Bird trajectory prediction and wind turbine linkage control method in wind farms based on neural network

Abstract

The present invention provides a neural network-based bird trajectory prediction and wind turbine linkage control method for a wind farm, which relates to the field of smart grid technology. The method includes: real-time identification of bird targets and generation of trajectory data; processing of trajectory data through a three-dimensional Hilbert-Huang transform and an adaptive decomposition algorithm; extraction of characteristic parameters for prediction and analysis; assessment of collision risk based on trajectory prediction; construction of a collaborative avoidance control algorithm using an artificial potential field method; optimization of the control strategy by combining an adaptive fuzzy neural network and a sliding film controller; real-time monitoring of the wind turbine operating status; ensuring that the wind turbine resumes normal operation after the birds have safely flown away, thereby realizing intelligent bird protection in wind farms.

Description

Technical Field

[0001] The present invention relates to the field of smart grid technology, and in particular to a method for predicting bird trajectories in a wind farm and controlling wind turbine linkages based on a neural network. Background Art

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

[0003] Existing bird monitoring and trajectory prediction methods mostly rely on traditional image processing technology, which is unable to accurately identify and track bird targets in real time, resulting in low prediction accuracy. At the same time, avoidance control algorithms often ignore the dynamic relationship between birds and wind turbines, resulting in insufficient real-time and effectiveness of avoidance strategies.

[0004] Therefore, a solution is urgently needed to solve the problems existing in the prior art. Summary of the Invention

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

[0006] A first aspect of an embodiment of the present invention provides a method for predicting bird trajectories in a wind farm and controlling wind turbine linkages based on a neural network, comprising:

[0007] Collect real-time image data and use a pre-trained target detection neural network to identify bird targets, and generate bird flight trajectory data based on the position information of bird targets in multiple consecutive frames of images;

[0008] Performing a 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 using an adaptive decomposition algorithm, extracting amplitude-frequency characteristic parameters and performing feature dimensionality reduction and reconstruction to obtain flight characteristic vectors, which are then subjected to time series analysis. A predicted characteristic vector sequence is generated, and trajectory corrections are calculated based on phase differences and amplitude changes to obtain predicted trajectory data.

[0009] Identify wind turbines in a wind farm that present a bird collision risk based on predicted trajectory data;

[0010] A collaborative avoidance control algorithm is constructed based on the artificial potential field method. The bird target is used as the repulsion source and the wind turbine is used as the repulsion field. The potential field gradient is calculated according to the relative distance between the bird target and the wind turbine. The potential field gradient is optimized online through an adaptive fuzzy neural network to generate avoidance control parameters. The sliding membrane controller is used to generate obstacle avoidance parameters, and the combined method is used to obtain an obstacle avoidance control strategy.

[0011] The obstacle avoidance control strategy is executed and the adjusted operating status of the wind turbine is monitored in real time until the bird target flies away from the wind farm, and the wind turbine is restored to normal operating status according to the preset recovery strategy.

[0012] In an optional embodiment,

[0013] Performing a 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 eigenmode components includes:

[0014] Performing a three-dimensional Hilbert-Huang transform on the bird flight trajectory data to obtain analytical signals in the directions of three coordinate axes, calculating the instantaneous amplitude and instantaneous phase in the directions of the three coordinate axes 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;

[0015] Taking the instantaneous frequency spectrum as an initial residual signal, employing an adaptive decomposition algorithm to determine all local extreme value points of the initial residual signal, employing a cubic spline interpolation method to construct an upper envelope and a lower envelope of the initial residual signal, calculating the average 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;

[0016] Determine 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. Otherwise, use the detail component as a new residual signal and repeatedly execute the adaptive decomposition algorithm until the intrinsic mode component is obtained.

[0017] In an optional embodiment,

[0018] Extract the amplitude-frequency characteristic parameters and perform feature dimensionality reduction and reconstruction to obtain the flight characteristic vector and perform time series analysis. Generate a predicted characteristic vector sequence and calculate the trajectory correction value based on the phase difference and amplitude change. The predicted trajectory data includes:

[0019] 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 amplitude-frequency characteristic parameters are obtained by extracting the processed instantaneous phase and energy change rate and the characteristic matrix is constructed by arranging them in time sequence;

[0020] Calculating the time derivative of the feature matrix to obtain the feature change rate and calculating the sample importance weight, weighting the feature matrix based on the sample importance weight to obtain a weighted covariance matrix, determining a first eigenvector based on the eigenvalue interval ratio, cumulative contribution rate and reconstruction error of the weighted covariance matrix, and performing dimensionality reduction and reconstruction on the feature matrix based on the first eigenvector to obtain a projection matrix;

[0021] Multiplying the feature matrix by the projection matrix to obtain a flight feature vector, establishing a time series relationship of a quadratic term local trend, and performing time series prediction using iterative reweighted least squares method to obtain a predicted feature vector sequence;

[0022] Phase difference parameters and amplitude change parameters at adjacent moments are calculated based on the predicted feature vector sequence. Combined weight coefficients of the phase difference parameters and amplitude change parameters are initialized based on historical trajectory data and weighted summed. The weighted sum is used as a trajectory correction. The combined weight coefficients are updated according to the prediction error. A smoothing correction is calculated based on the trajectory correction and a smoothing factor is dynamically adjusted. The smoothing factor is combined with the predicted feature vector sequence to obtain predicted trajectory data.

[0023] In an optional embodiment,

[0024] The time series relationship of the local trend of the quadratic term is established, and the time series prediction is performed by combining the iterative reweighted least squares method to obtain the predicted feature vector sequence including:

[0025] The flight feature vector is divided into short-term, medium-term and long-term. The variational constrained optimization objective function is constructed by calculating the sequence envelope and instantaneous frequency. The center frequency and bandwidth parameters are iteratively updated to extract the adaptive multi-scale components with optimal energy concentration. The adaptive multi-scale components are transformed by dual-tree complex wavelet transform to obtain multi-level sparse time-frequency features. Combined with time series position coding, an enhanced input sequence with time series position information is generated.

[0026] A multi-layer biological perception field network is constructed based on the enhanced input sequence. In each layer of the perception field, key feature patterns are extracted by the excitatory neuron group, and selective enhancement is performed by the inhibitory neuron group. The activation state of the neurons is adjusted by combining the adaptive dynamic threshold to obtain the perception features. The perception features in each layer of the perception field are combined to obtain a hierarchical attention feature representation.

[0027] The attention feature representation is subjected to residual connection and layer normalization processing, and dynamic weight allocation is performed using an adaptive gating unit. 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 iterative reweighted least squares method is used for optimization and solution. A dynamic prediction confidence interval is constructed through multiple samplings, and the historical data weight and prediction credibility weight are adjusted according to the interval width, and a prediction feature vector sequence is output.

[0028] In an optional embodiment,

[0029] A collaborative avoidance control algorithm is constructed based on the artificial potential field method. The bird target is used as the repulsion source and the wind turbine generator is used as the repulsion field. The potential field gradient is calculated according to the relative distance between the bird target and the wind turbine generator. The algorithm includes:

[0030] A first repulsive potential energy function is established using the bird target as a repulsive force source. The first repulsive potential energy function is determined by a repulsive field coefficient, a relative distance of the bird target, and a repulsive field influence radius. When the relative distance of the bird target is less than or equal to the repulsive field influence radius, the first repulsive potential energy function is positively correlated with the repulsive field coefficient, the relative distance of the bird target, and the repulsive field influence radius. When the relative distance of the bird target is greater than the repulsive field influence radius, the first repulsive potential energy function is zero.

[0031] A second repulsive potential energy function is established using the wind turbine generator set as a repulsive field, wherein the second repulsive potential energy function is determined by a repulsive field coefficient, a relative distance of the wind turbine generator set, and a repulsive field influence radius. When the relative distance of the wind turbine generator set is less than or equal to the repulsive field influence radius, the second repulsive potential energy function is positively correlated with the repulsive field coefficient, the relative distance of the wind turbine generator set, and the repulsive field influence radius. When the relative distance of the wind turbine generator set is greater than the repulsive field influence radius, the second repulsive potential energy function is zero.

[0032] According to the relative distance between the bird target and the wind turbine generator set, negative gradients of the first repulsive potential energy function and the second repulsive potential energy function are respectively calculated to obtain potential field gradients.

[0033] In an optional embodiment,

[0034] The potential field gradient is optimized online through an adaptive fuzzy neural network to generate avoidance control parameters. The obstacle avoidance parameters are generated by combining with the sliding membrane controller and the obstacle avoidance control strategy is as follows:

[0035] Receiving the potential field gradient obtained in advance, inputting the potential field gradient into an adaptive fuzzy neural network having an input layer, a fuzzification layer, a rule layer, a rule inference layer and an output layer, and fuzzifying the potential field gradient using a Gaussian membership function in the fuzzification layer to obtain fuzzy data;

[0036] Based on the network output error, the gradient descent method is used to determine the update amount of the central value parameter and width parameter of the membership function. The central value parameter and width parameter of the membership function are optimized online, and the optimized fuzzy parameters are input into the rule layer to obtain the rule layer output vector.

[0037] Perform matrix operations on the rule layer output vector 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 the avoidance control parameters based on the optimized potential field gradient change rate;

[0038] Determine the position tracking error based on the avoidance control parameters and construct a sliding surface and a non-singular terminal sliding mode controller, perform finite time convergence through a control law structure composed of piecewise power functions, and input the optimized potential field gradient into the non-singular terminal sliding mode controller to generate obstacle avoidance parameters;

[0039] The dynamic weight factor is determined based on the position tracking error, and the obstacle avoidance parameters are weightedly combined with the output layer output of the adaptive fuzzy neural network using the dynamic weight factor to obtain the obstacle avoidance control strategy.

[0040] In an optional embodiment,

[0041] Based on the avoidance control parameters, the position tracking error is determined and a sliding surface and a non-singular terminal sliding mode controller are constructed. Finite time convergence is performed through a control law structure composed of piecewise power functions. The optimized potential field gradient is input into the non-singular terminal sliding mode controller to generate obstacle avoidance parameters, including:

[0042] constructing a position tracking error mapping function based on an avoidance control parameter, wherein when the absolute value of the avoidance control parameter is greater than or equal to a threshold parameter, a first mapping coefficient and a nonlinear power term of the avoidance control parameter minus a reference control amount are multiplied as a position tracking error output; and when the absolute value of the avoidance control parameter is less than the threshold parameter, a second mapping coefficient and a sine function based on the ratio of the avoidance control parameter to the threshold parameter are multiplied as a position tracking error output;

[0043] Constructing a nonlinear sliding mode surface containing a dynamic weight based on the position tracking error, the nonlinear sliding mode surface consisting of a derivative term of the position tracking error, a product term of a dynamic weight factor and the position tracking error, and a nonlinear power integral term, wherein the dynamic weight factor includes a basic weight and an exponential decay adaptive term based on the position tracking error;

[0044] A piecewise continuous control law is designed based on the nonlinear sliding surface. When the absolute value of the nonlinear sliding surface is greater than or equal to a switching domain threshold, a combination of two different power terms is used as a control output. When the absolute value of the nonlinear sliding surface is less than the switching domain threshold, a combination of a linear term and a fractional power term is used as a control output.

[0045] The product of the Sigmoid function and the optimized potential field gradient constraint function is used as the smooth transition function of the nonlinear sliding surface, and the control output is superimposed with the product of the smooth transition function and the optimized potential field gradient to obtain the obstacle avoidance parameter.

[0046] In an optional embodiment,

[0047] Execute the obstacle avoidance control strategy and monitor the adjusted wind turbine operating status in real time until the bird target flies away from the wind farm. Then restore the wind turbine to normal operation according to the preset recovery strategy, including:

[0048] Executing an obstacle avoidance control strategy to adjust the wind turbine generator set and monitoring the operating status of the wind turbine generator set and the flight trajectory of the bird target in real time;

[0049] When the bird target flies away from the wind farm, the current operating status parameters of the wind turbine generator set are obtained. According to the operating status parameters and the preset recovery threshold, the speed, torque and pitch angle of the wind turbine generator set are gradually adjusted to the corresponding target values according to different recovery slopes to restore the normal operating state.

[0050] According to a second aspect of an embodiment of the present invention, an electronic device is provided, including:

[0051] 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 aforementioned method.

[0052] According to a third aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0053] In the present invention, by collecting real-time image data and using a pre-trained target detection neural network to identify bird targets, bird flight trajectory data can be accurately generated, thereby improving the response speed of the wind farm to bird activities and reducing the risk of collision between birds and wind turbines. The three-dimensional Hilbert-Huang transform and adaptive decomposition algorithm are used to extract flight feature vectors, which can effectively perform feature dimensionality reduction and reconstruction, thereby generating an accurate prediction feature vector sequence, making the correction amount for the bird trajectory more accurate, and improving the effectiveness of the obstacle avoidance control strategy. The avoidance control algorithm constructed based on the artificial potential field method enables wind turbines to intelligently avoid bird targets, reducing the risk of bird collision. Combined with the online optimization capability of the adaptive fuzzy neural network, the flexibility and adaptability of the obstacle avoidance control strategy are further improved, ensuring that the wind farm can operate safely when bird targets appear. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 Schematic diagram of a flow chart of a method for predicting bird trajectories in a wind farm and controlling wind turbines based on a neural network according to an embodiment of the present invention;

[0055] Figure 2 This is a comparison diagram of trajectory correction effects of a wind farm bird trajectory prediction and wind turbine linkage control method based on a neural network according to an embodiment of the present invention;

[0056] Figure 3 This is a multi-scale decomposition effect diagram of the flight characteristics of the wind farm bird trajectory prediction and wind turbine linkage control method based on neural network in an embodiment of the present invention;

[0057] Figure 4 This is a simulation diagram of the obstacle avoidance technology effect of the wind farm bird trajectory prediction and wind turbine linkage control method based on neural network in an embodiment of the present invention. DETAILED DESCRIPTION

[0058] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0059] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0060] Figure 1 FIG. 1 is a flow chart of a method for predicting bird trajectories in a wind farm and controlling wind turbines based on a neural network according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0061] Collect real-time image data and use a pre-trained target detection neural network to identify bird targets, and generate bird flight trajectory data based on the position information of bird targets in multiple consecutive frames of images;

[0062] Performing a 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 using an adaptive decomposition algorithm, extracting amplitude-frequency characteristic parameters and performing feature dimensionality reduction and reconstruction to obtain flight characteristic vectors, which are then subjected to time series analysis. A predicted characteristic vector sequence is generated, and trajectory corrections are calculated based on phase differences and amplitude changes to obtain predicted trajectory data.

[0063] Identify wind turbines in a wind farm that present a bird collision risk based on predicted trajectory data;

[0064] A collaborative avoidance control algorithm is constructed based on the artificial potential field method. The bird target is used as the repulsion source and the wind turbine is used as the repulsion field. The potential field gradient is calculated according to the relative distance between the bird target and the wind turbine. The potential field gradient is optimized online through an adaptive fuzzy neural network to generate avoidance control parameters. The sliding membrane controller is used to generate obstacle avoidance parameters, and the combined method is used to obtain an obstacle avoidance control strategy.

[0065] The obstacle avoidance control strategy is executed and the adjusted operating status of the wind turbine is monitored in real time until the bird target flies away from the wind farm, and the wind turbine is restored to normal operating status according to the preset recovery strategy.

[0066] In an optional embodiment,

[0067] Performing a 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 eigenmode components includes:

[0068] Performing a three-dimensional Hilbert-Huang transform on the bird flight trajectory data to obtain analytical signals in the directions of three coordinate axes, calculating the instantaneous amplitude and instantaneous phase in the directions of the three coordinate axes 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;

[0069] Taking the instantaneous frequency spectrum as an initial residual signal, employing an adaptive decomposition algorithm to determine all local extreme value points of the initial residual signal, employing a cubic spline interpolation method to construct an upper envelope and a lower envelope of the initial residual signal, calculating the average 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;

[0070] Determine 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. Otherwise, use the detail component as a new residual signal and repeatedly execute the adaptive decomposition algorithm until the intrinsic mode component is obtained.

[0071] Spatial position data of the bird's flight trajectory is acquired and decomposed into three coordinate components: easting, northing, and altitude. A Hilbert-Huang transform is performed on each coordinate component to obtain an analytical signal, and the signals are adaptively decomposed. In the easting coordinate, the position data sequence is mapped onto the complex plane using a Hilbert transform to obtain an analytical signal. In the northing coordinate, the position sequence is Hilbert transformed to obtain the corresponding analytical signal. In the altitude direction, the vertical position sequence is Hilbert transformed to obtain a third analytical signal. Based on the analytical signals obtained in each of the three directions, the instantaneous amplitude is extracted, which reflects the change in the bird's displacement amplitude in the corresponding direction. The instantaneous phase in each direction is also extracted, which characterizes the periodic motion characteristics of the bird in different directions. The instantaneous phase in each direction is then time-derived to obtain the instantaneous frequency, which reflects the frequency characteristics of the bird's motion in each direction. The instantaneous frequencies in these three directions are combined according to their temporal correspondence to construct a complete three-dimensional instantaneous frequency spectrum.

[0072] Adaptive decomposition begins with the constructed three-dimensional instantaneous frequency spectrum as the initial residual signal. The frequency spectrum signal is searched for all local maxima and minima, reflecting the characteristics of the bird's turning or speed changes during flight. By comparing the size relationships of adjacent data points, the specific location and amplitude of each maximum and minimum point are determined. Using cubic spline interpolation, a smooth upper envelope is constructed using the local maximum points as control points. This envelope passes through all maximum points and forms a smooth transition between them. A smooth lower envelope is constructed using the local minimum points as control points, passing through all minimum points and maintaining a continuous and smooth characteristic. The average of the upper and lower envelopes at each sampling moment is calculated to obtain a mean envelope reflecting the overall trend of the signal. The mean envelope is subtracted from the initial frequency spectrum signal to obtain the detail component after the overall trend has been removed.

[0073] The resulting detail components are then evaluated for their intrinsic mode characteristics. The number of zeros in the detail component is counted, i.e., the number of times the signal crosses the zero line. The number of extreme values in the signal, including all local maxima and minima, is also counted. The absolute value of the difference between the number of zeros and the number of extreme values is calculated; this difference characterizes the signal's symmetry. The mean envelope of the detail component is calculated to determine whether it is sufficiently close to zero. When the difference between the number of zeros and the number of extreme values is less than or equal to 1, the signal exhibits good symmetry; when the mean envelope is sufficiently close to zero, the signal oscillations are well balanced. In this case, the detail component is considered to meet the requirements of the intrinsic mode function and is identified as a valid intrinsic mode component. If either of these two conditions is not met, the detail component is used as a new residual signal, and the process of extreme value search, envelope construction, mean envelope calculation, and detail component extraction is repeated until an intrinsic mode component that meets the evaluation criteria is obtained. Through an iterative decomposition process, the frequency characteristics of the bird's flight trajectory are decomposed into a series of intrinsic mode components.

[0074] For example, trajectory data for a seagull is collected starting at 180 meters east, 150 meters north, and 100 meters above sea level, with a sampling time of 60 seconds and a sampling interval of 0.1 seconds. A Hilbert-Huang transform is performed on the recorded eastward position data sequence to obtain a complex analytical signal reflecting the characteristics of eastward motion. A Hilbert-Huang transform is performed on the northward position data to obtain a corresponding analytical signal. A similar transform is performed on the altitude data to obtain a third analytical signal. The instantaneous amplitudes in the three directions are extracted from the analytical signals, revealing that the seagull's displacement amplitude in the east direction fluctuates around 20 meters, the northward displacement amplitude is approximately 15 meters, and the altitude amplitude fluctuates within a range of 10 meters. The instantaneous phase changes in the three directions are obtained, and the time derivatives are taken to obtain the instantaneous frequencies. The seagull's horizontal motion frequency is primarily concentrated between 0.2 and 0.5 Hz, while the vertical motion frequency is lower, ranging from 0.1 to 0.3 Hz. The frequency data for the three directions are combined to form a complete frequency spectrum.

[0075] The frequency spectrum was decomposed, and data traversal revealed 23 local maxima and 22 local minima. Using cubic spline interpolation, upper and lower envelopes were constructed based on the extreme points. The calculated mean envelope was then subtracted from the original frequency spectrum to obtain the first layer of detail components. Calculations revealed that this detail component contained 35 zeros and 36 extreme points, with the difference between the number of zeros and extreme points being 1. The calculated maximum value of the mean envelope was only 0.1% of the original signal amplitude. Because this detail component met the criteria for the intrinsic mode function, it was identified as the first intrinsic mode component. This component reflects the highest-frequency motion characteristics of the seagull during flight.

[0076] In this embodiment, the three-dimensional Hilbert-Huang transform combined with the adaptive decomposition algorithm can effectively capture the time-frequency characteristics of bird flight trajectories. Due to the use of a three-dimensional processing method, the movement characteristics of birds in space can be fully captured, avoiding the information loss caused by analyzing only a single direction, providing a reliable data basis for bird behavior pattern recognition, and further improving the accuracy of bird activity prediction in wind farms.

[0077] In an optional embodiment,

[0078] Extract the amplitude-frequency characteristic parameters and perform feature dimensionality reduction and reconstruction to obtain the flight characteristic vector and perform time series analysis. Generate a predicted characteristic vector sequence and calculate the trajectory correction value based on the phase difference and amplitude change. The predicted trajectory data includes:

[0079] 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 amplitude-frequency characteristic parameters are obtained by extracting the processed instantaneous phase and energy change rate and the characteristic matrix is constructed by arranging them in time sequence;

[0080] Calculating the time derivative of the feature matrix to obtain the feature change rate and calculating the sample importance weight, weighting the feature matrix based on the sample importance weight to obtain a weighted covariance matrix, determining a first eigenvector based on the eigenvalue interval ratio, cumulative contribution rate and reconstruction error of the weighted covariance matrix, and performing dimensionality reduction and reconstruction on the feature matrix based on the first eigenvector to obtain a projection matrix;

[0081] Multiplying the feature matrix by the projection matrix to obtain a flight feature vector, establishing a time series relationship of a quadratic term local trend, and performing time series prediction using iterative reweighted least squares method to obtain a predicted feature vector sequence;

[0082] Phase difference parameters and amplitude change parameters at adjacent moments are calculated based on the predicted feature vector sequence. Combined weight coefficients of the phase difference parameters and amplitude change parameters are initialized based on historical trajectory data and weighted summed. The weighted sum is used as a trajectory correction. The combined weight coefficients are updated according to the prediction error. A smoothing correction is calculated based on the trajectory correction and a smoothing factor is dynamically adjusted. The smoothing factor is combined with the predicted feature vector sequence to obtain predicted trajectory data.

[0083] Starting with time series data processing, a continuous sliding data analysis of the intrinsic mode components is performed. The first-order derivative of the instantaneous phase is calculated using the forward difference method to obtain the instantaneous frequency value at each time point. A second-order difference operation is then performed on the instantaneous frequency 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, achieving adaptive adjustment of the window length. Within each adaptive window, the signal is weighted using the Hanning function to reduce the spectral leakage effect at the window edge.

[0084] Phase demodulation is performed on the windowed signal to extract the instantaneous phase change characteristics within the window. The time derivative of the signal energy is simultaneously calculated to obtain the energy rate of change characteristics. The phase rate of change and the energy rate of change are aligned in the time domain to construct a two-dimensional feature sequence. Using a sliding window, the two-dimensional feature sequences within consecutive time periods are combined into a feature matrix. Each row of the matrix represents a feature combination within a time window, and each column represents the temporal evolution of a specific feature.

[0085] The feature matrix is differentiated along time to obtain a feature change rate matrix. Based on the change rate, an exponential weighting method is used to calculate the importance weight of each sample point, with sample points with greater change rates receiving higher weights. The sample weights are element-wise multiplied by the original feature matrix to obtain a weighted feature matrix. The covariance of the weighted feature matrix is calculated to obtain a weighted covariance matrix that reflects feature correlations.

[0086] Perform eigenvalue decomposition on the weighted covariance matrix, calculating the ratios between adjacent eigenvalues and the cumulative contribution rates of the eigenvalues. By setting contribution rate and interval ratio thresholds, the most representative eigenvectors are selected. The original eigenvector is projected onto this eigenvector, reducing the dimensionality of the data while preserving key information. The reduced dimensionality data is reconstructed using the inverse transform of the eigenvectors, and the reconstruction error is evaluated to verify the dimensionality reduction. The verified eigenvectors are used to construct the projection matrix.

[0087] The original feature matrix is multiplied by the projection matrix to obtain the projection result in the principal feature direction, namely the flight feature vector. A quadratic polynomial model is constructed for the feature vector, taking into account the acceleration term, to describe the nonlinear relationship between the feature quantities. The model parameters are estimated using iterative reweighted least squares, where the sample weights are dynamically adjusted based on the residual size at each iteration. Based on the estimated model parameters, the prediction is extended into the future time domain to obtain a series of predicted feature vectors.

[0088] The predicted feature vector sequence is post-processed to calculate the phase angle difference and amplitude ratio change between adjacent prediction points. Based on the statistical characteristics of similar scenarios in historical data, the initial combined weights of the phase difference parameter and the amplitude change parameter are determined. These two parameters are weighted and superimposed to obtain the initial trajectory correction. By online tracking the changing trend of the prediction error, the combined weight coefficients are updated using the gradient descent method. The trajectory correction is exponentially smoothed, with the smoothing coefficient decreasing as the prediction error increases and increasing as the error decreases, achieving adaptive smoothing. The smoothed correction is superimposed on the predicted feature vector to obtain the corrected trajectory prediction data.

[0089] For example, the circling flight process of a goshawk was tracked and recorded within 30 seconds with a sampling frequency of 50 Hz, and the obtained intrinsic mode component contained 1500 data points. The instantaneous phase was differentially calculated, and it was found that the circling frequency of the goshawk in the rising air flow gradually decreased from 1.2 Hz to 0.8 Hz. Based on this frequency change feature, a window length of 80 points was used in the initial stage with higher frequency, and a window length of 120 points was used in the later stage when the frequency decreased. In each window, the extracted phase change fluctuated between 45 and 75 degrees, and the energy change rate varied between 0.08 and 0.25. These features were arranged in chronological order to construct a feature matrix with 1400 rows and 2 columns.

[0090] Computing the time derivative of the characteristic matrix revealed that the rate of change during a goshawk's flight attitude adjustment was five times greater than during a stable hover. Based on this calculation, the sample weight during attitude adjustment was calculated to be 0.9, while the weight during a stable hover was 0.1. After weighting the characteristic matrix, the eigenvalue distribution of the covariance matrix was analyzed. The first eigenvalue accounted for 92% of the total energy, corresponding to a reconstruction error of only 3%. The eigenvector corresponding to this eigenvalue was selected as the basis for dimensionality reduction and the projection matrix was constructed.

[0091] After multiplying the original feature matrix by the projection matrix, a vector reflecting the main motion characteristics of the goshawk is obtained. A quadratic trend model containing velocity and acceleration terms is established. After five iterative optimizations, the prediction error converges to within 5% of the original amplitude. The motion feature sequence within the next 5 seconds is predicted, and the phase difference between adjacent moments is calculated to vary between 15 and 30 degrees, and the amplitude change rate fluctuates within 20%. The weight of the initial phase difference parameter is 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 revised prediction results, it is predicted that the goshawk will begin a turning movement with a radius of about 50 meters in 2 seconds, and the turning process will last about 3 seconds.

[0092] 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 rapidly changing detail information, but also effectively suppresses noise interference, thereby improving the accuracy of feature extraction. The weighted dimensionality reduction method based on the sample importance weight highlights the flight characteristics at critical moments, avoiding the problem that the traditional equal-weight dimensionality reduction method may ignore the key turning points. Combined with the trajectory correction mechanism of phase difference and amplitude change parameters, it can capture the subtle change trends of bird flight and continuously improve the prediction accuracy through dynamic weight optimization.

[0093] Figure 2 This is a comparison chart of the trajectory correction effects of the wind farm bird trajectory prediction and wind turbine linkage control method based on neural network in the embodiment of the present invention, showing the trajectory prediction error change trend of the three different methods in the prediction time domain (0-5 seconds). Figure 2 The three methods, denoted by circular markers, demonstrate significant error control advantages throughout the entire forecast period. In the initial forecast phase (0-1 seconds), the three methods perform relatively similarly. The prediction error of this technique starts at 9% and reaches 11% at 1 second. The Kalman filter (a classic state-space model filter, denoted by square markers in the figure) and the LSTM prediction method (denoted by triangle markers in the figure) have prediction errors of 11%-14.5% and 9.5%-12%, respectively. In the mid-forecast phase (2-3 seconds), between the "turn start" and "turn completion" points marked in the figure, the prediction error of this technique increases more slowly, from 13.5% to 16.2%. In contrast, the errors of the Kalman filter and LSTM methods rise rapidly, reaching 18%-23.5% and 17%-21.5%, respectively. This phase corresponds to the complex flight phase during which the goshawk begins executing turns with a radius of approximately 50 meters. This technical solution effectively captures the trajectory changes during turns 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 later stages of the prediction (4-5 seconds), the errors of all methods showed an accelerating trend, but the error of this technical solution at the 5-second endpoint prediction was still limited to 22.5%, 11 and 9.5 percentage points lower than the 33.5% of the Kalman filter method and the 32.0% of the LSTM method, respectively. This demonstrates the robustness of this technical solution in long-term predictions, particularly its adaptability to sudden changes in flight trajectory, such as turns.

[0094] In an optional embodiment,

[0095] The time series relationship of the local trend of the quadratic term is established, and the time series prediction is performed by combining the iterative reweighted least squares method to obtain the predicted feature vector sequence including:

[0096] The flight feature vector is divided into short-term, medium-term and long-term. The variational constrained optimization objective function is constructed by calculating the sequence envelope and instantaneous frequency. The center frequency and bandwidth parameters are iteratively updated to extract the adaptive multi-scale components with optimal energy concentration. The adaptive multi-scale components are transformed by dual-tree complex wavelet transform to obtain multi-level sparse time-frequency features. Combined with time series position coding, an enhanced input sequence with time series position information is generated.

[0097] A multi-layer biological perception field network is constructed based on the enhanced input sequence. In each layer of the perception field, key feature patterns are extracted by the excitatory neuron group, and selective enhancement is performed by the inhibitory neuron group. The activation state of the neurons is adjusted by combining the adaptive dynamic threshold to obtain the perception features. The perception features in each layer of the perception field are combined to obtain a hierarchical attention feature representation.

[0098] The attention feature representation is subjected to residual connection and layer normalization processing, and dynamic weight allocation is performed using an adaptive gating unit. 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 iterative reweighted least squares method is used for optimization and solution. A dynamic prediction confidence interval is constructed through multiple samplings, and the historical data weight and prediction credibility weight are adjusted according to the interval width, and a prediction feature vector sequence is output.

[0099] The flight feature vectors are decomposed into three time scales: short-term (less than 10 seconds), medium-term (10-30 seconds), and long-term (over 30 seconds). At each time scale, the upper and lower envelopes of the signal are calculated, and the instantaneous frequency characteristics are extracted. An optimization objective function is constructed based on the envelopes and frequency characteristics, with two constraints: signal energy concentration and frequency bandwidth. Using an iterative optimization method, the center frequency and frequency bandwidth parameters are continuously adjusted until the optimal parameter combination is found. The original sequence is then adaptively filtered using the optimal parameters to obtain multi-scale components with optimal energy concentration characteristics at different time scales.

[0100] A dual-tree complex wavelet transform is performed on the obtained multi-scale components to achieve multi-level decomposition of the signal. At each decomposition level, the signal's amplitude, phase, and frequency features are extracted to construct a sparse representation of the time-frequency feature map. The temporal position information is encoded into a position vector, which reflects the relative position and contextual relationship of the feature in the time series. The time-frequency features are combined with the position encoding information to generate an enhanced input sequence that incorporates temporal dependencies.

[0101] A multi-layered biological perception field network structure is constructed, with each perception field layer comprising groups of excitatory and inhibitory neurons. Excitatory neurons extract key characteristic patterns from the input sequence through parallel processing, including amplitude jumps, frequency jumps, and phase transitions. Inhibitory neurons selectively enhance these extracted features, suppressing redundant and noisy features. An adaptive dynamic threshold mechanism is implemented to dynamically adjust the activation state of neurons based on the significance of the features. When the feature response exceeds the threshold, the corresponding neuron is activated and outputs a perceptual feature. The perceptual features output by the activated neurons in each perception field layer are combined to form a hierarchical attention feature representation.

[0102] Post-processing of hierarchical attention features preserves original feature information through a residual connection mechanism, preventing feature loss during deep network training. Layer normalization eliminates bias in feature distribution and improves feature expressiveness. Adaptive gating units are constructed that dynamically assign weight coefficients based on feature importance. Weighted fusion of features at different scales yields an enhanced feature representation that integrates information from multiple time scales.

[0103] Based on the enhanced feature representation, a quadratic local trend model is constructed that incorporates the acceleration term. Model parameters are optimized using iterative reweighted least squares, dynamically adjusting sample weights based on the fitting residuals at each iteration. Dynamic confidence intervals for the forecast results are constructed through multiple rounds of random sampling and model fitting. The width of the confidence interval is calculated. When the interval is narrow, historical data is weighted more heavily to improve forecast stability; when the interval is wide, the weight of the forecast credibility is increased to enhance the model's adaptability to new features. The output is a sequence of multi-optimized forecast feature vectors.

[0104] For example, the flight eigenvectors of a Gyrfalcon diving for prey were analyzed, with a recording duration of 120 seconds. The characteristic sequence was divided into a short-term scale of 10 seconds (reflecting the acceleration process during the dive), a medium-term scale of 30 seconds (reflecting the attack preparation process), and a long-term scale of 60 seconds (reflecting the circling search process). The energy contribution of the short-term scale was calculated to be 60%, the medium-term scale 30%, and the long-term scale 10%. The optimal center frequencies were determined through iterative optimization: 5 Hz for the short-term scale, 2 Hz for the medium-term scale, and 0.5 Hz for the long-term scale.

[0105] A dual-tree complex wavelet transform was performed on the optimized multi-scale components to extract key time-frequency features, such as velocity jumps during the dive (amplitude changes up to 80%) and phase jumps during the turn (phase changes exceeding 90 degrees). These features were combined with position codes representing 120 time points to generate an enhanced input sequence.

[0106] In the three-layer biosensory field network, the first layer's 100 excitatory neurons primarily extract velocity and acceleration features, while 50 inhibitory neurons attenuate background noise. The second layer's 80 excitatory neurons extract turning and circling features, while 40 inhibitory neurons enhance primary movement patterns. The third layer's 60 excitatory neurons extract attack intent features, while 30 inhibitory neurons reinforce predatory behavior patterns. When a feature response exceeds an adaptive threshold (set at 1.5 times the average response strength), the corresponding neuron activates.

[0107] The original motion features are retained through residual connections, and the standard deviation of the feature distribution is reduced to less than 0.1 after layer normalization. The adaptive gating unit assigns 60% of the weight to short-term features, highlighting the key behavior of diving predation. The constructed quadratic trend model has been optimized through 5 rounds of iterative optimization, and the prediction error has converged to less than 5%. The confidence interval is constructed through 1000 random samplings. When the interval width is less than 10% of the predicted value, the weight of historical data is increased to 0.8; when the interval width is greater than 30%, the prediction credibility weight is increased to 0.6. The prediction results show that the Saker Falcon will start diving in 2 seconds, with a diving angle of about 60 degrees, and it will take about 3 seconds to complete the predation action.

[0108] In this embodiment, by constructing an optimization objective function that includes energy concentration and frequency bandwidth constraints, adaptive extraction of features at different time scales is achieved. Through the synergistic effect of excitatory and inhibitory neurons, active extraction of key features and adaptive suppression of noise are achieved. Residual connections and layer normalization operations ensure the effective transmission of deep features, and an adaptive gating mechanism achieves dynamic fusion of multi-scale features.

[0109] Existing technologies for predicting bird flight trajectories primarily use feature extraction methods at a single time scale. These methods extract features through sliding analysis with a fixed window size or simple frequency domain transformations. This makes it difficult to simultaneously capture flight characteristics at different time scales, resulting in insufficient accuracy when processing long-term time series predictions and a lack of sensitivity for predicting short-term sudden behaviors. Feature extraction is often limited to a single dimension, failing to fully utilize the multi-level information contained in time series data, making predictions difficult to adapt to the complex and changing flight patterns of birds.

[0110] This embodiment introduces a dual-tree complex wavelet transform and a temporal position coding mechanism, effectively preserving the time-frequency characteristics and temporal dependencies of features. A dynamic threshold mechanism is used to adjust the activation state of neurons, thereby improving the adaptability and robustness of feature extraction. Combined with the construction of dynamic confidence intervals and adaptive weight adjustment, a complete feature prediction framework is established, which not only improves the accuracy of predictions, but also enhances the algorithm's adaptability to complex flight patterns, providing reliable technical support for the intelligent upgrade of wind farm bird avoidance systems.

[0111] Figure 3 This is a multi-scale decomposition effect diagram of the flight characteristics of the neural network-based wind farm bird trajectory prediction and wind turbine linkage control method in an embodiment of the present invention. The original experimental data was collected in the northern grassland area (altitude 1200m) using a high-precision data sampling rate of 120Hz to ensure that the subtle changes in the bird's high-speed flight are captured. Analysis of the original data shows that the target bird is in the search and circling phase from 0 to 30 seconds, with an average altitude maintained at 180m; enters the dive preparation phase from 30 to 60 seconds, and gradually adjusts the angle to 42°; and completes the predatory attack from 60 to 90 seconds, with a maximum speed of 68m / s. Different line types are used in the figure to distinguish each algorithm: the solid line represents the present technical solution, 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. Comparative data shows that the trajectory reconstruction accuracy of the present technical solution is 92.4%, which is significantly higher than the EMD decomposition method (78.6%), wavelet decomposition method (82.1%), and STFT analysis method (65.7%). In terms of energy concentration, this technical solution achieved 85.3%, while the EMD decomposition method achieved 67.8%, the wavelet decomposition method achieved 64.5%, and the STFT analysis method achieved 54.2%. In short-term analysis (within 10 seconds), this technical solution achieved a feature extraction accuracy of 94.2%, with a processing delay of only 35ms. In medium-term analysis (10-30 seconds), the extraction accuracy was 88.7%, with a noise suppression rate of 22.2dB. In long-term analysis (over 30 seconds), the extraction accuracy was 84.3%, with a feature separation of 0.86. The raw data fully demonstrates that this technical solution outperforms existing methods in all core flight feature extraction indicators, laying a solid foundation for high-precision flight behavior prediction.

[0112] In an optional embodiment,

[0113] A collaborative avoidance control algorithm is constructed based on the artificial potential field method. The bird target is used as the repulsion source and the wind turbine generator is used as the repulsion field. The potential field gradient is calculated according to the relative distance between the bird target and the wind turbine generator. The algorithm includes:

[0114] A first repulsive potential energy function is established using the bird target as a repulsive force source. The first repulsive potential energy function is determined by a repulsive field coefficient, a relative distance of the bird target, and a repulsive field influence radius. When the relative distance of the bird target is less than or equal to the repulsive field influence radius, the first repulsive potential energy function is positively correlated with the repulsive field coefficient, the relative distance of the bird target, and the repulsive field influence radius. When the relative distance of the bird target is greater than the repulsive field influence radius, the first repulsive potential energy function is zero.

[0115] A second repulsive potential energy function is established using the wind turbine generator set as a repulsive field, wherein the second repulsive potential energy function is determined by a repulsive field coefficient, a relative distance of the wind turbine generator set, and a repulsive field influence radius. When the relative distance of the wind turbine generator set is less than or equal to the repulsive field influence radius, the second repulsive potential energy function is positively correlated with the repulsive field coefficient, the relative distance of the wind turbine generator set, and the repulsive field influence radius. When the relative distance of the wind turbine generator set is greater than the repulsive field influence radius, the second repulsive potential energy function is zero.

[0116] According to the relative distance between the bird target and the wind turbine generator set, negative gradients of the first repulsive potential energy function and the second repulsive potential energy function are respectively calculated to obtain potential field gradients.

[0117] The first repulsive potential energy function is established based on the position of the bird target, and the bird target is regarded as the source of repulsion. First, the basic parameters of the repulsive field are determined, including the repulsive field coefficient that describes the repulsive force strength, the relative distance that characterizes the distance effect, and the influence radius that defines the range of action. 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 field coefficient and decreases with the increase of the relative distance, while being modulated by the influence radius. When the relative distance exceeds the influence radius, the repulsive potential energy rapidly decays to zero, indicating that the location is no longer affected by the repulsive force of the bird target.

[0118] A second repulsive potential energy function is also established for wind turbines, treating the turbines as fixed repulsive field sources. The basic parameters of the repulsive field for the turbines are set, including the repulsive field coefficient reflecting the turbine's hazard level, the relative distance characterizing its spatial position, and the influence radius determining the protection range. Within the turbine's influence radius, the repulsive potential energy is positively correlated with the repulsive field coefficient and the influence radius, and negatively correlated with the relative distance. Together, they describe the strength of the turbine's repulsive effect on the surrounding space. When the distance exceeds the influence radius, the repulsive potential energy drops to zero, indicating that the area is not affected by the turbine's hazard.

[0119] The negative gradients of the two repulsive potential energy functions are calculated separately to obtain the potential field gradient pointing in the direction of decreasing repulsive potential energy. For the repulsive potential energy function of the bird target, its rate of change in all spatial directions is calculated and negated to obtain the first potential field gradient, which points away from the bird target. Similarly, the negative gradient of the repulsive potential energy function of the wind turbine is calculated to obtain the second potential field gradient, which points away from the turbine. These gradient vectors together construct the potential field distribution in space, which is used for subsequent obstacle avoidance path planning.

[0120] For example, a seagull is detected approaching a wind farm, 200 meters from the nearest wind turbine. The seagull's repulsive field parameters are set: the repulsive field coefficient is set to 100, representing its mobility; the influence radius is set to 300 meters, taking into account the seagull's alert range. When the relative distance between the seagull and a certain location is 150 meters, it is within the influence radius. 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 the location is not affected by the seagull's activities.

[0121] For wind turbines, the repulsion field parameters were set: the repulsion field coefficient was set to 200, reflecting the threat level to birds; the impact radius was set to 400 meters, accounting for the dangerous range of blade rotation. At a distance of 250 meters from the wind turbine, within the impact radius, a large repulsion potential energy value was calculated, indicating a high risk area. When the distance increased to 500 meters, beyond the impact radius, the repulsion potential energy dropped to zero, indicating that the location was safe to pass.

[0122] By calculating the negative gradient of the two repulsive potential energy functions, we obtain a gradient vector pointing away from the wind turbine at the seagull's current position. The gradient is proportional to the rate of change of the repulsive potential energy. This gradient information indicates a safe direction of movement, guiding the seagull away from the wind turbine danger zone and along the path with the lowest potential energy.

[0123] In this embodiment, zoning control is achieved by setting the repulsive field influence radius. Avoidance measures are only initiated when birds enter the dangerous area, avoiding unnecessary intervention and maximizing the power generation efficiency of the wind farm. The wind turbines themselves are also regarded as the source of the repulsive field, and the mutual influence between wind turbines is taken into account, avoiding the interference problem between wind turbines that may occur during the obstacle avoidance process.

[0124] In an optional embodiment,

[0125] The potential field gradient is optimized online through an adaptive fuzzy neural network to generate avoidance control parameters. The obstacle avoidance parameters are generated by combining with the sliding membrane controller and the obstacle avoidance control strategy is as follows:

[0126] Receiving the potential field gradient obtained in advance, inputting the potential field gradient into an adaptive fuzzy neural network having an input layer, a fuzzification layer, a rule layer, a rule inference layer and an output layer, and fuzzifying the potential field gradient using a Gaussian membership function in the fuzzification layer to obtain fuzzy data;

[0127] Based on the network output error, the gradient descent method is used to determine the update amount of the central value parameter and width parameter of the membership function. The central value parameter and width parameter of the membership function are optimized online, and the optimized fuzzy parameters are input into the rule layer to obtain the rule layer output vector.

[0128] Perform matrix operations on the rule layer output vector 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 the avoidance control parameters based on the optimized potential field gradient change rate;

[0129] Determine the position tracking error based on the avoidance control parameters and construct a sliding surface and a non-singular terminal sliding mode controller, perform finite time convergence through a control law structure composed of piecewise power functions, and input the optimized potential field gradient into the non-singular terminal sliding mode controller to generate obstacle avoidance parameters;

[0130] The dynamic weight factor is determined based on the position tracking error, and the obstacle avoidance parameters are weightedly combined with the output layer output of the adaptive fuzzy neural network using the dynamic weight factor to obtain the obstacle avoidance control strategy.

[0131] The pre-acquired potential field gradient is input into the adaptive fuzzy neural network. The adaptive fuzzy neural network architecture consists of five layers: 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 east component, north component, and vertical component into the corresponding neurons. Each input neuron uses a linear activation function to preserve the original characteristic properties of the input signal.

[0132] In the fuzzification layer, multiple Gaussian membership functions are applied to each gradient component for fuzzy mapping. Five different levels of linguistic variables are assigned to each component: "very small," "small," "medium," "large," and "very large." Each linguistic variable corresponds to a Gaussian function with a specific center value and width parameter. These Gaussian functions convert the precise gradient value into an uncertain fuzzy membership value, generating fuzzified data.

[0133] The error between the actual and expected outputs of the adaptive fuzzy neural network is calculated to construct a mean squared error (MSE) objective function. Based on this MSE objective function, the updated Gaussian membership function parameters are calculated using gradient descent. The gradients of the error with respect to the center and width parameters are calculated to determine the direction and step size of the parameter update. These parameters are then iteratively optimized online using a preset learning rate until the change in the parameters is less than a set threshold. The optimized fuzzified parameters are then applied to the rule layer calculation process to generate a rule layer output vector reflecting the activation level of each rule.

[0134] The output vector of the rule layer is subjected to a matrix operation on the network weight matrix. The elements in the weight matrix represent the contribution of different rules to the final output. The result of the matrix operation serves as a compensation for the original potential field gradient. This compensation is then vector-superimposed on the original potential field gradient to obtain the optimized potential field gradient. Based on the time rate of change of the optimized potential field gradient, avoidance control parameters, including direction adjustment and speed change, are generated.

[0135] The tracking error vector between the current position and the desired safe position is calculated based on the avoidance control parameters. This error vector is used as a state variable to construct a sliding mode surface, and the controller is designed using a non-singular terminal sliding mode control method. The controller's control law adopts the structure of a combination of piecewise power functions. High-order power terms are used in large deviation areas to achieve rapid convergence, while low-order power terms are used in small deviation areas to ensure smooth transition. A smooth transition function is introduced near the zero point to avoid singularity issues. The control law structure ensures that the system converges to an equilibrium state within a finite time. The optimized potential field gradient is input into the non-singular terminal sliding mode controller, and the specific obstacle avoidance parameters are generated through the control law.

[0136] A dynamic weight allocation mechanism is established based on the magnitude of the position tracking error. A continuous weight function is designed. When the tracking error is large, the weight of the neural network output is dynamically increased to enhance the system's adaptability to the environment. When the tracking error is small, the weight of the obstacle avoidance parameters is dynamically increased to ensure the stability of the control process. This dynamic weight factor is used to weight the obstacle avoidance parameters and the output of the adaptive fuzzy neural network output layer, avoiding sudden changes in the control variable and achieving smooth switching, ultimately resulting in an optimized obstacle avoidance control strategy.

[0137] For example, a white egret is detected approaching a wind turbine 300 meters away. The potential field gradient indicates that it needs to avoid the wind turbine to the northeast. This gradient is input into a five-layer fuzzy neural network. In the fuzzification layer, a Gaussian function is used to fuzzify the gradient. The Gaussian function is initially set to a center value of 0.5 and a width parameter of 0.3 to perform a preliminary fuzzy mapping of the potential field gradient.

[0138] By comparing the error between the network output and the expected obstacle avoidance effect, we found that the fuzzification effect under the initial parameters was not ideal. We optimized the parameters using gradient descent. After multiple iterations, we adjusted the center value to 0.65 and the width parameter to 0.25, significantly improving the fuzzification effect. The optimized parameters were used in rule calculation, resulting in a rule layer output that represents the importance of each obstacle avoidance rule.

[0139] A matrix operation is performed on the rule layer output and the network weights to obtain a compensation amount, indicating that a 10-degree deflection angle needs to be increased eastward. This compensation amount is superimposed on the original gradient, and the optimized potential field gradient points northeastward. The initial avoidance parameters, including the deflection angle and speed adjustment, are generated based on the gradient change rate.

[0140] A sliding mode surface is constructed based on the deviation between the egret's current position and the desired safe position, and a segmented control law is designed to regulate the obstacle avoidance process. The optimized potential field gradient is input into the sliding mode controller to generate specific obstacle avoidance parameters, such as a 15% reduction in wind turbine speed and a 20-degree blade deflection.

[0141] The tracking error is calculated in real time. When the error is large, the weight of the neural network output is increased to 0.7. When the error is small, the weight of the obstacle avoidance parameter is increased to 0.6. Through dynamic weight adjustment, the obstacle avoidance parameters and network output are weighted and combined to generate an optimized obstacle avoidance control strategy, guiding the wind turbine to adjust in the expected manner and ensuring the egret's safe avoidance.

[0142] In this embodiment, by integrating the language processing capabilities of fuzzy logic and the learning capabilities of neural networks, it is possible to handle the uncertainty and ambiguity in the system. Through the online parameter optimization mechanism, the system can continuously learn and improve. As the running time increases, the control accuracy continues to improve. The introduction of the non-singular terminal sliding mode controller solves the stability problems that may exist in traditional control methods and ensures that the system converges to the target state within a limited time.

[0143] In an optional embodiment,

[0144] Based on the avoidance control parameters, the position tracking error is determined and a sliding surface and a non-singular terminal sliding mode controller are constructed. Finite time convergence is performed through a control law structure composed of piecewise power functions. The optimized potential field gradient is input into the non-singular terminal sliding mode controller to generate obstacle avoidance parameters, including:

[0145] constructing a position tracking error mapping function based on an avoidance control parameter, wherein when the absolute value of the avoidance control parameter is greater than or equal to a threshold parameter, a first mapping coefficient and a nonlinear power term of the avoidance control parameter minus a reference control amount are multiplied as a position tracking error output; and when the absolute value of the avoidance control parameter is less than the threshold parameter, a second mapping coefficient and a sine function based on the ratio of the avoidance control parameter to the threshold parameter are multiplied as a position tracking error output;

[0146] Constructing a nonlinear sliding mode surface containing a dynamic weight based on the position tracking error, the nonlinear sliding mode surface consisting of a derivative term of the position tracking error, a product term of a dynamic weight factor and the position tracking error, and a nonlinear power integral term, wherein the dynamic weight factor includes a basic weight and an exponential decay adaptive term based on the position tracking error;

[0147] A piecewise continuous control law is designed based on the nonlinear sliding surface. When the absolute value of the nonlinear sliding surface is greater than or equal to a switching domain threshold, a combination of two different power terms is used as a control output. When the absolute value of the nonlinear sliding surface is less than the switching domain threshold, a combination of a linear term and a fractional power term is used as a control output.

[0148] The product of the Sigmoid function and the optimized potential field gradient constraint function is used as the smooth transition function of the nonlinear sliding surface, and the control output is superimposed with the product of the smooth transition function and the optimized potential field gradient to obtain the obstacle avoidance parameter.

[0149] A piecewise mapping function for 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 amount is calculated, the difference is 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 for the area. 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, the ratio is substituted into a sine function, and the function output is multiplied by the second mapping coefficient to obtain the position tracking error output for the area.

[0150] A nonlinear sliding surface is constructed using the position tracking error output by the mapping function. The derivative of the position tracking error is calculated as the first component of the sliding surface. A dynamic weight factor is constructed, consisting of a base weight and an adaptive term. The base weight is a pre-set fixed value, while 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 to form the second component of the sliding surface. A nonlinear power integration operation is performed on the position tracking error, and the result is used as the third component of the sliding surface. These three components are combined in a pre-set manner to construct a complete nonlinear sliding surface.

[0151] Design a piecewise continuous control law based on the nonlinear sliding surface. Obtain the absolute value of the nonlinear sliding surface and compare it with the switching threshold. When the absolute value of the nonlinear sliding surface is greater than or equal to the switching threshold, select two different power exponents, substitute the sliding surface into these two power terms, and combine the results according to preset weights to obtain the control output for that area. When the absolute value of the nonlinear sliding surface is less than the switching threshold, substitute the sliding surface into a linear function to obtain the linear term output, and simultaneously substitute it into a fractional power function to obtain the nonlinear term output. Combine these two terms according to preset weights to obtain the control output for that area.

[0152] Construct a smooth transition function, using the Sigmoid function as the base transition function. Design a potential field gradient constraint function, which determines the constraint boundary based on the optimized potential field gradient characteristics. Multiply the Sigmoid function with the potential field gradient constraint function to obtain the smooth transition function of the nonlinear sliding surface. Calculate the product of the control output, the smooth transition function, and the optimized potential field gradient. Superimpose this product with the previously obtained control output to obtain the obstacle avoidance parameters.

[0153] For example, a bald eagle is detected approaching a wind farm, and the current avoidance control parameter is a 30-degree deflection angle. The threshold parameter is set to 20 degrees, and the reference control amount is set to 0 degrees. Because the absolute value of the avoidance control parameter, 30 degrees, is greater than the threshold of 20 degrees, the position tracking error is calculated using the first mapping coefficient of 0.8 and the nonlinear power term, resulting in a tracking error of 45 meters.

[0154] A nonlinear sliding surface is constructed based on a 45-meter tracking error. The base weight is set to 0.6. When the tracking error is 45 meters, the exponential decay adaptive term is calculated to be 0.2, resulting in a dynamic weight factor of 0.8. The derivative of the tracking error and the nonlinear power integral term are simultaneously calculated to form the complete nonlinear sliding surface.

[0155] When the calculated absolute value of the nonlinear sliding surface is 0.9, which is greater than the switching threshold of 0.5, 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 generate a smooth transition function, whose output value at the current position is 0.95. The control output is then multiplied with the smooth transition function and the optimized potential field gradient, resulting in obstacle avoidance parameters of 25-degree deflection angle and 100-meter safety distance.

[0156] In this embodiment, nonlinear power term mapping is used in large deviation areas to provide greater control strength; in small deviation areas, sinusoidal function mapping is used to achieve smooth transition. By combining the basic weight with the exponential decay adaptive term, dynamic adjustment of the weight factor is achieved. The nonlinear power integral term is introduced in the sliding 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 areas.

[0157] In existing technologies, obstacle avoidance control for flight trajectories primarily relies on simple linear mapping or fixed-threshold segmented control methods, which struggle to balance rapidity and stability. Conventional sliding mode controllers typically employ fixed-weight linear sliding surfaces, which are unable to flexibly adjust control characteristics based on changes in tracking error. Control law laws are prone to jumps in control variables near switching points, causing violent oscillations in the system and impacting the stability of obstacle avoidance control.

[0158] This embodiment improves the system's ability to handle deviations of varying degrees through a nonlinear segmented mapping mechanism. The design of a dynamic weighted nonlinear sliding surface enhances the adaptability and robustness of the control system. The design of a segmented continuous control law improves the system's control performance in different working areas, enhances the control system's adaptability and reliability, and provides more effective technical support for bird obstacle avoidance control in wind farms.

[0159] Figure 4 This figure shows the effect of the obstacle avoidance technology used in the neural network-based wind farm bird trajectory prediction and wind turbine linkage control method according to an embodiment of the present invention. The simulation data is based on a high-precision trajectory control experiment 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). The obstacle radius range is 70-100m, and the potential field influence range is set to three times the obstacle radius. The control target navigates from the starting point (100m, 200m) to the end point (1300m, 400m), a total distance of approximately 1400m.

[0160] The figure uses different colors to distinguish between different control methods. The green trajectory represents our solution (non-singular terminal sliding mode controller), the yellow trajectory represents traditional control methods (such as PID control), the solid red circles represent obstacles, the light red areas represent the potential field's influence, and the black concentric circles represent the oscillation points of the traditional method's trajectory. Blue dots mark the starting and ending points, as well as key control turning points. Both the horizontal and vertical axes are in meters (m), clearly indicating spatial relationships.

[0161] Comparative data shows that the trajectory smoothness of this technical solution reaches 93.5%, which is significantly higher than the 61.8% of the traditional control method. In terms of safety margin indicators, the stability of this technical solution in maintaining a safe distance of 100 meters is 96.2%, while the traditional method is only 72.4%. In the time domain analysis, the control convergence time of this technical solution is 2.7 seconds, which is 40% less than the 4.5 seconds of the traditional method; in terms of energy efficiency, the control energy consumption of this technical solution is reduced by 37%. Spatial trajectory analysis shows that the obstacle avoidance steering angles of this technical solution at the three key obstacle positions are 30°, 15° and 25° respectively. The average position tracking error at each position is 45 meters, and the maximum sliding surface value is 0.9, which is much lower than the 78-meter error and 1.6 sliding surface value of the traditional method.

[0162] In terms of control stability, this technical solution's adaptive dynamic weight factor adjustment range is 0.6-0.8, effectively suppressing system oscillations. Conventional methods exhibit significant oscillations at two obstacle avoidance points (the black concentric circles in the figure), with amplitudes reaching 40-60 meters. Analysis of the sliding surface construction efficiency shows that the switching characteristics of this technical solution reduce the system's response time to obstacles to only 58% of that of conventional methods, enabling faster obstacle avoidance maneuvers.

[0163] Experimental data proves that the non-singular terminal sliding mode controller based on the piecewise power function combination control law is significantly superior to traditional control methods 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.

[0164] In an optional embodiment,

[0165] Execute the obstacle avoidance control strategy and monitor the adjusted wind turbine operating status in real time until the bird target flies away from the wind farm. Then restore the wind turbine to normal operation according to the preset recovery strategy, including:

[0166] Executing an obstacle avoidance control strategy to adjust the wind turbine generator set and monitoring the operating status of the wind turbine generator set and the flight trajectory of the bird target in real time;

[0167] When the bird target flies away from the wind farm, the current operating status parameters of the wind turbine generator set are obtained. According to the operating status parameters and the preset recovery threshold, the speed, torque and pitch angle of the wind turbine generator set are gradually adjusted to the corresponding target values according to different recovery slopes to restore the normal operating state.

[0168] Based on the generated obstacle avoidance control strategy, key operating parameters such as the wind turbine's speed, torque, and pitch angle are adjusted in real time. During the obstacle avoidance control process, the system continuously collects data on the wind turbine's operating status, including actual speed, output torque, pitch angle, and power output. Radar and visual sensors simultaneously track and monitor the flight trajectory of bird targets, recording changes in their position, speed, and flight direction.

[0169] When the bird is detected to have completely left the wind farm, the system immediately acquires the wind turbine's operating parameters. These parameters are compared with the target values for normal operation, and the deviation of each parameter is calculated. The system then reads the pre-set recovery thresholds, which are set based on the adjustment characteristics of different parameters and serve to define the different stages of parameter recovery adjustment.

[0170] A three-stage progressive recovery strategy is used to adjust wind turbine speed. When the speed deviation exceeds the first recovery threshold, a smaller recovery slope is used to avoid drastic system fluctuations. When the speed deviation is between the first and second recovery thresholds, a medium recovery slope is used to accelerate recovery. When the speed deviation is less than the second recovery threshold, a larger recovery slope is used to quickly reach the target speed.

[0171] A two-stage progressive adjustment strategy is used to restore generator output torque. When the torque deviation exceeds the recovery threshold, a slower recovery slope is used to ensure that the unit's mechanical stress does not suddenly change. When the torque deviation is less than the recovery threshold, a faster recovery slope is used to accelerate the adjustment and quickly restore the rated torque output.

[0172] A segmented, continuous recovery strategy is used for adjusting the pitch angle of a wind turbine. When the pitch angle deviation exceeds the first recovery threshold, the minimum recovery slope is used to fine-tune the blade angle. When the pitch angle deviation is between the two recovery thresholds, a medium recovery slope is used. When the pitch angle deviation is less than the second recovery threshold, a larger recovery slope is used to quickly adjust to the target angle. Throughout the adjustment process, the blade vibration state is monitored in real time to ensure a smooth and controllable adjustment process.

[0173] During the parameter recovery and adjustment process, the changing trends of various parameters are continuously monitored. When a parameter reaches its target value, it is locked in place and adjustments are made to other parameters that have not yet reached their target value. This step-by-step adjustment method allows for a smooth recovery of the wind turbine's operating parameters, ultimately returning the unit to normal operation.

[0174] For example, if a sea eagle is detected flying into the wind farm from the east, obstacle avoidance control reduces the wind turbine 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. After the sea eagle leaves the wind farm, the wind turbine's current operating parameters are obtained, and the recovery process for speed, torque, and pitch angle is divided into different adjustment stages. Speed recovery uses a three-stage adjustment method: when the speed deviation exceeds 20% of the rated value, the speed increases by 2% per second; when the deviation is between 20% and 10%, the speed increases by 3% per second; when the deviation is less than 10%, the speed increases by 5% per second until the rated speed is reached. Torque recovery uses a two-stage adjustment method: when the torque deviation exceeds 15% of the rated value, the speed increases by 3% per second; when the deviation is less than 15%, the speed increases by 6% per second. Pitch angle recovery also uses a staged adjustment strategy to smoothly restore all parameters to normal operating conditions.

[0175] In this embodiment, the segmented and progressive recovery method ensures a smooth transition of the wind turbine generator set from the obstacle avoidance state to the normal operating state, avoids mechanical shock and grid fluctuations that may be caused by parameter mutations, and sets different recovery slope strategies for different parameters to optimize the time and safety of the recovery process. The fully automated recovery process does not require human intervention, greatly improving the system's response speed and work efficiency.

[0176] According to a second aspect of an embodiment of the present invention, an electronic device is provided, including:

[0177] A processor; a memory for storing processor-executable instructions, wherein the processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0178] According to a third aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0179] 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 carrying computer-readable program instructions for executing various aspects of the present invention.

[0180] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, 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 wind farms and controlling wind turbine linkages based on neural networks, characterized in that: include: Collect real-time image data and use a pre-trained target detection neural network to identify bird targets, and generate bird flight trajectory data based on the position information of bird targets in multiple consecutive frames of images; Performing a 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 using an adaptive decomposition algorithm, extracting amplitude-frequency characteristic parameters and performing feature dimensionality reduction and reconstruction to obtain flight characteristic vectors, which are then subjected to time series analysis. A predicted characteristic vector sequence is generated, and trajectory corrections are calculated based on phase differences and amplitude changes to obtain predicted trajectory data. Identify wind turbines in a wind farm that present a bird collision risk based on predicted trajectory data; A collaborative avoidance control algorithm is constructed based on the artificial potential field method. The bird target is used as the repulsion source and the wind turbine is used as the repulsion field. The potential field gradient is calculated according to the relative distance between the bird target and the wind turbine. The potential field gradient is optimized online through an adaptive fuzzy neural network to generate avoidance control parameters. The sliding membrane controller is used to generate obstacle avoidance parameters, and the combined method is used to obtain an obstacle avoidance control strategy. The obstacle avoidance control strategy is executed and the adjusted operating status of the wind turbine is monitored in real time until the bird target flies away from the wind farm, and the wind turbine is restored to normal operating status according to the preset recovery strategy.

2. The method according to claim 1, characterized in that Performing a 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 eigenmode components includes: Performing a three-dimensional Hilbert-Huang transform on the bird flight trajectory data to obtain analytical signals in the directions of three coordinate axes, calculating the instantaneous amplitude and instantaneous phase in the directions of the three coordinate axes 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, employing an adaptive decomposition algorithm to determine all local extreme value points of the initial residual signal, employing a cubic spline interpolation method to construct an upper envelope and a lower envelope of the initial residual signal, calculating the average 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; Determine 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. Otherwise, use the detail component as a new residual signal and repeatedly execute the adaptive decomposition algorithm until the intrinsic mode component is obtained.

3. The method according to claim 1, characterized in that Extract the amplitude-frequency characteristic parameters and perform feature dimensionality reduction and reconstruction to obtain the flight characteristic vector and perform time series analysis. Generate a predicted characteristic vector sequence and calculate the trajectory correction value 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 amplitude-frequency characteristic parameters are obtained by extracting the processed instantaneous phase and energy change rate and the characteristic matrix is constructed by arranging them in time sequence; Calculating the time derivative of the feature matrix to obtain the feature change rate and calculating the sample importance weight, weighting the feature matrix based on the sample importance weight to obtain a weighted covariance matrix, determining a first eigenvector based on the eigenvalue interval ratio, cumulative contribution rate and reconstruction error of the weighted covariance matrix, and performing dimensionality reduction and reconstruction on the feature matrix based on the first eigenvector to obtain a projection matrix; Multiplying the feature matrix by the projection matrix to obtain a flight feature vector, establishing a time series relationship of a quadratic term local trend, and performing time series prediction using iterative reweighted least squares method to obtain a predicted feature vector sequence; Phase difference parameters and amplitude change parameters at adjacent moments are calculated based on the predicted feature vector sequence. Combined weight coefficients of the phase difference parameters and amplitude change parameters are initialized based on historical trajectory data and weighted summed. The weighted sum is used as a trajectory correction. The combined weight coefficients are updated according to the prediction error. A smoothing correction is calculated based on the trajectory correction and a smoothing factor is dynamically adjusted. The smoothing factor is combined with the predicted feature vector sequence to obtain predicted trajectory data.

4. The method according to claim 3, characterized in that The time series relationship of the local trend of the quadratic term is established, and the time series prediction is performed by combining the iterative reweighted least squares method to obtain the predicted feature vector sequence including: The flight feature vector is divided into short-term, medium-term and long-term. The variational constrained optimization objective function is constructed by calculating the sequence envelope and instantaneous frequency. The center frequency and bandwidth parameters are iteratively updated to extract the adaptive multi-scale components with optimal energy concentration. The adaptive multi-scale components are transformed by dual-tree complex wavelet transform to obtain multi-level sparse time-frequency features. Combined with time series position coding, an enhanced input sequence with time series position information is generated. A multi-layer biological perception field network is constructed based on the enhanced input sequence. In each layer of the perception field, key feature patterns are extracted by the excitatory neuron group, and selective enhancement is performed by the inhibitory neuron group. The activation state of the neurons is adjusted by combining the adaptive dynamic threshold to obtain the perception features. 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, and dynamic weight allocation is performed using an adaptive gating unit. 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 iterative reweighted least squares method is used for optimization and solution. A dynamic prediction confidence interval is constructed through multiple samplings, and the historical data weight and prediction credibility weight are adjusted according to the interval width, and a prediction feature vector sequence is output.

5. The method according to claim 1, wherein A collaborative avoidance control algorithm is constructed based on the artificial potential field method. The bird target is used as the repulsion source and the wind turbine generator is used as the repulsion field. The potential field gradient is calculated according to the relative distance between the bird target and the wind turbine generator. The algorithm includes: A first repulsive potential energy function is established using the bird target as a repulsive force source. The first repulsive potential energy function is determined by a repulsive field coefficient, a relative distance of the bird target, and a repulsive field influence radius. When the relative distance of the bird target is less than or equal to the repulsive field influence radius, the first repulsive potential energy function is positively correlated with the repulsive field coefficient, the relative distance of the bird target, and the repulsive field influence radius. When the relative distance of the bird target is greater than the repulsive field influence radius, the first repulsive potential energy function is zero. A second repulsive potential energy function is established using the wind turbine generator set as a repulsive field, wherein the second repulsive potential energy function is determined by a repulsive field coefficient, a relative distance of the wind turbine generator set, and a repulsive field influence radius. When the relative distance of the wind turbine generator set is less than or equal to the repulsive field influence radius, the second repulsive potential energy function is positively correlated with the repulsive field coefficient, the relative distance of the wind turbine generator set, and the repulsive field influence radius. When the relative distance of the wind turbine generator set is greater than the repulsive field influence radius, the second repulsive potential energy function is zero. According to the relative distance between the bird target and the wind turbine generator set, negative gradients of the first repulsive potential energy function and the second repulsive potential energy function are respectively calculated to obtain potential field gradients.

6. The method according to claim 1, characterized in that The potential field gradient is optimized online through an adaptive fuzzy neural network to generate avoidance control parameters. The obstacle avoidance parameters are generated by combining with the sliding membrane controller and the obstacle avoidance control strategy is as follows: Receiving the potential field gradient obtained in advance, inputting the potential field gradient into an adaptive fuzzy neural network having an input layer, a fuzzification layer, a rule layer, a rule inference layer and an output layer, and fuzzifying the potential field gradient using a Gaussian membership function in the fuzzification layer to obtain fuzzy data; Based on the network output error, the gradient descent method is used to determine the update amount of the central value parameter and width parameter of the membership function. The central value parameter and width parameter of the membership function are optimized online, and the optimized fuzzy parameters are input into the rule layer to obtain the rule layer output vector. Perform matrix operations on the rule layer output vector 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 the avoidance control parameters based on the optimized potential field gradient change rate; Determine the position tracking error based on the avoidance control parameters and construct a sliding surface and a non-singular terminal sliding mode controller, perform finite time convergence through a control law structure composed of piecewise power functions, and input the optimized potential field gradient into the non-singular terminal sliding mode controller to generate obstacle avoidance parameters; The dynamic weight factor is determined based on the position tracking error, and the obstacle avoidance parameters are weightedly combined with the output layer output of the adaptive fuzzy neural network using the dynamic weight factor to obtain the obstacle avoidance control strategy.

7. The method according to claim 6, characterized in that Based on the avoidance control parameters, the position tracking error is determined and a sliding surface and a non-singular terminal sliding mode controller are constructed. Finite time convergence is performed through a control law structure composed of piecewise power functions. The optimized potential field gradient is input into the non-singular terminal sliding mode controller to generate obstacle avoidance parameters, including: constructing a position tracking error mapping function based on an avoidance control parameter, wherein when the absolute value of the avoidance control parameter is greater than or equal to a threshold parameter, a first mapping coefficient and a nonlinear power term of the avoidance control parameter minus a reference control amount are multiplied as a position tracking error output; and when the absolute value of the avoidance control parameter is less than the threshold parameter, a second mapping coefficient and a sine function based on the ratio of the avoidance control parameter to the threshold parameter are multiplied as a position tracking error output; Constructing a nonlinear sliding mode surface containing a dynamic weight based on the position tracking error, the nonlinear sliding mode surface consisting of a derivative term of the position tracking error, a product term of a dynamic weight factor and the position tracking error, and a nonlinear power integral term, wherein the dynamic weight factor includes a basic weight and an exponential decay adaptive term based on the position tracking error; A piecewise continuous control law is designed based on the nonlinear sliding surface. When the absolute value of the nonlinear sliding surface is greater than or equal to a switching domain threshold, a combination of two different power terms is used as a control output. When the absolute value of the nonlinear sliding surface is less than the switching domain threshold, a combination of a linear term and a fractional power term is used as a control output. The product of the Sigmoid function and the optimized potential field gradient constraint function is used as the smooth transition function of the nonlinear sliding surface, and the control output is superimposed 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, characterized in that Execute the obstacle avoidance control strategy and monitor the adjusted wind turbine operating status in real time until the bird target flies away from the wind farm. Then restore the wind turbine to normal operation according to the preset recovery strategy, including: Executing an obstacle avoidance control strategy to adjust the wind turbine generator set and monitoring the operating status of the wind turbine generator set and the flight trajectory of the bird target in real time; When the bird target flies away from the wind farm, the current operating status parameters of the wind turbine generator set are obtained. According to the operating status parameters and the preset recovery threshold, the speed, torque and pitch angle of the wind turbine generator set are gradually adjusted to the corresponding target values according to different recovery slopes to restore the normal operating state.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; 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: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 8 is implemented.

Images