A real-time super-resolution method for urban wind fields based on sparse data

Abstract

The present invention relates to the technical field of urban environmental monitoring and prediction, and specifically, to a real-time super-resolution method for an urban wind field environment, comprising CFD technology, a residual graph convolutional neural network classifier algorithm, a graph feature encoder, a physical information graph auxiliary autoencoder, and a physical loss generator. The method comprises a computational fluid dynamics simulation technology to obtain spatiotemporal data at various wind attack angles within a study area, a residual graph convolutional neural network classifier algorithm to extract high-dimensional features from sensor and peripheral data, and on this basis to determine the wind attack angle to which the current data belongs and to provide auxiliary information to the physical information graph auxiliary autoencoder. The graph feature encoder restructures the auxiliary information provided by the difference graph convolutional neural network classifier and transmits it to the physical information graph auxiliary autoencoder. The physical information graph auxiliary autoencoder predicts the wind speed field at the corresponding moment after integrating the high-dimensional auxiliary information with the sensor wind speed data. The physical loss generator is used to enhance the accuracy of the model.

Description

Technical Field

[0001] The present invention relates to the technical field of urban environment monitoring and prediction, and in particular to a real-time super-resolution method for urban wind fields based on sparse data. Background Art

[0002] Research on urban wind environments is a crucial area of ​​wind engineering research, closely linked to the quality of life for a large number of urban residents. With increasing urbanization and a growing urban population, the impact of urban wind environments is also increasing. Therefore, high-resolution monitoring of the entire urban wind field is crucial.

[0003] Problems such as the urban heat island effect, urban pollutant diffusion, urban low-altitude drone transportation, and urban wind energy collection are all closely related to urban wind farms. High-resolution urban wind farm data can provide great data convenience for the above research, but due to the numerous obstacles in the city and the cost of the anemometer itself, at this stage, only sparse sensors can be deployed in the city and provide sparse wind speed monitoring data at the corresponding locations. Existing full-field wind speed data construction technologies, such as computational fluid dynamics simulation and wind tunnel testing, often require huge research costs to obtain high-resolution wind farm data, or the data acquisition time is long and has obvious lags, making it difficult to achieve real-time monitoring of urban wind farm data. Therefore, timely and rapid high-resolution monitoring of urban wind farms has become an urgent and necessary requirement for a large number of wind engineering-related research and engineering problems. Summary of the Invention

[0004] The purpose of the present invention is to overcome the shortcomings of existing technologies, adapt to actual needs, and provide a real-time super-resolution method for urban wind fields based on sparse data.

[0005] The present invention is achieved through the following technical solutions:

[0006] A real-time super-resolution method for urban wind fields based on sparse data uses five parts: computational fluid dynamics (CFD) simulation technology, residual graph convolution neural network classifier algorithm (RGCNNC), graph feature encoders (Graph Feature Encoders), physics-informed graph-assisted auto-encoder (PGI-AE), and physical loss generator (Physical Loss Generator). The working order of each algorithm is: computational fluid dynamics simulation technology, residual graph convolution neural network classifier, graph feature encoder, physics-informed graph-assisted auto-encoder, and physical loss generator.

[0007] The functions and basic principles of each algorithm of the present invention are as follows:

[0008] Computational fluid dynamics (CFD) simulation technology can simulate the wind environment based on digital geometric models and the statistical patterns of wind speeds in the study area. The simulated data can be discretized using Python programs, and post-processing programs can generate mask data and SDF data to further characterize the environmental information of the study area.

[0009] The Residual Graph Convolution Neural Network Classifier (RGCNNC) algorithm can extract and analyze the graph data structure composed of sparse sensor data and its surrounding environment data, fully capture the relationship between node attributes and edge attributes, and further fully characterize these relationships with high-dimensional features. Ultimately, it can determine the wind angle type of the current data through high-dimensional features.

[0010] The Graph Feature Encoders are composed of a module consisting of a two-dimensional convolutional layer, a batch normalization layer, and an activation function layer. The number of modules used varies for different dimensional features extracted at different depths in the residual graph convolutional neural network classifier algorithm. The higher the dimensionality of the data features, the more modules are needed. This part uses the nonlinear mapping function of multiple modules to restructure the data while keeping the information richness of the high-dimensional data almost unchanged, making it consistent with the data structure at the same depth within the physical information graph-assisted autoencoder. This allows for data splicing and provides the physical information graph-assisted autoencoder with richer high-dimensional information.

[0011] The Physics-informed Graph-assisted Auto-Encoder (PGI-AE) consists of five serially connected convolutional downsampling modules (including a two-dimensional convolutional layer, a batch normalization layer, and an activation function layer), five skip connection layers, and five serially connected integrated decoders (including a two-dimensional convolutional layer, a batch normalization layer, and an activation function layer). The five convolutional downsampling modules perform layer-by-layer feature extraction on the sparse sensor wind speed data. After each extraction, the skip connection layer directly inputs the data at that depth into the integrated decoder at the corresponding depth. Furthermore, the five integrated decoders simultaneously receive high-dimensional data from the graph feature encoder at the corresponding depth and combine it with the sparse sensor wind speed data transmitted by the skip connection. Ultimately, based on this information, learning is performed with the wind field at the corresponding moment as the target, acquiring the ability to reconstruct the wind field data at that moment from the sparse data.

[0012] The Physical Loss Generator verifies the x- and y-direction wind speed fields generated by the physical information graph-assisted autoencoder based on the two-dimensional fluid continuity equation, and uses the resulting difference as part of the loss function to correct the model parameters, thereby enhancing the model's prediction accuracy and robustness.

[0013] Specifically, the following steps are included:

[0014] Step 1: Using computational fluid dynamics (CFD) technology, we simulate the wind field in the target urban area, obtain spatiotemporal data at multiple wind angles, discretize the data into regular grids, and perform further post-processing to obtain the Signed Distance Function (SDF) and obstacle distribution mask data to provide support for subsequent steps.

[0015] Step 2: The Get Graph Feature Layer (GGF) neural network layer constructs and restructures the rasterized sensor wind speed data and its surrounding environment data into a three-dimensional data tensor constructed from three dimensions: "number of sensor and surrounding environment data points," "number of sensors," and "number of data features," providing support for subsequent steps.

[0016] Step 3: The Neighboring Points Assemble Layer (NPA) extracts spatial data features from the graph data structure processed in Step 1. This compresses the data dimension of "number of sensor and surrounding data points" and transforms the graph data structure into a two-dimensional data matrix constructed from two dimensions: "number of sensors" and "number of data features." This provides support for subsequent steps.

[0017] Step 4: The Point Cloud and Graph Features Enrichment Layer (PCGFE) extracts representative data features from the graph data structure processed in Step 3. The mean and maximum values ​​of each data feature are calculated in the data dimension of "number of sensors," resulting in two one-dimensional data vectors, each with a length equal to the number of data features. These two data vectors are then concatenated to obtain a one-dimensional vector twice the length of the number of data features, providing support for subsequent steps.

[0018] Step 5: The Residual Graph Convolution Neural Network Classifier (RGCNNC) connects the neural network layers in steps 2 to 4 in series and trains RGCNNC to classify the wind angles corresponding to the data based on the processed graph structure data. This guides RGCNNC to collaborate with the neural network layers in steps 2 to 4 to fully extract the mutual influence relationship between the sensor wind speed data and its surrounding environment data, as well as the high-dimensional data information characteristics.

[0019] Step 6: Graph Feature Encoders re-extract graph feature information and restructure the data at five different depths between the GGF and NPA of RGCNNC. The extracted information of each sensor and its surrounding environment is rearranged according to its spatial position in the wind field, and other positions in the space are padded with zeros to form a high-dimensional sparse numerical matrix, which provides auxiliary information for the subsequent main neural network structure.

[0020] Step 7: The Physics-informed Graph-assisted Auto-Encoder (PGI-AE), the main neural network structure, simultaneously receives wind speed data from the sparse sensor and the high-dimensional sparse numerical matrix provided by the graph feature encoder. Based on this information, the model is trained with the wind field data at the corresponding time as the target. This enables PGI-AE to predict the wind field data at the corresponding time in real time based on the sparse data.

[0021] Step 8: The physical loss generator (Physical Loss Generator) performs physical correction on the flow field results generated by PGI-AE based on the fluid continuity equation to improve the accuracy of the model prediction results.

[0022] In the above technical solution, the specific process of step one is: constructing a geometric model and digital calculation domain of the study area, obtaining the spatiotemporal wind field data at various wind attack angles in the area under the simulation of the incoming wind that approximates the local wind field, and discretizing the wind field data in a high-resolution regular manner. The data is further post-processed to obtain the spatial distance function value and obstacle distribution mask data required for subsequent steps.

[0023] The specific process of step 2 is as follows: extract the wind speed components u and v in the x and y directions from the sensor position, extract the horizontal spatial coordinates x and y, the time t corresponding to the wind speed data, the obstacle mask data binary, and the spatial distance function value SDF from the rasterized data of the sensor and its surroundings (a total of 9 data points for each sensor), and assign u and v at the sensor position to the 8 surrounding data points, thereby forming 7 data attribute dimensions for each data node; on this basis, subtract the surrounding rasterized data from the sensor position data in each attribute dimension to obtain the unidirectional edge attributes from the surrounding points to the central sensor point in each attribute dimension, and fill the edge attribute values ​​at the sensor position with 0. The point attribute data and the edge attribute data are superimposed in the attribute dimension to obtain local graph structure data with 14 attribute dimensions. At this time, the three-dimensional data tensor formed for each sensor and its surrounding data points is 3x3x14. After that, the tensor is flattened into a two-dimensional tensor (i.e., 9x14), and the data of all sensors (assuming there are N) are integrated to form a three-dimensional tensor in the form of N (number of sensors) x 9 (number of data points of sensors and their surrounding environment) x 14 (number of data features), providing support for subsequent steps.

[0024] The specific process of step three is as follows: for the three-dimensional tensor (Nx9x14) generated in step two, it is regarded as a plane composed of "number of sensors" and "number of data features", and "number of sensor and surrounding environment data points" as the data attribute dimension of the three-dimensional tensor, and its data attribute dimension is extracted. That is, the two-dimensional convolution layer, pooling layer, batch normalization layer, and activation function are used to reduce the data dimension of "number of sensor and surrounding environment data points". Finally, this dimension is expressed with a single value, which integrates the connection between the sensor and the surrounding environment and reduces the data structure to two dimensions (Nx14), providing a guarantee for subsequent steps.

[0025] The specific process of step four is as follows: for the two-dimensional numerical matrix (Nx14) generated in step three, calculate the maximum value and mean of the matrix in its "number of sensors" dimension respectively, and obtain two one-dimensional vectors (14x1) respectively, so as to further extract data features. Then, the two one-dimensional vectors are vertically spliced ​​to obtain a new one-dimensional vector of 28x2 to provide a guarantee for subsequent steps.

[0026] The specific process of step five is as follows: The residual graph convolutional neural network classifier consists of a graph feature capture neural network layer, five sets of convolution-batch normalization-activation modules, a neighboring point data integration layer, a point cloud and graph feature enhancement layer, two sets of linear mapping-batch normalization modules-activation-Dropout modules, and a final layer of linear mapping units. During training, the residual graph convolutional neural network classifier determines the wind angle corresponding to the current wind location based on input sensor and surrounding environmental data. During this process, the model will continuously attempt to extract patterns from the input data and generate corresponding high-dimensional representations within the neural network.

[0027] The specific process of step six is ​​as follows: the graph feature encoder extracts high-dimensional graph feature information from the five end positions of the five groups of convolution-batch normalization-activation modules in step five, and uses an appropriate number of convolution-batch normalization-activation modules to re-extract the high-dimensional graph feature information and reshape the data structure so that it can be spliced ​​with the intermediate data at the corresponding depth position of the main neural network, thereby providing auxiliary information for the main network.

[0028] The specific process of step seven is as follows: the physical information graph-assisted autoencoder receives sparse wind speed data from each sensor at the starting point of its neural network, and receives high-dimensional graph feature information from the graph feature encoder at different depths within its network; during the training phase, the physical information graph-assisted autoencoder uses the wind field data corresponding to the sparse data as the target, and performs training based on the sparse data and high-dimensional graph feature information to reconstruct the wind field data at the corresponding moment; after the training is completed, it can have the real-time reconstruction capability of the urban wind field.

[0029] The specific process of step eight is: apply the two-dimensional fluid continuity equation to the wind field of the x and y direction wind speed components predicted by the physical information graph assisted autoencoder, that is, The generated wind field is constrained, and the model is induced to learn the revised prediction results, thereby further improving the prediction accuracy and robustness of the model.

[0030] The beneficial effects of the present invention are as follows: the present invention includes CFD technology, a residual graph convolutional neural network classifier algorithm, a graph feature encoder, a physical information graph assisted autoencoder, and a physical loss generator; wherein, the computational fluid dynamics simulation technology obtains the spatiotemporal data at various wind attack angles in the study area, the residual graph convolutional neural network classifier algorithm performs high-dimensional feature extraction on the sensor and its surrounding data, and uses this as a basis to judge the wind attack angle to which the current data belongs and provide auxiliary information to the physical information graph assisted autoencoder, the graph feature encoder restructures the auxiliary information provided by the difference graph convolutional neural network classifier and transmits it to the physical information graph assisted autoencoder, the physical information graph assisted autoencoder predicts the wind speed field at the corresponding moment after integrating the high-dimensional auxiliary information with the sensor wind speed data, and the physical loss generator is used to enhance the model accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a visualization result diagram of the computational fluid dynamics wind field simulation, mask data, and SDF value of the present invention.

[0032] Figure 2 It is a schematic diagram of the data operation of the neural network layer for capturing graph features in the present invention.

[0033] Figure 3 It is a schematic diagram of the algorithm of the adjacent point data integration layer in the present invention.

[0034] Figure 4 This is a schematic diagram of the point cloud and graph feature enhancement layer data operations in the present invention.

[0035] Figure 5 It is a schematic diagram of the residual graph convolutional neural network classifier algorithm structure in the present invention.

[0036] Figure 6 It is a schematic diagram of the structure of the graph feature encoder algorithm in the present invention.

[0037] Figure 7 It is a schematic diagram of the structure of the physical information graph assisted autoencoder algorithm in the present invention.

[0038] Figure 8 It is a schematic diagram of the principle of the physical loss generator in the present invention. DETAILED DESCRIPTION

[0039] To make the purpose, technical solutions and advantages of the embodiments of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present application and the corresponding drawings. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of them.

[0040] Embodiment: A real-time super-resolution method for urban wind farms based on sparse data uses five parts: computational fluid dynamics simulation technology (CFD), residual graph convolution neural network classifier algorithm (RGCNNC), graph feature encoders (Graph Feature Encoders), physics-informed graph-assisted auto-encoder (PGI-AE), and physical loss generator (Physical Loss Generator). The working order of each algorithm is: computational fluid dynamics simulation technology, residual graph convolution neural network classifier, graph feature encoder, physics-informed graph-assisted auto-encoder (PGI-AE), and physical loss generator. Specifically, the method includes the following steps:

[0041] Step 1: Based on the computational fluid dynamics (CFD) simulation technology, the wind field of the target urban area is simulated to obtain the spatiotemporal data of the area under multiple wind attack angles. The data is discretized into regular grids and further post-processed to obtain the spatial distance function value (Signed Distance Function, SDF) and obstacle distribution mask data to provide support for subsequent steps, such as Figure 1 As shown;

[0042] Step 2: The Get Graph Feature Layer (GGF) is used to construct and restructure the rasterized sensor wind speed data and its surrounding environment data into a three-dimensional data tensor constructed from three dimensions: "number of sensor and surrounding environment data points", "number of sensors", and "number of data features", providing support for subsequent steps, such as Figure 2 As shown;

[0043] Step 3: The Neighboring Points Assemble Layer (NPA) extracts spatial dimension data features from the graph data structure processed in Step 1, compresses the data dimension of "number of sensor and surrounding environment data points", and transforms the graph data structure into a two-dimensional data matrix constructed by two dimensions: "number of sensors" and "number of data features", providing support for subsequent steps, such as Figure 3 As shown;

[0044] Step 4: The Point Cloud and Graph Features Enrichment Layer (PCGFE) extracts representative data features from the graph data structure processed in Step 3, and solves the mean and maximum values ​​of each "data feature" in the data dimension of "number of sensors", thereby obtaining two one-dimensional data vectors, both of which have a length of "number of data features". These two data vectors are then concatenated to obtain a one-dimensional vector with a length twice that of "number of data features", which provides a guarantee for subsequent steps, such as Figure 4 As shown;

[0045] Step 5: The Residual Graph Convolution Neural Network Classifier (RGCNNC) connects the neural network layers in steps 2 to 4 in series, and trains RGCNNC to classify the wind angle of attack corresponding to the data based on the processed graph structure data, thereby guiding RGCNNC to cooperate with the neural network layers in steps 2 to 4 to fully extract the mutual influence relationship between the sensor wind speed data and its surrounding environment data and the high-dimensional data information characteristics, such as Figure 5 As shown;

[0046] Step 6: Graph Feature Encoders re-extract graph feature information and restructure the data at five different depth positions between the GGF and NPA of RGCNNC. The extracted information of each sensor and its surrounding environment is rearranged according to its spatial position in the wind field, and other positions in the space are filled with 0 to form a high-dimensional sparse numerical matrix, which provides auxiliary information for the subsequent main neural network structure, such as Figure 6 As shown;

[0047] Step 7: The Physics-informed Graph-assisted Auto-Encoder (PGI-AE), the main neural network structure, simultaneously receives the wind speed data from the sparse sensor and the high-dimensional sparse numerical matrix provided by the graph feature encoder. Based on the information from both, the model is trained with the wind field data at the corresponding moment as the target, so that PGI-AE has the ability to predict the wind field data at the corresponding moment in real time based on the sparse data, such as Figure 7 As shown;

[0048] Step 8: Physical Loss Generator performs physical correction on the flow field results generated by PGI-AE based on the fluid continuity equation to improve the accuracy of the model prediction results, such as Figure 8 shown.

[0049] The above description is merely a preferred embodiment of the present invention and does not limit the present invention in any way. Any person skilled in the art who, without departing from the scope of the present invention, makes any equivalent substitution, modification, or other variation to the technical solution and technical contents disclosed in the present invention shall be deemed to be within the scope of the present invention and still fall within the scope of protection of the present invention.

Claims

1. A real-time super-resolution method for urban wind fields based on sparse data, characterized by: The algorithm uses five parts: computational fluid dynamics simulation technology, residual graph convolutional neural network classifier algorithm, graph feature encoder, physical information graph assisted autoencoder, and physical loss generator. The working order of each algorithm is as follows: computational fluid dynamics simulation technology, residual graph convolutional neural network classifier, graph feature encoder, physical information graph assisted autoencoder, and physical loss generator. The specific steps include: Step 1: Use computational fluid dynamics simulation technology to simulate the wind field in the target urban area, obtain spatiotemporal data at multiple wind attack angles in the area, discretize the data into regular grids, and perform further post-processing to obtain spatial distance function values ​​and obstacle distribution mask data to provide support for subsequent steps; Step 2: The graph feature capture neural network layer constructs and restructures the rasterized sensor wind speed data and its surrounding environment data into a three-dimensional data tensor constructed from three dimensions: "number of sensor and surrounding environment data points," "number of sensors," and "number of data features," providing support for subsequent steps. Step 3: The neighboring point data integration layer extracts spatial dimension data features from the graph data structure processed in Step 1. This compresses the data dimension of "number of sensor and surrounding environment data points" and transforms the graph data structure into a two-dimensional data matrix constructed from the dimensions of "number of sensors" and "number of data features," providing support for subsequent steps. Step 4: The point cloud and graph feature enhancement layer extracts representative data features from the graph data structure processed in Step 3. The mean and maximum values ​​of each "data feature" are calculated in the "number of sensors" data dimension, resulting in two one-dimensional data vectors, each with a length equal to the "number of data features." These two data vectors are then concatenated to obtain a one-dimensional vector with a length twice that of the "number of data features," providing support for subsequent steps. Step 5: The residual graph convolutional neural network classifier connects the neural network layers in steps 2 to 4 in series, and trains the residual graph convolutional neural network classifier algorithm to classify the wind angle of attack corresponding to the data based on the processed graph structure data, thereby guiding the residual graph convolutional neural network classifier algorithm to cooperate with the neural network layers in steps 2 to 4 to fully extract the mutual influence relationship between the sensor wind speed data and its surrounding environment data and the high-dimensional data information characteristics; Step 6: The graph feature encoder re-extracts the graph feature information and restructures the data at five different depths between the graph feature capture neural network layer and the neighboring point data integration layer of the residual graph convolutional neural network classifier algorithm. The extracted information of each sensor and its surrounding environment is rearranged according to its spatial position in the wind farm, and other positions in the space are padded with zeros to form a high-dimensional sparse numerical matrix, which provides auxiliary information for the subsequent main neural network structure. Step 7: The physical information graph-assisted autoencoder simultaneously receives the wind speed data from the sparse sensor and the high-dimensional sparse numerical matrix provided by the graph feature encoder. Based on the information from both, the model is trained with the wind field data at the corresponding moment as the target, so that the physical information graph-assisted autoencoder has the ability to predict the wind field data at the corresponding moment in real time based on the sparse data; Step 8: The physical loss generator performs physical correction on the flow field results generated by the physical information graph-assisted autoencoder based on the fluid continuity equation to improve the accuracy of the model prediction results.

2. The real-time super-resolution method for urban wind fields based on sparse data according to claim 1 is characterized in that: The specific process of step one is as follows: constructing a geometric model and a digital calculation domain of the study area, obtaining the spatiotemporal wind field data of the area at various wind attack angles under the simulation of the incoming wind that approximates the local wind field, discretizing the wind field data in a high-resolution regularized manner, and further post-processing the data to obtain the spatial distance function value and obstacle distribution mask data required for subsequent steps.

3. The real-time super-resolution method for urban wind fields based on sparse data according to claim 2 is characterized in that: The specific process of step 2 is as follows: extract the x and y direction wind speed components u and v from the sensor position, extract the horizontal spatial coordinates x and y, the time t corresponding to the wind speed data, the obstacle mask data binary, and the spatial distance function value SDF from the sensor and its surrounding rasterized data, and assign u and v at the sensor position to the surrounding 8 data points, thereby forming 7 data attribute dimensions for each data node. On this basis, the surrounding rasterized data is subtracted from the sensor position data in each attribute dimension to obtain the direction from each surrounding point to the central sensor in each attribute dimension. The unidirectional edge attributes of the points are padded with 0 at the sensor location. The point attribute data and the edge attribute data are superimposed in the attribute dimension to obtain local graph structure data with 14 attribute dimensions. The three-dimensional data tensor formed for each sensor and its surrounding data points is 3x3x14. This tensor is flattened into a two-dimensional tensor of 9x14, and all sensor data are integrated to form a three-dimensional tensor in the form of Nx9x14, where N is the number of sensors, 9 is the number of sensor and surrounding environment data points, and 14 is the number of data features, which provides a guarantee for subsequent steps.

4. The real-time super-resolution method for urban wind fields based on sparse data according to claim 3 is characterized in that: The specific process of step three is as follows: for the three-dimensional tensor Nx9x14 generated in step two, it is regarded as a plane composed of "number of sensors" and "number of data features", and "number of sensor and surrounding environment data points" as a three-dimensional tensor of the data attribute dimension, and data feature extraction is performed on its data attribute dimension, that is, a two-dimensional convolution layer, a pooling layer, a batch normalization layer, and an activation function are used to reduce the data dimension of the dimension of "number of sensor and surrounding environment data points". Finally, this dimension is expressed with a single value, that is, the connection between the sensor and the surrounding environment is integrated, and the data structure is reduced to the two-dimensional Nx14 to provide a guarantee for subsequent steps.

5. The real-time super-resolution method for urban wind farms based on sparse data according to claim 4 is characterized in that: The specific process of step four is as follows: for the two-dimensional numerical matrix Nx14 generated in step three, the maximum value and mean value of the matrix are calculated in its "number of sensors" dimension, respectively, to obtain two one-dimensional vectors 14x1, thereby further extracting data features. The two one-dimensional vectors are then vertically spliced ​​to obtain a new one-dimensional vector of 28x2, providing a guarantee for subsequent steps.

6. The real-time super-resolution method for urban wind farms based on sparse data according to claim 5 is characterized in that: The specific process of step five is as follows: the residual graph convolutional neural network classifier consists of a graph feature capture neural network layer, 5 groups of convolution-batch normalization-activation modules, a neighboring point data integration layer, a point cloud and graph feature enhancement layer, 2 groups of linear mapping-batch normalization modules-activation-Dropout modules and a final layer of linear mapping units. During the training process, the residual graph convolutional neural network classifier judges the wind attack angle corresponding to the current wind location based on the input sensor and its surrounding environment data. During this process, the model will continuously try to extract the patterns in the input data and generate corresponding high-dimensional expressions within the neural network.

7. The real-time super-resolution method for urban wind farms based on sparse data according to claim 6 is characterized in that: The specific process of step six is ​​as follows: the graph feature encoder extracts high-dimensional graph feature information from the five end positions of the five groups of convolution-batch normalization-activation modules in step five, and uses an appropriate number of convolution-batch normalization-activation modules to re-extract the high-dimensional graph feature information and reshape the data structure so that it can be spliced ​​with the intermediate data at the corresponding depth position of the main neural network, thereby providing auxiliary information for the main network.

8. The real-time super-resolution method for urban wind farms based on sparse data according to claim 7 is characterized in that: The specific process of step seven is as follows: the physical information graph-assisted autoencoder receives sparse wind speed data from each sensor at the starting point of its neural network, and receives high-dimensional graph feature information from the graph feature encoder at different depths within its network; During the training phase, the physical information graph-assisted autoencoder uses the wind field data corresponding to the sparse data as the target and conducts training to reconstruct the wind field data at the corresponding moment based on the sparse data and high-dimensional graph feature information; After the training, the system will have the ability to reconstruct the urban wind farm in real time.

9. The real-time super-resolution method for urban wind farms based on sparse data according to claim 8, characterized in that: The specific process of step eight is: applying the two-dimensional fluid continuity equation to the wind field of the x and y direction wind speed components predicted by the physical information graph assisted autoencoder, that is, The generated wind field is constrained, and the model is induced to learn the revised prediction results, thereby further improving the prediction accuracy and robustness of the model.

Images