A fan wake flow inversion dynamic balance and life prediction method based on multi-source PINN

Abstract

The application discloses a fan tail flow inversion dynamic balance and life prediction method based on a multi-source PINN, relates to the technical field of wind turbine operation state monitoring and intelligent operation and maintenance, and comprises the following steps: S100, multi-source data acquisition, S200, construction of a physical information neural network, S300, construction of an aerodynamic physical information constraint loss function, S400, construction of an observation sequence physical information constraint loss function, S500, construction of a tail flow evolution physical information constraint loss function, S600, construction of a total loss function and completion of PINN training, and S700, residual life prediction. The application provides a quantitative basis for fan state monitoring, fault early warning, maintenance plan formulation and life extension decision-making, and improves the intelligent operation and maintenance level of the wind turbine and the operation economy in the whole life cycle.

Description

Technical Field

[0001] This invention relates to the field of wind turbine generator operation status monitoring and intelligent operation and maintenance technology, and in particular to a wind turbine wake inversion dynamic balance and life prediction method based on multi-source PINN. Background Technology

[0002] Wind turbines operate in open atmospheric environments for extended periods. Their rotors are influenced by a multitude of factors, including incoming wind speed and direction, turbulence intensity, yaw rate, and flow fields induced by complex terrain. This results in significant time-varying and spatially non-uniform aerodynamic force distribution before and after the blades. When blades experience problems such as icing, fouling, leading-edge erosion, localized damage, uneven mass distribution, or aerodynamic degradation, deviations in aerodynamic torque between different blades occur. This leads to increased alternating loads on the main shaft, gearbox, nacelle support, and tower structure, causing a decline in the unit's dynamic balance performance. In severe cases, it can accelerate fatigue damage to critical components and shorten the overall service life of the turbine. Therefore, accurately identifying the dynamic balance state of wind turbine blades and the rotation system, and further conducting life assessments, has always been a crucial technical challenge in the field of wind turbine condition monitoring and intelligent operation and maintenance.

[0003] In existing technologies, the identification of wind turbine dynamic balance status typically relies on vibration sensors, strain gauges, accelerometers, nacelle monitoring devices, or blade load testing systems. These systems collect vibration, strain, or acceleration signals from components such as the main shaft, gearbox, and tower to analyze whether unbalanced excitation exists within the unit. While these methods can reflect the equipment's operating status to some extent, they are usually localized measurements, only obtaining structural response results and failing to directly characterize the actual aerodynamic changes before and after the blades. Furthermore, sensor installation locations are limited, deployment costs are high, and maintenance is difficult. Moreover, under complex wind conditions, they are easily affected by environmental noise, structural transmission paths, and operating condition fluctuations, resulting in insufficient robustness and generalization ability of the diagnostic results.

[0004] On the other hand, existing technologies employ CFD (Computational Fluid Dynamics) numerical simulations, wake models, or pure data-driven algorithms to analyze wind turbine wakes in an attempt to indirectly identify blade stress states or turbine operational anomalies. However, traditional CFD methods are computationally expensive and have strict requirements for boundary conditions and mesh accuracy, making them difficult to meet the needs of long-term online monitoring of wind farms. Empirical wake models typically have limited adaptability to complex terrain, multi-wind directions, and unsteady operating conditions, making it difficult to accurately describe the wake deflection, diffusion, and recovery processes in real wind farm environments. While pure data-driven models have some predictive ability, they often lack clear physical constraints and are prone to physical distortion when measurement points are sparse, samples are insufficient, or operating conditions shift, making it difficult to establish a stable mapping relationship between wake characteristics and blade aerodynamics, dynamic balance of the rotating system, and lifetime evolution. Especially in complex sites such as mountainous wind farms, hilly wind farms, and coastal wind farms, terrain slope, surface roughness, wind shear, and local thermal effects can significantly alter the wind turbine's incoming flow and wake structure. If the above environmental factors are ignored, the wake anomaly caused by external wind field disturbances may be misjudged as blade failure or dynamic balance deviation, thereby reducing the accuracy of diagnosis.

[0005] Therefore, those skilled in the art are dedicated to developing a method for wind turbine wake inversion dynamic balance and lifetime prediction based on multi-source PINN. Summary of the Invention

[0006] In view of the above-mentioned deficiencies of the prior art, the technical problem to be solved by the present invention is how to comprehensively utilize external wake observation information, basic laws of fluid mechanics and environmental evolution factors to conduct a unified analysis of wind turbine blade dynamic balance, rotation system excitation and life of key components.

[0007] Existing technologies lack a unified method for analyzing the dynamic balance of wind turbine blades, the excitation of the rotating system, and the lifespan of key components in complex wind farm environments by comprehensively utilizing external wake observation information, fundamental laws of fluid mechanics, and environmental evolution factors. During the operation of wind turbine generators in large wind farm environments, the identification of dynamic balance states is difficult due to non-uniform incoming flow, unsteady wake evolution, inconsistent blade aerodynamic forces, and flow deviations induced by complex terrain. Furthermore, it is difficult to comprehensively reflect the stress state of the entire turbine. This application analyzes the airflow state before and after the wind turbine impeller as the starting point, using the wake structure as an externally observable carrier. By integrating physical mechanism constraints, radar measurement data constraints, and the coupling conditions of complex terrain and wind direction, a PINN (Physics-Informed Neural Networks) is constructed to establish a unified mapping relationship from wake characteristics to blade aerodynamic torque, rotating system excitation, and structural fatigue life, thereby enabling the prediction of the wind turbine generator's operating life.

[0008] In one embodiment of the present invention, a method for wind turbine wake inversion dynamic balancing and lifetime prediction based on multi-source PINN is provided, comprising the following steps: S100. Multi-source data acquisition: Establish the local coordinate system of the wind turbine, acquire multi-source data, and perform preprocessing. S200. Construct a physical information neural network, construct PINN, define the input and output of PINN, and construct a unified velocity field; S300, Construct an aerodynamic physical information constraint loss function. Based on the principle of mass and momentum conservation of incompressible fluids, construct the aerodynamic physical information constraint loss function of PINN. S400. Construct the physical information constraint loss function for the observation sequence. Based on the time derivative and wake wind speed prediction value of the observation sequence of the pulse Doppler wind radar, define the physical information constraint loss function for the radar observation sequence. S500. Construct a physical information constraint loss function for wake evolution, considering the influence of complex terrain and wind direction, define the effective wake propagation direction angle, calculate the wake propagation speed, define the wake attenuation coefficient, and construct a physical information constraint loss function for wake wind speed. S600. Construct the total loss function and complete PINN training. Define the multi-constraint weighted total loss function, construct the training sample set, iteratively optimize the PINN network parameters, and complete PINN training. S700, Remaining Life Prediction: Establish blade dynamic balance deviation index, calculate equivalent fatigue load and equivalent fatigue load, and estimate remaining life.

[0009] Optionally, in the wind turbine wake inversion dynamic balancing and lifetime prediction method based on multi-source PINN in the above embodiments, step S100 includes: S110. Establish a local coordinate system for the wind turbine, with the impeller center as the origin. Establish the local coordinate system of the wind turbine Geographic coordinate system As a unified reference for the entire site, the local coordinate system of the wind turbine is linked to the geographic coordinate system through a rotation transformation, as shown in the following formula: , , in, The coordinates of the wind turbine hub center in the geographic coordinate system; This refers to the azimuth angle of the cabin. The coordinates represent the main incoming flow direction of the wind turbine; The horizontal offset coordinate; The vertical coordinates are relative to the center of the wheel hub; S120, Multi-source data acquisition, acquires multi-source data for wind turbine wake inversion, dynamic balance identification and life prediction; S130. Data preprocessing: preprocessing multi-source data, including: timestamp unification, coordinate mapping, outlier removal, noise filtering, missing value imputation, invalid observation removal, and resampling.

[0010] Optionally, in the wind turbine wake inversion dynamic balance and lifetime prediction method based on multi-source PINN in any of the above embodiments, the multi-source data includes pulse Doppler wind measurement radar data, wind turbine SCADA operation data, yaw angle data, pitch angle data, impeller speed data, and environmental wind direction and speed data.

[0011] Furthermore, in the wind turbine wake inversion dynamic balance and lifetime prediction method based on multi-source PINN in the above embodiments, pulse Doppler wind radar data is used to obtain wind speed observation information at different distances and heights in the inflow area in front of the wind turbine impeller and the wake area behind the impeller.

[0012] Furthermore, in the wind turbine wake inversion dynamic balancing and lifetime prediction method based on multi-source PINN in the above embodiments, the wind turbine SCADA operation data includes active power, unit operating status, unit start-up and shutdown status, and fault codes.

[0013] Furthermore, in the wind turbine wake inversion dynamic balance and lifetime prediction method based on multi-source PINN in the above embodiments, the yaw angle data characterizes the nacelle orientation.

[0014] Furthermore, in the wind turbine wake inversion dynamic balance and lifetime prediction method based on multi-source PINN in the above embodiments, the pitch angle data characterizes the blade angle of attack adjustment state.

[0015] Furthermore, in the wind turbine wake inversion dynamic balancing and life prediction method based on multi-source PINN in the above embodiments, the impeller speed data characterizes the impeller rotation condition.

[0016] Furthermore, in the wind turbine wake inversion dynamic balance and lifetime prediction method based on multi-source PINN in the above embodiments, the environmental wind direction and speed data are used to characterize the boundary inflow conditions of the local area of ​​the wind farm where the wind turbine is located.

[0017] Optionally, in the wind turbine wake inversion dynamic balance and lifetime prediction method based on multi-source PINN in any of the above embodiments, in the case of a wind farm with complex terrain, the multi-source data also includes a digital elevation model (DEM), surface roughness parameters, landform aspect information and underlying surface thermal characteristic parameters, which are used to construct wake evolution constraints under the coupling conditions of terrain and wind direction.

[0018] Furthermore, in the wind turbine wake inversion dynamic balancing and lifetime prediction method based on multi-source PINN in the above embodiments, timestamp unification includes aligning multi-source data to a unified time series. ,in It represents the time variable in a unified time series, characterizing the sampling time of multi-source data under the same time reference; This indicates the total duration of the preset analysis time window, which is the upper limit of the time range corresponding to the unified time series.

[0019] Furthermore, in the wind turbine wake inversion dynamic balancing and lifetime prediction method based on multi-source PINN in the above embodiments, the total duration of the preset analysis time window is... The value range is [10, 3600] seconds.

[0020] Preferably, in the wind turbine wake inversion dynamic balancing and lifetime prediction method based on multi-source PINN in the above embodiments, the total duration of the preset analysis time window is... The value is 2000 seconds.

[0021] Furthermore, in the wind turbine wake inversion dynamic balance and lifetime prediction method based on multi-source PINN in the above embodiments, coordinate mapping includes mapping multi-source data to the local coordinate system of the wind turbine.

[0022] Furthermore, in the wind turbine wake inversion dynamic balance and lifetime prediction method based on multi-source PINN in the above embodiments, missing value completion includes using threshold discrimination, interpolation completion or sliding window smoothing to correct for observation samples with abnormal jumps, missing records or low signal-to-noise ratio.

[0023] Furthermore, in the wind turbine wake inversion dynamic balancing and lifetime prediction method based on multi-source PINN in the above embodiments, the resampling process includes using a uniform step size for data sources with different sampling frequencies. Resampling is performed to form a time-synchronized multi-source dataset.

[0024] Optionally, in the wind turbine wake inversion dynamic balancing and lifetime prediction method based on multi-source PINN in any of the above embodiments, step S200 includes: S210. Construct PINN: Construct a PINN (Physical Information Neural Network) for wind turbine wake inversion. PINN adopts a deep artificial neural network structure, including an input layer, several hidden layers, and an output layer. Define the basic input variables of PINN. , The formula is as follows: , in, Indicates the inflow cross section in front of the impeller. Spatial coordinates of the upsampling point; Indicates the wake section behind the impeller Spatial coordinates of the upsampling point; It represents a time variable, characterizing the time dependence of unsteady wind conditions, yaw adjustment, pitch control, and wake evolution during wind turbine operation; S220, Forming the extended input vector The formula is as follows: , in, The dominant wind direction angle time series; Integrated terrain parameter vector; This refers to the cabin yaw angle. For variable pitch angle; The impeller speed, T Indicates transpose; S230, Calculate the input of PINN The formula is as follows: , in, Indicates the input to PINN; S240. Define the output of PINN using the following formula: , in, This is the output of PINN (Physical Information Neural Network). The predicted aerodynamic pressure difference between the blade and the front and back is output by PINN at a given spatial location and time, forming the aerodynamic pressure difference field between the blade and the front and back in the entire spatiotemporal domain; This is the predicted wind speed of the incoming flow in front of the impeller. This is the predicted wind speed of the wake behind the impeller. S250. Construct a unified velocity field and an interpolation weighting function between the front and rear cross sections. Assume the incoming flow cross section is located at... The wake section is located , The coordinates for the main incoming flow direction of the wind turbine are given by the following formula: , , in, Let be the coordinates of the inflow section in the direction of the main inflow. The coordinates of the rear wake section in the direction of the main incoming flow are given. Let represent the contribution weights of the incoming flow state and the backward wake state to the current spatial coordinate position, respectively, satisfying: , when Sometimes, , And when Sometimes, , ; The unified velocity field is represented as follows: , in, To ensure a uniform velocity field in the areas before and after the wind turbine. The coordinates represent the main incoming airflow direction of the wind turbine. The horizontal offset coordinates are... The vertical coordinates are relative to the center of the wheel hub. It is a time variable; S260 Automatic Differentiation: Based on PINN, an automatic differentiation function is established to calculate the number of partial derivatives in the physical information constraint loss function with respect to spatial and temporal variables.

[0025] Optionally, in the wind turbine wake inversion dynamic balance and lifetime prediction method based on multi-source PINN in any of the above embodiments, the deep artificial neural network includes a multi-layer feedforward fully connected neural network or a residual neural network.

[0026] Optionally, in the wind turbine wake inversion dynamic balance and lifetime prediction method based on multi-source PINN in any of the above embodiments, the number of hidden layers of PINN and the number of neurons in each hidden layer are adjusted according to the complexity of the wind turbine wake field, the number of observation samples, and the training convergence.

[0027] Furthermore, in the wind turbine wake inversion dynamic balance and lifetime prediction method based on multi-source PINN in the above embodiments, the number of hidden layers of PINN is 3 to 12.

[0028] Preferably, in the wind turbine wake inversion dynamic balance and lifetime prediction method based on multi-source PINN in the above embodiments, the number of hidden layers of PINN is 8.

[0029] Furthermore, in the wind turbine wake inversion dynamic balance and lifetime prediction method based on multi-source PINN in the above embodiments, the number of neurons in each hidden layer is 16 to 256.

[0030] Furthermore, in the wind turbine wake inversion dynamic balance and lifetime prediction method based on multi-source PINN in the above embodiments, the number of neurons in each hidden layer is 64.

[0031] Optionally, in the wind turbine wake inversion dynamic balance and lifetime prediction method based on multi-source PINN in any of the above embodiments, the mapping relationship between network layers is expressed as follows: , in, For the first The state vector of the hidden layer. For the first Layer weight matrix; For the first Layer bias vector; For activation functions; This refers to the network layer number.

[0032] Optionally, in the wind turbine wake inversion dynamic balancing and lifetime prediction method based on multi-source PINN in the above embodiments, the activation function includes the Tanh function, the Sine function or other differentiable activation functions to ensure that the network output can be automatically differentiated with respect to time and space variables.

[0033] Optionally, in the wind turbine wake inversion dynamic balance and lifetime prediction method based on multi-source PINN in the above embodiments, the number of partial derivatives includes the wake wind speed predicted by PINN. PINN output predicted aerodynamic pressure difference before and after the blade And the velocity components of the unified velocity field in the direction of the main flow and the transverse direction. .

[0034] Optionally, in the wind turbine wake inversion dynamic balancing and lifetime prediction method based on multi-source PINN in any of the above embodiments, step S300 includes: S310. Define the momentum equation for the main flow direction. Based on the momentum conservation equation for incompressible fluids, the momentum equation for the main flow direction is obtained as follows: , in, air density, It represents the change of wake velocity over time, reflecting the unsteady effects caused by gusts, yaw adjustments, and pitch control; This represents the self-convection term in the direction of the main inflow, reflecting the momentum transport of the wake along the downstream direction. To unify the velocity components of the velocity field in the direction of the main incoming flow; This indicates the lateral transport of momentum in the direction of the main flow caused by the lateral velocity. To unify the velocity components of the velocity field in the lateral direction; This indicates the vertical transport of momentum in the direction of the main flow caused by the vertical velocity. To unify the velocity components of the velocity field in the vertical direction; This represents the pressure gradient driving force along the mainstream direction; S320. Modify the momentum equation in the main incoming flow direction, and use the predicted aerodynamic pressure difference before and after the blades output by PINN. Constructing the pressure driving term in the direction of the main inflow, the momentum equation in the direction of the main inflow is modified as follows: ; S330, Define the aerodynamic physical information constraint loss function The formula is as follows: , in, It is the square of the mass conservation error; It is the square of the momentum conservation error in the direction of the main flow, subscript Indicates the physical constraint sampling point number. The number of sampling points constrained by aerodynamics.

[0035] Optionally, in the wind turbine wake inversion dynamic balance and lifetime prediction method based on multi-source PINN in any of the above embodiments, the number of aerodynamically constrained sampling points is... The range is .

[0036] Preferably, in the wind turbine wake inversion dynamic balancing and lifetime prediction method based on multi-source PINN in the above embodiments, the number of aerodynamically constrained sampling points is... .

[0037] Optionally, in the wind turbine wake inversion dynamic balancing and lifetime prediction method based on multi-source PINN in any of the above embodiments, step S400 includes: S410. Define the time derivative of the radar observation sequence for the first... Preprocessing sequence of radar observation points Its time derivative is defined as: , in, For the first The rate of change of wake wind speed at each radar observation point over time, i.e., the time derivative; S420, Calculate the predicted wake velocity, calculate the first... Radar observation points place, time Wake wind speed prediction The formula is as follows: , in, The wake wind speed field output by PINN; For the first The location of each radar observation point in the local coordinate system of the wind turbine. For the first The position coordinates of each radar observation point along the main flow direction in the local coordinate system of the wind turbine; For the first Lateral offset coordinates of each radar observation point in the local coordinate system of the wind turbine; For the first The vertical coordinates of each radar observation point relative to the hub center in the local coordinate system of the wind turbine; It is a time variable; For optional environmental and control inputs, As the prevailing wind angle, ; To synthesize the topographic parameter vector, which characterizes environmental factors, including topographic slope, aspect, roughness, and underlying surface thermal properties; The cabin yaw angle, ; Pitch angle, ; Impeller angular velocity, ; S430, Define the radar differential consistency residual, PINN for the... The predicted time series for each radar observation point is represented as follows: , Calculate at time The radar differential consistency residual is: , in, For the first Each radar observation point at time The observed trend residuals; This represents the time rate of change of the PINN prediction of the wake wind speed; This represents the rate of change of the radar-measured wake wind speed over time. S440. Define the physical information constraint loss function for the radar observation sequence, as follows: , in, This is an effective observation mask used to exclude invalid or missing observations; the values ​​in square brackets are the first and second observations. The radar observation point at the ... The differential residual at each moment, The total number of effective radar observation points is [number missing]. The number of time sampling moments.

[0038] Optionally, in the wind turbine wake inversion dynamic balancing and lifetime prediction method based on multi-source PINN in any of the above embodiments, step S500 includes: S510. Define the effective wake propagation direction angle, as follows: , in, The effective wake propagation direction angle; The dominant wind direction angle time series represents the main direction of the incoming flow in the horizontal plane at the location of the wind turbine; This is the wake direction correction amount caused by the combined effects of terrain and yaw; Yaw error, representing the difference between the prevailing wind angle and the cabin yaw angle. The comprehensive topographic parameter vector characterizes the environmental factors at the wind turbine site, including topographic slope, aspect, surface roughness, and underlying thermal properties. S520. Calculate the wake propagation velocity. The wake propagation velocity characterizes how quickly the wake deficit structure migrates downstream. The calculation formula is as follows: , in, As the baseline propagation term; The direct impact of incoming wind speed on wake propulsion speed. This is the correction factor for the incoming airflow velocity. For PINN, the predicted inflow velocity in front of the impeller; This represents the effect of the relative slope angle between wind direction and propagation direction on the uphill and downhill slope effects. The angle of the main slope of the terrain. The angle between the wind direction and the slope aspect. This is the wind direction-slope aspect coupling correction factor; The effect of terrain slope intensity on propagation speed. This is the slope correction factor. For terrain slope intensity; The effect of surface roughness on the propulsion velocity of near-surface wake. This is the roughness correction factor. Length represents the surface roughness. This is a correction for wake centerline drift caused by yaw error. This is the yaw correction factor. This refers to yaw error; These are model parameters, determined through training; S530. Define the wake attenuation coefficient, as follows: , in, Basic recovery item; Roughness refers to the amount of material in the wake. The greater the roughness, the stronger the entrainment and mixing in the wake, and the faster the recovery. This is the roughness-induced hybrid reinforcement coefficient. , where is the surface roughness length, and together they characterize the enhancing effect of surface roughness on wake entrainment, turbulent mixing, and velocity recovery processes; The slope will enhance local shear and flow direction adjustment, thus affecting the recovery rate. This is the slope-induced recovery adjustment coefficient. Both topographic slope intensity and topographic slope intensity together characterize the moderating effect of topographic relief on local flow shear, wake diffusion and recovery intensity. When the wind direction and slope aspect are inconsistent, lateral disturbances are more pronounced. The wind direction-slope aspect mismatch recovery influence coefficient. As the prevailing wind angle, The angle of the main slope of the terrain. It represents the magnitude of the lateral deviation of the wind direction relative to the slope aspect, and is used to characterize the degree of enhancement of the lateral disturbance and asymmetric recovery of the wake when the wind direction and slope aspect are inconsistent. The greater the intensity of wind direction fluctuations, the more pronounced the wake instability recovery. The intensity of wind direction change, The wind direction fluctuation recovery response coefficient; For offline calibration parameters, historical radar wake data, SCADA data, and known operating condition data are used to perform offline calibration through least squares fitting, Bayesian estimation, or grid search. ; S540. Construct a first-order evolution equation for wake wind speed. To describe the comprehensive evolution of wake wind speed over time, in the main propagation direction, and in the lateral offset direction, a first-order evolution equation for wake wind speed is constructed, as follows: , in, Predict wake wind speed for PINN This indicates that PINN predicts the incoming wind speed. The principal propagation coordinates are established along the effective wake propagation direction. To and Orthogonal lateral offset coordinates The rate of change of wake wind speed with respect to time represents the unsteady response caused by gusts, control actions, and environmental fluctuations. The term represents the wake translation along the main propagation direction. Let be the wake propagation velocity function, representing the effective propagation velocity of the wake-depleted structure migrating downstream under the combined influence of prevailing wind direction, topographic parameters, and time conditions. For the wake wind speed predicted by PINN, This represents the spatial gradient of the wake wind speed along the effective propagation direction; it also represents the wake structure as it advances downstream. The lateral deflection velocity function, The lateral deflection term represents the lateral drift of the wake under the influence of complex terrain or changes in wind direction. This is the wake recovery term, representing how quickly the wake velocity recovers to the incoming flow velocity. This is the wake attenuation coefficient, used to describe how quickly the wake wind speed recovers from a deficit state to an incoming flow state. This represents the deviation of the wake velocity from the predicted incoming velocity. Predict the incoming wind speed for PINN; S550, Constructing the physical information constraint loss function for wake wind speed The formula is as follows: , in, This represents the number of physical sampling points for wake evolution.

[0039] Furthermore, in the wind turbine wake inversion dynamic balancing and lifetime prediction method based on multi-source PINN in the above embodiments, the number of wake evolution physical sampling points... The value range is [5000, 50000].

[0040] Preferably, in the wind turbine wake inversion dynamic balance and lifetime prediction method based on multi-source PINN in the above embodiments, the number of wake evolution physical sampling points is... .

[0041] Optionally, in the wind turbine wake inversion dynamic balancing and lifetime prediction method based on multi-source PINN in any of the above embodiments, step S600 includes: S610. Define the total loss function, as follows: , in, These represent the aerodynamic loss weight, radar observation loss weight, and complex terrain and wind direction coupled wake loss weight, respectively, and their function is to adjust the relative importance of different types of constraints during the training process. S620, Constructing the training sample set The formula is as follows: , in, The set of sampling points for the aerodynamic physical information constraint loss function; For radar observation point set; A set of sampling points constrained by the wake evolution in complex environments; For the set of boundary condition sampling points; S630 and PINN network parameters are iteratively optimized so that PINN gradually approaches the optimal solution under limited observation and multiple physical constraints. S640 and PINN training: Perform PINN forward propagation and calculate the residuals of the aerodynamic physical information constraint loss function, the radar observation sequence physical information constraint loss function, and the wake evolution physical information constraint loss function. Then, use the training sample set... Input into PINN, perform forward propagation, and obtain the predicted values ​​of the pressure difference before and after the blade, the predicted values ​​of the incoming wind speed, and the predicted values ​​of the wake wind speed; Once S650 and PINN training are complete, parameter updates are stopped and PINN training is finished when the total loss function meets the convergence condition.

[0042] Optionally, in the wind turbine wake inversion dynamic balance and lifetime prediction method based on multi-source PINN in the above embodiments, the aerodynamic loss weight... The value range is [0.3, 0.5].

[0043] Preferably, in the wind turbine wake inversion dynamic balance and lifetime prediction method based on multi-source PINN in the above embodiments, the aerodynamic loss weight... .

[0044] Optionally, in the wind turbine wake inversion dynamic balance and lifetime prediction method based on multi-source PINN in the above embodiments, the radar observation loss weight is... The value range is [0.2, 0.4].

[0045] Preferably, in the wind turbine wake inversion dynamic balancing and lifetime prediction method based on multi-source PINN in the above embodiments, the radar observation loss weight is... .

[0046] Optionally, in the wind turbine wake inversion dynamic balance and lifetime prediction method based on multi-source PINN in the above embodiments, the wake loss weight is coupled with complex terrain and wind direction. The value range is [0.2, 0.4].

[0047] Preferably, in the wind turbine wake inversion dynamic balance and lifetime prediction method based on multi-source PINN in the above embodiments, the complex terrain and wind direction are coupled with the wake loss weight. .

[0048] Furthermore, in the wind turbine wake inversion dynamic balance and lifetime prediction method based on multi-source PINN in the above embodiments, in order to improve training efficiency and key area modeling capabilities, a partitioned encrypted sampling strategy is adopted, as follows: using the full-field benchmark sampling density... For reference, in the impeller front and rear action zones and near the wake zone Inside, the aerodynamic constraint point set The sampling density was increased to Furthermore, the number of sampling points in this area accounts for 40% to 60% of the total number of points; in the area near the centerline of the wake covered by radar. , Inside, the set of observation constraint points will be... Sampling density increased to When the slope angle is greater than The lateral drift of the wake is greater than Within the region, the set of wake evolution constraint points The sampling density was increased to .

[0049] Furthermore, in the wind turbine wake inversion dynamic balancing and lifetime prediction method based on multi-source PINN in any of the above embodiments, step S630 includes: S631. Determine the training objective. The parameter set for PINN is as follows: , in, Represented as the first Layer weight matrix; Represented as the first Layer bias vector; The formula for calculating the training objective is as follows: (where represents the number of network layers) , in, These are the optimal network parameters after training. For network parameters The total loss function is as follows; S632, Parameter Iterative Update: Gradient descent-like methods are used to iteratively update the parameters. , in, For training iteration rounds; For the first The learning rate of each training round; This represents the gradient of the total loss function with respect to the parameters.

[0050] Furthermore, in the wind turbine wake inversion dynamic balance and lifetime prediction method based on multi-source PINN in the above embodiments, the PINN training uses the Adam optimizer for global pre-training, and then uses L-BFGS for fine optimization, taking into account both global search capability and local convergence accuracy.

[0051] Furthermore, in the wind turbine wake inversion dynamic balance and lifetime prediction method based on multi-source PINN in the above embodiments, the convergence condition is any of the following conditions: 1. In 1000 consecutive iterations, the relative change in the total loss function is less than ; 2. The gradient norm of the total loss function with respect to the network parameters is less than... ; 3. The maximum number of training rounds in the Adam pre-training stage is 1,000 to 50,000, preferably 5,000 to 20,000; the maximum number of iterations in the L-BFGS fine-tuning stage is 100 to 5,000, preferably 500 to 2,000.

[0052] Preferably, in the wind turbine wake inversion dynamic balance and lifetime prediction method based on multi-source PINN in the above embodiments, the maximum number of training rounds in the Adam pre-training stage is 15,000, and the maximum number of iterations in the L-BFGS fine optimization stage is 1,500.

[0053] Optionally, in the wind turbine wake inversion dynamic balancing and lifetime prediction method based on multi-source PINN in any of the above embodiments, step S700 includes: S710. Establish blade dynamic balance deviation index. Based on the local pressure difference distribution of the blade, calculate the total aerodynamic torque of each blade at a given moment and establish blade dynamic balance deviation index. S720. Calculate the equivalent unbalanced excitation of the rotating system. The formula for calculating the equivalent unbalanced excitation of the rotating system is as follows: , in, For the rotating system at time The equivalent unbalanced excitation amplitude; The effective radius of the impeller; This represents the deviation of the blade relative to the average aerodynamic torque; For the first The instantaneous azimuth angle of each blade; To map the unbalanced effects of each blade to a complex phase factor in the rotating coordinate system, the torque deviations of different blades are not simply added as scalars during rotation, but rather superimposed as rotating loads of different phases, ultimately forming a periodic synthetic unbalanced excitation that acts on the main shaft and tower. It is a natural constant, approximately equal to 2.71828. The imaginary unit, In the first Each leaf at time Instantaneous azimuth; S730. Calculate the equivalent fatigue load using the following formula: , in, Equivalent fatigue load; Fatigue statistics time window; Fatigue index; This indicates the nonlinear amplification effect of high-amplitude loads on fatigue damage; S740. Estimate remaining life, and quantitatively assess the fatigue damage degree and remaining life of key components based on material fatigue characteristics and damage criteria.

[0054] Furthermore, in the wind turbine wake inversion dynamic balancing and lifetime prediction method based on multi-source PINN in the above embodiments, step S710 includes: S711. Calculate the aerodynamic torque for the first... Each leaf, at time Aerodynamic torque The calculation formula is as follows: , in, For the first The effective force-bearing area of ​​each blade; For local pressure difference, The lever arm of the local force-bearing point relative to the axis of rotation; The surface area of ​​the blade is a micro-element.

[0055] S712, Calculate the average aerodynamic torque, in In a wind turbine composed of blades, the average aerodynamic torque is defined by the following formula: , in, For all blades at time The average aerodynamic torque; The total number of blades is usually taken as follows for three-bladed fans. ; S713. Calculate the dynamic balance deviation, define the first... The dynamic balance deviation of each blade relative to the average aerodynamic moment is calculated using the following formula: , in, for Dynamic balance deviation of each blade This indicates that the blade's condition at that moment is consistent with the overall average state. A larger value indicates that the blade has significant aerodynamic imbalance.

[0056] Furthermore, in the wind turbine wake inversion dynamic balancing and lifetime prediction method based on multi-source PINN in the above embodiments, step S740 includes: S741. Determine the damage criterion, using the Miner linear cumulative damage criterion, as shown in the following formula: , in, The number of segments for the load range or operating condition; No. The actual number of cycles that have been applied at the load level; The failure cycle number obtained from the SN curve at this load level, when cumulative fatigue damage When the value is close to 1, it indicates that the component is approaching the critical state of fatigue failure. S742. Estimate the remaining lifetime. The remaining lifetime can be approximately calculated as follows: , The current cumulative damage is The damage growth rate per unit time is , Remaining useful life; S743, Maintenance Priority Decision: For multiple key components, including wind turbine blades, main shaft, and tower, calculate the equivalent fatigue load, cumulative damage value, and remaining life respectively, and sort them according to the cumulative damage value from largest to smallest or the remaining life from shortest to longest to form a multi-component maintenance priority list for maintenance priority decision.

[0057] Furthermore, in the wind turbine wake inversion dynamic balancing and lifetime prediction method based on multi-source PINN in the above embodiments, the operation and maintenance priority decision refers to the highest operation and maintenance priority corresponding to the component with the largest cumulative damage value or the shortest remaining lifetime.

[0058] This application integrates pulse Doppler wind radar observation information, aerodynamic conservation laws, and the wake evolution laws coupled with complex terrain and wind direction into a physical information neural network, achieving a unified mapping between wind turbine wake characteristics, blade aerodynamic torque deviation, unbalanced excitation of the rotating system, and fatigue life of key components. This application not only improves the accuracy and robustness of wind turbine dynamic balance diagnosis in complex wind farm environments but also effectively distinguishes between environmentally induced wake anomalies and equipment condition-induced wake anomalies, reducing the risk of false alarms and missed alarms. Without the need for directly deploying blade load sensors, this application can identify the health status and estimate the remaining life of key components such as blades, main shafts, and towers based on external wake observations, offering non-contact, interpretable, and online deployment advantages. This application provides quantitative basis for wind turbine condition monitoring, fault early warning, maintenance planning, and life extension decisions, improving the intelligent operation and maintenance level and the economic efficiency of wind turbine operation throughout its entire life cycle.

[0059] The following will further explain the concept, specific structure, and technical effects of the present invention in conjunction with the accompanying drawings, so as to fully understand the purpose, features, and effects of the present invention. Attached Figure Description

[0060] Figure 1 This is a flowchart illustrating a wind turbine wake inversion dynamic balancing and lifetime prediction method based on a multi-source PINN according to an exemplary embodiment. Detailed Implementation

[0061] The following description, with reference to the accompanying drawings, illustrates several preferred embodiments of the present invention to make its technical content clearer and easier to understand. The present invention can be embodied in many different forms, and the scope of protection of the present invention is not limited to the embodiments mentioned herein.

[0062] In the accompanying drawings, components with the same structure are indicated by the same numerical designation, and components with similar structures or functions are indicated by similar numerical designations. The dimensions and thicknesses of each component shown in the drawings are arbitrary, and the present invention does not limit the dimensions and thicknesses of each component. To make the illustrations clearer, the thickness of components is schematically exaggerated in some places in the drawings.

[0063] This application designs a method for wind turbine wake inversion dynamic balance and lifetime prediction based on multi-source PINN, such as... Figure 1 As shown, it includes the following steps: S100. Multi-source data acquisition: Establishing a local coordinate system for the wind turbine, acquiring multi-source data, and performing preprocessing; specifically including: S110. Establish a local coordinate system for the wind turbine, with the impeller center as the origin. Establish the local coordinate system of the wind turbine Geographic coordinate system As a unified reference for the entire site, the local coordinate system of the wind turbine is linked to the geographic coordinate system through a rotation transformation, as shown in the following formula: , , in, The coordinates of the wind turbine hub center in the geographic coordinate system; This refers to the azimuth angle of the cabin. The coordinates represent the main incoming flow direction of the wind turbine; The horizontal offset coordinate; The vertical coordinates are relative to the center of the wheel hub; S120 Multi-source data acquisition: This involves acquiring multi-source data for wind turbine wake inversion, dynamic balance identification, and lifespan prediction. This includes pulse-Doppler wind radar data, wind turbine SCADA operation data, yaw angle data, pitch angle data, rotor speed data, and ambient wind direction and speed data. Pulse-Doppler wind radar data is used to obtain wind speed observation information at different distances and heights in the inflow area in front of the turbine rotor and the wake area behind the rotor. Wind turbine SCADA operation data includes active power, unit operating status, unit start / stop status, and fault codes. Yaw angle data represents the nacelle orientation, pitch angle data represents the blade angle of attack adjustment status, rotor speed data represents the rotor rotation condition, and ambient wind direction and speed data represents the boundary inflow conditions of the local area of ​​the wind farm where the wind turbine is located. In complex terrain wind farm scenarios, the multi-source data also includes a digital elevation model (DEM), surface roughness parameters, topographic slope information, and underlying surface thermal characteristic parameters, used to construct wake evolution constraints under terrain and wind direction coupling conditions. S130. Data preprocessing: Preprocessing multi-source data, including: timestamp unification, coordinate mapping, outlier removal, noise filtering, missing value imputation, invalid observation removal, and resampling; timestamp unification includes aligning multi-source data to a unified time series. ,in It represents the time variable in a unified time series, characterizing the sampling time of multi-source data under the same time reference; This indicates the total duration of the preset analysis time window, which is the upper limit of the time range corresponding to the unified time series. The value is set to 2000 seconds; coordinate mapping includes mapping multi-source data to the local coordinate system of the wind turbine; missing value imputation includes using threshold discrimination, interpolation imputation, or sliding window smoothing to correct for observation samples with abnormal jumps, missing records, or low signal-to-noise ratios; resampling processing includes using a uniform step size for data sources with different sampling frequencies. Resampling is performed to form a time-synchronized multi-source dataset.

[0064] S200. Construct a physical information neural network, build PINN, define the input and output of PINN, and construct a unified velocity field; specifically including: S210. Construct PINN: Construct a PINN (Physical Information Neural Network) for wind turbine wake inversion. PINN adopts a multi-layer feedforward fully connected neural network structure, including an input layer, several hidden layers, and an output layer. The number of hidden layers and the number of neurons in each hidden layer are adjusted according to the complexity of the wind turbine wake field, the number of observation samples, and the training convergence. PINN has 8 hidden layers, and each hidden layer has 64 neurons. The mapping relationship between network layers is as follows: , in, For the first The state vector of the hidden layer. For the first Layer weight matrix, For the first Layer bias vector, For activation function, Network layers are numbered; the Tanh function is used as the activation function to ensure that the network output can be automatically differentiated with respect to time and space variables; the number of partial derivatives includes those used by PINN to predict wake wind speed. PINN output predicted aerodynamic pressure difference before and after the blade And the velocity components of the unified velocity field in the direction of the main flow and the transverse direction. ; Define the basic input variables of PINN The formula is as follows: , in, Indicates the inflow cross section in front of the impeller. Spatial coordinates of the upsampling point; Indicates the wake section behind the impeller Spatial coordinates of the upsampling point; It represents a time variable, characterizing the time dependence of unsteady wind conditions, yaw adjustment, pitch control, and wake evolution during wind turbine operation; S220, Forming the extended input vector The formula is as follows: , in, The dominant wind direction angle time series; Integrated terrain parameter vector; This refers to the cabin yaw angle. For variable pitch angle; The impeller speed, T Indicates transpose; S230, Calculate the input of PINN The formula is as follows: , in, Indicates the input to PINN; S240. Define the output of PINN using the following formula: , in, This is the output of PINN (Physical Information Neural Network). The predicted aerodynamic pressure difference between the blade and the front and back is output by PINN at a given spatial location and time, forming the aerodynamic pressure difference field between the blade and the front and back in the entire spatiotemporal domain; This is the predicted wind speed of the incoming flow in front of the impeller. This is the predicted wind speed of the wake behind the impeller. S250. Construct a unified velocity field and an interpolation weighting function between the front and rear cross sections. Assume the incoming flow cross section is located at... The wake section is located , The coordinates for the main incoming flow direction of the wind turbine are given by the following formula: , , in, Let be the coordinates of the inflow section in the direction of the main inflow. The coordinates of the rear wake section in the direction of the main incoming flow are given. Let represent the contribution weights of the incoming flow state and the backward wake state to the current spatial coordinate position, respectively, satisfying: , when Sometimes, , And when Sometimes, , ; The unified velocity field is represented as follows: , in, To ensure a uniform velocity field in the areas before and after the wind turbine. The coordinates represent the main incoming airflow direction of the wind turbine. The horizontal offset coordinates are... The vertical coordinates are relative to the center of the wheel hub. It is a time variable; S260 Automatic Differentiation: Based on PINN, an automatic differentiation function is established to calculate the number of partial derivatives in the physical information constraint loss function with respect to spatial and temporal variables.

[0065] S300. Constructing an aerodynamic physical information constraint loss function: Based on the principles of mass and momentum conservation of incompressible fluids, construct the aerodynamic physical information constraint loss function for PINN; specifically including: S310. Define the momentum equation for the main flow direction. Based on the momentum conservation equation for incompressible fluids, the momentum equation for the main flow direction is obtained as follows: , in, air density, It represents the change of wake velocity over time, reflecting the unsteady effects caused by gusts, yaw adjustments, and pitch control; This represents the self-convection term in the direction of the main inflow, reflecting the momentum transport of the wake along the downstream direction. To unify the velocity components of the velocity field in the direction of the main incoming flow; This indicates the lateral transport of momentum in the direction of the main flow caused by the lateral velocity. To unify the velocity components of the velocity field in the lateral direction; This indicates the vertical transport of momentum in the direction of the main flow caused by the vertical velocity. To unify the velocity components of the velocity field in the vertical direction; This represents the pressure gradient driving force along the mainstream direction; S320. Modify the momentum equation in the main incoming flow direction, and use the predicted aerodynamic pressure difference before and after the blades output by PINN. Constructing the pressure driving term in the direction of the main inflow, the momentum equation in the direction of the main inflow is modified as follows: ; S330, Define the aerodynamic physical information constraint loss function The formula is as follows: , in, It is the square of the mass conservation error; It is the square of the momentum conservation error in the direction of the main flow, subscript Indicates the physical constraint sampling point number. The number of sampling points is constrained by aerodynamics. .

[0066] S400. Construct the physical information constraint loss function for the observation sequence. Based on the time derivative and wake wind speed prediction values ​​of the pulse Doppler wind measurement radar observation sequence, define the physical information constraint loss function for the radar observation sequence; specifically including: S410. Define the time derivative of the radar observation sequence for the first... Preprocessing sequence of radar observation points Its time derivative is defined as: , in, For the first The rate of change of wake wind speed at each radar observation point over time, i.e., the time derivative; S420, Calculate the predicted wake velocity, calculate the first... Radar observation points place, time Wake wind speed prediction The formula is as follows: , in, The wake wind speed field output by PINN; For the first The location of each radar observation point in the local coordinate system of the wind turbine. For the first The position coordinates of each radar observation point along the main flow direction in the local coordinate system of the wind turbine; For the first Lateral offset coordinates of each radar observation point in the local coordinate system of the wind turbine; For the first The vertical coordinates of each radar observation point relative to the hub center in the local coordinate system of the wind turbine; It is a time variable; For optional environmental and control inputs, As the prevailing wind angle, ; To synthesize the topographic parameter vector, which characterizes environmental factors, including topographic slope, aspect, roughness, and underlying surface thermal properties; The cabin yaw angle, ; Pitch angle, ; Impeller angular velocity, ; S430, Define the radar differential consistency residual, PINN for the... The predicted time series for each radar observation point is represented as follows: , Calculate at time The radar differential consistency residual is: , in, For the first Each radar observation point at time The observed trend residuals; This represents the time rate of change of the PINN prediction of the wake wind speed; This represents the rate of change of the radar-measured wake wind speed over time. S440. Define the physical information constraint loss function for the radar observation sequence, as follows: , in, This is an effective observation mask used to exclude invalid or missing observations; the values ​​in square brackets are the first and second observations. The radar observation point at the ... The differential residual at each moment, The total number of effective radar observation points is [number missing]. The number of time sampling moments.

[0067] S500. Construct a wake evolution physical information constraint loss function, considering the influence of complex terrain and wind direction, define the effective wake propagation direction angle, calculate the wake propagation velocity, define the wake attenuation coefficient, and construct a wake wind speed physical information constraint loss function; specifically including: S510. Define the effective wake propagation direction angle, as follows: , in, The effective wake propagation direction angle; The dominant wind direction angle time series represents the main direction of the incoming flow in the horizontal plane at the location of the wind turbine; This is the wake direction correction amount caused by the combined effects of terrain and yaw; Yaw error, representing the difference between the prevailing wind angle and the cabin yaw angle. The comprehensive topographic parameter vector characterizes the environmental factors at the wind turbine site, including topographic slope, aspect, surface roughness, and underlying thermal properties. S520. Calculate the wake propagation velocity. The wake propagation velocity characterizes how quickly the wake deficit structure migrates downstream. The calculation formula is as follows: , in, As the baseline propagation term; The direct impact of incoming wind speed on wake propulsion speed. This is the correction factor for the incoming airflow velocity. For PINN, the predicted inflow velocity in front of the impeller; This represents the effect of the relative slope angle between wind direction and propagation direction on the uphill and downhill slope effects. The angle of the main slope of the terrain. The angle between the wind direction and the slope aspect. This is the wind direction-slope aspect coupling correction factor; The effect of terrain slope intensity on propagation speed. This is the slope correction factor. For terrain slope intensity; The effect of surface roughness on the propulsion velocity of near-surface wake. This is the roughness correction factor. Length represents the surface roughness. This is a correction for wake centerline drift caused by yaw error. This is the yaw correction factor. This refers to yaw error; These are model parameters, determined through training; S530. Define the wake attenuation coefficient, as follows: , in, Basic recovery item; Roughness refers to the amount of material in the wake. The greater the roughness, the stronger the entrainment and mixing in the wake, and the faster the recovery. This is the roughness-induced hybrid reinforcement coefficient. , where is the surface roughness length, and together they characterize the enhancing effect of surface roughness on wake entrainment, turbulent mixing, and velocity recovery processes; The slope will enhance local shear and flow direction adjustment, thus affecting the recovery rate. This is the slope-induced recovery adjustment coefficient. Both topographic slope intensity and topographic slope intensity together characterize the moderating effect of topographic relief on local flow shear, wake diffusion and recovery intensity. When the wind direction and slope aspect are inconsistent, lateral disturbances are more pronounced. The wind direction-slope aspect mismatch recovery influence coefficient. As the prevailing wind angle, The angle of the main slope of the terrain. It represents the magnitude of the lateral deviation of the wind direction relative to the slope aspect, and is used to characterize the degree of enhancement of the lateral disturbance and asymmetric recovery of the wake when the wind direction and slope aspect are inconsistent. The greater the intensity of wind direction fluctuations, the more pronounced the wake instability recovery. The intensity of wind direction change, The wind direction fluctuation recovery response coefficient; For offline calibration parameters, historical radar wake data, SCADA data, and known operating condition data are used to perform offline calibration through least squares fitting, Bayesian estimation, or grid search. ; S540. Construct a first-order evolution equation for wake wind speed. To describe the comprehensive evolution of wake wind speed over time, in the main propagation direction, and in the lateral offset direction, a first-order evolution equation for wake wind speed is constructed, as follows: , in, Predict wake wind speed for PINN This indicates that PINN predicts the incoming wind speed. The principal propagation coordinates are established along the effective wake propagation direction. To and Orthogonal lateral offset coordinates The rate of change of wake wind speed with respect to time represents the unsteady response caused by gusts, control actions, and environmental fluctuations. The term represents the wake translation along the main propagation direction. Let be the wake propagation velocity function, representing the effective propagation velocity of the wake-depleted structure migrating downstream under the combined influence of prevailing wind direction, topographic parameters, and time conditions. For the wake wind speed predicted by PINN, This represents the spatial gradient of the wake wind speed along the effective propagation direction; it also represents the wake structure as it advances downstream. The lateral deflection velocity function, The lateral deflection term represents the lateral drift of the wake under the influence of complex terrain or changes in wind direction. This is the wake recovery term, representing how quickly the wake velocity recovers to the incoming flow velocity. This is the wake attenuation coefficient, used to describe how quickly the wake wind speed recovers from a deficit state to an incoming flow state. This represents the deviation of the wake velocity from the predicted incoming velocity. Predict the incoming wind speed for PINN; S550, Constructing the physical information constraint loss function for wake wind speed The formula is as follows: , in, The number of physical sampling points for wake evolution. .

[0068] S600. Construct the total loss function and complete PINN training. Define the multi-constraint weighted total loss function, construct the training sample set, iteratively optimize the PINN network parameters, and complete PINN training; specifically including: S610. Define the total loss function, as follows: , in, These represent the aerodynamic loss weight, radar observation loss weight, and wake loss weight coupled with complex terrain and wind direction, respectively. Their function is to adjust the relative importance of different types of constraints during the training process. , , ; S620, Constructing the training sample set The formula is as follows: , in, The set of sampling points for the aerodynamic physical information constraint loss function; For radar observation point set; A set of sampling points constrained by the wake evolution in complex environments; This is the set of sampling points for boundary conditions. To improve training efficiency and the ability to model key regions, a partitioned encrypted sampling strategy is adopted, as follows: using the overall baseline sampling density... For reference, in the impeller front and rear action zones and near the wake zone Inside, the aerodynamic constraint point set The sampling density was increased to Furthermore, the number of sampling points in this area accounts for 40% to 60% of the total number of points; in the area near the centerline of the wake covered by radar. , Inside, the set of observation constraint points will be... Sampling density increased to When the slope angle is greater than The lateral drift of the wake is greater than Within the region, the set of wake evolution constraint points The sampling density was increased to ; Iterative optimization of S630 and PINN network parameters enables PINN to gradually approach the optimal solution under limited observations and multiple physical constraints; specifically including: S631. Determine the training objective. The parameter set for PINN is as follows: , in, Represented as the first Layer weight matrix; Represented as the first Layer bias vector; The formula for calculating the training objective is as follows: (where represents the number of network layers) , in, These are the optimal network parameters after training. For network parameters The total loss function is as follows; S632, Parameter Iterative Update: Gradient descent-like methods are used to iteratively update the parameters. , in, For training iteration rounds; For the first The learning rate of each training round; This represents the gradient of the total loss function with respect to the parameters. S640 and PINN training: Perform PINN forward propagation and calculate the residuals of the aerodynamic physical information constraint loss function, the radar observation sequence physical information constraint loss function, and the wake evolution physical information constraint loss function. Then, use the training sample set... Input into PINN, perform forward propagation to obtain the predicted values ​​of the pressure difference before and after the blade, the predicted values ​​of the incoming wind speed and the predicted values ​​of the wake wind speed. Use the Adam optimizer for global pre-training, and then use L-BFGS for fine optimization to balance global search capability and local convergence accuracy. Once S650 and PINN training are complete, parameter updates will stop and PINN training will be finished when the total loss function meets the convergence condition. The convergence condition is any of the following: 1. In 1000 consecutive iterations, the relative change in the total loss function is less than ; 2. The gradient norm of the total loss function with respect to the network parameters is less than... ; 3. The maximum number of training rounds in the Adam pre-training phase is 15,000, and the maximum number of iterations in the L-BFGS fine-tuning phase is 1,500.

[0069] S700, Remaining service prediction, establishing blade dynamic balance deviation index, calculating equivalent fatigue load and equivalent fatigue load, and estimating remaining service life; including: S710. Establish blade dynamic balance deviation index. Based on the local pressure difference distribution of the blade, calculate the total aerodynamic torque of each blade at a given moment, and establish the blade dynamic balance deviation index; including: S711. Calculate the aerodynamic torque for the first... Each leaf, at time Aerodynamic torque The calculation formula is as follows: , in, For the first The effective force-bearing area of ​​each blade; For local pressure difference, The lever arm of the local force-bearing point relative to the axis of rotation; The surface area of ​​the blade is a micro-element.

[0070] S712, Calculate the average aerodynamic torque, in In a wind turbine composed of blades, the average aerodynamic torque is defined by the following formula: , in, For all blades at time The average aerodynamic torque; The total number of blades is usually taken as follows for three-bladed fans. ; S713. Calculate the dynamic balance deviation, define the first... The dynamic balance deviation of each blade relative to the average aerodynamic moment is calculated using the following formula: , in, for Dynamic balance deviation of each blade This indicates that the blade's condition at that moment is consistent with the overall average state. A larger value indicates that the blade has significant aerodynamic imbalance; S720. Calculate the equivalent unbalanced excitation of the rotating system. The formula for calculating the equivalent unbalanced excitation of the rotating system is as follows: , in, For the rotating system at time The equivalent unbalanced excitation amplitude; The effective radius of the impeller; This represents the deviation of the blade relative to the average aerodynamic torque; For the first The instantaneous azimuth angle of each blade; To map the unbalanced effects of each blade to a complex phase factor in the rotating coordinate system, the torque deviations of different blades are not simply added as scalars during rotation, but rather superimposed as rotating loads of different phases, ultimately forming a periodic synthetic unbalanced excitation that acts on the main shaft and tower. It is a natural constant, approximately equal to 2.71828. The imaginary unit, In the first Each leaf at time Instantaneous azimuth; S730. Calculate the equivalent fatigue load using the following formula: , in, Equivalent fatigue load; Fatigue statistics time window; Fatigue index; This indicates the nonlinear amplification effect of high-amplitude loads on fatigue damage; S740. Estimate remaining life: Based on material fatigue characteristics and damage criteria, quantitatively assess the fatigue damage level and remaining life of critical components; specifically including: S741. Determine the damage criterion, using the Miner linear cumulative damage criterion, as shown in the following formula: , in, The number of segments for the load range or operating condition; No. The actual number of cycles that have been applied at the load level; The failure cycle number obtained from the SN curve at this load level, when cumulative fatigue damage When the value is close to 1, it indicates that the component is approaching the critical state of fatigue failure. S742. Estimate the remaining lifetime. The remaining lifetime can be approximately calculated as follows: , The current cumulative damage is The damage growth rate per unit time is , Remaining useful life; S743, Maintenance Priority Decision: For multiple key components, including wind turbine blades, main shaft, and tower, calculate the equivalent fatigue load, cumulative damage value, and remaining life respectively, and sort them according to the cumulative damage value from largest to smallest or the remaining life from shortest to longest to form a multi-component maintenance priority list. Maintenance priority decision means that the component with the largest cumulative damage value or the shortest remaining life corresponds to the highest maintenance priority.

[0071] To verify the technical effectiveness of the above embodiments, under the same wind turbine model, the same observation time window, the same pulse Doppler wind radar deployment conditions, and the same wind field environmental parameters, the above embodiments were compared with traditional single-sensor monitoring methods, pure data-driven wake prediction methods, and wake analysis methods that do not consider the coupling effect of complex terrain and wind direction. The comparison mainly included wake wind speed field reconstruction capability, wake dynamic evolution trend tracking capability, blade dynamic balance state identification capability, environmental adaptability in complex terrain scenarios, and the rationality of key component life assessment.

[0072] Comparative test results show that the above embodiments demonstrate better physical consistency and interpretability in identifying the location of the wind turbine wake wind speed deficit zone, characterizing the wake deflection trend, responding to wake dynamic changes, and inverting the aerodynamic imbalance of the wind turbine blades. Especially under complex terrain, multi-wind direction, and unsteady wind conditions, the above embodiments, by simultaneously introducing aerodynamic physical constraints, radar observation sequence differential constraints, and complex terrain and wind direction coupled wake evolution constraints, make the predicted wake structure more consistent with the actual observation results, effectively reducing misjudgments caused by local noise, sparse observations, or environmental disturbances.

[0073] Furthermore, compared to existing methods that rely solely on vibration signals, nacelle monitoring signals, or simple wake speed fitting results, the above embodiments can establish a unified link between externally observable wake changes and the pressure difference before and after the blades, blade aerodynamic torque deviations, unbalanced excitation of the rotating system, and the evolution of fatigue loads. Therefore, they provide stronger mechanistic support for identifying wind turbine dynamic balance states and assessing remaining life. Comparative tests show that the above embodiments can not only more accurately distinguish between external wake anomalies caused by complex terrain and wind direction changes and internal anomalies caused by blade damage, adhesion, icing, or aerodynamic inconsistencies, but also provide a more reasonable basis for determining the maintenance priorities of key components such as blades, main shafts, and towers.

[0074] Furthermore, under the same data input conditions, the above embodiments outperform traditional methods in terms of model training convergence stability, cross-condition adaptability, and multi-source information fusion capability. This demonstrates that the above embodiments can achieve unified analysis of wind turbine wake characteristics, dynamic balance state, and life evolution without directly deploying blade load sensors.

[0075] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for wind turbine wake inversion dynamic balancing and lifetime prediction based on multi-source PINN, characterized in that, Includes the following steps: S100. Multi-source data acquisition: Establish the local coordinate system of the wind turbine, acquire multi-source data, and perform preprocessing. S200, construct a physical information neural network, construct PINN, and construct a unified velocity field; S300, Constructing the aerodynamic physical information constraint loss function, wherein S300 includes: S310. Define the momentum equation for the main flow direction. Based on the momentum conservation equation for incompressible fluids, the momentum equation for the main flow direction is obtained as follows: ， in, air density, This represents the change in wake velocity with respect to time; Indicates the direction of the main inflow from the convection term. To unify the velocity components of the velocity field in the direction of the main incoming flow; This indicates the lateral transport of momentum in the direction of the main flow caused by the lateral velocity. To unify the velocity components of the velocity field in the lateral direction; This indicates the vertical transport of momentum in the direction of the main flow caused by the vertical velocity. To unify the velocity components of the velocity field in the vertical direction; This represents the pressure gradient driving force along the mainstream direction; S320. Modify the momentum equation in the main incoming flow direction, and use the predicted aerodynamic pressure difference before and after the blade output by PINN. The pressure driving term in the main inflow direction is constructed as follows: ； S330, Define the aerodynamic physical information constraint loss function The formula is as follows: ， in, It is the square of the mass conservation error; It is the square of the momentum conservation error in the direction of the main flow, subscript Indicates the physical constraint sampling point number. The number of aerodynamically constrained sampling points; S400, Constructing the physical information constraint loss function for the observation sequence, defining the radar observation sequence physical information constraint loss function, wherein S400 includes: S410. Define the time derivative of the radar observation sequence for the first... Preprocessing sequence of radar observation points Its time derivative is defined as: ， in, For the first The rate of change of wake wind speed at each radar observation point over time; S420, Calculate the predicted wake velocity, calculate the first... Radar observation points place, time Wake wind speed prediction The formula is as follows: ， in, The wake wind speed field output by PINN; For the first The location of each radar observation point in the local coordinate system of the wind turbine. For the first The position coordinates of each radar observation point along the main flow direction in the local coordinate system of the wind turbine; For the first Lateral offset coordinates of each radar observation point in the local coordinate system of the wind turbine; For the first The vertical coordinates of each radar observation point relative to the hub center in the local coordinate system of the wind turbine; It is a time variable; For optional environmental and control inputs, The prevailing wind angle; To synthesize the topographic parameter vector, which characterizes environmental factors, including topographic slope, aspect, roughness, and underlying surface thermal properties; This refers to the cabin yaw angle. Pitch angle; Impeller angular velocity; S430, Define the radar differential consistency residual, PINN for the... The predicted time series for each radar observation point is represented as follows: ， Calculate at time The radar differential consistency residual is: ， in, For the first Each radar observation point at time The observed trend residuals; This represents the time rate of change of the PINN prediction of the wake wind speed; This represents the rate of change of the radar-measured wake wind speed over time. S440. Define the physical information constraint loss function for the radar observation sequence, as follows: ， in, For effective observation masking, the part in square brackets is the first... The radar observation point at the ... The differential residual at each moment, The total number of effective radar observation points is [number missing]. The number of time sampling moments; S500: Construct a wake evolution physical information constraint loss function, define a wake attenuation coefficient, and construct a wake wind speed physical information constraint loss function. S500 includes: S510. Define the effective wake propagation direction angle, as follows: ， in, For the effective wake propagation direction angle, For the prevailing wind direction angle time series, This is the amount of wake direction correction caused by the combined effects of terrain and yaw. For yaw error, This is a vector of integrated terrain parameters; S520. Calculate the wake propagation velocity using the following formula: ， in, As the baseline propagation term; The direct impact of incoming wind speed on wake propulsion speed. This is the correction factor for the incoming airflow velocity. For PINN, the predicted inflow velocity in front of the impeller; This represents the effect of the relative slope angle between wind direction and propagation direction on the uphill and downhill slope effects. The angle of the main slope of the terrain. The angle between the wind direction and the slope aspect. This is the wind direction-slope aspect coupling correction factor; The effect of terrain slope intensity on propagation speed. This is the slope correction factor. For terrain slope intensity; The effect of surface roughness on the propulsion velocity of near-surface wake. This is the roughness correction factor. Length represents the surface roughness. This is a correction for wake centerline drift caused by yaw error. This is the yaw correction factor. This refers to yaw error; These are model parameters; S530. Define the wake attenuation coefficient, as follows: ， in, Basic recovery items, For roughness, This is the roughness-induced hybrid reinforcement coefficient. The length of the surface roughness. The slope will enhance local shear and flow direction adjustment. This is the slope-induced recovery adjustment coefficient. For terrain slope intensity, When the wind direction and slope aspect are inconsistent, lateral disturbances are more pronounced. The wind direction-slope aspect mismatch recovery influence coefficient. As the prevailing wind angle, The angle of the main slope of the terrain. This indicates the magnitude of the lateral deviation of the wind direction relative to the slope aspect. The greater the intensity of wind direction fluctuations, the more pronounced the wake instability recovery. The intensity of wind direction change, The wind direction fluctuation recovery response coefficient; For offline calibration parameters, ; S540. Construct the first-order evolution equation for the wake wind speed. The formula is as follows: ， in, Predict wake wind speed for PINN This indicates that PINN predicts the incoming wind speed. The principal propagation coordinates are established along the effective wake propagation direction. To and Orthogonal lateral offset coordinates The rate of change of wake wind speed with respect to time represents the unsteady response caused by gusts, control actions, and environmental fluctuations. The term represents the wake translation along the main propagation direction. Let be the wake propagation velocity function. For the wake wind speed predicted by PINN, This represents the spatial gradient of the wake wind speed along the effective propagation direction. The lateral deflection velocity function, For lateral deflection, For wake recovery, The wake attenuation coefficient is... This represents the deviation of the wake velocity from the predicted incoming velocity. Predict the incoming wind speed for PINN; S550, Constructing the physical information constraint loss function for wake wind speed The formula is as follows: ， in, This represents the number of physical sampling points for wake evolution. S600. Construct the total loss function and complete PINN training. Define the multi-constraint weighted total loss function, construct the training sample set, iteratively optimize the PINN network parameters, and complete the PINN training. S700, Remaining Life Prediction: Establish blade dynamic balance deviation index, calculate equivalent fatigue load and equivalent fatigue load, and estimate remaining life.

2. The method for wind turbine wake inversion dynamic balancing and lifetime prediction based on multi-source PINN as described in claim 1, characterized in that, S100 includes: S110. Establish a local coordinate system for the wind turbine, with the impeller center as the origin. Establish the local coordinate system of the wind turbine Geographic coordinate system As a unified reference for the entire field, the local coordinate system of the wind turbine and the geographic coordinate system are linked through a rotation transformation, as shown in the following formula: ， ， in, The coordinates of the wind turbine hub center in the geographic coordinate system; This refers to the azimuth angle of the cabin. The coordinates represent the main incoming flow direction of the wind turbine; The horizontal offset coordinate; The vertical coordinates are relative to the center of the wheel hub; S120, Multi-source data acquisition, acquires multi-source data for wind turbine wake inversion, dynamic balance identification and life prediction; S130. Data preprocessing: preprocess the multi-source data.

3. The wind turbine wake inversion dynamic balancing and lifetime prediction method based on multi-source PINN as described in claim 1, characterized in that, S200 includes: S210. Construct PINN: Construct PINN for wind turbine wake inversion, and define the basic input variables of PINN. The formula is as follows: ， in, Indicates the inflow cross section in front of the impeller. Spatial coordinates of the upsampling point; Indicates the wake section behind the impeller Spatial coordinates of the upsampling point; Represents a time variable; S220, Forming the extended input vector The formula is as follows: ， in, The dominant wind direction angle time series; Integrated terrain parameter vector; This refers to the cabin yaw angle. For variable pitch angle; The impeller speed, T Indicates transpose; S230, Calculate the input of PINN The formula is as follows: ， in, This represents the input of PINN. This represents the basic input variable of the PINN. This represents an expanded input vector; S240. Define the output of PINN using the following formula: ， in, For the output of PINN, The predicted aerodynamic pressure difference before and after the blade is output by the PINN at a given spatial location and time. This is the predicted wind speed of the incoming flow in front of the impeller. This is the predicted wind speed of the wake behind the impeller. S250. Construct a unified velocity field and an interpolation weighting function between the front and rear cross sections. Assume the incoming flow cross section is located at... The wake section is located The formula is as follows: ， ， in, Let be the coordinates of the inflow section in the direction of the main inflow. The coordinates of the rear wake section in the direction of the main incoming flow are given. These represent the contribution weights of the incoming flow state and the wake state to the current spatial coordinate position, respectively. The unified velocity field is represented as follows: ， in, To ensure a uniform velocity field in the areas before and after the wind turbine. The coordinates represent the main incoming airflow direction of the wind turbine. The horizontal offset coordinates are... The vertical coordinates are relative to the center of the wheel hub. It is a time variable; S260, Automatic Differentiation: Establishes automatic differentiation of the output with respect to spatial and temporal variables, and calculates the number of partial derivatives in the physical information constraint loss function.

4. The wind turbine wake inversion dynamic balancing and lifetime prediction method based on multi-source PINN as described in claim 1, characterized in that, The S600 includes: S610. Define the total loss function, as follows: ， in, These represent the weights for aerodynamic losses, radar observation losses, and wake losses caused by the coupling of complex terrain and wind direction, respectively. S620, Constructing the training sample set The formula is as follows: ， in, The set of sampling points for the aerodynamic physical information constraint loss function; For radar observation point set; A set of sampling points constrained by the wake evolution in complex environments; For the set of boundary condition sampling points; S630 and PINN network parameters are iteratively optimized so that the PINN gradually approaches the optimal solution under limited observation and multiple physical constraints; S640 and PINN training: Perform PINN forward propagation and calculate the residuals of the aerodynamic physical information constraint loss function, the radar observation sequence physical information constraint loss function, and the wake evolution physical information constraint loss function. Then, use the training sample set... By inputting the PINN, the predicted values ​​of the pressure difference before and after the blade, the predicted values ​​of the incoming wind speed, and the predicted values ​​of the wake wind speed are obtained. S650. When the PINN training is complete and the total loss function meets the convergence condition, parameter updates are stopped, and the PINN training is completed.

5. The wind turbine wake inversion dynamic balancing and lifetime prediction method based on multi-source PINN as described in claim 4, characterized in that, The S630 includes: S631. Determine the training objective. The parameter set of PINN is as follows: ， in, Represented as the first Layer weight matrix; Represented as the first Layer bias vector; The formula for calculating the training objective is as follows: (where represents the number of network layers) ， in, These are the optimal network parameters after training. For network parameters The total loss function is as follows; S632, Parameter Iterative Update: Gradient descent-like methods are used to iteratively update the parameters. ， in, For training iteration rounds; For the first The learning rate of each training round; This represents the gradient of the total loss function with respect to the parameters.

6. The method for wind turbine wake inversion dynamic balancing and lifetime prediction based on multi-source PINN as described in claim 1, characterized in that, The S700 includes: S710. Establish blade dynamic balance deviation index, calculate the total aerodynamic torque of each blade at a given time, and establish blade dynamic balance deviation index. S720. Calculate the equivalent unbalanced excitation of the rotating system. The formula for calculating the equivalent unbalanced excitation of the rotating system is as follows: ， in, For the rotating system at time The equivalent unbalanced excitation amplitude, The effective radius of the impeller is 1. This represents the deviation of the blade relative to the average aerodynamic torque. For the first Instantaneous azimuth angle of each blade To map the unbalance effects of each blade to a complex phase factor in a rotating coordinate system, It is a natural constant. The imaginary unit, For the first Each leaf, at time Aerodynamic torque; S730. Calculate the equivalent fatigue load using the following formula: ， in, Equivalent fatigue load; Fatigue statistics time window; Fatigue index; This indicates the nonlinear amplification effect of high-amplitude loads on fatigue damage; S740. Estimate remaining life and quantitatively assess the fatigue damage level and remaining life of key components based on damage criteria.

7. The wind turbine wake inversion dynamic balancing and lifetime prediction method based on multi-source PINN as described in claim 6, characterized in that, The S710 includes: S711. Calculate the aerodynamic torque for the first... Each leaf, at time Aerodynamic torque The calculation formula is as follows: ， in, For the first The effective force-bearing area of ​​each blade For local pressure difference, The lever arm of the local force point relative to the axis of rotation. blade surface area micro-element; S712, Calculate the average aerodynamic torque, in In a wind turbine composed of blades, the average aerodynamic torque is defined by the following formula: ， in, For all blades at time The average aerodynamic torque, This represents the total number of leaves; S713. Calculate the dynamic balance deviation, define the first... The dynamic balance deviation of each blade relative to the average aerodynamic moment is calculated using the following formula: ， in, For the first The dynamic balance deviation of each blade.

Images