A method for predicting blade structural stress driven by the fusion of multi-source data and reduced-order models

Abstract

This invention discloses a blade structure stress prediction method driven by the fusion of multi-source data and a reduced-order model. Through a reduced-order surrogate model combining unidirectional fluid-structure interaction (FSI) simulation of the blade, it achieves rapid prediction of the blade structure stress field under different wind speeds, rotational speeds, and pitch angles. The method includes: constructing a unidirectional FSI simulation model of the wind turbine blade; extracting reduced-order modes and coefficients from high-fidelity blade stress field data to reduce the order of the full-order simulation results; constructing and training a surrogate model based on the reduced-order modes and coefficients, and obtaining the mapping relationship between different wind speeds and reduced-order mode coefficients by optimizing the model hyperparameters; the surrogate model rapidly predicts the modal coefficients of the blade stress field based on the real-time wind speed of the wind turbine, and reconstructs the blade stress field results with the reduced-order modes. This invention can rapidly predict the blade stress field under real-time wind speed conditions based on a small number of high-fidelity simulation samples, which is significant for the intelligent operation and maintenance of offshore wind turbines.

Description

Technical Field

[0001] This invention belongs to the field of wind turbine monitoring and digital twin technology, specifically involving a method for predicting blade structural stress driven by the fusion of multi-source data and reduced-order models. Background Technology

[0002] As offshore wind power develops towards deeper waters and larger scales, the structural health monitoring and assessment of wind turbine blades, as key load-bearing components, is crucial for ensuring the safe operation of wind turbine units and reducing maintenance costs. In actual operation, offshore wind turbine blades are subjected to complex and variable wind loads over long periods, making them highly susceptible to stress concentration and fatigue damage, leading to stress defects and even structural failure. Traditional sensor-based monitoring methods struggle to capture the full stress distribution across the blade, and sensor placement is easily affected by harsh sea conditions. While high-fidelity numerical simulation methods for stress analysis provide high accuracy, their computational costs are extremely high, failing to meet real-time monitoring requirements.

[0003] To overcome the computational bottleneck of high-fidelity numerical simulation, existing technologies often employ surrogate models as a substitute for numerical simulation. However, the stress distribution of wind turbine blades under wind loads exhibits significant high-dimensionality and physical complexity: on the one hand, the blade geometry is highly distorted and features thin walls and a large aspect ratio, resulting in high-fidelity finite element models typically containing a large number of mesh nodes; on the other hand, the stress field of blades under wind loads exhibits strong spatial non-uniformity, such as stress concentration in the blade root region and complex stress gradients at the leading and trailing edges. Data-driven surrogate models often face the "curse of dimensionality" when dealing with high-dimensional, strongly nonlinear problems, requiring massive amounts of training samples to ensure accuracy, and have poor generalization ability to operating conditions outside the training data range, making it difficult to accurately capture stress details under complex loads.

[0004] Therefore, in the field of offshore wind turbine blade condition monitoring, there is an urgent need for a technical solution that can efficiently utilize limited high-fidelity simulation data, predict the stress field of the entire blade in real time and with high accuracy, and identify stress defects accordingly, so as to support intelligent operation and maintenance decisions for wind turbine blades. Summary of the Invention

[0005] To address the shortcomings of existing technologies and achieve the goal of improving the real-time prediction accuracy of the global stress field of offshore wind turbine blades while reducing computational costs, this invention adopts the following technical solution:

[0006] A method for predicting blade structural stress driven by the fusion of multi-source data and a reduced-order model includes the following steps:

[0007] A three-dimensional model of a wind turbine blade is constructed. Through unidirectional fluid-structure interaction simulation of the blade, stress field data of different sub-regions of the blade under different wind speeds are extracted. The snapshot matrix assembled from these data is then processed by a block intrinsic orthogonal decomposition algorithm to reduce the order of the stress field data. The reduced-order modes and corresponding modal coefficients of the stress field data of different sub-regions are obtained, forming a reduced-order basis for reconstructing the original stress field.

[0008] A proxy model is constructed that maps different wind speeds to modal coefficients. Using the first k reduced-order modes and their corresponding modal coefficients, the blade stress field data under the corresponding wind speed is reconstructed and trained. Based on the real-time wind speed obtained by the anemometer installed in the actual wind turbine nacelle, the trained proxy model is used to predict the blade stress field data in real time.

[0009] Furthermore, keeping the control strategy unchanged, unidirectional fluid-structure interaction simulations of wind turbine blades at different wind speeds and corresponding rotational speeds and pitch angles are established. The reduced-order modes and corresponding coefficients of the stress field data in different sub-regions of the wind turbine blades at different wind speeds are extracted using a block-based intrinsic orthogonal decomposition algorithm, forming a reduced-order basis for reconstructing the original stress field. The specific steps include the following:

[0010] Based on the three-dimensional model of the wind turbine blade, a CFD (Computational Fluid Dynamics) model of the flow field and a FEM (Finite Element Method) model of the blade are constructed. The wind load is calculated through the CFD model of the flow field and transferred unidirectionally to the FEM model of the blade, forming a unidirectional fluid-structure interaction simulation.

[0011] Numerical simulations were performed on wind turbine blades under multiple wind speeds. Based on the spanwise geometric characteristics and stress distribution characteristics of the blades, the physical domain of the blades was divided into multiple sub-regions. The numerical simulation results of the sub-regions were assembled into a matrix and then decentered to obtain a snapshot matrix. The row vectors represent different spatial coordinate dimensions of the space, and the column vectors represent the changes of data in the same spatial dimension relative to a certain time. The covariance matrix of the snapshot matrix was calculated.

[0012] By orthogonally diagonalizing the covariance matrix, the reduced-order modes of the stress field in the blade sub-region are obtained. The orthogonal diagonalization formula is as follows:

[0013]

[0014] Where, Φ (d) Each column is C (d) One of the feature vectors, and Λ (d) The element on the diagonal is C. (d) The eigenvalues ​​are sorted from the largest eigenvalue to the smallest eigenvalue to obtain n. d 1 eigenvalue, and arranged in n d ×nd n in matrix Φ d The set of eigenvectors represents the reduced-order modes of the stress field in the d-th sub-region of the blade.

[0015] Based on the snapshot matrix and the reduced-order modes, the modal coefficients corresponding to the reduced-order modes of the stress field in the blade sub-region are calculated using the following formula:

[0016]

[0017]

[0018] Among them, each a ij For the projection of data obtained on time i and spatial mode j, A (d) The columns are the modal coefficients of the reduced-order modes;

[0019] Based on modal coefficients, the snapshot matrix of the blade sub-region is represented as the sum of contributions from multiple modes at multiple time points. By selecting multiple sub-matrices as principal components and superimposing them, the snapshot matrix is ​​reconstructed, thereby reducing the order of the data. The formula for the snapshot matrix is ​​as follows:

[0020]

[0021] By selecting the first k (d) By superimposing the principal components of each submatrix, the original snapshot matrix U of the leaf subregion can be reconstructed. (d) :

[0022] .

[0023] Furthermore, combining the Latin hypercube sampling method, numerical simulations were performed on wind turbine blades under multiple wind speeds to obtain stress field results based on the spatial coordinates of the blades at multiple wind speeds. According to the blade's spanwise geometric characteristics and stress distribution characteristics, the blade's physical domain was divided into multiple sub-regions. The numerical simulation results of the sub-regions were assembled into a matrix, and after centering, a snapshot matrix was obtained. The covariance matrix of the snapshot matrix was calculated using the following formula:

[0024]

[0025]

[0026] in, This is the high-fidelity stress field result of the numerical simulation of the d-th sub-region after centering, where x represents the spatial coordinates on the blade, v represents the wind speed, and U... d Represents a snapshot matrix, C d Let represent the covariance matrix.

[0027] Furthermore, the flow field CFD model adopts the SST k-ω turbulence model, dividing the fluid domain into a rotating domain and a discrete domain. The rotating domain encloses the blade body, and the rotational speed of the rotating domain is matched with the actual operating speed of the wind turbine to simulate the blade rotation condition. The stationary domain covers the flow field region surrounding the rotating domain, and boundary conditions such as velocity inlet, pressure outlet, periodic boundary, and no-slip wall are set. The blade FEM model sets remote displacement constraints at the blade root, and the loading conditions include the wind load transmitted by the flow field CFD model and the centrifugal load generated by the blade rotation. The structural response data such as blade stress field and deformation are obtained by solving.

[0028] Furthermore, select the first k (d) The d-th sub-matrix inevitably introduces truncation error into the result of the d-th sub-region. Therefore, by selecting a specific order, the relative truncation error is calculated to determine whether the truncation error meets the set order range. If it does not meet the range, the order is incremented by one. The truncation error formula is as follows:

[0029]

[0030] Where I(p) represents the relative cutoff error when the order is p, and the order range is [1, min(m,n)]. The order is chosen to be the minimum number of modes that makes I(p) ≤ 0.1% or 0.01%.

[0031] Furthermore, the high-fidelity blade stress field was obtained by using the Fluent and Static Structural modules in ANSYS software, employing the finite element method and finite volume method.

[0032] Furthermore, a mapping from different wind speeds to modal coefficients is constructed using a surrogate model. The blade stress field data at the corresponding wind speed is reconstructed using reduced-order modes and corresponding coefficients. Based on real-time wind speed data measured by an anemometer installed in the wind turbine nacelle, the blade stress field data is predicted in real time, including the following steps:

[0033] Train a surrogate model to construct a mapping from different wind speeds to corresponding modal coefficients; mapping methods include, but are not limited to: deep neural networks, kriging, radial basis function methods, etc.

[0034]

[0035] Where f represents the mapping relationship, W and B represent the neuron weights and biases of the neural network, and A is the modal coefficient vector formed by concatenating all local modal coefficients.

[0036] During the training phase of the surrogate model, a singular value-weighted adaptive loss is constructed. For the physical sub-region of the blade, local singular values ​​of all truncated modes are extracted and locally normalized within each sub-region to obtain local penalty weights, which are used to guide training. The formula for the local penalty weights is as follows:

[0037]

[0038] The weights of all sub-regions of the blade are then concatenated to form an improved weighted loss; the loss function formula is as follows:

[0039]

[0040] Where N is the number of samples in the batch; and These are the actual and predicted values ​​of the modal coefficients, respectively.

[0041] By combining the reduced-order modes and the modal coefficients obtained by mapping, the blade stress field data is reconstructed, and the evaluation index is quantified based on the surrogate model. The hyperparameters of the surrogate model are optimized based on the grid search algorithm.

[0042] Real-time wind speed data is obtained by using an anemometer installed in the wind turbine nacelle, and the measured data is corrected using the nacelle transfer function to obtain the true free-flow wind speed.

[0043]

[0044] Where U represents the true free-flow wind speed, U nacelle This indicates the real-time wind speed data measured by the anemometer;

[0045] The corrected wind speed is input into the trained surrogate model to quickly predict the blade stress field data at the corresponding wind speed.

[0046] Furthermore, the surrogate model's quantitative evaluation metrics include, but are not limited to, relative error (RE) and coefficient of determination (R²). 2 Root mean square error (RMSE) and normalized maximum absolute error (NMAE) are given. The formulas are as follows:

[0047]

[0048]

[0049]

[0050]

[0051] Furthermore, multiple sets of model hyperparameter combinations are obtained through the grid search algorithm. The average value of the evaluation index of each parameter combination model on the validation set is calculated and standardized to calculate the weighted score S, and the optimal hyperparameters of the model are selected accordingly.

[0052]

[0053]

[0054] A blade structure stress prediction method driven by the fusion of multi-source data and reduced-order model is applied to offshore wind turbines. The method is used to predict the stress field data of offshore wind turbine blade structure and perform real-time analysis based on the predicted data.

[0055] The advantages and beneficial effects of this invention are as follows:

[0056] This invention achieves rapid prediction of the global stress field of blades under different wind speeds by constructing a reduced-order surrogate model based on reduced-order modes and coefficients extracted from high-fidelity fluid-structure interaction simulation. This overcomes the limitations of traditional high-fidelity fluid-structure interaction simulations, which are computationally time-consuming and cannot meet real-time requirements, providing support for online monitoring of blade stress state and early defect warning. The surrogate model constructed in this invention directly learns the low-dimensional mapping of "wind speed-reduced-order mode coefficients," requiring only a limited number of high-fidelity simulation samples to complete high-precision training, significantly reducing modeling costs and improving the extrapolation capability of data-driven models, ensuring the reliability of predicting blade stress fields under complex operating conditions. Attached Figure Description

[0057] Figure 1 This is a flowchart of a method according to an embodiment of the present invention.

[0058] Figure 2 This is a schematic diagram of the CFD model of the NREL 5MW wind turbine blade in an embodiment of the present invention.

[0059] Figure 3 This is a schematic diagram of the intrinsic orthogonal decomposition order reduction algorithm in an embodiment of the present invention.

[0060] Figure 4 This is a diagram showing the blade stress field predicted by the POD-MLP reduced-order surrogate model at the first wind speed in this embodiment of the invention.

[0061] Figure 5 This is a simulation calculation result diagram of the POD-MLP reduced-order surrogate model prediction under the first wind speed in this embodiment of the invention.

[0062] Figure 6 This is a relative error cloud map of the POD-MLP reduced-order surrogate model prediction at the first wind speed in this embodiment of the invention.

[0063] Figure 7 This is a diagram showing the blade stress field predicted by the POD-MLP reduced-order surrogate model at the second wind speed in this embodiment of the invention.

[0064] Figure 8 This is a simulation calculation result diagram of the POD-MLP reduced-order surrogate model prediction under the second wind speed in this embodiment of the invention.

[0065] Figure 9 This is a relative error cloud map of the POD-MLP reduced-order surrogate model prediction at the second wind speed in this embodiment of the invention. Detailed Implementation

[0066] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0067] This invention addresses the shortcomings of existing offshore wind turbine blade stress monitoring technologies, such as high computational cost, poor real-time performance, and low local accuracy. It proposes a blade structure stress prediction method driven by the fusion of multi-source data and a reduced-order model. Taking the NREL 5MW wind turbine blade as an example, this invention first constructs a 3D model of the wind turbine blade using SOLIDWORKS, and then builds CFD and FEM models using ANSYS. The CFD model calculates wind loads and transmits them to the FEM model to achieve unidirectional fluid-structure interaction simulation. Multiple wind speed samples are obtained using Latin hypercube sampling. The unidirectional fluid-structure interaction model obtains the blade stress field (stress values ​​of all grid nodes of the blade) under multiple wind speeds (blade rotational speed and pitch angle are obtained according to the optimal tip speed ratio control strategy in the NREL report). After order reduction using block intrinsic orthogonal decomposition (block POD), this data is used as training data and input into the surrogate model to construct the mapping between wind speed and blade stress field. Finally, combined with real-time wind speed data collected by an anemometer, rapid prediction of the blade stress field under different wind speeds is achieved. Figure 1 As shown, the method of the present invention specifically includes the following steps:

[0068] Step S1: Using the optimal power generation efficiency control strategy in the NREL report, establish a unidirectional fluid-structure interaction simulation of the NREL 5MW wind turbine blade at different wind speeds and corresponding rotational speeds and pitch angles. Extract the reduced-order modes and corresponding coefficients of the stress field data of different sub-regions of the wind turbine blade at different wind speeds through the block intrinsic orthogonal decomposition algorithm to form a reduced-order basis for reconstructing the original stress field.

[0069] Step S1.1: Construct a 3D model of the NREL 5MW wind turbine blade using SOLIDWORKS, and import it into ANSYS to construct a flow field CFD model and a blade FEM model. The CFD model is as follows: Figure 2 As shown, wind loads are calculated using a CFD model and unidirectionally transferred to the FEM model, achieving unidirectional fluid-structure interaction simulation. The CFD model uses the finite volume method for discretizing the computational domain, employing the SSTk-ω turbulence model. The fluid domain is divided into a rotating domain and a discrete domain. The rotating domain encompasses the blade body, and the rotational speed is matched to the actual operating speed of the wind turbine to simulate blade rotation. The stationary domain covers the flow field region surrounding the rotating domain, with boundary conditions including velocity inlet, pressure outlet, periodic boundaries, and no-slip walls. The FEM model uses the finite element method for discretizing the computational domain, setting constraints as blade root remote displacement constraints. Loading conditions include wind loads transferred from the CFD model and centrifugal loads generated by blade rotation. The structural response data, such as blade stress field and deformation, are obtained by solving for these parameters.

[0070] Step S1.2: Using the Latin hypercube sampling method, numerical simulations are performed on 30 sets of wind turbine blades under different wind speeds to obtain 30 sets of blade stress fields u. Based on the spanwise geometric characteristics and stress distribution characteristics of the blade, the physical domain of the blade is divided into D sub-regions. In this embodiment, D=3 is selected, and the blade is divided into three sub-regions: blade root, blade middle, and blade tip. The numerical simulation results of the d-th sub-region are assembled into a matrix and centered to obtain the snapshot matrix U. d Calculate U d The covariance matrix C d ;

[0071]

[0072]

[0073] in, The result of high-fidelity numerical simulation of the d-th sub-region after centering is given, where x represents the spatial coordinates on the blade and v represents the wind speed.

[0074] Step S1.3: Calculate the covariance matrix C using the following formula. (d) Perform orthogonal diagonalization;

[0075]

[0076] Where Φ (d) Each column is C (d) One of the feature vectors, and Λ (d) The element on the diagonal is C. (d) The eigenvalues ​​of n are then sorted from the largest to the smallest eigenvalue to obtain n. d 1 eigenvalue, and arranged in n d ×n d n in matrix Φ dThe set of eigenvectors represents the reduced-order modes of the stress field in the d-th sub-region of the blade.

[0077] Step S1.4: Calculate the modal coefficient A corresponding to the reduced-order mode of the stress field in the d-th sub-region of the blade using the following formula. (d) ;

[0078]

[0079]

[0080] A (d) The columns are the modal coefficients of the reduced-order modes.

[0081] Step S1.5: The original snapshot matrix U of the d-th sub-region (d) It can be represented as n d The sum of contributions of each mode at 30 wind speeds;

[0082]

[0083] By selecting the first k (d) By superimposing the principal components of each submatrix, the original snapshot matrix U of the d-th subregion can be reconstructed. (d) To reduce the order of the data, the algorithm flow is as follows: Figure 3 As shown, row vectors represent space in different dimensions (such as spatial coordinates), while column vectors represent the changes of data in the same dimension relative to a certain variable (such as time).

[0084] Step S1.6: Select the first k (d) The d-th sub-matrix will inevitably introduce truncation error to the result of the d-th sub-region. The following formula is used to determine whether the truncation error satisfies the condition, with the order ranging from [1, min(m,n)]. d If the condition is not met, the order is incremented by one. Usually, the order is chosen to be the minimum number of modes that makes I(p) ≤ 0.1% or 0.01%.

[0085]

[0086] Where I(p) represents the relative truncation error when the order is p.

[0087] Step S2: Construct a mapping from different wind speeds to modal coefficients using a surrogate model, reconstruct the blade stress field data at the corresponding wind speed using reduced-order modes and corresponding coefficients, and predict the blade stress field data in real time based on the real-time wind speed data measured by the anemometer installed in the wind turbine nacelle.

[0088] Step S2.1: Train a multilayer perceptron (MLP) surrogate model according to the following formula to construct the modal coefficient mapping of the blade under different wind speeds;

[0089]

[0090] Where f represents the mapping relationship, W and B represent the neuron weights and biases of the neural network, and A is the modal coefficient vector formed by concatenating all local modal coefficients.

[0091] An adaptive loss function based on singular value weighting was used during the model training phase. For the d-th physical sub-region, the local singular values ​​of all its truncated modes were extracted, and local normalization was performed within each sub-region to obtain local penalty weights, which were used to guide network training.

[0092]

[0093] Subsequently, the weights of all sub-regions are concatenated into an improved weighted loss function;

[0094]

[0095] Where N is the number of samples in the batch; and These are the actual and predicted values ​​of the modal coefficients, respectively.

[0096] Step S2.2: Combine the reduced-order modes with the modal coefficients obtained by mapping to reconstruct the blade stress field data. Then, optimize the surrogate model hyperparameters based on the surrogate model quantification evaluation index and the grid search algorithm. The surrogate model quantification index used includes relative error RE and coefficient of determination R. 2 Global root mean square error (RMSE), standardized maximum absolute error (NMAE), etc.

[0097]

[0098]

[0099]

[0100]

[0101] in, The result is obtained from the prediction of the reduced-order model. This refers to the original numerical solution or the result obtained from numerical simulation.

[0102] Multiple hyperparameter combinations of models are obtained through a grid search algorithm. The average evaluation index of each parameter combination on the validation set is calculated and standardized. The weighted score S is calculated and the optimal hyperparameters of the model are selected accordingly.

[0103]

[0104]

[0105] Step S2.3: Taking real free-flow wind speeds of 8.80 m / s and 22.08 m / s as examples, the real-time wind speed is input to the surrogate model. The surrogate model outputs the corresponding modal coefficients, and the blade stress field data is quickly reconstructed from the reduced-order modes, such as... Figures 4 to 9 As shown, the first wind speed v = 8.80 m / s, corresponding to a rotational speed n = 10.03 rpm, a pitch angle β = 0º, and a relative error of 0.14%; the second wind speed v = 22.08 m / s, corresponding to a rotational speed n = 12.10 rpm, a pitch angle β = 20.08°, and a relative error of 0.27%.

[0106] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting blade structural stress driven by the fusion of multi-source data and a reduced-order model, characterized in that: A three-dimensional model of a wind turbine blade is constructed. Through unidirectional fluid-structure interaction simulation of the blade, stress field data of different sub-regions of the blade based on the physical domain are extracted under different wind speeds and the order is reduced to obtain the reduced-order modes and corresponding modal coefficients of the stress field data of different sub-regions of the blade. Among them, the relative truncation error is calculated by selecting a specific order to determine whether the truncation error meets the set order range. A proxy model mapping different wind speeds to modal coefficients is constructed. Using the reduced-order modes and corresponding modal coefficients, the blade stress field data under the corresponding wind speed is reconstructed and trained. For the physical sub-region of the blade, the local singular values ​​of all truncated modes are extracted and local normalization is performed within each sub-region to obtain local penalty weights. The weights of all sub-regions of the blade are then concatenated to form an improved weighted loss to guide the prediction training of modal coefficients. Based on real-time wind speed, the trained proxy model is used to predict the blade stress field data in real time. Based on the three-dimensional model of the wind turbine blade, a flow field model and a blade model are constructed. The wind load is calculated through the flow field model and transmitted unidirectionally to the blade model, forming a unidirectional fluid-structure interaction simulation. Numerical simulations were performed on wind turbine blades under multiple wind speeds. The physical domain of the blade was divided into multiple sub-regions based on the blade spanwise geometric characteristics and stress distribution characteristics using a block-based intrinsic orthogonal decomposition algorithm. The numerical simulation results of the sub-regions were assembled into a matrix and centered to obtain a snapshot matrix. The covariance matrix of the snapshot matrix was then calculated. By orthogonally diagonalizing the covariance matrix, the reduced-order modes of the stress field in the blade sub-region are obtained; Based on the snapshot matrix and the reduced-order modes, calculate the modal coefficients corresponding to the reduced-order modes of the stress field in the blade sub-region; Based on the modal coefficients, the snapshot matrix of the blade sub-region is represented as the sum of the contributions of multiple modes at multiple time points. The snapshot matrix is ​​reconstructed by superimposing multiple sub-matrices as principal components.

2. The blade structure stress prediction method driven by the fusion of multi-source data and reduced-order model according to claim 1, characterized in that: By combining the Latin hypercube sampling method, numerical simulations were performed on wind turbine blades under multiple wind speeds to obtain stress field results based on the spatial coordinates of the blades at multiple wind speeds. The stress field results were then divided into physical sub-regions of the blades based on stress distribution characteristics. The numerical simulation results of the blade sub-regions were centered and assembled into a snapshot matrix, and the covariance matrix of the snapshot matrix was calculated.

3. The blade structure stress prediction method driven by the fusion of multi-source data and reduced-order model according to claim 1, characterized in that: The flow field model adopts a turbulence model, which divides the fluid domain into a rotating domain and a discrete domain. The rotating domain surrounds the blade body. By setting the rotational speed of the rotating domain to match the actual operating speed of the wind turbine, the blade rotation condition is simulated. The stationary domain covers the flow field area outside the rotating domain, and boundary conditions are set. The blade model sets remote displacement constraints at the blade root. The loading conditions include the wind load transmitted by the flow field model and the centrifugal load generated by the blade rotation. The blade structural response data is obtained by solving.

4. The blade structure stress prediction method driven by the fusion of multi-source data and reduced-order model according to claim 1, characterized in that: The relative truncation error is calculated by selecting a specific order to determine whether the truncation error meets the set order range. If it does not meet the range, the order is incremented by one.

5. The blade structure stress prediction method driven by the fusion of multi-source data and reduced-order model according to claim 1, characterized in that: The stress field of the blade was obtained by analyzing the finite element method and the finite volume method.

6. The method for predicting blade structural stress driven by the fusion of multi-source data and reduced-order model according to claim 1, characterized in that: Train the surrogate model to construct a mapping from different wind speeds to corresponding modal coefficients; During the training phase, a singular value-weighted adaptive loss function is constructed. For the current physical sub-region of the blade, local singular values ​​of all truncated modes are extracted and local normalization is performed within each sub-region to obtain local penalty weights. The difference between the predicted and true values ​​of the modal coefficients is combined to guide the training. The weights of all sub-regions of the blade are concatenated into an improved weighted loss function. By minimizing the weighted loss function, the modal coefficients of the blade mapped to each sub-region under different wind speeds are trained. By combining the reduced-order modes and the modal coefficients obtained by mapping, the blade stress field data is reconstructed, and the evaluation index is quantified based on the surrogate model. The hyperparameters of the surrogate model are optimized based on the grid search algorithm. The measured data is corrected using the nacelle transfer function by real-time wind speed data to obtain the true free-flow wind speed. The corrected wind speed is input into the trained surrogate model to predict the blade stress field data at the corresponding wind speed.

7. The blade structure stress prediction method driven by the fusion of multi-source data and reduced-order model according to claim 6, characterized in that: The quantitative evaluation indicators of the proxy model include, but are not limited to, relative error (RE) and coefficient of determination (R). 2 Global root mean square error (RMSE) and standardized maximum absolute error (NMAE) are also mentioned.

8. The method for predicting blade structural stress driven by the fusion of multi-source data and a reduced-order model according to claim 6, characterized in that: The grid search algorithm is used to obtain multiple sets of model hyperparameter combinations. The average value of the evaluation index of each parameter combination model on the validation set is calculated and standardized to calculate the weighted score S, and the optimal hyperparameters of the model are selected accordingly.

9. A method for predicting blade structural stress driven by the fusion of multi-source data and reduced-order models, applied to offshore wind turbines, characterized by: The blade structure stress prediction method driven by the fusion of multi-source data and reduced-order model as described in any one of claims 1 to 8 is used to predict the stress field data of offshore wind turbine blade structure and perform real-time analysis based on the predicted data.

Images