A pinn-based fan blade load prediction method

Abstract

The application discloses a fan blade load prediction method based on PINN, relates to the technical fields of wind power generation, wind turbine blade aerodynamics and physical information neural network modeling, and comprises the following steps: S100, blade parameter construction; S200, blade standard section division; S300, PINN modeling and wall surface hard constraint construction; S400, sub-domain sample set construction; S500, total loss function construction; S600, sub-stage hybrid training; S700, consistency checking; and S800, load prediction. The application realizes high-precision prediction of a blade velocity field, a pressure field and a spanwise load, and finally generated standard section data not only has good local flow field precision, but also can guarantee continuity and smoothness of overall load distribution.

Description

Technical Field

[0001] This invention relates to the field of wind power generation, wind turbine blade aerodynamics and physical information neural network modeling technology, and in particular to a wind turbine blade load prediction method based on PINN. Background Technology

[0002] As wind turbine generators develop towards larger capacity, longer blades, and greater flexibility, the aerodynamic flow field distribution of the blades under the combined effects of complex incoming flow, rotational induced loads, and unsteady loads becomes increasingly complex. The pressure distribution on the blade surface, the variation in spanwise aerodynamic loads, and the resulting first- and second-order flapping responses directly affect the blade structural safety, fatigue life, and overall turbine operational reliability. Therefore, establishing an aerodynamic flow field modeling method that can accurately characterize the local flow field of the blade while reflecting the continuity of the overall load and the structural availability has become a crucial technical requirement in the field of wind turbine aerodynamic analysis.

[0003] Existing aerodynamic analysis methods for wind turbine blades mainly fall into two categories: simplified engineering models and high-precision numerical simulations. While commonly used in engineering, blade element momentum methods offer high computational efficiency and are suitable for preliminary design and rapid evaluation, these methods typically discretize the blade into several relatively independent two-dimensional sections. This makes it difficult to fully describe the local aerodynamic characteristics of the blade root, tip, and complex three-dimensional flow regions. Furthermore, their ability to represent near-wall flow, abrupt pressure gradient changes, wake evolution, and unsteady coupling effects is limited. Especially under conditions of yaw inflow, shear wind, complex turbulence, and large flexible blades, relying solely on simplified models often fails to yield high-precision results for surface pressure distribution and spanwise loads.

[0004] Another type of method is high-fidelity numerical simulation based on computational fluid dynamics. This type of method can solve the flow field around the blade in greater detail and obtain the velocity field, pressure field, and surface load distribution, but it has high requirements for mesh quality, computational resources, and solution time. When it is necessary to perform repeated analysis on multiple standard sections, multiple operating conditions, or multiple design states, the computational cost is high, which is not conducive to rapid modeling and engineering promotion.

[0005] In recent years, physical information neural networks (PINNs) have provided new ideas for predicting complex flow fields. However, existing PINN methods still have significant shortcomings in wind turbine blade applications: First, wall boundary conditions are often applied through soft constraints using loss functions, which can easily lead to error accumulation on the blade surface and near-wall region. Second, existing methods focus on point-value fitting of local velocity and pressure fields, lacking explicit constraints on structurally sensitive indicators such as spanwise integral loads and flapping generalized moments. Third, for the generation of intermediate standard segment data, traditional methods often use simple interpolation or low-precision extrapolation, which makes it difficult to guarantee the continuity and smoothness of load distribution between segments and the structural availability required for subsequent fatigue life analysis.

[0006] Therefore, those skilled in the art are dedicated to developing a PINN-based method for predicting wind turbine blade loads. Summary of the Invention

[0007] In view of the above-mentioned deficiencies of the prior art, the technical problem to be solved by the present invention is how to model the aerodynamic flow field while taking into account the wall boundary accuracy, global physical consistency, spanwise load continuity and structural dynamic constraints.

[0008] Existing technologies suffer from problems such as complex three-dimensional flow in wind turbine blades, insufficient wall boundary accuracy, and difficulty in directly constraining structural loads. Unlike geometrically uniform and uniformly operating airfoils, wind turbine blades typically exhibit the following characteristics along their span: continuously varying chord length distribution, with a larger chord length at the blade root and a gradually decreasing chord length at the blade tip; continuously varying twist angle distribution, with significant differences in effective angle of attack and local aerodynamic characteristics at different radii; significant variations in thickness and airfoil family, with thicker airfoils at the blade root prioritizing load-bearing capacity and structural strength, while those at the blade mid-section and tip prioritize aerodynamic efficiency; non-uniform structural dynamic sensitivity, with certain spanwise coordinates exhibiting higher levels of involvement in first- and second-order flapping modes; non-uniform local load gradients, with more complex aerodynamic changes often observed in the leading-edge separation sensitive area, the high-lift area in the blade mid-section, and the three-dimensional loss area at the blade tip; and significant differences in the geometric shapes of the blade root, mid-section, and tip regions, with the flow near the wall simultaneously affected by curvature effects, rotational effects, and changes in incoming flow conditions.

[0009] First, a mapping relationship is established between blade geometric parameters and a rotating coordinate system, enabling the chord length, twist angle, thickness, and local airfoil information of different spanwise sections of the blade to be transformed into spatial geometric constraints recognizable by PINN. Next, descriptions of the inflow conditions, rotation conditions, and relative velocity field of the wind turbine under specific operating conditions are established, providing a consistent operating condition basis for the construction of fluid control equations, boundary conditions, and training samples. First and second-order flapping mode functions of the blade are extracted, providing a theoretical basis for embedding structural dynamics information into the aerodynamic modeling process and establishing first / second-order flapping generalized moment constraints based on integral loads. By dividing the blade into standard sections, the originally continuous and complex spanwise variations are transformed into segmented, controllable, modeled, and constrained technical objects. Then, through key section identification, limited high-fidelity computing resources are prioritized for the regions with the greatest impact on the overall flow field accuracy and structural load continuity. Finally, structural dynamics requirements are... This application integrates data generation and PINN training processes to achieve the modeling objectives of guiding a small number of high-fidelity segments, generating constrained data across the entire span, and ensuring continuous availability of overall loads. It establishes a boundary-fitting finite element mesh consistent with the actual geometry of the wind turbine blade on the blade wall and near-wall region, providing a geometric basis for subsequent wall distance function construction, boundary normal extraction, near-wall collocation organization, and PINN output hard constraint embedding. Furthermore, it selects high-fidelity samples based on spatial selection principles and introduces low-fidelity samples to improve training efficiency while maintaining accuracy in key areas. Finally, it establishes a PINN multi-objective loss function and performs staged mixed training to obtain the constrained model.

[0010] This application constructs a complete technical route consisting of standard segment division, hard boundary constraints, phased hybrid training, and torque consistency verification. The blade surface boundary conditions are directly embedded into the network output. Combined with high-fidelity data guidance, physical conservation constraints, and first / second-order flapping generalized torque constraints, it enables high-precision prediction of the blade velocity field, pressure field, and spanwise load. The final generated standard segment data not only has good local flow field accuracy but also ensures the continuity and smoothness of the overall load distribution. It can be directly used for aerodynamic analysis of wind turbine blades and subsequent operation and maintenance.

[0011] In this application, "blade" refers to wind turbine blades.

[0012] In one embodiment of the present invention, a method for predicting wind turbine blade loads based on PINN is provided, comprising the following steps: S100, Blade parameter construction, collecting blade geometric parameters, and establishing the actual operating environment and boundary conditions of the wind turbine blade; S200, blade standard segment division, divides the blade spanwise region into multiple standard segments, and identifies key segments that affect the overall aerodynamic flow field accuracy, spanwise load continuity and structural dynamic response. S300 and PINN are modeled and hard constraints are constructed on the wall. A boundary fitting finite element surface mesh consistent with the three-dimensional shape is generated on the blade surface. Finite element elements in the near-wall region are constructed. Finite element shape functions are used for interpolation. PINN is constructed and hard constraint velocity output is defined. S400, constructing a domain-specific sample set: select high-fidelity samples, physical collocations, and low-fidelity samples to construct a high-fidelity sample set, a physical collocation set, and a low-fidelity sample set, respectively; S500, Construction of Total Loss Function: Establish physical equation residual loss, high-fidelity supervision loss, low-fidelity supervision loss, torque loss, and construct the total loss function; S600 uses a phased hybrid training method, performing first-stage and second-stage training. The continuous weight scheduling function ensures a smooth transition between the first-stage and second-stage training processes, reducing oscillations caused by sudden changes in the loss term. S700, consistency check, generate intermediate standard section load, predict velocity field and pressure field, form and output standard section dataset directly for wind turbine blade aerodynamic analysis and operation and maintenance assessment. S800 Load Prediction: Based on the trained PINN parameters, load prediction is performed on intermediate standard segments not included in the high-fidelity sample set. Consistency and inter-segment continuity checks are also performed. Data that pass the consistency and inter-segment continuity checks are encapsulated into standard segment results and output.

[0013] Optionally, in the PINN-based wind turbine blade load prediction method in the above embodiments, step S100 includes: S110. Blade geometric parameter acquisition: Acquire blade geometric parameters and convert them into a parametric expression suitable for PINN modeling. Define the local section coordinate system and the global rotating coordinate system of the wind turbine. S120. Establish the actual operating environment and boundary conditions. Establish the actual operating environment and boundary conditions of the wind turbine blades at the inlet boundary. Define the incoming flow velocity vector and the blade spanwise displacement.

[0014] Optionally, in the PINN-based wind turbine blade load prediction method in any of the above embodiments, the blade geometric parameters include the blade's three-dimensional geometric shape data and the blade's total length. , span of chord spanwise twist , span thickness , airfoil parameters at preset section locations, and typical section location parameters, among which, Represents the spanwise coordinates of the blade, with the domain being: , when Corresponding to the leaf root, Corresponding to the leaf tip; Indicates the blade in spanwise coordinates The span of the chord at that point; Indicates the blade in spanwise coordinates The spanwise twist angle at the location; Indicates the blade in spanwise coordinates The span of thickness at that location.

[0015] Furthermore, in the PINN-based wind turbine blade load prediction method in the above embodiments, the three-dimensional geometric shape data of the blade includes the three-dimensional coordinate point set of the blade suction surface and pressure surface, the spatial coordinates of the leading edge line and trailing edge line, the blade root flange mounting surface parameters, the blade tip profile data, and the blade reference axis data.

[0016] Furthermore, in the PINN-based wind turbine blade load prediction method in the above embodiments, the airfoil parameters at the preset cross-section position include the upper and lower surface coordinates of the blade, relative thickness, camber, maximum thickness position, leading edge radius, and trailing edge thickness.

[0017] Furthermore, in the PINN-based wind turbine blade load prediction method in the above embodiments, the preset cross-sectional positions include the blade root transition zone cross-section, the blade middle main aerodynamic working zone cross-section, and the blade tip transition zone cross-section.

[0018] Furthermore, in the PINN-based wind turbine blade load prediction method in the above embodiments, typical cross-sectional locations include the blade root, blade middle, and blade tip.

[0019] Optionally, in the PINN-based wind turbine blade load prediction method of any of the above embodiments, the parameterized expression is as follows: , in, Indicates the blade in spanwise coordinates The set of geometric parameters at that location.

[0020] Optionally, in the PINN-based wind turbine blade load prediction method of any of the above embodiments, the local section coordinate system is represented as follows: ,in, Represents local chord coordinates. Indicates the spanwise thickness normal coordinate. The local spanwise coordinates are represented by the global rotational coordinate system of the wind turbine. .

[0021] Optionally, in the PINN-based wind turbine blade load prediction method in any of the above embodiments, the incoming flow velocity vector It is expressed as follows: , in, The magnitude of the free flow velocity; This is the unit vector representing the direction of the incoming flow.

[0022] Furthermore, in the PINN-based wind turbine blade load prediction method in the above embodiments, the operating environment parameters include free-flow wind speed, inflow direction, air density, dynamic viscosity, impeller speed, pitch angle, and yaw angle.

[0023] Furthermore, in the PINN-based wind turbine blade load prediction method in the above embodiments, the boundary conditions include inlet velocity boundary, outlet pressure boundary, far-field boundary, and blade wall no-slip boundary condition.

[0024] Optionally, in the PINN-based wind turbine blade load prediction method in any of the above embodiments, the blade spanwise displacement Using modal superposition, it is represented as follows: , in, for First wave mode function; For the first Generalized coordinates of wave mode, Indicates time; For the wave mode order, , hour For the first-order waving mode function, hour It is a second-order waving mode function.

[0025] Optionally, in the PINN-based wind turbine blade load prediction method in any of the above embodiments, the first-order flapping mode function and the second-order flapping mode function can be obtained through finite element modal analysis of the blade structure, or through experimental mode identification or by calling existing blade structure design data.

[0026] Optionally, in the PINN-based wind turbine blade load prediction method in the above embodiments, step S200 includes: S210, Standard blade section division, dividing the blade spanwise interval Divided into Each standard segment meets the following conditions: , And it must satisfy the condition that adjacent standard segments are connected end to end, as shown in the following formula: , in, The number of standard segments is a positive integer. For the first The right boundary of the first standard segment; the first Each standard segment is denoted as This indicates that the entire blade is completely decomposed into continuous spanwise sub-intervals without overlap or omission; S220. Key segment identification: Classify the segmented standard segments into key segments and intermediate standard segments.

[0027] Furthermore, in the PINN-based wind turbine blade load prediction method in the above embodiments, the number of standard sections... Greater than or equal to 3.

[0028] Furthermore, in the PINN-based wind turbine blade load prediction method in the above embodiments, for conventional megawatt-class wind turbine blades, the number of standard sections... The range is 6 to 12.

[0029] Preferably, in the PINN-based wind turbine blade load prediction method in the above embodiments, the number of standard sections... It is 8 to 10.

[0030] Furthermore, in the PINN-based wind turbine blade load prediction method in the above embodiments, step S220 includes: S211. Criticality scoring: Define criticality scoring for the standard segment using the following formula: , in, For the first Keyness score for each standard segment; For the first The average overall sensitivity of each standard segment The average reference load for the standard section; For the first The average reference load level of each standard section; For the first Structural dynamic sensitivity index of each standard segment; They are respectively Weighting coefficients; S212, Calculation of generalized swing torque for standard segment: Calculate the first-order and second-order generalized swing torque for each standard segment; S213. Standard segment classification: Classify standard segments based on criticality scores. The standard segment for the critical segment threshold is the critical segment, and the intermediate segment threshold is... The standard segment for the critical segment threshold is the intermediate standard segment.

[0031] Furthermore, in the PINN-based wind turbine blade load prediction method described in the above embodiments, The range is [0.2, 0.3]. The range is [0.2, 0.35]. The range is [0.2, 0.3]. The range is [0.2, 0.35].

[0032] Preferably, in the PINN-based wind turbine blade load prediction method in the above embodiments, .

[0033] Furthermore, in the PINN-based wind turbine blade load prediction method in the above embodiments, for the first... Standard Section The formulas for calculating the first-order and second-order swing generalized moments are as follows: , , in, The first The first-order swing generalized torque of each standard segment and the... The second-order swing generalized torque of a standard segment These are the first-order waving mode function and the second-order waving mode function, respectively. for span coordinates The aerodynamic load per unit length in the flapping direction at a given point is given by the following formula: , in, for span coordinates The blade cross section at that location has the following chordal coordinates: The pressure difference between the upper and lower surfaces of the blade cross-section is calculated using the following formula: , in, This refers to the pressure on the upper surface of the blade cross section. This represents the pressure on the lower surface of the blade cross-section.

[0034] Furthermore, in the PINN-based wind turbine blade load prediction method in the above embodiments, the critical segment threshold is greater than or equal to 0.7 and less than 1.0.

[0035] Preferably, in the PINN-based wind turbine blade load prediction method in the above embodiments, the critical segment threshold is equal to 0.8.

[0036] Furthermore, in the PINN-based wind turbine blade load prediction method in the above embodiments, the threshold for the middle section is greater than or equal to 0.4 and less than 0.7.

[0037] Preferably, in the PINN-based wind turbine blade load prediction method in the above embodiments, the intermediate section threshold is equal to 0.55.

[0038] Optionally, in the PINN-based wind turbine blade load prediction method in any of the above embodiments, step S300 includes: S310, finite element mesh fitting, blade surface Generate a boundary-fitted finite element surface mesh that matches the 3D shape, and obtain the wall node set. With wall unit set The formulas are as follows: , , in, This represents the total number of wall nodes; For the first Spatial coordinates of each wall node This represents the total number of wall units. For the first Spatial coordinates of each wall node; S320, Finite Element Element Construction: Finite element elements are constructed along the normal direction on the outer side of the blade wall to form a near-wall subdomain. The near-wall subfield is defined as follows: , in, The near-wall layer thickness threshold. As a venue to the blade wall The shortest distance is calculated using the following formula: , in, Let be any point in the computational domain; Let be any point on the blade wall; Point With point The Euclidean distance; This indicates taking the infimum, i.e., the minimum distance; S330. Finite element shape function interpolation: Finite element shape functions are used for interpolation within near-wall elements, as shown in the following formula: , in, This is the wall distance function approximated by the finite element method; This represents the number of unit nodes. For the first The shape function corresponding to each node; For the first The distance values ​​from each node to the wall; S340. Construct PINN and define the hard-constraint velocity output. Construct PINN, which adopts a deep artificial neural network structure, including an input layer, several hidden layers, and an output layer. Define the hard-constraint velocity output of PINN as follows: , in, This refers to the output speed under hard constraints, i.e., the speed output after applying hard constraints. The boundary velocity of the blade relative to the reference coordinate system. For the original output velocity of PINN, in the relative rotating reference frame of the blades, the blade wall satisfies the no-slip boundary condition, that is, the wall velocity relative to the fluid is zero. ; S350, Hard Constraint Velocity Output Gradientization: To calculate the first and second spatial derivatives of the hard constraint velocity, the gradient of the hard constraint velocity is calculated using the following formula: , in, Represents the hard-constrained velocity gradient. This represents the boundary velocity gradient of the blade relative to the reference coordinate system. Indicates field point to the blade wall The shortest distance gradient, PINN's original output velocity gradient.

[0039] Furthermore, in the PINN-based wind turbine blade load prediction method described in the above embodiments, the total number of wall nodes... Scope .

[0040] Preferably, in the PINN-based wind turbine blade load prediction method in the above embodiments, the total number of wall nodes is... .

[0041] Furthermore, in the PINN-based wind turbine blade load prediction method in the above embodiments, the total number of wall elements... The range is .

[0042] Preferably, in the PINN-based wind turbine blade load prediction method in the above embodiments, the total number of wall elements is... .

[0043] Furthermore, in the PINN-based wind turbine blade load prediction method in the above embodiments, the near-wall thickness threshold... The value is the span chord length. of ; Preferably, in the PINN-based wind turbine blade load prediction method in the above embodiments, the near-wall thickness threshold is... The value is the span chord length. 6%.

[0044] Optionally, in the PINN-based wind turbine blade load prediction method in any of the above embodiments, the deep artificial neural network includes a multilayer feedforward fully connected neural network or a residual neural network.

[0045] Optionally, in the PINN-based wind turbine blade load prediction method in any of the above embodiments, the number of hidden layers in 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.

[0046] Furthermore, in the PINN-based wind turbine blade load prediction method in the above embodiments, the number of hidden layers in PINN is 3 to 12.

[0047] Preferably, in the PINN-based wind turbine blade load prediction method in the above embodiments, the number of hidden layers in PINN is 8.

[0048] Furthermore, in the PINN-based wind turbine blade load prediction method in the above embodiments, the number of neurons in each hidden layer is 16 to 256.

[0049] Furthermore, in the PINN-based wind turbine blade load prediction method in the above embodiments, the number of neurons in each hidden layer is 64.

[0050] Optionally, in the PINN-based wind turbine blade load prediction method in any of the above embodiments, step S400 includes: S410. High-fidelity sample selection: Select high-fidelity samples according to spatial selection principles to generate a high-fidelity sample set. High-fidelity samples are selected from key segments and supplemented from intermediate standard segments through PINN prediction. S420. Physical point selection: Select physical points according to the principle of layered layout to generate a set of physical points; S430. Low-fidelity sample selection: Select samples from areas outside the critical segment, intermediate standard segments, and regular outflow areas to generate a low-fidelity sample set.

[0051] Furthermore, in the PINN-based wind turbine blade load prediction method in the above embodiments, the spatial selection principle includes: 1. Sample points are extracted in the near-wall high gradient region, blade wall surface, and near-wall layer to reflect changes in velocity boundary layer and wall pressure. The near-wall high gradient region is defined as the area located within the near-wall layer where the pressure gradient exceeds a pressure threshold, or the velocity gradient exceeds a velocity threshold. The pressure threshold is selected from the top 70%–95% quantile of the statistical distribution of the pressure gradient, and the velocity threshold is selected from the top 70%–95% quantile of the statistical distribution of the velocity gradient. The near-wall layer is defined as the area where the shortest distance from the field point to the blade wall surface is not greater than the near-wall layer thickness threshold. ; 2. Sampling points are drawn in the leading edge region, where the local pressure gradient is large, making it sensitive to lift formation and separation initiation; 3. Sampling points were taken in the trailing edge region, which are related to wake detachment and pressure recovery; 4. Sample points are extracted from the load-sensitive area, which is the critical segment; 5. Sample points are extracted in the region near the blade tip to reflect the three-dimensional induced loss and wake vortex initiation characteristics. The region near the blade tip has normalized spanwise coordinates that satisfy... The area.

[0052] Furthermore, in the PINN-based wind turbine blade load prediction method described in the above embodiments, the high-fidelity sample set... It is expressed as follows: , in, This represents the total number of high-fidelity samples. For the first A high-fidelity sample location; For high-fidelity speed tags; For high-fidelity pressure labels.

[0053] Furthermore, in the PINN-based wind turbine blade load prediction method in the above embodiments, the layered arrangement principle includes: 1. In the near-wall region Extract physical collocation points to strengthen equation constraints near the boundary layer; 2. Aerodynamic gradient sensitive region Strengthen residual control at the leading edge, trailing edge, and separation-sensitive area; 3. Whealing Domain Strengthen the constraints on wake recovery and velocity loss propagation; 4. Conventional external watersheds Provides global background physical constraints; And the following conditions must be met: , in, This is the computational domain for the external flow field of the wind turbine blades. The union symbol represents the regions mentioned above that together constitute the computational domain of the external flow field of the wind turbine blades. Each region can intersect with the others but does not completely overlap.

[0054] Furthermore, in the PINN-based wind turbine blade load prediction method in the above embodiments, the physical point set... It is expressed as follows: , in, The total number of physical points. For the first A physical coordinate point.

[0055] Furthermore, in the PINN-based wind turbine blade load prediction method in the above embodiments, the total number of physical points... for .

[0056] Preferably, in the PINN-based wind turbine blade load prediction method in the above embodiments, the total number of physical points... .

[0057] Furthermore, in the PINN-based wind turbine blade load prediction method described in the above embodiments, low-fidelity data... It is expressed as follows: , in, This represents the total number of low-fidelity samples. For the first A low-fidelity sample location; Low-fidelity speed tag; This is a low-fidelity pressure label.

[0058] Furthermore, in the PINN-based wind turbine blade load prediction method described in the above embodiments, the total number of low-fidelity samples... for .

[0059] Preferably, in the PINN-based wind turbine blade load prediction method in the above embodiments, the total number of low-fidelity samples is... .

[0060] Optionally, in the PINN-based wind turbine blade load prediction method in any of the above embodiments, step S500 includes: S510. Establish the physical equation residual loss for the physical collocation set, as follows: , in, For the residual loss of the physical equation, For physical point indexing, For the first One physical coordinate point; For the first The momentum equation residual vector at each physical collocation point The square of the second norm of the momentum residual; For the first Continuous residual scalar at each physical collocation point; For the square of continuous residuals, The weights for continuous residuals are used to balance the importance of the momentum equation and incompressible constraints. This represents the total number of physical points. S520, Defining High-Fidelity Surveillance Loss The formula is as follows: , in, For the first The position coordinates of a high-fidelity sample; The total number of high-fidelity samples. For the first High-fidelity samples with high-speed labels; For the first A high-fidelity sample of high-fidelity pressure label; For the first The overall weight of each high-fidelity sample; For high-fidelity samples, the PINN prediction speed is improved. For the first PINN prediction pressure for a high-fidelity sample For the first Error weights for each high-fidelity sample; S530, Definition of Low-Fidelity Surveillance Loss The formula is as follows: , in, For the first The overall weight of low-fidelity samples, For low-fidelity pressure error weighting, For the first The location coordinates of a low-fidelity sample For the total number of low-fidelity samples, For the first Speed ​​labels for low-fidelity samples For the first Pressure labels of low-fidelity samples For the first PINN prediction speed for low-fidelity samples For the first PINN prediction pressure for a low-fidelity sample; S540, Construction moment loss, combined with the predicted value of the generalized swing moment of the standard section, construction moment loss. : , in, The number of standard segments is a positive integer. For the first The weights of the first-order swing generalized torque and the weights of the second-order swing generalized torque for each standard segment; For the first Predicted values ​​of the first-order swing generalized torque for the intermediate standard segment. For the first Predicted values ​​of the second-order swing generalized torque for the intermediate standard segment; S550, Construct the total loss function The formula is as follows: , in, These are the weighting coefficients for the physical equation residual loss, high-fidelity supervision loss, low-fidelity supervision loss, and torque loss, when no low-fidelity samples or boundary additions are introduced. .

[0061] Furthermore, in the PINN-based wind turbine blade load prediction method described in the above embodiments, the continuity residual weight... The range is .

[0062] Preferably, in the PINN-based wind turbine blade load prediction method in the above embodiments, the continuity residual weight... .

[0063] Furthermore, in the PINN-based wind turbine blade load prediction method described in the above embodiments, the total number of high-fidelity samples... The range is .

[0064] Preferably, in the PINN-based wind turbine blade load prediction method in the above embodiments, the total number of high-fidelity samples is... .

[0065] Furthermore, in the PINN-based wind turbine blade load prediction method in the above embodiments, the first... The overall weighting of a high-fidelity sample .

[0066] Preferably, in the PINN-based wind turbine blade load prediction method in the above embodiments, the first... The overall weighting of a high-fidelity sample .

[0067] Furthermore, in the PINN-based wind turbine blade load prediction method described in the above embodiments, the high-fidelity pressure error weight is set to... .

[0068] Preferably, in the PINN-based wind turbine blade load prediction method in the above embodiments, the high-fidelity pressure error weight is set to a value of... .

[0069] Furthermore, in the PINN-based wind turbine blade load prediction method in the above embodiments, the first... The overall weight value of each low-fidelity sample .

[0070] Preferably, in the PINN-based wind turbine blade load prediction method in the above embodiments, the first... The overall weight value of each low-fidelity sample .

[0071] Furthermore, in the PINN-based wind turbine blade load prediction method described in the above embodiments, the low-fidelity pressure error weight is set to... .

[0072] Preferably, in the PINN-based wind turbine blade load prediction method in the above embodiments, the low-fidelity pressure error weight is set to a specific value. .

[0073] Furthermore, in the PINN-based wind turbine blade load prediction method described in the above embodiments, the total number of low-fidelity samples... The value is .

[0074] Preferably, in the PINN-based wind turbine blade load prediction method in the above embodiments, the total number of low-fidelity samples is... .

[0075] Optionally, in the PINN-based wind turbine blade load prediction method in any of the above embodiments, step S600 includes: S610, First stage training, namely high-fidelity warm-up and guided training, setting the number of training rounds and initial parameters for the first stage; S620, the second stage of training, namely physical-dominated and low-fidelity extended training, completes PINN training.

[0076] Furthermore, in the PINN-based wind turbine blade load prediction method in the above embodiments, step S610 includes: S611. Set the number of training rounds for the first stage as follows: , in, This represents the total number of training rounds. This refers to the number of training rounds in the first phase. Indicates rounding up; S612, Initial parameter setting, initialize the PINN parameter to Call the output form that has been embedded with hard wall constraints to make PINN satisfy the blade wall conditions on the velocity boundary; S613, High-fidelity-led updates, using high-fidelity sample loss in each training round. As the main driving factor, and at the same time using Maintaining basic physical plausibility, using The formula for applying light traction to the integral load of the standard section is as follows: , in, , where is the number of training rounds; For the first The PINN parameters for each training round; for The PINN parameters are updated during each training round. For learning rate, The gradient of the first-stage loss with respect to the parameters; S614. Output the network parameters for the first stage of initial convergence. The PINN parameters, which are obtained through high-fidelity guided warm-up training, are expressed as follows: .

[0077] Furthermore, in the PINN-based wind turbine blade load prediction method described in the above embodiments, the learning rate is set to a value of... .

[0078] Furthermore, in the PINN-based wind turbine blade load prediction method in the above embodiments, step S620 includes: S621. Update the network parameters that have initially converged in the first stage, using the following formula: , in, The gradient of the second-stage loss with respect to the parameters; S622, Enable low-fidelity sample set Introducing trend supervision into low-fidelity samples can compensate for the insufficient coverage caused by the sparsity of high-fidelity samples. S623. Transformation of the available load model for the structure based on the moment loss function. This transforms PINN from a point-value flow field model into a structurally usable load model. S624. Output the PINN parameters trained in the second stage. The input model for generating and verifying the load consistency of the intermediate standard section is as follows: .

[0079] Furthermore, in the PINN-based wind turbine blade load prediction method in the above embodiments, the judgment condition for the available load model of the structure is as follows: For all standard segments to be output: , Adjacent segment interfaces satisfy: , in, Indicates the first The first-order swing generalized torque deviation of the standard segment This represents the first-order swing generalized torque deviation threshold. Indicates the first The deviation of the second-order generalized swing torque in the standard segment. This represents the second-order swing generalized torque deviation threshold. Indicates the first Load jump variables in a standard segment Indicates the load jump threshold. Indicates the first The jump variable of the load derivative of each standard segment This represents the threshold for the load derivative to jump.

[0080] Furthermore, in the PINN-based wind turbine blade load prediction method described in the above embodiments, the first-order flapping generalized torque deviation threshold is... For the first The absolute value of the first-order swing generalized torque in each standard segment is 1% to 20%.

[0081] Preferably, in the PINN-based wind turbine blade load prediction method in the above embodiments, the first-order flapping generalized torque deviation threshold is... For the first 10% of the absolute value of the first-order swing generalized torque of each standard segment.

[0082] Furthermore, in the PINN-based wind turbine blade load prediction method described in the above embodiments, the second-order flapping generalized moment deviation threshold... For the first The absolute value of the second-order swing generalized torque in each standard segment is 3% to 15%.

[0083] Preferably, in the PINN-based wind turbine blade load prediction method in the above embodiments, the second-order flapping generalized moment deviation threshold is... For the first The absolute value of the second-order swing generalized torque of the standard segment is 8%.

[0084] Furthermore, in the PINN-based wind turbine blade load prediction method described in the above embodiments, Load jump threshold For the first The absolute value of the load jump variable in each standard section is 2% to 8%.

[0085] Preferably, in the PINN-based wind turbine blade load prediction method in the above embodiments, Load jump threshold For the first The absolute value of the load jump variable in each standard segment is 5%.

[0086] Furthermore, in the PINN-based wind turbine blade load prediction method in the above embodiments, the threshold value representing the load derivative jump is... This corresponds to 8% to 15% of the absolute value of the derivative jump variable of the load.

[0087] Preferably, in the PINN-based wind turbine blade load prediction method in the above embodiments, the threshold value representing the load derivative jump is... This is 12% of the absolute value of the jump variable corresponding to the load derivative.

[0088] Furthermore, in the PINN-based wind turbine blade load prediction method in the above embodiments, a continuous weight scheduling function is used to make the training process smoother and reduce oscillations caused by sudden changes in loss terms.

[0089] Furthermore, in the PINN-based wind turbine blade load prediction method in the above embodiments, the formula for the continuous weight scheduling function is as follows: , , in, This is the current training round number; For high-fidelity supervision, the initial weights, The maximum target weight for low-fidelity supervision, Let be the scheduling rate coefficient, when hour, This indicates that it is still in the first phase; when hour, Exponential decay, Monotonous increase, This represents the number of training rounds in the first phase, obtained by rounding up one-third of the total number of training rounds.

[0090] Furthermore, in the PINN-based wind turbine blade load prediction method described in the above embodiments, the high-fidelity supervised initial weights... .

[0091] Preferably, in the PINN-based wind turbine blade load prediction method in the above embodiments, the high-fidelity supervised initial weights are used. .

[0092] Furthermore, in the PINN-based wind turbine blade load prediction method described in the above embodiments, the maximum target weight of low-fidelity supervision is... .

[0093] Preferably, in the PINN-based wind turbine blade load prediction method in the above embodiments, the maximum target weight of low-fidelity supervision is... .

[0094] Furthermore, in the PINN-based wind turbine blade load prediction method described in the above embodiments, the scheduling rate coefficient... .

[0095] Preferably, in the PINN-based wind turbine blade load prediction method in the above embodiments, the scheduling rate coefficient... .

[0096] Optionally, in the PINN-based wind turbine blade load prediction method in any of the above embodiments, step S700 includes: S710, intermediate standard section load generation: For intermediate standard sections not included in the high-fidelity sample set, call the trained PINN to predict the load and obtain the velocity field, pressure field, upper and lower surface pressure difference distribution and spanwise unit load distribution under the given working conditions. S720, predicting velocity and pressure fields for any intermediate standard segment. Under given operating conditions and spatial sampling points Next, the trained PINN is used to predict the velocity and pressure fields: , , in, For the first The predicted velocity field of the intermediate standard segment, No. Predicted pressure field for an intermediate standard segment, For the PINN speed output after embedding hard constraints on the wall, For PINN pressure output; S730. Calculate the pressure difference distribution for spanwise coordinates. At the blade cross section, let the pressure on the upper surface be... The pressure on the lower surface is The pressure differential distribution is then defined as: , in, chord position Pressure difference; ; ; S740. Calculate the unit spanwise load, relative to the spanwise coordinates. For the blade section at a given location, the unit spanwise load in the flapping direction is calculated using the following formula: , in, for span coordinates The unit at the location is waving the load in the direction of the wave; The projection coefficient of local pressure onto the waving direction; S750, Calculate the predicted value of the swinging generalized torque, for the first... For each intermediate standard segment, based on its spanwise load distribution, the predicted values ​​of the first-order and second-order generalized swing moments are calculated using the following formulas: , in, The first The predicted values ​​of the first-order and second-order swing generalized moments for the intermediate standard segment; S760, Calculate torque deviation, calculate the first... The torque deviation of the intermediate standard section is calculated using the following formula: , And determine whether it meets the threshold requirement: , in, The first The prediction deviation thresholds for the first-order swing generalized moment and the prediction deviation thresholds for the second-order swing generalized moment in the intermediate standard segment; S770, Inter-segment continuity check: Calculate the jump variables of unit spanwise load values ​​and derivative jump variables at the interface of adjacent standard segments, and determine whether they meet the prediction value deviation thresholds of the first-order swinging generalized moment and the second-order swinging generalized moment. If the threshold requirements are met, the intermediate standard segment is determined to have passed the check and is included in the standard segment result set; if the threshold requirements are not met, the intermediate standard segment is determined to have failed the check and is not directly included in the standard segment result set. S780. Generate a standard segment dataset. Unify and encapsulate the high-fidelity results of the key segments and the data of the intermediate segments generated by the PINN model and verified through inter-segment continuity to form and output a standard segment dataset that can be directly used for wind turbine blade aerodynamic analysis and operation and maintenance assessment.

[0097] Furthermore, in the PINN-based wind turbine blade load prediction method in the above embodiments, the projection coefficient of local pressure towards the flapping direction... Values .

[0098] Preferably, in the PINN-based wind turbine blade load prediction method in the above embodiments, the projection coefficient of local pressure towards the flapping direction is... .

[0099] Furthermore, in the PINN-based wind turbine blade load prediction method in the above embodiments, the first... The predicted deviation threshold of the first-order swing generalized torque in the intermediate standard segment The value is the first-order swing generalized torque. , No. The predicted deviation threshold of the second-order swing generalized torque in the intermediate standard segment The value is the second-order swing generalized torque. .

[0100] Preferably, in the PINN-based wind turbine blade load prediction method in the above embodiments, the first... The predicted deviation threshold of the first-order swing generalized torque in the intermediate standard segment The value is the first-order swing generalized torque. , No. The predicted deviation threshold of the second-order swing generalized torque in the intermediate standard segment The value is the second-order swing generalized torque. .

[0101] This application addresses the challenges of complex three-dimensional flow in wind turbine blades, insufficient wall boundary accuracy, and difficulty in directly constraining structural loads. It constructs a complete technical approach comprising standard segment division, hard boundary constraints, phased hybrid training, and torque consistency verification. This application directly embeds blade surface boundary conditions into the PINN output, combining high-fidelity data guidance, physical conservation constraints, and first / second-order flapping generalized torque constraints to achieve high-precision prediction of the blade velocity field, pressure field, and spanwise loads. The resulting standard segment data not only possesses good local flow field accuracy but also ensures the continuity and smoothness of the overall load distribution, making it directly applicable to wind turbine blade aerodynamic analysis and subsequent operation and maintenance.

[0102] 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

[0103] Figure 1 This is a flowchart illustrating a PINN-based wind turbine blade load prediction method according to an exemplary embodiment. Detailed Implementation

[0104] 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.

[0105] This invention provides a method for predicting wind turbine blade loads based on PINN, such as... Figure 1 As shown, it includes the following steps: S100, Blade Parameter Construction: This involves collecting blade geometric parameters and establishing the actual operating environment and boundary conditions of the wind turbine blades; specifically including: S110. Blade geometric parameter acquisition: Acquire blade geometric parameters, including three-dimensional geometric shape data of the blade and total blade length. , span of chord spanwise twist , span thickness The system includes: airfoil parameters at preset section locations; typical section location parameters; and three-dimensional geometric data of the blade, including the three-dimensional coordinate point set of the blade's suction and pressure surfaces, the spatial coordinates of the leading and trailing edges, blade root flange mounting surface parameters, blade tip profile data, and blade reference axis data. The airfoil parameters at preset section locations include the coordinates of the upper and lower surfaces, relative thickness, camber, maximum thickness location, leading edge radius, and trailing edge thickness. Preset section locations include the blade root transition zone section, the main aerodynamic working zone section in the middle of the blade, and the blade tip transition zone section. Typical section locations include the blade root, middle of the blade, and blade tip. Represents the spanwise coordinates of the blade, with the domain being: , when Corresponding to the leaf root, Corresponding to the leaf tip; Indicates the blade in spanwise coordinates The span of the chord at that point; Indicates the blade in spanwise coordinates The spanwise twist angle at the location; Indicates the blade in spanwise coordinates The spanwise thickness at the location is determined and transformed into a parametric expression suitable for PINN modeling. A local section coordinate system and a global rotating coordinate system for the wind turbine are defined. The parametric expression is as follows: , in, Indicates the blade in spanwise coordinates The set of geometric parameters at the location; the local section coordinate system is represented as ,in, Represents local chord coordinates. Indicates the spanwise thickness normal coordinate. The local spanwise coordinates are represented by the global rotational coordinate system of the wind turbine. ; S120. Establish the actual operating environment and boundary conditions. Establish the actual operating environment parameters and boundary conditions of the wind turbine blades at the inlet boundary. Define the incoming flow velocity vector and the blade spanwise displacement; Incoming flow velocity vector It is expressed as follows: , in, The magnitude of the free flow velocity; The incoming flow direction is a unit vector; operating environment parameters include free-flow wind speed, incoming flow direction, air density, dynamic viscosity, impeller speed, pitch angle, and yaw angle; boundary conditions include inlet velocity boundary, outlet pressure boundary, far-field boundary, and blade wall no-slip boundary condition; blade spanwise displacement. Using modal superposition, it is represented as follows: , in, for First wave mode function; For the first Generalized coordinates of wave mode, Indicates time; For the wave mode order, , hour For the first-order waving mode function, hour The second-order flapping mode function is obtained by calling existing blade structure design data.

[0106] S200, Blade Standard Segment Division: The blade spanwise region is divided into multiple standard segments to identify key segments affecting overall aerodynamic flow field accuracy, spanwise load continuity, and structural dynamic response; specifically including: S210, Standard blade section division, dividing the blade spanwise interval Divided into Each standard segment meets the following conditions: , And it must satisfy the condition that adjacent standard segments are connected end to end, as shown in the following formula: , in, The number of standard segments is a positive integer. ; For the first The right boundary of the first standard segment; the first Each standard segment is denoted as This indicates that the entire blade is completely decomposed into continuous spanwise sub-intervals without overlap or omission; S220. Key segment identification: Classifying the segmented standard segments into key segments and intermediate standard segments, specifically including: S211. Criticality scoring: Define criticality scoring for the standard segment using the following formula: , in, For the first Keyness score for each standard segment; For the first The average overall sensitivity of each standard segment The average reference load for the standard section; For the first The average reference load level of each standard section; For the first Structural dynamic sensitivity index of each standard segment They are respectively The weighting coefficients, ; S212, Calculation of the generalized swing moment for the standard segment: Calculate the first-order and second-order generalized swing moments for each standard segment; for the... Standard Section The formulas for calculating the first-order and second-order swing generalized moments are as follows: , , in, The first The first-order swing generalized torque of each standard segment and the... The second-order generalized swing torque of the standard segment, These are the first-order waving mode function and the second-order waving mode function, respectively. for span coordinates The aerodynamic load per unit length in the flapping direction at a given point is given by the following formula: , in, for span coordinates The blade cross section at that location has the following chordal coordinates: The pressure difference between the upper and lower surfaces of the blade cross-section is calculated using the following formula: , in, This refers to the pressure on the upper surface of the blade cross section. This represents the pressure on the lower surface of the blade cross-section.

[0107] S213. Standard segment classification: Classify standard segments based on criticality scores. The standard segment for the critical segment threshold is the critical segment, and the intermediate segment threshold is... The standard segment for the critical segment threshold is the middle standard segment. The critical segment threshold is 0.8, and the middle segment threshold is 0.55.

[0108] S300 and PINN modeling and construction of wall hard constraints, generation of boundary-fitted finite element surface meshes consistent with the 3D shape on the blade surface, construction of near-wall finite element elements, interpolation using finite element shape functions, construction of PINN and definition of hard constraint velocity output; including: S310, finite element mesh fitting, blade surface Generate a boundary-fitted finite element surface mesh that matches the 3D shape, and obtain the wall node set. With wall unit set The formulas are as follows: , , in, This represents the total number of wall nodes. ; For the first Spatial coordinates of each wall node This represents the total number of wall units. , For the first Spatial coordinates of each wall node; S320, Finite Element Element Construction: Finite element elements are constructed along the normal direction on the outer side of the blade wall to form a near-wall subdomain. The near-wall subfield is defined as follows: , in, The near-wall layer thickness threshold. The value is the span chord length. 6%, As a venue to the blade wall The shortest distance is calculated using the following formula: , in, Let be any point in the computational domain; Let be any point on the blade wall; Point With point The Euclidean distance; This indicates taking the infimum, i.e., the minimum distance; S330. Interpolation is performed using finite element shape functions. The finite element shape function is used for interpolation within the near-wall element, as shown in the following formula: , in, This is the wall distance function approximated by the finite element method; This represents the number of unit nodes. For the first The shape function corresponding to each node; For the first The distance values ​​from each node to the wall; S340. Construct PINN and define the hard-constraint velocity output. Construct PINN, which uses a multi-layer feedforward fully connected neural network, including an input layer, 8 hidden layers, and an output layer. Each hidden layer has 64 neurons. Define the hard-constraint velocity output of PINN as follows: , in, This refers to the output speed under hard constraints, i.e., the speed output after applying hard constraints. The boundary velocity of the blade relative to the reference coordinate system. For the original output velocity of PINN, in the relative rotating reference frame of the blades, the blade wall satisfies the no-slip boundary condition, that is, the wall velocity relative to the fluid is zero. ; S350, Hard Constraint Velocity Output Gradientization: To calculate the first and second spatial derivatives of the hard constraint velocity, the gradient of the hard constraint velocity is calculated using the following formula: , in, Represents the hard-constrained velocity gradient. This represents the boundary velocity gradient of the blade relative to the reference coordinate system. Indicates field point to the blade wall The shortest distance gradient, PINN's original output velocity gradient.

[0109] S400, Domain-Specific Sample Construction: High-fidelity samples, physical collocations, and low-fidelity samples are selected to construct high-fidelity sample sets, physical collocation sets, and low-fidelity sample sets, respectively; specifically including: S410. High-fidelity sample selection: High-fidelity samples are selected according to spatial selection principles to generate a high-fidelity sample set. High-fidelity samples are selected from key segments and supplemented from intermediate standard segments through PINN prediction. Spatial selection principles include: 1. Sample points are extracted in the near-wall high gradient region, blade wall surface, and near-wall layer to reflect changes in velocity boundary layer and wall pressure. The near-wall high gradient region is defined as the area located within the near-wall layer where the pressure gradient exceeds a pressure threshold, or the velocity gradient exceeds a velocity threshold. The pressure threshold is selected from the top 70%–95% quantile of the statistical distribution of the pressure gradient, and the velocity threshold is selected from the top 70%–95% quantile of the statistical distribution of the velocity gradient. The near-wall layer is defined as the area where the shortest distance from the field point to the blade wall surface is not greater than the near-wall layer thickness threshold. ; 2. Sampling points are drawn in the leading edge region, where the local pressure gradient is large, making it sensitive to lift formation and separation initiation; 3. Sampling points were taken in the trailing edge region, which are related to wake detachment and pressure recovery; 4. Sample points are extracted from the load-sensitive area, which is the critical segment; 5. Sample points are extracted in the region near the blade tip to reflect the three-dimensional induced loss and wake vortex initiation characteristics. The region near the blade tip has normalized spanwise coordinates that satisfy... The area; High-fidelity sample set It is expressed as follows: , in, The total number of high-fidelity samples. ; For the first A high-fidelity sample location; For high-fidelity speed tags; For high-fidelity pressure labels; S420. Physical point selection: Select physical points according to the hierarchical layout principle to generate a physical point set; the hierarchical layout principle includes: 1. In the near-wall region Extract physical collocation points to strengthen equation constraints near the boundary layer; 2. Aerodynamic gradient sensitive region Strengthen residual control at the leading edge, trailing edge, and separation-sensitive area; 3. Whealing Domain Strengthen the constraints on wake recovery and velocity loss propagation; 4. Conventional external watersheds Provides global background physical constraints; And the following conditions must be met: , in, This is the computational domain for the external flow field of the wind turbine blades. The union symbol represents the above regions, which together constitute the computational domain of the external flow field of the wind turbine blades. Each region can intersect with the others but does not completely overlap. Physics point set It is expressed as follows: , in, The total number of physical points. , For the first One physical coordinate point; S430. Low-fidelity sample selection: Low-fidelity sample sets are generated from areas outside the critical segment, intermediate standard segments, and regular outflow areas; low-fidelity data... It is expressed as follows: , in, For the total number of low-fidelity samples, ; For the first A low-fidelity sample location; Low-fidelity speed tag; This is a low-fidelity pressure label.

[0110] S500, Construction of the Total Loss Function: Establishing the physical equation residual loss, high-fidelity supervision loss, low-fidelity supervision loss, and torque loss, and constructing the total loss function; including: S510. Establish the physical equation residual loss for the physical collocation set, as follows: , in, For the residual loss of the physical equation, For physical point indexing, For the first One physical coordinate point; For the first The momentum equation residual vector at each physical collocation point The square of the second norm of the momentum residual; For the first Continuous residual scalar at each physical collocation point; For the square of continuous residuals, For continuous residual weights, The importance of using the momentum equation and incompressible constraints to balance the momentum equation; S520, Defining High-Fidelity Surveillance Loss The formula is as follows: , in, For the first The position coordinates of a high-fidelity sample; The total number of high-fidelity samples. , For the first High-fidelity samples with high-speed labels; For the first A high-fidelity sample of high-fidelity pressure label; For the first The overall weight of each high-fidelity sample. ; For high-fidelity samples, the PINN prediction speed is improved. For the first PINN prediction pressure for a high-fidelity sample For the first Error weights for each high-fidelity sample. ; S530, Definition of Low-Fidelity Surveillance Loss The formula is as follows: , in, For the first The overall weight of low-fidelity samples, , For low-fidelity pressure error weighting, , For the first The location coordinates of a low-fidelity sample Total number of low-fidelity samples , For the first Speed ​​labels for low-fidelity samples For the first Pressure labels of low-fidelity samples For the first PINN prediction speed for low-fidelity samples For the first PINN prediction pressure for a low-fidelity sample; S540, Construction moment loss, combined with the predicted value of the generalized swing moment of the standard section, construction moment loss. : , in, The number of standard segments is a positive integer. For the first The weights of the first-order swing generalized torque and the weights of the second-order swing generalized torque for each standard segment; For the first Predicted values ​​of the first-order swing generalized torque for the intermediate standard segment. For the first Predicted values ​​of the second-order swing generalized torque for the intermediate standard segment; S550, Construct the total loss function The formula is as follows: , in, For the weighting coefficients of the physical equation residual loss, high-fidelity supervision loss, low-fidelity supervision loss, and torque loss, when no low-fidelity samples or boundary additions are introduced, .

[0111] S600 employs a phased hybrid training approach, performing both the first and second phases of training. A continuous weight scheduling function ensures a smooth transition between the two phases, reducing oscillations caused by sudden changes in the loss term. The formula for the continuous weight scheduling function is as follows: , , in, This is the current training round number; For high-fidelity supervision, the initial weights, ; The maximum target weight for low-fidelity supervision, ; The scheduling rate coefficient, ;when hour, This indicates that it is still in the first phase; when hour, Exponential decay, Monotonous increase, This is the number of training rounds in the first phase, obtained by rounding up one-third of the total number of training rounds. Specifically, it includes: S610, Phase 1 Training, namely high-fidelity warm-up and guided training, sets the number of training rounds and initial parameters for Phase 1; specifically including: S611. Set the number of training rounds for the first stage as follows: , in, This represents the total number of training rounds. This refers to the number of training rounds in the first phase. Indicates rounding up; S612, Initial parameter setting, initialize the PINN parameter to Call the output form that has been embedded with hard wall constraints to make PINN satisfy the blade wall conditions on the velocity boundary; S613, High-fidelity-led updates, using high-fidelity sample loss in each training round. As the main driving factor, and at the same time using Maintaining basic physical plausibility, using The formula for applying light traction to the integral load of the standard section is as follows: , in, , where is the number of training rounds; For the first The PINN parameters for each training round; for The PINN parameters are updated during each training round. For learning rate, , The gradient of the first-stage loss with respect to the parameters; S614. Output the network parameters for the first stage of initial convergence. The PINN parameters, which are obtained through high-fidelity guided warm-up training, are expressed as follows: .

[0112] S620, the second phase of training, namely physics-driven and low-fidelity extended training, completes PINN training; specifically including: S621. Update the network parameters that have initially converged in the first stage, using the following formula: , in, The gradient of the second-stage loss with respect to the parameters; S622, Enable low-fidelity sample set Introducing trend supervision into low-fidelity samples can compensate for the insufficient coverage caused by the sparsity of high-fidelity samples. S623. Transformation of the available load model for the structure based on the moment loss function. This transforms PINN from a point-value flow field model into a structurally available load model. The criteria for determining the structurally available load model are as follows: For all standard segments to be output: , Adjacent segment interfaces satisfy: , in, Indicates the first The first-order swing generalized torque deviation of the standard segment This represents the first-order swing generalized torque deviation threshold. For the first 10% of the absolute value of the first-order swing generalized torque of each standard segment. Indicates the first The deviation of the second-order generalized swing torque in the standard segment. This represents the second-order swing generalized torque deviation threshold. For the first The absolute value of the second-order generalized torque of the standard segment is 8%. Indicates the first Load jump variables in a standard segment Indicates the load jump threshold. For the first 5% of the absolute value of the load jump variable in each standard segment Indicates the first The jump variable of the load derivative of each standard segment This represents the threshold for the load derivative to jump. This corresponds to 12% of the absolute value of the jump variable of the load derivative; S624. Output the PINN parameters trained in the second stage. The input model for generating and verifying the load consistency of the intermediate standard section is as follows: .

[0113] S700 consistency check, generating intermediate standard section loads, predicting velocity and pressure fields, generating standard section datasets, forming and outputting standard section datasets directly for wind turbine blade aerodynamic analysis and operation and maintenance assessment; specifically including: S710, intermediate standard section load generation: For intermediate standard sections not included in the high-fidelity sample set, call the trained PINN to predict the load and obtain the velocity field, pressure field, upper and lower surface pressure difference distribution and spanwise unit load distribution under the given working conditions. S720, predicting velocity and pressure fields for any intermediate standard segment. Under given operating conditions and spatial sampling points Next, the trained PINN is used to predict the velocity and pressure fields: , , in, For the first The predicted velocity field of the intermediate standard segment, No. Predicted pressure field for an intermediate standard segment, For the PINN speed output after embedding hard constraints on the wall, For PINN pressure output; S730. Calculate the pressure difference distribution for spanwise coordinates. At the blade cross section, let the pressure on the upper surface be... The pressure on the lower surface is The pressure differential distribution is then defined as: , in, chord position Pressure difference; ; ; S740. Calculate the unit spanwise load, relative to the spanwise coordinates. For the blade section at a given location, the unit spanwise load in the flapping direction is calculated using the following formula: , in, for span coordinates The unit at the location is waving the load in the direction of the wave; The projection coefficient of local pressure onto the waving direction. ; S750, Calculate the predicted value of the swinging generalized torque, for the first... For each intermediate standard segment, based on its spanwise load distribution, the predicted values ​​of the first-order and second-order generalized swing moments are calculated using the following formulas: , in, The first The predicted values ​​of the first-order and second-order swing generalized moments for the intermediate standard segment; S760, Calculate torque deviation, calculate the first... The torque deviation of the intermediate standard section is calculated using the following formula: , And determine whether it meets the threshold requirement: , in, The first The prediction deviation thresholds for the first-order and second-order generalized swing moments in the intermediate standard segment. The value is the first-order swing generalized torque. , The value is the second-order swing generalized torque. ; S770, Inter-segment continuity check: Calculate the jump variables of unit spanwise load values ​​and derivative jump variables at the interface of adjacent standard segments, and determine whether they meet the prediction value deviation thresholds of the first-order swinging generalized moment and the second-order swinging generalized moment. If the threshold requirements are met, the intermediate standard segment is determined to have passed the check and is included in the standard segment result set; if the threshold requirements are not met, the intermediate standard segment is determined to have failed the check and is not directly included in the standard segment result set. S780. Generate a standard segment dataset. Unify and encapsulate the high-fidelity results of the key segments and the data of the intermediate segments generated by the PINN model and verified through inter-segment continuity to form and output a standard segment dataset that can be directly used for wind turbine blade aerodynamic analysis and operation and maintenance assessment.

[0114] S800, S800, Load Prediction: Based on the trained PINN parameters, load prediction is performed on intermediate standard segments not included in the high-fidelity sample set. Consistency and inter-segment continuity checks are also performed. Data that pass the consistency and inter-segment continuity checks are encapsulated into standard segment results and output.

[0115] To verify the technical effectiveness of the above embodiments, under the same blade geometry parameters, the same operating conditions, and the same or comparable number of high-fidelity samples, the above embodiments were compared and analyzed with traditional blade element momentum methods and conventional PINN methods using soft boundary conditions. Traditional blade element momentum methods are mainly used to characterize the ability of existing simplified engineering models in predicting overall load trends, while conventional PINN methods are mainly used to characterize the ability of existing physical information neural networks in handling complex wall boundaries, structural load constraints, and standard segment continuity control. The above embodiments employ a technical approach combining standard segment division and key segment identification, boundary fitting finite element hard constraints, multi-source sample domain construction, phased hybrid training, and a combination of first-order / second-order flapping generalized moments and inter-segment continuous smooth constraints.

[0116] Comparative results show that, compared with traditional blade element momentum methods, the above embodiments can more fully reflect the spanwise geometric changes, local pressure distribution changes, and complex near-wall flow characteristics of wind turbine blades. They provide a more detailed description of the aerodynamic state in the blade root, mid-blade, and tip regions, and exhibit better adaptability and consistency in predicting spanwise load distribution and structurally relevant generalized moments. Especially in regions where blade three-dimensional effects are significant, local load gradients are large, and structural dynamic responses are sensitive, the method of this invention can more accurately reflect the influence of the local flow field on the overall structural load compared to traditional simplified models.

[0117] Furthermore, compared to the conventional PINN method using soft boundary conditions, the above embodiments, by directly embedding the blade wall boundary conditions into the network output, result in more stable predictions of the velocity and pressure fields on the blade surface and near the wall, significantly reducing error accumulation near the wall. Simultaneously, by introducing first-order and second-order flapping generalized moments constraints, as well as load continuity and smoothness constraints between adjacent standard segments, the above embodiments ensure that the network output not only maintains good consistency with the reference results at local point values ​​but also exhibits better usability in terms of structural load indices in the integral sense. Compared to the conventional PINN method, the intermediate standard segment results generated by the above embodiments show smaller load jumps and smoother derivative changes at the interface between adjacent segments.

[0118] In summary, the above embodiments demonstrate superior performance in terms of boundary processing accuracy, overall physical consistency, spanwise load continuity, and structural usability. They can reduce the cost of high-fidelity CFD calculations across the entire blade while balancing the accuracy of aerodynamic flow field modeling for wind turbine blades with engineering application requirements, thus providing more reliable technical support for wind turbine blade load analysis.

[0119] 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 predicting wind turbine blade loads based on PINN, characterized in that, Includes the following steps: S100, Blade parameter construction, collecting blade geometric parameters, and establishing the actual operating environment and boundary conditions of the wind turbine blade; S200, blade standard segment division, divides the blade spanwise region into multiple standard segments, and identifies key segments that affect the overall aerodynamic flow field accuracy, spanwise load continuity and structural dynamic response. S300 and PINN are modeled and hard constraints are constructed on the wall. A boundary fitting finite element surface mesh consistent with the three-dimensional shape is generated on the blade surface. Finite element elements in the near-wall region are constructed. Finite element shape functions are used for interpolation. PINN is constructed and hard constraint velocity output is defined. S400, constructing a domain-specific sample set: select high-fidelity samples, physical collocations, and low-fidelity samples to construct a high-fidelity sample set, a physical collocation set, and a low-fidelity sample set, respectively; S500, Construction of Total Loss Function: Establish physical equation residual loss, high-fidelity supervision loss, low-fidelity supervision loss, torque loss, and construct the total loss function; S600 uses a phased hybrid training method, performing first-stage and second-stage training. The continuous weight scheduling function ensures a smooth transition between the first-stage and second-stage training processes, reducing oscillations caused by sudden changes in the loss term. S700, consistency check, generate intermediate standard section load, predict velocity field and pressure field, form and output standard section dataset directly for wind turbine blade aerodynamic analysis and operation and maintenance assessment. S800 Load prediction: Based on the trained PINN parameters, load prediction is performed on intermediate standard segments not included in the high-fidelity sample set, and consistency and inter-segment continuity checks are performed. Data that pass the consistency and inter-segment continuity checks are encapsulated into standard segment results and output.

2. The wind turbine blade load prediction method based on PINN as described in claim 1, characterized in that, S100 includes: S110. Blade geometric parameter acquisition: Acquire blade geometric parameters and convert them into a parametric expression suitable for PINN modeling. Define the local section coordinate system and the global rotating coordinate system of the wind turbine. S120. Establish the actual operating environment and boundary conditions. Establish the actual operating environment and boundary conditions of the wind turbine blades at the inlet boundary. Define the incoming flow velocity vector and the blade spanwise displacement.

3. The wind turbine blade load prediction method based on PINN as described in claim 2, characterized in that, The blade spanwise displacement Using modal superposition, it is represented as follows: ， in, for First wave mode function; for Generalized coordinates of wave mode, Indicates time; For the wave mode order, , hour For the first-order waving mode function, hour It is a second-order waving mode function.

4. The wind turbine blade load prediction method based on PINN as described in claim 1, characterized in that, S200 includes: S210, Standard blade section division, dividing the blade spanwise interval Divided into Each standard segment meets the following conditions: ， And it must satisfy the condition that adjacent standard segments are connected end to end, as shown in the following formula: ， in, The number of standard segments is a positive integer. For the first The right boundary of the first standard segment; the first Each standard segment is denoted as This indicates that the entire blade is completely decomposed into continuous spanwise sub-intervals without overlap or omission; S220. Key segment identification: Classify the segmented standard segments into key segments and intermediate standard segments.

5. The wind turbine blade load prediction method based on PINN as described in claim 4, characterized in that, S220 includes: S211. Criticality Score: Define a criticality score for the standard segment using the following formula: ， in, For the first Keyness score for each standard segment; For the first The average overall sensitivity of each standard segment The average reference load for the standard section; For the first The average reference load level of each standard section; For the first Structural dynamic sensitivity index of each standard segment; They are respectively Weighting coefficients; S212, Calculation of generalized swing torque for standard segment: Calculate the first-order and second-order generalized swing torque for each standard segment; S213. Standard segment classification: Classify standard segments based on the aforementioned criticality scores. The standard segment for the critical segment threshold is the critical segment, and the intermediate segment threshold is... The standard segment for the critical segment threshold is the intermediate standard segment.

6. The wind turbine blade load prediction method based on PINN as described in claim 1, characterized in that, The S300 includes: S310, finite element mesh fitting, blade surface Generate a boundary-fitted finite element surface mesh that matches the 3D shape, and obtain the wall node set. With wall unit set The formulas are as follows: ， ， in, This represents the total number of wall nodes. For the first Spatial coordinates of each wall node This represents the total number of wall units. For the first Spatial coordinates of each wall node; S320, Finite Element Unit Construction: Finite element units are constructed along the normal direction on the outer side of the blade wall to form a near-wall subdomain. The near-wall subdomain is defined as follows: ， in, The near-wall layer thickness threshold. As a venue to the blade wall surface The shortest distance is calculated using the following formula: ， in, Let be any point in the computational domain; Let be any point on the blade wall; Point With point The Euclidean distance; This indicates taking the infimum, i.e., the minimum distance; S330. Interpolation is performed using finite element shape functions. The interpolation within the near-wall element is calculated using the following formula: ， in, This is the wall distance function approximated by the finite element method; This represents the number of unit nodes. For the first The shape function corresponding to each node; For the first The distance values ​​from each node to the wall; S340. Construct PINN and define the hard-constraint velocity output. Construct PINN, which adopts a deep artificial neural network structure, including an input layer, several hidden layers, and an output layer. Define the hard-constraint velocity output of PINN as follows: ， in, The hard-constrained speed output is the speed output after applying the hard constraint. The boundary velocity of the blade relative to the reference coordinate system. For the original output velocity of PINN, in the global rotating coordinate system, the blade wall satisfies the no-slip boundary condition, that is, the wall velocity relative to the fluid is zero. ; S350, Hard Constraint Velocity Output Gradientization: To calculate the first and second spatial derivatives of the hard constraint velocity, the gradient of the hard constraint velocity is calculated using the following formula: ， in, Represents the hard-constrained velocity gradient. This represents the boundary velocity gradient of the blade relative to the reference coordinate system. Indicates field point to the blade wall The shortest distance gradient, PINN's original output velocity gradient.

7. The wind turbine blade load prediction method based on PINN as described in claim 1, characterized in that, The S400 includes: S410. High-fidelity sample selection: Select high-fidelity samples according to spatial selection principles to generate a high-fidelity sample set. The high-fidelity samples are selected from the key segments and supplemented from the intermediate standard segments through PINN prediction. S420. Physical point selection: Select physical points according to the principle of layered layout to generate a set of physical points; S430. Low-fidelity sample selection: Select samples from areas other than the key segment, the intermediate standard segment, and the regular outflow area to generate a low-fidelity sample set.

8. The wind turbine blade load prediction method based on PINN as described in claim 1, characterized in that, The S500 includes: S510. Establish the physical equation residual loss for the physical collocation set, as follows: ， in, For the residual loss of the physical equation, For physical point indexing, For the first One physical coordinate point; For the first The momentum equation residual vector at each physical collocation point The square of the second norm of the momentum residual; For the first Continuous residual scalar at each physical collocation point; For the square of continuous residuals, The continuous residual weight is used to balance the importance of the momentum equation and the incompressible constraint; This represents the total number of physical points. S520, Defining High-Fidelity Surveillance Loss The formula is as follows: ， in, For the first The position coordinates of a high-fidelity sample; The total number of high-fidelity samples. For the first High-fidelity samples with high-speed labels; For the first A high-fidelity sample of a high-fidelity pressure label; For the first The overall weight of each high-fidelity sample; For high-fidelity samples, the PINN prediction speed is improved. For the first PINN prediction pressure for a high-fidelity sample For the first Error weights for each high-fidelity sample; S530, Definition of Low-Fidelity Surveillance Loss The formula is as follows: ， in, For the first The overall weight of low-fidelity samples, For low-fidelity pressure error weighting, For the first The location coordinates of a low-fidelity sample Total number of low-fidelity samples For the first Speed ​​labels for low-fidelity samples For the first Pressure labels of low-fidelity samples For the first PINN prediction speed for low-fidelity samples For the first PINN prediction pressure for a low-fidelity sample; S540, Construction moment loss, combined with the predicted value of the generalized swing moment of the standard section, construction moment loss. : ， in, The number of standard segments is a positive integer. For the first The weights of the first-order swing generalized torque and the weights of the second-order swing generalized torque for each standard segment; For the first Predicted values ​​of the first-order swing generalized torque for the intermediate standard segment. For the first Predicted values ​​of the second-order swing generalized torque for the intermediate standard segment; S550, Construct the total loss function The formula is as follows: ， in, The weighting coefficients are those for the physical equation residual loss, the high-fidelity monitoring loss, the low-fidelity monitoring loss, and the torque loss.

9. The wind turbine blade load prediction method based on PINN as described in claim 1, characterized in that, The S600 includes: S610, First stage training, namely high-fidelity warm-up and guided training, setting the number of training rounds and initial parameters for the first stage; S620, the second stage of training, namely physical-dominated and low-fidelity extended training, completes PINN training.

10. The wind turbine blade load prediction method based on PINN as described in claim 1, characterized in that, The S700 includes: S710. Load generation of intermediate standard segments: For intermediate standard segments not included in the high-fidelity sample set, the trained PINN is called to predict the load, and the velocity field, pressure field, pressure difference distribution of upper and lower surfaces and spanwise unit load distribution are obtained under the given working conditions. S720, predicting velocity and pressure fields for any intermediate standard segment. Under given working conditions and spatial sampling points Next, the trained PINN is used to predict the velocity and pressure fields: ， ， in, For the first The predicted velocity field of the intermediate standard segment, No. Predicted pressure field for an intermediate standard segment, For the PINN speed output after embedding hard constraints on the wall, For PINN pressure output; S730. Calculate the pressure difference distribution for spanwise coordinates. At the blade cross section, let the pressure on the upper surface be... The pressure on the lower surface is The pressure differential distribution is defined as: ， in, chord position Pressure difference; ; ; S740. Calculate the unit spanwise load, relative to the spanwise coordinates. For the blade section at a given location, the unit spanwise load in the flapping direction is calculated using the following formula: ， in, For the spanwise coordinates The unit at the location is waving the load in the direction of the wave; The projection coefficient of local pressure onto the waving direction; S750, Calculate the predicted value of the swinging generalized torque, for the first... For the intermediate standard section, based on its spanwise load distribution, the predicted values ​​of the first-order and second-order flapping generalized moments are calculated using the following formulas: ， in, The first The predicted values ​​of the first-order and second-order swing generalized moments for the intermediate standard segment; These are the first-order waving mode function and the second-order waving mode function, respectively; S760, Calculate torque deviation, calculate the first... The torque deviation of the intermediate standard section is calculated using the following formula: ， And determine whether it meets the threshold requirement: ， in, The first The prediction deviation thresholds for the first-order swing generalized moment and the prediction deviation thresholds for the second-order swing generalized moment in the intermediate standard segment; S770. Inter-segment continuity check: Calculate the jump variables of unit spanwise load values ​​and derivative jump variables at the interface of adjacent standard segments, and determine whether they meet the prediction value deviation thresholds of the first-order swinging generalized moment and the second-order swinging generalized moment. If the threshold requirements are met, the intermediate standard segment is determined to have passed the check and is included in the standard segment result set; if the threshold requirements are not met, the intermediate standard segment is determined to have failed the check and is not directly included in the standard segment result set. S780. Generate a standard segment dataset. Unify and encapsulate the high-fidelity results of the key segments and the data of the intermediate segments generated by the PINN model and verified through inter-segment continuity to form and output a standard segment dataset that can be directly used for wind turbine blade aerodynamic analysis and operation and maintenance assessment.

Images