A heliostat feature extraction method, a heliostat aerodynamic simulation method and equipment

Abstract

The application discloses a heliostat feature extraction method, a heliostat aerodynamic simulation method and equipment, and belongs to the technical field of heliostat simulation. The existing heliostat point cloud processing method lacks feature mining means, cannot extract the geometric features of the heliostat, and cannot be used for the aerodynamic simulation of the heliostat. The heliostat feature extraction method of the application is characterized in that: a heliostat pretreatment model, a feature mining model and a feature fusion model are constructed, the heliostat point cloud scanning data is sampled to obtain heliostat key points; then the heliostat key points are subjected to feature mining to obtain multi-dimensional geometric features to capture the local features and global features of the heliostat; the multi-dimensional geometric features are subjected to nonlinear transformation to obtain structural feature quantities, which are coupled with environmental feature quantities and heliostat elevation angles to generate a fusion feature vector, so that the heliostat features can be accurately mined and precisely extracted, and then the aerodynamic simulation of the heliostat can be carried out.

Description

TECHNICAL FIELD

[0001] The application relates to a heliostat feature extraction method, a heliostat aerodynamic simulation method and equipment, and belongs to the technical field of heliostat simulation. BACKGROUND

[0002] As the core equipment of a tower type photo-thermal power station, the key feature extraction and accurate prediction of the aerodynamic coefficient of a heliostat play a key role in the stable operation and efficient power generation of the power station.

[0003] Further, a Chinese patent application (publication number: CN119756227A) discloses a heliostat back surface type detection method, which comprises the following steps: S1, collecting position information of a plurality of points on the back surface of a to-be-detected heliostat to form back surface point cloud data; the position information comprises a collection distance, a pitch angle and an azimuth angle; S2, sequentially performing coordinate conversion and rotation translation on the back surface point cloud data to obtain rotated point cloud data; the normal vector of the rotated point cloud data is parallel to the coordinate axis, and the center point of the rotated point cloud data coincides with the coordinate axis origin; S3, fitting the back surface type of the to-be-detected heliostat according to the rotated point cloud data to obtain a back surface fitting type, and determining a back surface type detection result of the to-be-detected heliostat according to the back surface fitting type.

[0004] The above-mentioned scheme provides a heliostat back surface point cloud processing method, but the scheme is mainly used for fitting the back surface structure of a heliostat, and therefore lacks feature mining means when processing point cloud data, cannot extract the geometric features of the heliostat, and cannot be used for aerodynamic simulation of the heliostat.

[0005] The information disclosed in the background section merely serves to enhance the understanding of the background of the present inventive concept, and therefore it can include information that does not constitute the prior art. SUMMARY

[0006] In view of the above problems or one of the above problems, the purpose of the present application is to provide a heliostat feature extraction method, a heliostat aerodynamic simulation method and equipment, which can accurately mine and accurately extract the features of a heliostat by constructing a heliostat preprocessing model, a feature mining model and a feature fusion model, sampling heliostat point cloud scanning data to obtain heliostat key points, mining the features of the heliostat key points to obtain multi-dimensional geometric features to capture the local features and global features of the heliostat, performing nonlinear transformation on the multi-dimensional geometric features to obtain structure feature quantities, coupling the structure feature quantities with environmental feature quantities and a heliostat elevation angle to generate a fusion feature vector, and thus the features of the heliostat can be accurately mined and accurately extracted, and the heliostat can be used for aerodynamic simulation, and the scheme is scientific, reasonable and feasible.

[0007] To solve the above problems or one of the above problems, the second object of the present application is to provide a heliostat aerodynamic simulation method, which fully considers the mutual influence and interaction between the multi-modal variables of the heliostat, can effectively model the multi-physical field interaction of the heliostat, can accurately reflect the aerodynamic characteristics of the heliostat under complex environment, so that the aerodynamic coefficient prediction result is close to the actual situation; and through the simulation calculation of the algorithm model, the modeling cost can be effectively reduced, the simulation time is shorter, the different design schemes of the heliostat can be quickly evaluated and optimized, the real-time requirement is met, the scheme is scientific, reasonable and feasible.

[0008] To achieve one of the above objects, the first technical solution of the present application is:

[0009] A heliostat feature extraction method, comprising the following steps:

[0010] Step one, obtaining heliostat point cloud scanning data;

[0011] Step two, using a pre-constructed heliostat preprocessing model, based on curvature weight, sampling the heliostat point cloud scanning data to obtain heliostat key points;

[0012] Step three, using a pre-constructed feature mining model, performing feature processing on the heliostat key points to obtain multi-dimensional geometric features to capture local features and global features of the heliostat;

[0013] Step four, using a pre-constructed feature fusion model, performing nonlinear transformation on the multi-dimensional geometric features to obtain structural feature quantities, and coupling the structural feature quantities with the environmental feature quantities of the heliostat and the elevation angle of the heliostat to generate a fusion feature vector, realizing feature extraction of the heliostat.

[0014] The present application constructs a heliostat preprocessing model, a feature mining model and a feature fusion model, samples the heliostat point cloud scanning data to obtain heliostat key points, then mines the features of the heliostat key points to obtain multi-dimensional geometric features to capture local features and global features of the heliostat, then performs nonlinear transformation on the multi-dimensional geometric features to obtain structural feature quantities, and couples the structural feature quantities with the environmental feature quantities and the elevation angle of the heliostat to generate a fusion feature vector, so that the features of the heliostat can be accurately mined and precisely extracted, which can be used for aerodynamic simulation of the heliostat, and the scheme is scientific, reasonable and feasible.

[0015] Further, the application can significantly reduce the point cloud data volume while effectively preserving key geometric structure information, thereby greatly improving data processing efficiency; and by using curvature weight and nonlinear transformation mechanism, the application can fully capture the aerodynamic response characteristics of the heliostat surface; meanwhile, through a multi-modal feature fusion strategy, the coupling and expression ability between different modalities are effectively enhanced, making the extracted features more discriminative and representative, so that more efficient and more accurate data processing and feature expression ability can be achieved.

[0016] As a preferred technical measure:

[0017] Step two, based on the curvature weight, the heliostat point cloud scanning data is sampled by using the pre-constructed heliostat preprocessing model to obtain the method of heliostat key points as follows:

[0018] According to the heliostat point cloud scanning data, a point cloud set is established, which includes a plurality of discrete points, and a selected point set is constructed;

[0019] Based on the plurality of discrete points, the maximum distance and the maximum curvature value between the discrete points are calculated;

[0020] The minimum distance of a discrete point to all discrete points in the selected point set is calculated;

[0021] Based on the covariance matrix, the curvature weight of the discrete point is determined;

[0022] According to the curvature weight, the minimum distance, the maximum distance and the maximum curvature value of the discrete point, the sampling value of the discrete point is calculated;

[0023] Based on the sampling values of the plurality of discrete points, a sampling distribution is established, and the maximum value and the discrete point corresponding to the maximum value are selected from the sampling distribution;

[0024] The discrete point corresponding to the maximum value is taken as the heliostat key point to represent the structure of the high curvature area of the heliostat.

[0025] As a preferred technical measure:

[0026] Based on the covariance matrix, the curvature weight of the discrete point is determined as follows:

[0027] According to the discrete point, the points in the neighborhood and the neighborhood point centroid are obtained;

[0028] According to the points in the neighborhood and the neighborhood point centroid, a covariance matrix of the points in the neighborhood of the discrete point is established;

[0029] The covariance matrix is subjected to eigenvalue decomposition to obtain three eigenvalues;

[0030] According to the three eigenvalues, the curvature is calculated;

[0031] The curvature is taken as the curvature weight of a certain discrete point.

[0032] As a preferred technical measure:

[0033] Step three, using a pre-constructed feature mining model, the key points of the heliostat are processed to obtain multi-dimensional geometric features as follows:

[0034] Calculate the covariance matrix of the points in the neighborhood of a certain heliostat key point;

[0035] The eigenvalue decomposition of the covariance matrix obtains the curvature;

[0036] The eigenvector corresponding to the smallest eigenvalue of the covariance matrix is taken as the normal feature;

[0037] The curvature and the normal feature are taken as the geometric feature, and the feature vector of a certain heliostat key point is constructed based on the geometric feature;

[0038] The feature vectors of all heliostat key points are combined to obtain a geometric feature matrix;

[0039] The geometric feature matrix is sampled, grouped and feature extracted to obtain multi-dimensional geometric features.

[0040] As a preferred technical measure:

[0041] The method for sampling, grouping and feature extracting the geometric feature matrix to obtain multi-dimensional geometric features is as follows:

[0042] The geometric feature matrix is processed according to a fixed time period or a sliding window sampling method to obtain several representative features;

[0043] According to each representative feature, a local coordinate system is constructed, and based on the local feature coordinate system, a plurality of representative features are found in the vicinity;

[0044] Based on the plurality of representative features, a plurality of neighborhood feature sets are established to capture the local structural features of the heliostat surface;

[0045] Using the plurality of neighborhood feature sets, the corresponding heliostat key points are determined;

[0046] According to the heliostat key point, the representative features in the neighborhood feature set are spliced with the original geometric features to form an enhanced feature representation, which is used to retain the basic geometric information and fuse the local context semantics;

[0047] Based on the enhanced feature representation, a field feature area is constructed;

[0048] The field feature area is subjected to a pooling operation to extract the most salient feature representing the field feature area;

[0049] The plurality of most significant features are aggregated to obtain a multi-dimensional geometric feature to describe the global feature.

[0050] As a preferred technical measure:

[0051] Step four, using a pre-constructed feature fusion model, the multi-dimensional geometric feature is nonlinearly transformed to obtain a structure feature quantity as follows:

[0052] Obtain a multi-dimensional geometric feature, and the data dimension is ;

[0053] Using a nonlinear activation function and a normalization operation, the multi-dimensional geometric feature is nonlinearly transformed to obtain a first simplified feature vector of dimension, which is used to mine potential information hidden in the multi-dimensional geometric feature;

[0054] Using a nonlinear activation function and a normalization operation, the first simplified feature vector is nonlinearly transformed to obtain a second simplified feature vector of dimension;

[0055] Using a nonlinear activation function and a normalization operation, the second simplified feature vector is nonlinearly transformed to obtain a third simplified feature vector of dimension, so as to form a structure feature quantity that can retain discriminability and be fused.

[0056] Further, The value of d is 312 or 256 or 128.

[0057] As a preferred technical measure:

[0058] The structure feature quantity is coupled with an environment feature quantity of the heliostat and a heliostat elevation angle to generate a fusion feature vector, and the method for extracting the feature of the heliostat is as follows

[0059] Obtain an environment feature quantity, which includes a 3-dimensional wind speed feature and a 1-dimensional air density feature;

[0060] Construct a multi-layer perception network for the wind speed feature, and nonlinearly transform the 3-dimensional wind speed feature to obtain a wind speed feature vector of dimension;

[0061] Construct a multi-layer perception network for the air density feature, and nonlinearly transform the 1-dimensional air density feature to obtain an air density feature vector of dimension;

[0062] Obtain a heliostat elevation angle, and encode the heliostat elevation angle to form a 2-dimensional elevation angle feature vector for representing a spatial posture of the heliostat;

[0063] A multi-layer perceptron network is constructed for the heliostat elevation feature, a 2-dimensional elevation feature vector is nonlinearly transformed, and the multi-layer perceptron network is guided to extract discriminative representations from the periodicity and directionality of the heliostat elevation, so as to obtain a heliostat elevation feature vector in 2 dimensions;

[0064] Then, the structural feature quantity, the wind speed feature vector, the air density feature vector and the heliostat elevation feature vector are spliced in rows to obtain a fusion feature vector.

[0065] As a preferred technical measure:

[0066] The method for obtaining the air density feature is as follows:

[0067] Air density data of the environment around the heliostat is collected;

[0068] The air density data is sorted to obtain a minimum density value and a maximum density value;

[0069] Based on the minimum density value and the maximum density value, the air density data is normalized to obtain an air density feature with a uniform order of magnitude;

[0070] Or / and, the method for obtaining the heliostat elevation feature is as follows:

[0071] First, based on the points of the heliostat point cloud scanning data, a plane equation is fitted to determine a heliostat plane equation and a reference plane equation;

[0072] Then, based on the heliostat plane equation and the reference plane equation, a normal vector one of the heliostat plane equation and a normal vector two of the reference plane equation are calculated;

[0073] According to the vector angle formula, the angle between the normal vector one and the normal vector two is calculated;

[0074] Based on the vector angle, the elevation of the heliostat is determined;

[0075] The elevation of the heliostat is converted into a numerical feature by using a trigonometric function coding to obtain a 2-dimensional heliostat elevation feature.

[0076] To achieve one of the above purposes, a second technical solution of the present application is:

[0077] A heliostat aerodynamic simulation method applies the above-mentioned heliostat feature extraction method to construct an aerodynamic coefficient prediction model, which includes the following steps:

[0078] Step one, processing the fusion feature vector of the heliostat to generate a plurality of modal features;

[0079] Step two, mining the interaction between the modal features to obtain modal coupling data;

[0080] Step three, introduce graph convolution operation, represent the modal coupling data as a graph structure, and perform message propagation on the graph structure to represent the dependency relationship between modal features, and form a graph embedding representation information;

[0081] Step four, aggregate and map the graph embedding representation information to obtain the predicted heliostat aerodynamic coefficient, and realize the heliostat aerodynamic simulation.

[0082] The present application fully considers the mutual influence and interaction between the multi-modal variables of the heliostat, can effectively model the multi-physical field interaction of the heliostat, can accurately reflect the aerodynamic characteristics of the heliostat under complex environment, so that the aerodynamic coefficient prediction result is close to the actual situation; and through the simulation calculation of the algorithm model, the modeling cost can be effectively reduced, the simulation time is shorter, the different design schemes of the heliostat can be quickly evaluated and optimized, the real-time requirement is met, the scheme is scientific, reasonable and feasible.

[0083] Further, while reducing the calculation cost, the adaptability to complex aerodynamic conditions and the prediction accuracy are significantly improved, and the present application has good engineering application value and popularization prospect.

[0084] To achieve one of the above purposes, the third technical scheme of the present application is:

[0085] A heliostat aerodynamic simulation device comprises:

[0086] One or more processing units;

[0087] Storage device for storing one or more programs;

[0088] When the one or more programs are executed by the one or more processing units, the one or more processing units implement the above-mentioned heliostat feature extraction method.

[0089] Compared with the prior art, the present application has the following beneficial effects:

[0090] The present application can accurately mine and extract the features of the heliostat by constructing a heliostat preprocessing model, a feature mining model and a feature fusion model, sampling the heliostat point cloud scanning data to obtain heliostat key points, mining the features of the heliostat key points to obtain multi-dimensional geometric features to capture the local features and global features of the heliostat, and coupling the structure feature quantity obtained by nonlinear transformation of the multi-dimensional geometric features with the environment feature quantity and the heliostat elevation angle to generate a fusion feature vector, so that the features of the heliostat can be accurately mined and extracted, and the aerodynamic simulation of the heliostat can be used, and the scheme is scientific, reasonable and feasible.

[0091] Meanwhile, the application can effectively model the multi-physical field interaction of the heliostat by fully considering the interaction and interaction between the multi-modal variables of the heliostat, can accurately reflect the aerodynamic characteristics of the heliostat in a complex environment, so that the aerodynamic coefficient prediction result is close to the actual situation; and through simulation calculation by the algorithm model, the modeling cost can be effectively reduced, the simulation time is shorter, the heliostats of different design schemes can be quickly evaluated and optimized, the real-time requirement is met, the scheme is scientific, reasonable and feasible.

[0092] Further, while effectively preserving key geometric structure information, the application can significantly reduce the amount of point cloud data, thereby greatly improving data processing efficiency; and by using curvature weight and nonlinear transformation mechanism, the aerodynamic response characteristics of the heliostat surface can be fully captured; at the same time, through the multi-modal feature fusion strategy, the coupling and expression ability between different modalities are effectively enhanced, so that the extracted features are more discriminative and representative, and therefore more efficient and accurate data processing and feature expression ability can be achieved.

[0093] Further, while reducing the calculation cost, the application significantly improves the adaptability and prediction accuracy of complex aerodynamic conditions, and has good engineering application value and promotion prospect. BRIEF DESCRIPTION OF DRAWINGS

[0094] Figure 1 Fig. 1 is a first flowchart of the heliostat feature extraction method of the application;

[0095] Figure 2 Fig. 2 is a second flowchart of the heliostat feature extraction method of the application;

[0096] Figure 3 Fig. 3 is a structure diagram of the deep cross-attention network of the application. DETAILED DESCRIPTION

[0097] In order to enable the personnel in the technical field to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by the personnel in the field without creative labor should belong to the scope of protection of the present application. The present application covers any alternative, modification, equivalent method and scheme made within the essence and scope of the present application defined by the claims.

[0098] As shown in Figure 1 Fig. 1, the first specific embodiment of the heliostat feature extraction method of the application is as follows:

[0099] A heliostat feature extraction method, comprising the following steps:

[0100] Step one, obtain heliostat point cloud scanning data;

[0101] Step two, use the pre-constructed heliostat preprocessing model to sample the heliostat point cloud scanning data based on the curvature weight, and obtain the heliostat key points;

[0102] Step three, use the pre-constructed feature mining model to process the features of the heliostat key points, and obtain multi-dimensional geometric features to capture the local features and global features of the heliostat;

[0103] Step four, use the pre-constructed feature fusion model to perform nonlinear transformation on the multi-dimensional geometric features to obtain structural feature quantities, and couple the structural feature quantities with the environmental feature quantities of the heliostat and the elevation angle of the heliostat to generate a fusion feature vector, thereby realizing feature extraction of the heliostat.

[0104] As shown in Figure 2 , a second specific embodiment of the heliostat feature extraction method of the application is as follows:

[0105] A heliostat feature extraction method, comprising the following steps:

[0106] First, obtain multi-modal features and pre-process them, which includes the following contents:

[0107] Step 11, use a laser scanning device to perform omnidirectional scanning on the heliostat to obtain detailed heliostat point cloud data.

[0108] Step 12, the traditional farthest point sampling (FPS) only considers the distance between points, and the embodiment introduces curvature weight optimization sampling distribution. Let the point cloud set be , the set of selected points be S, and the selected point be , then the improved FPS sampling formula is:

[0109]

[0110] wherein represents the selected point, P is the point cloud set, S is the set of selected points, represents the minimum distance from the point to all points in the set S, and are the maximum distance and the maximum curvature value in the current stage point cloud, is the curvature weight, and the optimal balance is obtained when is the index used to find the maximum value in a given function or array.

[0111] In this embodiment, the curvature weight K( The calculation method of the curvature weight K is as follows:

[0112] First, calculate the point Covariance matrix of points in the neighborhood The calculation formula is as follows:

[0113]

[0114] Wherein, is the point in the neighborhood, is the centroid of the neighborhood point.

[0115] Then, the eigenvalue decomposition is performed on the covariance matrix , and three eigenvalues λ1≥λ2≥λ3 are obtained, then the curvature is:

[0116]

[0117] Further, the curvature weight K( ) can be defined as K( )= .

[0118] By this method, 1024 key points are selected, reducing the data amount and retaining key information, especially better retaining the details of high curvature areas (such as mirror edges), which is crucial for aerodynamic characteristic analysis.

[0119] Step 13, normalize the air density data and map it to [0, 1]. Let the original air density data be x, and the normalized air density data be The normalization formula is:

[0120]

[0121] Wherein, and are the minimum and maximum values of the original air density data, respectively.

[0122] Normalization can eliminate the influence of different orders of magnitude data on the model, so that the model can learn features more stably.

[0123] Step 14, use computer vision and geometric analysis algorithms to identify the elevation angle of the heliostat from the three-dimensional model of the heliostat. The specific process is as follows:

[0124] First, determine the equations of the heliostat plane and the reference plane, which can be obtained by fitting the plane equation from the point cloud data.

[0125] Then, calculate the normal vectors of the two planes, set the normal vector of the heliostat plane as , and the normal vector of the reference plane as .

[0126] Then, using the formula for the angle between vectors, the normal vector of the heliostat plane is calculated. Normal vector to the reference plane The angle between the two points is used to determine the elevation angle θ of the heliostat. The calculation formula is as follows:

[0127]

[0128] Considering the potential deformation of the mirror surface, this algorithm processes the deformed point cloud data to restore it to its original shape as much as possible, thereby reducing the elevation angle recognition error caused by deformation.

[0129] Finally, trigonometric function encoding is used to convert the elevation angle of the heliostat into numerical features, that is, the elevation angle θ is encoded into two values, sin(θ) and cos(θ), which facilitates subsequent model processing.

[0130] The second step is to extract and encode geometric features, which includes the following:

[0131] Step 21: Extract multi-dimensional geometric features from 1024 key points, including curvature and normal features.

[0132] As mentioned earlier, curvature can be obtained by calculating the covariance matrix of points in the neighborhood of keypoints and performing eigenvalue decomposition. The curvature characteristics reflect the local bending of the heliostat surface and are closely related to the airflow patterns on the surface. For example, a larger curvature may lead to airflow separation and the generation of vortices, thus affecting the aerodynamic coefficients.

[0133] The normal vector is the eigenvector corresponding to the smallest eigenvalue of the covariance matrix. Let the covariance matrix be... The smallest eigenvalue is The corresponding feature vector is ,but That is the key point The normal characteristics of the heliostat surface. The normal characteristics indicate the orientation of the heliostat surface at that point and have a significant impact on the direction of airflow and pressure distribution.

[0134] The extracted geometric features are combined into a feature matrix. ,in For the first The feature vectors of each key point contain curvature and normal features.

[0135] Step 22: Encode the geometric features using the PointNet++ algorithm. The PointNet++ algorithm comprises three core processes: sampling, grouping, and feature extraction.

[0136] Sampling process, including the following:

[0137] The representative points are selected by using the improved sampling algorithm FPS. This sampling algorithm effectively reduces the data amount while preserving the key information, laying a foundation for subsequent efficient processing.

[0138] Grouping process, including the following:

[0139] Find the neighborhood point set for each sampling point. Use the k-neighbor method for grouping. The grouping operation enables the network to capture the local features of the point cloud data, thereby better understanding the geometric structure of the heliostat surface.

[0140] Feature extraction process, including the following:

[0141] Apply the small point network algorithm PointNet to hierarchical sampling and local feature aggregation on the neighborhood point set, and perform stitching and pooling operations to finally obtain a 256-dimensional feature vector, effectively capturing the local and global features of the point cloud. It mainly includes the following three steps:

[0142] For each sampling point, the point set in its neighborhood is denoted as which contains the point and its kkk nearest neighbors. In each neighborhood, a local coordinate system is constructed, and the geometric features of the point (such as curvature, normal) are taken as the input feature vector. These features are input into a small point network algorithm PointNet network with shared weights for feature extraction.

[0143] In this embodiment, the small point network algorithm PointNet network includes the following structure:

[0144] First layer: fully connected layer FC + activation function ReLU + batch normalization BN;

[0145] Second layer: fully connected layer FC + activation function ReLU + batch normalization BN;

[0146] Third layer: use the fully connected layer FC to output local features.

[0147] This structure can effectively model the geometric features of the local structure of the point cloud, and is especially suitable for irregular heliostat edges or curved surface deformation areas.

[0148] In this embodiment, the stitching operation includes the following:

[0149] For each point, the local features extracted by the small point network algorithm PointNet are stitched with the original geometric features (such as point coordinates, normal, curvature) to form an enhanced feature representation In this way, not only the basic geometric information is reserved, but also the local context semantics is fused, which is helpful for subsequent global feature modeling.

[0150] In this embodiment, the pooling operation includes the following:

[0151] The pooling operation is performed within each region to extract the most significant features representing the region, thereby eliminating the influence of the number of local points. Then the global feature description is formed by aggregating the features of multiple regions. The final output is a 256-dimensional feature vector, which is used for subsequent fusion and modeling.

[0152] In the third step, multiple feature vectors are fused to obtain a fused feature vector, which includes the following:

[0153] In order to further enhance the interaction and fusion effect between features, a nonlinear transformation can be performed on each feature vector before splicing. For example, a multi-layer perceptron (MLP) is used to transform each feature vector.

[0154] The expression of point cloud feature transformation is as follows:

[0155]

[0156] wherein is a multi-layer perceptron network for point cloud features, is the transformed point cloud feature, is the original point cloud feature, which contains local / global geometric features (such as curvature, normal direction) encoded by the point network enhancement algorithm PointNet++, and its distribution structure usually has nonlinearity and spatial dispersion.

[0157] In order to avoid the problem of feature scale imbalance and semantic inconsistency caused by direct splicing, the following MLP network is designed:

[0158] The input dimension is the original point cloud feature dimension (such as 256 dimensions);

[0159] The hidden layer is set to three perceptrons, each equipped with a nonlinear activation function (ReLU) and normalization (BatchNorm); the transformation process is as follows:

[0160] Layer Layer1: 256→128;

[0161] Layer Layer2: 128→64;

[0162] Layer Layer3: 64→32.

[0163] Output dimension: unified to 32 dimensions to align and fuse with other modal features.

[0164] The network can automatically learn high-order nonlinear relationships in point cloud features, so that the original geometric information is retained in a low-dimensional space while reducing redundant information.

[0165] The expression of the wind speed feature transformation is as follows:

[0166]

[0167] wherein is a multi-layer perceptron network for wind speed features, is the transformed wind speed feature, is the original wind speed feature.

[0168] The expression of the air density feature transformation is as follows:

[0169]

[0170] wherein is a multi-layer perceptron network for air density features, is the transformed air density feature, is the original air density feature.

[0171] Air density and wind speed belong to environmental characteristic quantities, and usually have low dimensions (such as wind speed: 3-dimensional vector, density: 1-dimensional scalar), but their changes have nonlinear effects on aerodynamic characteristics. For example, the flow state on the surface of the heliostat may undergo a critical transition at different air densities.

[0172] Therefore, an independent MLP network is designed for each physical quantity to extract its implicit nonlinear influencing factors. The structure is as follows:

[0173] Input dimension: 3 dimensions of wind speed or 1 dimension of air density;

[0174] Hidden layer structure:

[0175] Layer Layer1: input dimension → 16;

[0176] Layer Layer2: 16 → 32;

[0177] Output dimension: unified to 32 dimensions.

[0178] Thus, the original low-dimensional continuous variables are embedded into a unified feature space, facilitating splicing with point cloud features, and learning high-order effects of variables on aerodynamic effects through the network.

[0179] The expression of the heliostat elevation angle feature transformation is as follows:

[0180]

[0181] wherein,​ is a multi-layer perceptron network for heliostat elevation features, is the transformed heliostat elevation feature, is the original heliostat elevation feature.

[0182] The heliostat elevation is encoded by in(θ), cos(θ) to form a 2-dimensional vector, which is used to represent the spatial pose of the heliostat. Although the dimension is low, it has a great impact on the aerodynamic characteristics (such as changes in windward area, migration of flow separation point, etc.).

[0183] In order to make the model better learn the coupling relationship between the elevation feature and the aerodynamic performance, a dedicated MLP is designed, and its structure is as follows:

[0184] Input dimension: 2;

[0185] Structure: 2→16→32;

[0186] Output dimension: 32 dimensions, aligned with other modalities.

[0187] And guide the model to extract discriminative representation from the periodicity and directionality of the elevation, avoiding the "elevation information dilution" caused by direct concatenation.

[0188] Finally, the transformed feature vector is then concatenated to obtain the fused feature vector , whose expression is as follows:

[0189]

[0190] Through nonlinear transformation, the embodiment can learn more complex relationships between features, and mine potential information hidden in the features, thereby further improving the quality of the fused features and the prediction performance.

[0191] The nonlinear transformation of the present application enables the model to have the ability to fit high-order, nonlinear, and multivariate coupling relationships by introducing an activation function. For example, when the point cloud normal changes dramatically and the elevation is close to the critical value, the influence of wind speed or air density on the aerodynamic response will no longer be linear growth, but may exhibit complex behaviors such as jumping, saturation, and gating. Traditional linear mapping cannot capture such changes, but the multi-layer perceptron (MLP) can model these potential nonlinear patterns under the action of multiple activation units.

[0192] Furthermore, although the features of each modality (point cloud, wind speed, density, and elevation angle) are transformed separately through an MLP network in the early stages, the core purpose of this transformation is to project various features into a unified embedding space, giving them similar scale and distribution characteristics, thus providing a semantically aligned representation basis for subsequent feature fusion and interactive modeling. After concatenating to form a fused feature vector, a unified aerodynamic performance prediction model (heliostat aerodynamic simulation method) is used to further model the deep coupling relationships between modalities, thereby improving the overall prediction performance and model generalization ability.

[0193] A specific embodiment of the heliostat aerodynamic simulation method of the present invention:

[0194] A heliostat aerodynamic simulation method, namely an aerodynamic coefficient prediction model, includes the following steps:

[0195] Step 1. Apply the heliostat feature extraction method described above to process and extract features from the multi-source data of the heliostat to obtain a fused feature vector.

[0196] Step 2. Mine the interaction relationships between modal features to obtain modal coupling data; then introduce graph convolution operation to represent the modal coupling data as a graph structure, and perform message propagation on the graph structure to characterize the dependencies between modal features and form graph embedding representation information.

[0197] In this embodiment, the method for mining the interaction relationships between modal features to obtain modal coupling data is as follows:

[0198] A deep cross-attention network is constructed to achieve mutual attention between different modalities at multiple levels. This network iterates multiple times, calculating attention scores between features of different modalities and assigning dynamically changing weights to each feature.

[0199] A deep cross-attention network is constructed based on the cross-attention mechanism to more deeply explore the interaction relationships between different modalities of data. Let the input multimodal features be as follows: Each of them The data representing the i-th modal feature (e.g., point cloud, air density, elevation angle, etc.) consists of M modalities, each containing... There are positions (or nodes), each position being... 3D eigenvectors.

[0200] First, a linear transformation is performed on each modal feature to obtain the query matrix, key matrix, and value matrix. For the first... Modal features Its query matrix Key matrix Value matrix ,in , The weight matrix is ​​a learnable matrix. For the attention dimension.

[0201] The learnable weight matrix in this embodiment , All features were automatically optimized during the training phase of a deep cross-attention network, with the training objective being to minimize the designed physical constraint loss function. This training process fully considers the modal differences and physical consistency constraints of the input features, ensuring that the fused features retain physical interpretability while possessing strong fitting and generalization abilities.

[0202] Then, the attention scores between different modal features are calculated, for modal feature i and modal feature i. Attention score The calculation formula is as follows:

[0203]

[0204] in, It is the dimension of the key vector. For activation function, Modal features The query matrix, Modal features The bond matrix.

[0205] In this embodiment The value is set to 64. Experiments have shown that 64 is more suitable for fine-grained matching and attention calculation between point clouds and physical quantities.

[0206] Activation function The expression for row-wise calculation is as follows:

[0207]

[0208] in, The output value of the i-th node, where C is the number of output nodes, i.e., the number of categories; and e is the natural constant. The output value of the c-th node.

[0209] Then, the value matrix of modality feature j is weighted and summed according to the attention weights to obtain the fused feature of modality feature i, and the calculation formula is as follows:

[0210]

[0211] in, Modal features Each location feature is the result of a weighted fusion of all location features of modality feature j.

[0212] In the deep cross-attention network, multi-layer attention calculation is performed. Layer modal feature The fusion feature is The query, key and value matrices of the first layer continue to be generated based on the fusion feature of the previous layer, and the calculation formula is as follows:

[0213]

[0214] Similarly, define and The attention score and the fusion feature are updated iteratively, so as to realize the deep interaction between modalities at multiple semantic levels and improve the modeling ability of complex physical coupling relationships.

[0215] In this embodiment, a graph convolution operation is introduced, the modal coupling data is represented as a graph structure, and message propagation is performed on the graph structure to represent the dependency relationship between modal features, and the method of forming a graph embedding representation information is as follows:

[0216] The data of different modalities is represented as a graph structure , wherein is a node set, representing the features of multi-modal data; is an edge set, representing the relationship between features.

[0217] In this embodiment, the node set V represents the observation points on the surface of the heliostat, each node contains the geometric modal feature (such as coordinates, normal vectors, curvature, etc.) and the physical modal feature (such as local velocity, pressure, etc.) of its corresponding position, and the fusion forms a node feature vector; the edge set E represents the adjacency relationship (such as Euclidean distance, flow field correlation) between nodes, which is used for information propagation and aggregation in subsequent graph neural networks.

[0218] In the present application, the relationship between features is modeled by constructing edges in the graph structure, and the relationship between features includes geometric adjacency relationship, feature similarity relationship, and physical prior relationship.

[0219] The geometric adjacency relationship includes the following contents:

[0220] According to the positions of the point cloud or structural elements in space, the K-nearest neighbor algorithm is used to establish edges between nodes to reflect local physical interaction.

[0221] The feature similarity relationship includes the following contents:

[0222] The cosine similarity of the modal features between nodes is calculated, and the high similarity is connected to capture the coupling characteristics of similar behavior at a distance.

[0223] The physical prior relationship includes the following contents:

[0224] The physical dependent relationship such as airflow path and heat transfer path is defined in combination with the CFD simulation data or structural connection relationship, so as to enhance the physical interpretability of the graph structure.

[0225] Finally, the three relationships are fused to construct the graph structure G=(V, E), which provides a basic topology support for the graph neural network and supports deep fusion between modalities and physical information propagation.

[0226] Further, the graph convolution operation is introduced to update the features of the nodes , and the update formula is as follows:

[0227]

[0228] wherein: is the neighbor node set of the node ; , are the degrees of the nodes and v respectively; is the learnable weight matrix of the i-th layer; is an activation function (such as ReLU); is the updated feature after fusing the neighbor information; is the neighbor information.

[0229] In this embodiment, the nodes in the graph structure represent the spatial observation points on the heliostat surface, and the feature vectors thereof are composed of the features of the points in multiple modalities, including geometric features (such as coordinates, normal vectors, and curvatures) and physical features (such as local wind speed, air pressure, and density). These modality features are linearly transformed to obtain the query matrix (Query), the key matrix (Key), and the value matrix (Value) required by the attention network, and are cross-fused and graph-convoluted in the graph structure.

[0230] Further, the feature representation of the nodes is constantly updated through multiple graph convolution operations and the introduction of the information of the neighbor nodes, so as to further enhance the modeling capability of the inter-modal physical correlation and realize the capability of propagating the local modality features to the global. Meanwhile, the deep cross-attention network and the graph convolution network are comprehensively fused to model the interaction relationship between the multi-modal features and the structural correlation.

[0231] Step 3. The model network is trained and optimized, which includes the following contents:

[0232] A physical constraint loss function is established, and the expression thereof is as follows:

[0233]

[0234] wherein: ​denotes the loss value finally used for training optimization, and is a weighting coefficient and + =1, satisfying a normalization constraint, and are mean square errors of drag coefficient and pressure coefficient respectively, and is a penalty coefficient. is determined by hyperparameter grid search , , , , guiding the model network to learn features consistent with physical laws; denotes the predicted value of the drag coefficient corresponding to the i-th sample or prediction point; denotes the predicted value of the pressure coefficient corresponding to the i-th sample or prediction point.

[0235] The true values (benchmark values) of the drag coefficient and the pressure coefficient are provided by computational fluid dynamics (CFD) simulation data as supervisory signals. The mean square error (MSE) between the predicted values output by the model and these true labels is included in the physical constraint loss function to ensure that the built model not only fits well at the data level, but also meets the basic physical consistency requirements.

[0236] Step 4. Use the trained model network for prediction, which includes the following contents:

[0237] A fully connected neural network is used as the drag coefficient prediction head, and the input fusion feature vector is used. The network contains multiple hidden layers , and the calculation formula of each layer is as follows:

[0238]

[0239] wherein, is a weight matrix, is a bias vector, and σ is an activation function.

[0240] After multiple layers of calculation, the predicted drag coefficient is output.

[0241] The specific value of the bias vector is automatically learned during the model training process, and the learning process is as follows:

[0242] S1. In the initialization phase, i.e. network establishment, the bias vector of each layer is assigned by all-zero mode, and the initialization operation is completed. S2. In the forward propagation phase, the bias is only a vector with the same dimension as the output, which is added to the output variable of each node. S3. In the back propagation phase, i.e. in the training process, according to the loss between the network output and the real target, the gradient of the loss to the bias is calculated using the chain rule; these gradients are updated by the optimizer Adam. After multiple iterations of forward propagation and back propagation, the value of the bias vector gradually converges to a set of parameter values that minimize the loss function.

[0243] In this embodiment, the method of multi-layer calculation is as follows:

[0244] Step one, obtain the input data, the input data is a plurality of modal feature vectors with physical meaning, and the expression is as follows:

[0245]

[0246] Where each represents the data feature of the th mode, for example: is the point cloud shape feature, is the air density value, is the elevation angle parameter, is the wind field feature.

[0247] Step two, perform deep cross-attention mechanism fusion on the input data, which includes the following contents:

[0248] (1) For each modal feature , use a learnable matrix to perform query matrix (Query), key matrix (Key) and value matrix (Value) conversion, and the expression of the learnable matrix is as follows:

[0249]

[0250] (2) For modes and , calculate the cross-attention score, i.e. the attention weight, and the calculation formula of the attention weight of the th layer is as follows:

[0251]

[0252] (3) Perform feature fusion to obtain the fused feature , and the expression is as follows:

[0253]

[0254] After Layer attention mechanism iteratively updates to obtain the fused feature representation , i.e., modal coupling data.

[0255] Step three, modeling the graph structure and graph convolution processing, including the following:

[0256] (1) Construct a graph structure G = (V, E), where the node set Each node represents a modal fused feature vector (i.e., the part obtained in the previous step). The edge set E is obtained from the physical correlation or Euclidean distance / structural dependence between features, and the edge weight can be set based on the correlation coefficient or prior knowledge.

[0257] (2) Perform graph convolution operation to update the graph embedding representation information of each layer of the node , which is expressed as follows:

[0258]

[0259] Where: is the set of adjacent nodes; is the layer graph convolution learnable matrix; is the initial input of the graph convolution, which is the cross-attention output in the previous step, is the fused feature of the layer node i; is the activation function ReLU.

[0260] After layer graph convolution operation, the final graph embedding representation information is obtained.

[0261] Step four, map all node final features (graph embedding representation information) to resistance coefficient prediction value through aggregation and nonlinear regression layer , which is expressed as follows:

[0262]

[0263] Where: is the feature aggregation method, which is calculated by weighted summation method; is the regression function, which can be a fully connected network or a linear layer.

[0264] Step five, supervised training of the loss function, in the training stage, compare the predicted value with the true value , calculate the mean square error, which is calculated as follows:

[0265] ​

[0266] where the pressure coefficient is predicted in the same way as the drag coefficient , and the same process is used to calculate , which is finally combined with the physical constraint loss function as follows:

[0267]

[0268] In this embodiment, the pressure coefficient prediction head is also a fully connected neural network, with a structure similar to the drag coefficient prediction head. The input is the fusion feature vector, which is output after multiple layers of calculation as the predicted pressure coefficient .

[0269] A specific embodiment of applying the heliostat feature extraction method of the present application to heliostat aerodynamic simulation:

[0270] A large tower-type solar thermal power plant plans to upgrade and modify the existing heliostat system to improve power generation efficiency and stability. When designing a new heliostat scheme, it is necessary to quickly and accurately predict the aerodynamic coefficients of the heliostat in order to evaluate and optimize different design schemes.

[0271] The heliostat feature extraction method of the present application is applied to the aerodynamic simulation of the heliostat, and the fusion feature vector is input into the pre-constructed aerodynamic coefficient prediction model to obtain the aerodynamic coefficient prediction result of the heliostat. Compared with traditional wind tunnel test methods and CFD numerical simulation methods, the present application greatly shortens the data processing and feature extraction time from several hours or even several months to tens of minutes, while reducing the computational cost.

[0272] According to the predicted aerodynamic coefficients, the project team evaluated and optimized the heliostats of different design schemes. By adjusting the geometric shape, elevation angle and other parameters of the heliostat, the aerodynamic performance of the heliostat was improved, thereby improving the power generation efficiency and stability of the solar thermal power plant.

[0273] In this embodiment, the method of inputting the fusion feature vector into the aerodynamic coefficient prediction model to obtain the aerodynamic coefficient prediction result of the heliostat includes the following steps:

[0274] Step 1: Process the fusion feature vector of the heliostat to generate multiple modal features;

[0275] Step 2: Mine the interaction between modal features to obtain modal coupling data;

[0276] Step three, introducing graph convolution operation, representing the modal coupling data as a graph structure, and performing message propagation on the graph structure to represent the dependency relationship between modal features, forming a graph embedding representation information;

[0277] Step four, aggregating and mapping the graph embedding representation information to obtain the predicted heliostat aerodynamic coefficient, realizing the aerodynamic simulation of the heliostat.

[0278] In the embodiment, the method for processing the geometric data and air data of the heliostat to generate a plurality of modal features is as follows:

[0279] The geometric data of the heliostat is obtained, which includes three-dimensional point cloud data and elevation angle data of the heliostat;

[0280] The air data is obtained, which includes air density data, and the air density data is constructed in combination with historical meteorological data;

[0281] The three-dimensional point cloud data is processed to extract its geometric features to obtain the first modal feature;

[0282] The angle recognition and coding are performed on the elevation angle data of the heliostat to obtain the second modal feature;

[0283] The air density data is normalized to obtain the third modal feature;

[0284] The first modal feature, the second modal feature and the third modal feature are fused to form a plurality of modal features.

[0285] In the embodiment, the method for mining the interaction relationship between the modal features to obtain the modal coupling data is as follows:

[0286] The plurality of modal features are obtained, which at least include the first modal feature and the second modal feature, each modal feature includes a plurality of positions, and each position is a d-dimensional feature vector;

[0287] Considering the difference between the modal features and the physical consistency constraint, a learnable weight matrix corresponding to each modal feature is set; the learnable weight matrix is obtained by training a deep cross-attention network, and the training target is to minimize a physical constraint loss function;

[0288] Based on the learnable weight matrix, the modal features are linearly transformed to obtain a query matrix, a key matrix and a value matrix;

[0289] According to the query matrix and the key matrix, the attention score between the first modal feature and the second modal feature is calculated;

[0290] The value matrix of the second modality feature is weighted and summed based on the attention score to obtain a first fusion feature of the first modality feature; the value matrix of the first modality feature is weighted and summed based on the attention score to obtain a second fusion feature of the second modality feature;

[0291] According to the first fusion feature and the second fusion feature, a new query matrix, a key matrix and a value matrix are calculated;

[0292] Based on the new query matrix and the key matrix, the attention score between the first fusion feature and the second fusion feature is calculated;

[0293] The attention score and the fusion feature are continuously updated, so that the deep interaction between the modality features is realized at multiple semantic levels, and finally the first fusion feature and the second fusion feature are summarized to obtain the modality coupling data.

[0294] As shown in Figure 3 , in this embodiment, the method for training the deep cross-attention network is as follows:

[0295] A predicted heliostat aerodynamic coefficient is obtained, and compared with a reference aerodynamic coefficient value to calculate the mean square error of the two;

[0296] Based on the mean square error, the weighting coefficient and the penalty coefficient, a loss value is taken as the target to construct a physical constraint loss function;

[0297] Global interaction features between different modality features are extracted through the deep cross-attention network;

[0298] Then a graph structure between the modality features is constructed, and a local dependency relationship and topological association are modeled through graph convolution;

[0299] Then, based on the deep cross-attention network and the graph structure, an aerodynamic coefficient prediction result is output, and the physical constraint loss function is taken as an optimization target to guide the entire deep cross-attention network to fit the aerodynamic coefficient while maintaining physical consistency; a variable dependency and a back propagation mechanism form an end-to-end training joint learning network.

[0300] In step three of this embodiment, a graph convolution operation is introduced, the modality coupling data is represented as a graph structure, and message propagation is performed on the graph structure to represent the dependency relationship between the modality features, and the method for forming a graph embedding representation information is as follows:

[0301] Based on the observation points on the surface of the heliostat and the corresponding position information, a plurality of nodes are constructed;

[0302] According to the modality coupling data and the corresponding position information, the nodes are valued, so that each node has a geometric modality and a physical modality; and the geometric modality and the physical modality are fused to form a node feature vector;

[0303] Based on the modal coupling data, the relationship between the features is obtained; and according to the relationship between the features, a geometric edge is constructed;

[0304] According to the geometric edge, the node and the node feature vector, a graph structure is drawn;

[0305] Based on the deep cross-attention network and the graph convolutional network, and in combination with the node feature vector of the neighbor node, the node feature vector is updated, so that the local modal feature can be propagated to the global, and the dependence relationship and the structural correlation between the modal features can be represented.

[0306] In the embodiment, the relationship between the features includes a geometric adjacency relationship, a feature similarity relationship and a physical prior relationship;

[0307] The geometric adjacency relationship is established according to the positions of the point cloud or the structural elements in space and by using a nearest neighbor algorithm, and is used to reflect the local physical interaction;

[0308] The feature similarity relationship is established by calculating the cosine similarity of the modal features, and is used to capture the long-distance but similar behavior coupling characteristics;

[0309] The physical prior relationship is established based on the relationship between the structural connection relationship and the airflow path and the heat transfer path, and is used to enhance the physical interpretability of the graph structure.

[0310] Based on the deep cross-attention network and the graph convolutional network, and in combination with the node feature vector of the neighbor node, the node feature vector is updated as follows:

[0311] First, through the deep cross-attention network, the node feature vector is cross-attended layer by layer, and in combination with the attention mechanism, a multi-modal fusion feature is generated, which includes a plurality of fusion feature units;

[0312] The multi-modal fusion feature is taken as an input node of the graph convolutional network, and a graph structure is constructed, which includes a plurality of nodes and a plurality of edges, wherein the node represents the fusion feature unit, and the edge represents the physical dependence relationship, the spatial connection or the structural similarity between the feature units;

[0313] According to the node feature vector of the neighbor node, an adjacency matrix is established;

[0314] In the graph convolutional network, through the message propagation mechanism controlled by the adjacency matrix, the layer-by-layer update of the node feature vector is realized, so that the node feature vector can interact with the local modal feature, and can be integrated into the topological information of the global physical field.

[0315] In step four of the embodiment, the method for aggregating and mapping the graph embedding representation information to obtain the predicted heliostat aerodynamic coefficient is as follows:

[0316] Based on the weight matrix, the bias vector and the activation function, a coefficient calculation formula is constructed;

[0317] According to the coefficient calculation formula, a plurality of hidden layers are established;

[0318] The graph embedding representation information is input into the plurality of hidden layers for multilayer calculation to obtain node final features;

[0319] The node final features are subjected to feature aggregation and weighted summation, and then input into a regression function for calculation to obtain predicted heliostat aerodynamic coefficients; the heliostat aerodynamic coefficients include a drag coefficient and a pressure coefficient.

[0320] A server embodiment applying the method of the application:

[0321] A server comprises:

[0322] One or more processing units;

[0323] Storage means for storing one or more programs;

[0324] When the one or more programs are executed by the one or more processing units, the one or more processing units implement the above-mentioned heliostat feature extraction method.

[0325] The storage means are internal memory or external memory or cache memory or other special memory. The processing unit can be a general-purpose processor, a digital signal processor, an application-specific integrated circuit, a ready-made programmable gate array or other programmable logic device.

[0326] A device embodiment applying the method of the application:

[0327] An electronic device is provided with a computer readable storage medium, and the computer readable storage medium stores a computer program, which, when executed by a processing unit, implements the above-mentioned heliostat feature extraction method.

[0328] The computer readable storage medium refers to a physical carrier capable of storing computer-recognizable data, instructions or programs, which needs to meet the core characteristic of being "readable by a computer" (i.e. data exists in the form of electrical, magnetic, optical signals, which can be converted into binary information that can be processed by a computer through corresponding equipment). The physical carrier is a magnetic storage medium, an optical storage medium, a semiconductor storage medium or other storage medium.

[0329] The model in the present application is an object that constitutes an objective description of a morphological structure by means of a physical or virtual representation. The object is not equal to the object, and is not limited to the physical and virtual. It can be a data processing function, a software program, a processing mode, a use method, an operation mode, a work flow, an application process, an electronic hardware, a circuit module, a processing system, a system imitation product, or a simulation analog object.

[0330] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present application, which are used to illustrate the technical solutions of the present application, rather than limit the same. The protection scope of the present application is not limited thereto. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can modify the technical solutions recorded in the foregoing embodiments, or replace some of the technical features with equivalent replacements, within the technical scope disclosed by the present application, without departing from the spirit and scope of the technical solutions of the embodiments of the present application. Any modification or equivalent replacement that does not depart from the spirit and scope of the present application should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A heliostat feature extraction method, characterized in that: comprising the following steps: Step 1: Obtain heliostat point cloud scanning data; Step 2: Based on the curvature weight, use the pre-constructed heliostat preprocessing model to sample the heliostat point cloud scanning data to obtain heliostat key points; The method for obtaining heliostat key points is as follows: According to the heliostat point cloud scanning data, a point cloud set is established, which includes a plurality of discrete points, and a selected point set is constructed; Based on the plurality of discrete points, the maximum distance and the maximum curvature value between the discrete points are calculated; The minimum distance from a discrete point to all discrete points in the selected point set is calculated; Based on the covariance matrix, the curvature weight of the discrete point is determined; According to the curvature weight, the minimum distance, the maximum distance between the discrete points and the maximum curvature value of the discrete point, the sampling value of the discrete point is calculated; Based on the sampling values of the plurality of discrete points, a sampling distribution is established, and the maximum value and the discrete point corresponding to the maximum value are selected from the sampling distribution; The discrete point corresponding to the maximum value is taken as the heliostat key point to represent the structure of the high curvature area of the heliostat; Step 3: Use the pre-constructed feature mining model to process the heliostat key points to obtain multi-dimensional geometric features to capture the local and global features of the heliostat; Step 4: Use the pre-constructed feature fusion model to perform nonlinear transformation on the multi-dimensional geometric features to obtain structure feature quantities, and couple the structure feature quantities with the environmental feature quantities of the heliostat and the elevation angle of the heliostat to generate a fusion feature vector to realize the feature extraction of the heliostat; The method for obtaining the structure feature quantity is as follows: The multi-dimensional geometric features are obtained, and the data dimensions thereof are ; By using a nonlinear activation function and a normalization operation, the multidimensional geometric features are nonlinearly transformed to obtain a first simplified feature vector of the first dimension, which is used to mine potential information hidden in the multidimensional geometric features; By using a nonlinear activation function and normalization operation, a nonlinear transformation is performed on the first simplified feature vector to obtain... The second simplified eigenvector of dimension; The second simplified feature vector is nonlinearly transformed by using a nonlinear activation function and a normalization operation to obtain a third simplified feature vector of dimensionality The third simplified feature vector is of dimensionality, and thus a structural feature quantity that can retain discriminativeness and be fused in alignment is formed. 2.The heliostat feature extraction method of claim 1, characterized in that: The method for determining the curvature weight of a discrete point based on the covariance matrix is as follows: According to the discrete point, the points in the neighborhood and the neighborhood point centroid are obtained; According to the points in the neighborhood and the neighborhood point centroid, a covariance matrix of the points in the neighborhood of the discrete point is established; Eigenvalue decomposition is performed on the covariance matrix to obtain three eigenvalues; According to the three eigenvalues, the curvature is calculated; The curvature is taken as the curvature weight of the discrete point. 3.The heliostat feature extraction method of claim 1, characterized in that: The method for obtaining multi-dimensional geometric features by using the pre-constructed feature mining model to process the heliostat key points is as follows: The covariance matrix of the points in the neighborhood of a heliostat key point is calculated; Eigenvalue decomposition is performed on the covariance matrix to obtain the curvature; The eigenvector corresponding to the minimum eigenvalue of the covariance matrix is taken as the normal feature; The curvature and the normal feature are taken as the geometric feature, and the feature vector of the heliostat key point is constructed based on the geometric feature; The feature vectors of all heliostat key points are combined to obtain a geometric feature matrix; The geometric feature matrix is sampled, grouped and feature extracted to obtain multi-dimensional geometric features. 4.The heliostat feature extraction method of claim 3, characterized in that: The method for obtaining multi-dimensional geometric features by sampling, grouping and feature extracting the geometric feature matrix is as follows: The geometric feature matrix is processed according to a fixed time period or a sliding window sampling method to obtain a plurality of representative features; According to each representative feature, a local coordinate system is constructed, and a plurality of adjacent representative features are found based on the local feature coordinate system; Based on the plurality of representative features, a plurality of neighborhood feature sets are established to capture the local structural features of the heliostat surface; The plurality of neighborhood feature sets are used to determine corresponding heliostat key points; According to the heliostat key points, the representative features in the neighborhood feature set are spliced with the original geometric features to form an enhanced feature representation for retaining the basic geometric information and fusing the local context semantics; Based on the enhanced feature representation, a field feature region is constructed; The field feature region is subjected to a pooling operation to extract the most significant features representing the field feature region; The plurality of most significant features are aggregated to obtain multi-dimensional geometric features to describe the global features.

5. The heliostat feature extraction method according to claim 4, wherein: The structural feature quantity is coupled with the environmental feature quantity of the heliostat and the heliostat elevation angle to generate a fusion feature vector, and the feature extraction method of the heliostat is as follows The environmental feature quantity is obtained, which includes a 3-dimensional wind speed feature and a 1-dimensional air density feature; A multi-layer perception network is constructed for the wind speed characteristics, and the 3-dimensional wind speed characteristics are nonlinearly transformed to obtain a 2-dimensional wind speed characteristic vector. a 2-dimensional wind speed characteristic vector; A multilayer perceptron network is constructed to perform a nonlinear transformation on the 1D air density features, resulting in... A dimensional air density feature vector; The heliostat elevation angle is obtained, and the heliostat elevation angle is encoded to form a 2-dimensional elevation angle feature vector for representing the spatial posture of the heliostat; A multi-layer perceptron network is constructed for the heliostat elevation feature, a 2-dimensional elevation feature vector is nonlinearly transformed, and the multi-layer perceptron network is guided to extract discriminative representations from the periodicity and directionality of the heliostat elevation, obtaining an elevation feature vector of the heliostat in the 2-dimensional space. Then, the structural feature quantity, the wind speed feature vector, the air density feature vector and the heliostat elevation angle feature vector are spliced in rows to obtain the fusion feature vector.

6. The heliostat feature extraction method according to claim 5, wherein: The method for obtaining the air density feature is as follows: Air density data of the environment around the heliostat is collected; The air density data is sorted to obtain the minimum density value and the maximum density value; Based on the minimum density value and the maximum density value, the air density data is normalized to obtain the air density feature with uniform magnitude.

7. A heliostat feature extraction method as claimed in claim 5, characterized by: The method for obtaining the heliostat elevation angle feature is as follows: First, based on the points of the heliostat point cloud scanning data, a plane equation is fitted to determine the heliostat plane equation and the reference plane equation; Then, the normal vector one of the heliostat plane equation and the normal vector two of the reference plane equation are calculated; According to the vector included angle formula, the included angle of the normal vector one and the normal vector two is calculated; Based on the vector included angle, the elevation angle of the heliostat is determined; The elevation angle of the heliostat is converted into a numerical feature by using trigonometric function coding to obtain a 2-dimensional heliostat elevation angle feature.

8. A heliostat aerodynamic simulation method, characterized in that: A heliostat feature extraction method according to any one of claims 1-7 is applied to construct an aerodynamic coefficient prediction model, which includes the following steps: Step one, processing the fusion feature vector of the heliostat to generate a plurality of modal features; Step two, mining the interaction relationship between the modal features to obtain modal coupling data; Step three, introducing graph convolution operation to represent the modal coupling data as a graph structure, and performing message propagation on the graph structure to represent the dependency relationship between the modal features, forming a graph embedding representation information; Step four, the information of the graph embedding is aggregated and mapped to obtain the predicted heliostat aerodynamic coefficient, and the heliostat aerodynamic simulation is realized.

9. A heliostat aerodynamic simulation device, characterized in that: it comprises: one or more processing units; a storage device for storing one or more programs; when the one or more programs are executed by the one or more processing units, the one or more processing units realize the heliostat feature extraction method according to any one of claims 1-7.

Images