target in real - time; When the bird target flies out of the wind farm, obtain the current operating state parameters of the wind turbine, and gradually adjust the rotational speed, torque, and pitch angle of the wind turbine to the corresponding target values in segments with different restoration slopes according to the operating state parameters and the preset restoration threshold to restore the normal operating state.

9. An electronic device, characterized in that, Include: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 8.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, The computer program instructions, when executed by the processor, implement the method according to any one of claims 1 to 8.

Images