A wind turbine clearance prediction method based on evolution of blade aerodynamic characteristics

Abstract

The application discloses a wind turbine clearance prediction method based on blade aerodynamic characteristic evolution, and belongs to the technical field of safety monitoring of wind power generators. The method acquires multi-source monitoring data such as blade root strain, azimuth angle, wind speed and rotating speed; separates the strain signal into three physical driving components of aerodynamic force, inertial force and gravity through variational mode decomposition; extracts dynamic bending characteristics and static bending characteristics of the blade from the inertial force component and the gravity component respectively; performs angle domain conversion and continuous wavelet transformation on the aerodynamic force component, and extracts the main frequency and amplitude spectral density of the aerodynamic excitation; inputs the above characteristics into a sparse time sequence convolution network, and outputs a tip dynamic deflection prediction value, wherein the sparsity of the network is adaptively determined according to the main frequency of the aerodynamic excitation, and a physical constraint loss function is used in the training process; and the clearance distance is calculated in combination with the blade spatial azimuth angle and the deflection prediction value. The application does not need to additionally install hardware such as laser radar, and realizes all-weather and high-precision clearance prediction.

Description

Technical Field

[0001] This invention relates to the field of wind turbine safety monitoring technology, and in particular to a wind turbine headroom prediction method based on the evolution of blade aerodynamic characteristics. Background Technology

[0002] Wind turbine clearance refers to the minimum distance between the blade tip and the tower during wind turbine operation. Insufficient clearance can lead to blade tip collisions with the tower, causing serious safety accidents such as blade breakage and tower collapse. With the continuous increase in single-unit capacity and rotor diameter of wind turbines (for example, the rotor diameter of Envision Energy's EN-200 / 6.25 model has reached 200m), blade deflection during operation is becoming increasingly significant, making clearance safety a more prominent issue. Especially under extreme wind conditions, high turbulence intensity, or emergency pitch control, the dynamic deflection of the blades increases sharply, and the clearance distance decreases rapidly, posing a serious threat to the safe operation of the turbine.

[0003] Currently, the existing technologies for monitoring and predicting the airspace above wind turbines mainly employ the following three types of solutions: The first type is the direct clearance measurement scheme based on lidar. This scheme directly measures the distance between the blade tip and the tower by installing lidar equipment at the bottom of the nacelle or the top of the tower, emitting a laser beam towards the blade tip and receiving the reflected signal. However, this type of scheme has several inherent defects: (1) lidar equipment is extremely expensive, with the cost of a single unit reaching hundreds of thousands of yuan, which greatly increases the construction cost of wind farms; (2) the performance of lidar is severely degraded under severe weather conditions such as rain, fog, and snow, and the measurement accuracy and reliability cannot be guaranteed; (3) lidar equipment needs to be calibrated and maintained regularly, resulting in high operation and maintenance costs, and the equipment itself is subjected to harsh environmental conditions outside the nacelle, resulting in a limited service life.

[0004] The second type is the clearance measurement scheme based on video image analysis. This scheme involves installing industrial cameras on the tower or nacelle to collect images of the blades in operation, and then calculating the pixel distance between the blade tip and the tower through image recognition and visual ranging algorithms, which is then converted into the actual spatial distance. However, this type of scheme has the following technical bottlenecks: (1) Image quality is greatly affected by lighting conditions and weather conditions, and cannot work normally at night or in foggy conditions; (2) The visual ranging algorithm is extremely sensitive to camera calibration parameters, and parameter drift caused by camera vibration or temperature changes will directly affect the measurement accuracy; (3) The image processing algorithm has a large computational load, making it difficult to implement real-time processing on the edge computing equipment of the wind turbine.

[0005] The third category is the air clearance prediction scheme based on physical model simulation. This scheme establishes a multibody dynamics model of the wind turbine, inputs operating parameters such as wind speed, rotational speed, and pitch angle, and simulates and calculates the dynamic deflection of the blades and the air clearance. However, this type of scheme has the following shortcomings: (1) The physical model is highly dependent on prior information such as blade structural parameters and aerodynamic parameters, while the stiffness and damping characteristics of the blades will change due to fatigue accumulation during actual operation, resulting in a gradual decrease in the accuracy of the model prediction; (2) The computational load of the physical model is extremely large, making it difficult to meet the needs of real-time online prediction; (3) The model cannot be updated adaptively and cannot be corrected using actual operating data.

[0006] In summary, existing airspace monitoring and prediction schemes all have significant technical limitations: direct measurement schemes are costly and susceptible to environmental interference, while physical model simulation schemes involve large computational loads and cannot be updated adaptively. How to achieve high-precision, real-time airspace prediction using existing sensor signals from wind turbines without requiring additional expensive hardware is a pressing technical challenge in this field. Summary of the Invention

[0007] The purpose of this invention is to provide a wind turbine clearance prediction method based on the evolution of blade aerodynamic characteristics. This method solves the problems of existing technologies, such as reliance on expensive hardware for clearance monitoring, susceptibility to environmental interference, and large computational load of physical models that cannot be adaptively updated. It achieves high-precision real-time prediction of clearance distance without increasing hardware investment.

[0008] To achieve the above objectives, this invention provides a method for predicting the headroom of wind turbines based on the evolution of blade aerodynamic characteristics. include: S1. Acquire multi-source monitoring data of wind turbine blades under operating conditions. The multi-source monitoring data includes blade root strain time-series signal, blade azimuth angle time-series signal, nacelle wind speed time-series signal, and wind turbine speed time-series signal. S2. Perform variational mode decomposition on the strain time series signal at the root of the blade to separate it into aerodynamic strain components, inertial strain components, and gravity strain components. S3. Based on the blade azimuth time-series signal and the wind turbine rotation speed time-series signal, extract the blade dynamic bending characteristics and the blade static bending characteristics from the inertial force driven strain component and the gravity driven strain component, respectively. S4. Based on the nacelle wind speed time-series signal and the blade azimuth angle time-series signal, perform time-frequency analysis on the aerodynamic drive strain component and extract the blade aerodynamic excitation characteristics. S5. Input the dynamic bending characteristics, static bending characteristics and aerodynamic excitation characteristics of the blade into the pre-built blade deflection prediction model. The blade deflection prediction model is built based on a sparse temporal convolutional network and outputs the predicted dynamic deflection value of the blade at the tip position. S6. Determine the current spatial azimuth of the blade based on the wind turbine rotation speed timing signal and the blade azimuth timing signal, and calculate the predicted clearance distance between the blade tip and the tower based on the spatial azimuth and the predicted dynamic deflection value.

[0009] Preferably, in step S2, variational mode decomposition is performed on the strain time series signal at the blade root to separate it into aerodynamically driven strain components, inertial force-driven strain components, and gravity-driven strain components, specifically including: Construct the objective function for variational mode decomposition, which is: ; in, For the first One modal component, For the first The center frequency of each modal component For the Dirac function, For convolution operators, The imaginary unit, For time The partial derivative operator, for Norm; Set the number of modes , respectively corresponding to the aerodynamic strain components Inertial force-driven strain components and gravity-driven strain components ; The objective function is solved iteratively using the alternating direction multiplier method until the convergence condition is met, and the three separated modal components are output.

[0010] Preferably, in step S3, the dynamic bending characteristics of the blade are extracted from the inertial force-driven strain component based on the blade azimuth angle timing signal and the rotor speed timing signal, specifically including: Based on the timing signals of the blade azimuth angle and the wind turbine rotation speed, calculate the tangential acceleration and normal acceleration of the blade in the plane of rotation. Based on the tangential acceleration and the normal acceleration, the inertial force driven strain component is decoupled from the inertial force to extract the dynamic strain component caused by the blade bending deformation. The dynamic strain component is subjected to time-domain statistical feature extraction and frequency-domain feature extraction to generate the dynamic bending feature of the blade. The dynamic bending feature of the blade includes at least the dynamic strain amplitude, dynamic strain frequency and dynamic strain phase.

[0011] Preferably, in step S3, the static bending characteristics of the blade are extracted from the gravity-driven strain component based on the blade azimuth time-series signal and the rotor speed time-series signal, specifically including: Based on the timing signal of the blade azimuth angle, the spatial azimuth angle of the blade during the rotation process is determined. The spatial azimuth angle is the angle between the blade axis and the vertical direction. Based on the spatial azimuth, gravity decoupling is performed on the gravity-driven strain component to extract the static strain component caused by the blade's own weight. Statistical analysis of the static strain components is performed to generate the static bending characteristics of the blade. The static bending characteristics of the blade include at least the mean static strain, the maximum static strain, and the static strain change rate.

[0012] Preferably, in step S4, based on the nacelle wind speed time-series signal and the blade azimuth angle time-series signal, a time-frequency analysis is performed on the aerodynamic strain component to extract the blade aerodynamic excitation characteristics, specifically including: Based on the blade azimuth time sequence signal, the aerodynamic drive strain component is transformed from the time domain to the angle domain under the impeller rotation coordinate system to generate the angle domain aerodynamic drive strain signal. A continuous wavelet transform is performed on the aerodynamic strain signal in the angular domain to generate the time spectrum: ; in, These are wavelet coefficients. For scale parameters, For translation parameters, For angle domain variables, This represents the aerodynamic strain signal in the angular domain. It is the conjugate function of the mother wavelet function; Based on the time spectrum, the aerodynamic excitation characteristics of the blade are extracted. The aerodynamic excitation characteristics of the blade include at least the aerodynamic excitation main frequency, the aerodynamic excitation amplitude spectral density, and the aerodynamic excitation bandwidth.

[0013] Preferably, in S5, the blade deflection prediction model is constructed based on a sparse temporal convolutional network, which includes a series of sparse causal convolutional modules and a global regression output layer connected in sequence. Each sparse causal convolution module includes a sparse convolution kernel, a gated recurrent unit, and residual connections; The sparse convolution kernel is implemented by introducing a sparse mask matrix into the standard convolution kernel. The element values ​​of the sparse mask matrix are adaptively determined according to the main frequency of the blade's aerodynamic excitation characteristics, which is used to enhance the sparse temporal convolution network's ability to perceive the frequency components of aerodynamic excitation. During the training process of this blade deflection prediction model, a physical constraint loss function is used, which includes a data fitting term and a physical constraint term. The physical constraint term is constructed based on the linear physical relationship between blade deflection and blade root bending moment.

[0014] Preferably, the adaptive determination method for the sparse mask matrix is ​​as follows: Based on the dominant frequency of the aerodynamic excitation characteristics of this blade Calculate the sparsity of the sparse convolution kernel. The calculation formula is: ; in, As the baseline sparsity, For adjustment coefficients, For reference frequency; Based on this sparsity The proportion of random selections in the weight matrix of the standard convolution kernel is... The weight elements are determined, and the remaining weight elements are set to zero to generate the sparse mask matrix.

[0015] Preferably, the physical constraint loss function is: ; in, For the sample size, For the first The blade deflection label value for each sample. This is the predicted blade deflection value output by the blade deflection prediction model. These are the physical constraint weighting coefficients. This is the proportionality coefficient between blade deflection and blade root bending moment. For the first Measured values ​​of the bending moment at the root of the leaf for each sample; This physical constraint Based on the linear physical relationship between blade deflection and blade root bending moment, the output of the blade deflection prediction model is constrained to conform to the laws of blade structural mechanics.

[0016] Preferably, in step S6, the current spatial azimuth of the blade is determined based on the rotor speed timing signal and the blade azimuth timing signal, and the predicted clearance distance between the blade tip and the tower is calculated based on the spatial azimuth and the predicted dynamic deflection value. Specifically, this includes: Based on the wind turbine rotation speed timing signal, calculate the cumulative rotation angle of the wind turbine within a preset time window; Based on the cumulative rotation angle and the blade azimuth time-series signal, determine the current spatial azimuth of the blade. ; According to this spatial azimuth angle The predicted value of dynamic deflection In addition to the initial geometric parameters of the blade, the predicted clearance distance between the blade tip and the tower is calculated. The calculation formula is: ; in, This is the initial distance between the blade tip and the tower when the blade has not undergone deflection deformation. Where is the radius of the wind turbine. This is the azimuth angle in space. This is the predicted value for the dynamic deflection.

[0017] Preferred options also include: Acquire the historical operating data of the wind turbine, which includes historical signals of blade root strain, blade azimuth angle, nacelle wind speed, rotor speed, and corresponding blade tip clearance distance labels. A training sample set was constructed using the historical operating data, and the blade tip clearance distance label was used as a supervision signal to train the blade deflection prediction model. During training, the Bayesian hyperparameter optimization method is used to optimize the network hyperparameters of the blade deflection prediction model. The network hyperparameters include at least the sparse convolution kernel size, the number of network layers, and the sparsity ratio. After training, the blade deflection prediction model is deployed to the main control system or edge computing device of the wind turbine for online air clearance prediction.

[0018] Therefore, the wind turbine headroom prediction method based on the evolution of blade aerodynamic characteristics using the above-described structure has the following beneficial effects: (1) This invention does not require additional expensive hardware equipment such as lidar and industrial cameras. It can achieve clearance distance prediction by using only the existing blade root strain sensor, azimuth encoder, anemometer and speed encoder of the wind turbine, which greatly reduces hardware investment and operation and maintenance costs and avoids the impact of harsh environment on sensor performance.

[0019] (2) This invention separates the blade root strain signal into aerodynamic driving component, inertial driving component and gravity driving component by variational mode decomposition technology, realizing effective decoupling of multiple physical excitation sources. The dynamic bending feature, static bending feature and aerodynamic excitation feature of the blade are extracted from the strain components of different physical mechanisms respectively. The feature extraction path is clear and the physical meaning is clear, which significantly improves the accuracy and interpretability of deflection prediction.

[0020] (3) This invention uses a sparse temporal convolutional network to construct a blade deflection prediction model. By introducing sparse convolutional kernels and an adaptive sparsity mechanism, the computational load of the model is significantly reduced while ensuring prediction accuracy, enabling it to be deployed in real time on the edge computing device of the wind turbine. At the same time, the introduction of the physical constraint loss function makes the model output conform to the basic laws of blade structural mechanics, enhancing the model's generalization ability and reliability.

[0021] (4) The airspace prediction method proposed in this invention does not depend on meteorological conditions (rain, fog, sunlight, etc.) and can work normally under all weather conditions. It overcomes the inherent defects of lidar and video measurement schemes being greatly affected by weather, and has wider applicability and higher reliability.

[0022] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0023] Figure 1 This is an overall flowchart of a wind turbine headroom prediction method based on the evolution of blade aerodynamic characteristics according to the present invention. Figure 2 This is a schematic diagram of the variational mode decomposition and strain component separation process in this invention; Figure 3 This is a schematic diagram of the sparse temporal convolutional network in this invention. Detailed Implementation

[0024] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0025] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0026] Example like Figure 1 As shown, this invention provides a wind turbine headroom prediction method based on the evolution of blade aerodynamic characteristics. The method specifically includes the following steps: S1. Simultaneous acquisition of multi-source monitoring data: Acquire multi-source monitoring data of the wind turbine blades during operation. The sampling frequency of this multi-source monitoring data is set to 200Hz to ensure complete capture of the high-frequency components of blade vibration (the first-order flapping frequency of the blade is usually between 0.5Hz and 2Hz, and a sampling frequency of 200Hz is sufficient to cover its higher-order harmonic components). The preset time window length is set to 30 seconds to balance data integrity and real-time calculation.

[0027] The specific signals acquired include: (1) Strain timing signal at the root of the blade Strain readings are obtained by installing strain gauges at the blade root flange along the blade flapping direction, with units of microstrain (µs). ); (2) Blade azimuth timing signal The azimuth encoder at the impeller hub, in degrees (°), is used to determine the spatial position of the blades in the plane of rotation. (3) Cabin wind speed timing signal Data obtained from an ultrasonic anemometer on the top of the cabin, in meters per second (m / s). (4) Wind turbine rotation speed timing signal The speed is obtained through a speed encoder on the high-speed shaft of the gearbox or the end of the generator shaft, and the unit is revolutions per minute (rpm).

[0028] The above signals are all standard sensor signals from wind turbine units, and no additional hardware equipment is required.

[0029] S2. Variational mode decomposition and strain component separation: The strain signal at the blade root is a composite signal generated by the combined action of multiple physical excitation sources. Specifically: (1) Aerodynamic force-driven strain component, caused by aerodynamic loads acting on the blade surface, is related to parameters such as wind speed, blade pitch angle, and tip speed ratio, reflecting the aerodynamic excitation borne by the blade; (2) Inertial force-driven strain component, caused by centrifugal force and tangential inertial force generated by blade rotation, is related to wind turbine speed and blade azimuth angle, reflecting the dynamic load generated by blade rotation; (3) Gravity-driven strain component, caused by the blade's own gravity, is related to the blade azimuth angle, reflecting the change of gravity component of the blade in different spatial orientations.

[0030] The core of this step lies in separating the composite strain signal into the three strain components with distinct physical meanings, as described above. The process is as follows: Figure 2 As shown.

[0031] Specifically, we construct the objective function for variational mode decomposition. Variational mode decomposition is an adaptive, non-recursive signal decomposition method that can decompose a signal into several sparse mode components, each closely surrounding its center frequency. The objective function is: ; in, For the first Each modal component ( ), is the time-domain signal to be solved; For the first The center frequency of each modal component is the variable to be solved. For the Dirac function, in The value at one position is infinity, and the value at all other positions is 0. The imaginary unit satisfies ; This is the convolution operator; For time The first-order partial derivative operator; for Norm, also known as Euclidean norm.

[0032] The physical meaning of this objective function is: for each modal component First, construct its analytic signal using the Hilbert transform. Multiply by the compound exponent. Shift its spectrum to baseband, then calculate the time derivative of the shifted signal, and finally take... The square of the norm. Minimizing this objective function is equivalent to making each modal component as smooth as possible in the time domain and as compact as possible around its center frequency in the frequency domain. .

[0033] Set the number of modes , respectively corresponding to the aerodynamic strain components Inertial force-driven strain components and gravity-driven strain components .

[0034] The objective function is solved iteratively using the alternating direction multiplier method. This method decomposes the original optimization problem into several subproblems that are solved alternately, achieving high computational efficiency while ensuring convergence. The modal components are updated sequentially during the iteration process. Center frequency And Lagrange multipliers, until the convergence condition is met. (in For the number of iterations, As a preset convergence tolerance, this embodiment takes After iterative convergence, the output shows the three separated modal components. , , .

[0035] S3. Blade bending feature extraction: Based on the blade azimuth angle time-series signal and rotor speed time-series signal obtained in S1, the inertial force-driven strain component separated from S2 is extracted. and gravity-driven strain components Extract the leaf bending features.

[0036] (i) Extraction of dynamic bending features of blades, from strain components driven by inertial force Extracting dynamic bending features of the blade: First, based on the blade azimuth timing signal Wind turbine speed timing signal Calculate the tangential acceleration of the blade in the plane of rotation. and normal acceleration : ; ; in, R is the rotor radius (for the EN-200 / 6.25 model, R=100m). The rate of change of wind turbine rotation speed is calculated using the numerical difference method. Tangential acceleration. This reflects the acceleration or deceleration state of the blade in the direction of rotation; normal acceleration. It reflects the centripetal acceleration of the blades pointing towards the hub.

[0037] Based on the tangential acceleration and the normal acceleration, the strain components driven by the inertial force are... Decoupling of inertial forces is performed. This is because there is a linear relationship between inertial forces and acceleration. (in For leaf quality, (for acceleration), this step is achieved by... Divide by instantaneous acceleration amplitude Inertial force normalization was achieved, and dynamic strain components caused by blade bending deformation were extracted. : ; in, The preset reference acceleration is used to restore the normalized strain to a physically meaningful magnitude.

[0038] For the dynamic strain component Time-domain statistical feature extraction and frequency-domain feature extraction are performed to generate the dynamic bending features of the blade. Specifically, the time-domain statistical characteristics include the peak value, peak-to-peak value (the difference between the maximum and minimum values), root mean square value, and standard deviation of the dynamic strain signal; the frequency-domain characteristics are obtained by performing a fast Fourier transform on the dynamic strain signal to extract its amplitude spectral density and phase spectrum. This blade dynamic bending characteristic... It includes at least the dynamic strain amplitude, dynamic strain frequency, and dynamic strain phase.

[0039] (ii) Extraction of static bending features of blades, from gravity-driven strain components Extracting static bending features of the blade: First, based on the blade azimuth timing signal Determine the spatial azimuth angle of the blade during rotation. The spatial azimuth angle is defined as the angle between the blade axis and the vertically downward direction. When the blade is directly below... When it is directly above .

[0040] According to this spatial azimuth angle For gravity-driven strain components Gravity decoupling is performed. The bending moment generated by the blade under gravity is related to the sine of the blade azimuth angle. Proportional to the weight of the blade, therefore the static strain component caused by the blade's own weight is extracted. for: ; in, The preset reference azimuth angle is used to normalize the strain to a unified reference system.

[0041] For the static strain component Statistical analysis was performed to generate the static bending characteristics of the blade. The static bending characteristics of this blade At least include the mean static strain. Maximum static strain and static strain rate of change Among them, the mean static strain reflects the average degree of blade bending under gravity and is an indirect measure of blade stiffness; the maximum static strain reflects the maximum gravitational bending deformation of the blade in a specific orientation (usually horizontal); and the static strain change rate reflects the rate of change of gravitational bending deformation during blade rotation.

[0042] S4. Blade aerodynamic excitation feature extraction: Based on the nacelle wind speed time-series signal and blade azimuth angle time-series signal obtained in S1, the aerodynamic driven strain components separated in S2 are extracted. Time-frequency analysis was performed to extract the aerodynamic excitation characteristics of the blades.

[0043] First, based on the blade azimuth timing signal aerodynamic force drives strain component The transformation is performed from the time domain to the angular domain in the rotor rotation coordinate system. Since aerodynamic loads are related to the blade's azimuth in the plane of rotation (the wind speed and inflow angle vary depending on the blade's azimuth), the angular domain transformation eliminates the modulation effect of rotor rotation on the aerodynamic signal, allowing the aerodynamic excitation characteristics to exhibit stable periodicity in the angular domain. The specific method for angular domain transformation is as follows: using the blade azimuth angle... Using x as the x-axis, Resampling is performed at equal angular intervals to generate angle-domain aerodynamic drive strain signals. .

[0044] A continuous wavelet transform is performed on the aerodynamic strain signal u_1(\theta) in the angular domain to generate the time spectrum: ; in, These are wavelet coefficients, which are scaling parameters. Translation parameters The function reflects the time-frequency energy distribution of the signal at different scales and locations; The scaling parameter controls the scaling of the wavelet function and is inversely proportional to the frequency, i.e. The smaller the value, the higher the frequency of the component; The translation parameter controls the translation position of the wavelet function in the angular domain. For angle domain variables, the unit is degrees (°); This represents the aerodynamic strain signal in the angular domain. This is the conjugate function of the mother wavelet function. In this embodiment, the Morlet wavelet is selected as the mother wavelet function because it has good time-frequency localization characteristics and is suitable for extracting the frequency modulation features of aerodynamic signals.

[0045] According to the time spectrum Extracting the aerodynamic excitation features of the blade Specifically, the region of maximum energy is searched in the time-frequency spectrum, and the characteristic frequency corresponding to this region is extracted as the aerodynamic excitation frequency. Integrating the time spectrum along the frequency direction yields the aerodynamic excitation amplitude spectral density. Extract its maximum value and bandwidth. This blade aerodynamic excitation characteristic. At least including the aerodynamic excitation frequency Aerodynamic excitation amplitude spectral density and aerodynamic excitation bandwidth .

[0046] S5. Blade deflection prediction based on sparse temporal convolutional networks, using the dynamic bending features of the blade extracted in S3. 1. Static bending characteristics of blades aerodynamic excitation features of blades extracted from S4 The features are spliced ​​and merged to form a comprehensive feature vector. The input is fed into a pre-built blade deflection prediction model, which outputs the predicted dynamic deflection value of the blade at the blade tip position. .

[0047] The blade deflection prediction model is built based on a sparse temporal convolutional network, and its structure is as follows: Figure 3 As shown.

[0048] This sparse temporal convolutional network comprises multiple layers of sparse causal convolutional modules and a global regression output layer connected in sequence. This embodiment sets up four sparse causal convolutional modules, each containing a sparse convolutional kernel, a gated recurrent unit, and residual connections.

[0049] The core of each sparse causal convolution module is the sparse convolution kernel, which introduces a sparse mask matrix into the standard convolution kernel. Implementation. The standard convolutional kernel weight matrix is: ,in The kernel size is [size]. The number of input channels. Sparse mask matrix. The elements of are either 0 or 1, and are multiplied element-wise by W to obtain the sparse convolution kernel weights. ,in This is an element-wise product. The element values ​​of this sparse mask matrix are based on the dominant frequency of the blade aerodynamic excitation features extracted by S4. Adaptive determination.

[0050] Specifically, the adaptive determination method for this sparse mask matrix is ​​as follows: Based on the dominant frequency of the aerodynamic excitation characteristics of this blade Calculate the sparsity of the sparse convolution kernel. The calculation formula is: ; in, As the baseline sparsity, this embodiment takes This ensures a basic level of sparsity; As the adjustment coefficient, this embodiment takes... Control the adjustment range of the sparsity rate by the pneumatic excitation frequency; For reference frequency, this embodiment takes This corresponds to a typical value of the first-order flapping frequency of the blade; The absolute value of the deviation between the aerodynamic excitation frequency and the reference frequency is denoted as . When the aerodynamic excitation frequency is close to the reference frequency, the exponential term approaches 1 and the sparsity increases (becomes more sparse); when the deviation increases, the exponential term approaches 0 and the sparsity approaches the reference value.

[0051] Based on this sparsity The proportion of random selections in the weight matrix of the standard convolution kernel is... The weight elements are set to zero, and the remaining weight elements are set to zero to generate the sparse mask matrix. This adaptive sparsity mechanism enables the network to pay more attention to the dominant frequency component when the aerodynamic excitation frequency is close to the blade's natural frequency (at which point the risk of blade resonance is high and deflection prediction is more critical), thereby enhancing the network's predictive ability under critical operating conditions.

[0052] This gated recurrent unit is used to capture long-term dependencies in temporal data, alleviating the problem of limited receptive field in ordinary convolutional networks. Residual connections add the module input and output, mitigating the vanishing gradient problem during deep network training.

[0053] The global regression output layer of this sparse temporal convolutional network is a single-neuron fully connected layer, which outputs the predicted value of leaf tip dynamic deflection. The unit is meters (m).

[0054] During the training process of this blade deflection prediction model, a physical constraint loss function is used, which includes a data fitting term and a physical constraint term: ; in, This represents the number of training samples; For the first The blade deflection label value of each sample (the actual deflection value obtained by LiDAR or video measurement, used only in the model training stage); This is the predicted blade deflection value output by the blade deflection prediction model; For physical constraint weighting coefficients, this embodiment takes... To achieve a balance between data fitting and physical constraints; This is the proportionality coefficient between blade deflection and blade root bending moment, expressed in N / m, based on blade structural parameters (elastic modulus). Moment of inertia of cross section Blade length The calculation yielded the following result: ; For the first The measured values ​​of the blade root bending moment for each sample were obtained from the blade root strain signal in S1. Multiply by the section modulus at the blade root The calculation shows that: .

[0055] This physical constraint Based on the linear physical relationship between cantilever beam deflection and root bending moment in mechanics of materials (in This is used to construct a model for blade tip deflection. This physical relationship is a fundamental law of blade structural mechanics, and the model predicts the deflection value. Multiply by the proportionality factor The following should be compared with the measured bending moment. Consistency. By minimizing this physical constraint term, the model's predicted output is forced to conform to the basic laws of blade structural mechanics, significantly improving the model's generalization ability and prediction reliability beyond the training data coverage.

[0056] S6. Clearance distance prediction calculation: Based on the wind turbine speed timing signal and blade azimuth timing signal obtained from S1, determine the current spatial azimuth angle of the blade, and calculate the predicted clearance distance between the blade tip and the tower based on the spatial azimuth angle and the dynamic deflection prediction value output from S5.

[0057] First, based on the wind turbine speed timing signal Calculate the cumulative rotation angle of the wind turbine within a preset time window. : ; in, For a moment The rotor speed, expressed in rad / s. Based on the cumulative rotation angle and blade azimuth angle timing signal... Determine the current spatial azimuth of the blade. .

[0058] According to this spatial azimuth angle The predicted value of dynamic deflection In addition to the initial geometric parameters of the blade, the predicted clearance distance between the blade tip and the tower is calculated. The calculation formula is: ; in, This is the initial distance between the blade tip and the tower when the blade has not undergone deflection (i.e., static clearance), in meters. This value is determined by the geometric layout parameters of the wind turbine. For the EN-200 / 6.25 model, the rotor radius is... ; This is the spatial azimuth angle, in degrees (°), when the blade tip points directly downwards. When pointing directly upwards ; This is the predicted value of the dynamic deflection, in meters (m). A positive value is taken when the blade deflects towards the tower.

[0059] The physical meaning of this calculation formula is: the clearance between the blade tip and the tower is equal to the static clearance. Subtract the projection of the blade rotation in the horizontal direction Then subtract the projection of the blade dynamic deflection in the horizontal direction. When the blade rotates to directly downwards ( When the clearance distance reaches its minimum value, At this time, the airspace safety risk is the highest.

[0060] The first term in the above formula This represents the theoretical clearance distance for rigid blades (ignoring deflection); the second term... This represents the reduction in headroom due to blade deflection. It is calculated by predicting the dynamic deflection of the blade. By incorporating clearance calculations, indirect measurement of clearance distance based on blade structural response was achieved.

[0061] When the calculated predicted net distance value When the clearance is less than the preset safety threshold (set to 5m in this embodiment according to the unit safety specifications), the system automatically generates an airspace warning signal, prompting the main control system to perform pitch control or power reduction operation to increase the airspace clearance and ensure the safe operation of the unit.

[0062] In one application example of this invention, historical operating data from an Envision Energy EN-200 / 6.25 wind turbine at a wind farm were used to verify the method. The verification results show that the average absolute error between the clearance distance predicted by the present invention based on blade root strain signals and the clearance distance measured by lidar is 0.12m, with a maximum error not exceeding 0.28m. The prediction accuracy meets the requirements of engineering applications. Furthermore, under rainy and foggy weather conditions, lidar measurements fail, but the method of this invention remains unaffected and continues to operate normally, verifying the reliability and adaptability of the present invention under harsh environmental conditions.

[0063] Working Principle: This invention utilizes the strain signal at the root of wind turbine blades as an indirect indicator of blade deflection. First, variational mode decomposition (VMD) separates the composite strain signal into three strain components with different physical origins: aerodynamic-driven, inertial-driven, and gravity-driven components. Then, dynamic bending features, static bending features, and aerodynamic excitation features are extracted from each component to characterize the blade's deflection deformation from multiple physical dimensions. Finally, these multi-dimensional features are fused and input into a blade deflection prediction model based on a sparse temporal convolutional network. Under the supervision of a physical constraint loss function, a high-precision prediction of the blade tip dynamic deflection is output, and the clearance distance is calculated by combining this with the blade's spatial azimuth angle. This method requires no additional hardware such as lidar or industrial cameras, utilizing only existing sensor resources of the wind turbine to achieve high-precision all-weather clearance distance prediction.

[0064] Therefore, the wind turbine headroom prediction method based on the evolution of blade aerodynamic characteristics adopted in this invention has the advantages of high accuracy, all-weather operation, low cost and high reliability, and effectively solves many shortcomings in the prior art.

[0065] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for predicting the headroom of a wind turbine based on the evolution of blade aerodynamic characteristics, characterized in that, include: S1. Acquire multi-source monitoring data of wind turbine blades under operating conditions. The multi-source monitoring data includes blade root strain time-series signal, blade azimuth angle time-series signal, nacelle wind speed time-series signal, and wind turbine speed time-series signal. S2. Perform variational mode decomposition on the strain time series signal at the root of the blade to separate it into aerodynamic strain components, inertial strain components, and gravity strain components. S3. Based on the blade azimuth time-series signal and the wind turbine rotation speed time-series signal, extract the blade dynamic bending characteristics and the blade static bending characteristics from the inertial force driven strain component and the gravity driven strain component, respectively. S4. Based on the nacelle wind speed time-series signal and the blade azimuth angle time-series signal, perform time-frequency analysis on the aerodynamic drive strain component and extract the blade aerodynamic excitation characteristics. S5. Input the dynamic bending characteristics, static bending characteristics and aerodynamic excitation characteristics of the blade into the pre-built blade deflection prediction model. The blade deflection prediction model is built based on a sparse temporal convolutional network and outputs the predicted dynamic deflection value of the blade at the tip position. S6. Determine the current spatial azimuth of the blade based on the wind turbine rotation speed timing signal and the blade azimuth timing signal, and calculate the predicted clearance distance between the blade tip and the tower based on the spatial azimuth and the predicted dynamic deflection value.

2. The wind turbine headroom prediction method based on blade aerodynamic characteristic evolution according to claim 1, characterized in that, In S2, variational mode decomposition is performed on the strain time series signal at the blade root, separating it into aerodynamically driven strain components, inertial force-driven strain components, and gravity-driven strain components, specifically including: Construct the objective function for variational mode decomposition, which is: ； in, For the first One modal component, For the first The center frequency of each modal component For the Dirac function, For convolution operators, The imaginary unit, For time Partial derivative operators, for Norm; Set the number of modes , respectively corresponding to the aerodynamic strain components Inertial force-driven strain components and gravity-driven strain components ; The objective function is solved iteratively using the alternating direction multiplier method until the convergence condition is met, and the three separated modal components are output.

3. The wind turbine headroom prediction method based on blade aerodynamic characteristic evolution according to claim 1, characterized in that, In S3, based on the blade azimuth angle time-series signal and the rotor speed time-series signal, the dynamic bending characteristics of the blade are extracted from the inertial force-driven strain component, specifically including: Based on the timing signals of the blade azimuth angle and the wind turbine rotation speed, calculate the tangential acceleration and normal acceleration of the blade in the plane of rotation. Based on the tangential acceleration and the normal acceleration, the inertial force driven strain component is decoupled from the inertial force to extract the dynamic strain component caused by the blade bending deformation. The dynamic strain component is subjected to time-domain statistical feature extraction and frequency-domain feature extraction to generate the dynamic bending feature of the blade. The dynamic bending feature of the blade includes at least the dynamic strain amplitude, dynamic strain frequency and dynamic strain phase.

4. The wind turbine headroom prediction method based on blade aerodynamic characteristic evolution according to claim 1, characterized in that, S3 extracts the static bending characteristics of the blade from the gravity-driven strain component based on the blade azimuth angle time-series signal and the rotor speed time-series signal, specifically including: Based on the timing signal of the blade azimuth angle, the spatial azimuth angle of the blade during the rotation process is determined. The spatial azimuth angle is the angle between the blade axis and the vertical direction. Based on the spatial azimuth, gravity decoupling is performed on the gravity-driven strain component to extract the static strain component caused by the blade's own weight. Statistical analysis of the static strain components is performed to generate the static bending characteristics of the blade. The static bending characteristics of the blade include at least the mean static strain, the maximum static strain, and the static strain change rate.

5. The wind turbine headroom prediction method based on blade aerodynamic characteristic evolution according to claim 1, characterized in that, S4 performs time-frequency analysis on the aerodynamic strain component based on the nacelle wind speed time-series signal and the blade azimuth angle time-series signal to extract the blade aerodynamic excitation characteristics, specifically including: Based on the blade azimuth time sequence signal, the aerodynamic drive strain component is transformed from the time domain to the angle domain under the impeller rotation coordinate system to generate the angle domain aerodynamic drive strain signal. A continuous wavelet transform is performed on the aerodynamic strain signal in the angular domain to generate the time spectrum: ； in, These are wavelet coefficients. For scale parameters, For translation parameters, For angle domain variables, This represents the aerodynamic strain signal in the angular domain. It is the conjugate function of the mother wavelet function; Based on the time spectrum, the aerodynamic excitation characteristics of the blade are extracted. The aerodynamic excitation characteristics of the blade include at least the aerodynamic excitation main frequency, the aerodynamic excitation amplitude spectral density, and the aerodynamic excitation bandwidth.

6. The wind turbine headroom prediction method based on blade aerodynamic characteristic evolution according to claim 1, characterized in that, The blade deflection prediction model in S5 is built on a sparse temporal convolutional network, which contains a series of sparse causal convolutional modules and a global regression output layer connected in sequence. Each sparse causal convolution module includes a sparse convolution kernel, a gated recurrent unit, and residual connections; The sparse convolution kernel is implemented by introducing a sparse mask matrix into the standard convolution kernel. The element values ​​of the sparse mask matrix are adaptively determined according to the main frequency of the blade's aerodynamic excitation characteristics, which is used to enhance the sparse temporal convolution network's ability to perceive the frequency components of aerodynamic excitation. During the training process of this blade deflection prediction model, a physical constraint loss function is used, which includes a data fitting term and a physical constraint term. The physical constraint term is constructed based on the linear physical relationship between blade deflection and blade root bending moment.

7. The wind turbine headroom prediction method based on blade aerodynamic characteristic evolution according to claim 6, characterized in that, The adaptive method for determining the sparse mask matrix is ​​as follows: Based on the dominant frequency of the aerodynamic excitation characteristics of this blade Calculate the sparsity of the sparse convolution kernel. The calculation formula is: ； in, As the baseline sparsity, For adjustment coefficients, For reference frequency; Based on this sparsity The proportion of random selections in the weight matrix of the standard convolution kernel is... The weight elements are determined, and the remaining weight elements are set to zero to generate the sparse mask matrix.

8. The wind turbine headroom prediction method based on blade aerodynamic characteristic evolution according to claim 6, characterized in that, The physical constraint loss function is: ； in, For the sample size, For the first The blade deflection label value for each sample. This is the predicted blade deflection value output by the blade deflection prediction model. These are the physical constraint weighting coefficients. This is the proportionality coefficient between blade deflection and blade root bending moment. For the first Measured values ​​of the bending moment at the root of the leaf for each sample; This physical constraint Based on the linear physical relationship between blade deflection and blade root bending moment, the output of the blade deflection prediction model is constrained to conform to the laws of blade structural mechanics.

9. The wind turbine headroom prediction method based on blade aerodynamic characteristic evolution according to claim 1, characterized in that, In S6, the current spatial azimuth of the blade is determined based on the rotor speed timing signal and the blade azimuth timing signal. Then, based on the spatial azimuth and the predicted dynamic deflection value, the predicted clearance distance between the blade tip and the tower is calculated, specifically including: Based on the wind turbine rotation speed timing signal, calculate the cumulative rotation angle of the wind turbine within a preset time window; Based on the cumulative rotation angle and the blade azimuth time-series signal, determine the current spatial azimuth of the blade. ; According to the spatial azimuth angle The predicted value of dynamic deflection In addition to the initial geometric parameters of the blade, the predicted clearance distance between the blade tip and the tower is calculated. The calculation formula is: ； in, This is the initial distance between the blade tip and the tower when the blade has not undergone deflection deformation. Where is the radius of the wind turbine. This is the azimuth angle in space. This is the predicted value for the dynamic deflection.

10. The wind turbine headroom prediction method based on blade aerodynamic characteristic evolution according to claim 1, characterized in that, Also includes: Acquire the historical operating data of the wind turbine, which includes historical signals of blade root strain, blade azimuth angle, nacelle wind speed, rotor speed, and corresponding blade tip clearance distance labels. A training sample set was constructed using the historical operating data, and the blade tip clearance distance label was used as a supervision signal to train the blade deflection prediction model. During training, the Bayesian hyperparameter optimization method is used to optimize the network hyperparameters of the blade deflection prediction model. The network hyperparameters include at least the sparse convolution kernel size, the number of network layers, and the sparsity ratio. After training, the blade deflection prediction model is deployed to the main control system or edge computing device of the wind turbine for online air clearance prediction.

Images