Three-dimensional wind field prediction method and system for complex terrain based on graph neural network

Abstract

The present application provides a method and system for predicting three-dimensional wind fields in complex terrain based on graph neural networks, and relates to the field of wind field prediction. The method for predicting three-dimensional wind fields in complex terrain includes: obtaining measurement data and terrain data, performing simulation, sampling and data fusion, obtaining input information to train a graph diffusion model; determining a first wind field distribution map and performing segmentation to construct multiple sub-graphs, and adding a central node to each sub-graph and performing information exchange to obtain a second wind field distribution map; based on a Graphformer model, updating node data by autoregression to obtain wind field distribution prediction information at the next moment. In the process of predicting wind field distribution based on a trained graph diffusion model, the present application does not require simulation processing, and makes calculations faster and more efficient based on a pure neural network algorithm. It promotes the learning and exchange of long-range information and global information by establishing a central node, and realizes long-term prediction of complex three-dimensional wind fields through a Graphformer model.

Description

Technical Field

[0001] The present application relates to the technical field of wind field prediction, and in particular to a method and system for predicting three-dimensional wind fields in complex terrain based on a graph neural network. Background Art

[0002] Wind farms are often built in hilly, mountainous and coastal areas with varied terrain because these areas can provide ideal wind resources. In these complex terrains, the sampling of wind field spatial data will be affected by the undulations of the terrain, and the distribution of sampling points will change with the fluctuations of the terrain. In order to ensure that the sampled data can meet the Nyquist sampling theorem, that is, the signal information can be fully restored during the sampling process to avoid signal distortion, higher sampling density is required in areas with drastic terrain changes. However, laser wind radar is subject to some technical limitations when acquiring wind field data. For example, the distance attenuation effect will cause the signal strength to decrease at long distances, affecting the accuracy of the data; the coverage of the scanning mode and scanning angle will also limit the sampling of the data. These factors will lead to uneven sampling of laser wind radar data in space.

[0003] In order to better handle these complex terrains and the technical limitations of laser wind radar, three-dimensional wind field data are usually organized into a non-uniform grid structure. This structure can better adapt to the undulating changes in terrain and conform to the sampling characteristics of laser wind radar, thereby improving the accuracy of wind field data analysis and processing. Most of the existing wind field modeling technology solutions are based on numerical simulation based on computational fluid dynamics, which consumes a lot of calculations. In the face of large-scale wind field structures, such as large wind farms, the calculation consumption is relatively large. In addition, the non-uniform structure of the wind field will bring great computational complexity to the numerical calculation, and the computational consumption will be further increased.

[0004] In related technologies, in the aggregation process of graph neural networks, there are problems such as difficulty in capturing long-range dependencies between nodes and insufficient global perception. Nodes that are far away usually find it difficult to capture each other's data relationships, and it is also difficult for nodes to perceive overall environmental changes, which poses a challenge for graph neural networks to learn complex wind field representations. Existing methods usually focus on building current 3D wind field models and are unable to predict future wind fields. Numerical simulation methods cannot give effective predictions due to the huge consumption of computing resources and the limitations of the tiny errors in each step of the simulation. Summary of the invention

[0005] In response to the shortcomings of the prior art, the present application provides a method and system for predicting three-dimensional wind fields in complex terrain based on graph neural networks, which solves the problems of unsatisfactory accuracy in complex three-dimensional wind field modeling and inaccurate wind field forecasting.

[0006] To achieve the above objectives, this application is implemented through the following technical solutions:

[0007] In a first aspect, an embodiment of the present application provides a method for predicting a three-dimensional wind field in complex terrain based on a graph neural network, the method comprising: obtaining measurement data and terrain data of a three-dimensional wind field in complex terrain, the terrain data comprising a digital elevation map and surface roughness; performing simulation, sampling and data fusion in accordance with physical constraints based on the terrain data to obtain input information for model training; performing model training based on the input information to obtain a graph diffusion model; inputting the measurement data and terrain data into the graph diffusion model to determine a first wind field distribution map at the current moment; wherein the first wind field distribution map is a non-uniform network comprising nodes and edges; segmenting the first wind field distribution map to construct multiple sub-graphs, and adding corresponding central nodes to each sub-graph, exchanging information through an attention mechanism to obtain a second wind field distribution map; based on a preset Graphformer model, updating the node data in the second wind field distribution map in the hidden state space through an autoregressive method to obtain wind field distribution prediction information at the next moment.

[0008] According to the first aspect of the embodiment of the present application, the aforementioned simulation, sampling and data fusion based on terrain data that complies with physical constraints to obtain input information for model training may specifically include the following steps: using terrain data as input data for fluid mechanics simulation, randomly selecting boundary conditions and performing numerical simulation based on physical equations to obtain simulation results; based on the simulation results, sampling is performed according to actual sampling positions to simulate the sampling of lidar in practice and obtain sampling data; the sampling data and terrain data are combined, and the combined data is determined as input information for model training.

[0009] According to the first aspect of the embodiment of the present application, the aforementioned model training is performed based on the input information to obtain the graph diffusion model, which may specifically include: comparing the predicted output of the trained intermediate model with the simulation output, calculating the data loss and the physical loss, and gradient backpropagating the training model to obtain the final graph diffusion model.

[0010] According to the first aspect of the embodiment of the present application, the first wind field distribution map is segmented to construct multiple sub-graphs, and a corresponding central node is added to each sub-graph, and information is exchanged through the attention mechanism to obtain a second wind field distribution map. Specifically, the following steps may be included: the structure of the first wind field distribution map is grouped, and a central node is added to each group node to promote information exchange between nodes; the graph nodes are decomposed into multiple groups of sub-graphs by applying the minimum cut algorithm multiple times to represent different local areas of interest; the information of each regional node is extracted and integrated into the corresponding central node through the attention mechanism, and the information of each central node is exchanged and updated through the attention mechanism; based on the attention mechanism, the updated information of the central node is propagated to each graph node in the same group, so that the graph nodes can learn long-range relationships and global relationships and obtain a second wind field distribution map.

[0011] According to the first aspect of the embodiment of the present application, the aforementioned method based on the preset Graphformer model updates the node data in the second wind field distribution map in the latent state space by autoregression to obtain the wind field distribution prediction information at the next moment, which may specifically include the following steps: encoding the data corresponding to each node in the second wind field distribution map into a first latent state space vector through a preset encoder to achieve data mapping; determining a latent space distribution map in the latent space corresponding to the first latent state space vector; performing multi-layer propagation aggregation on the latent space distribution map through the preset Graphformer model to obtain a second latent state space vector, wherein the Graphformer model is a Transformer that integrates spatial coding, edge coding and center coding; decoding the second latent state space vector into corresponding wind field data through a decoder corresponding to the encoder to obtain the wind field distribution prediction information at the next moment.

[0012] According to the first aspect of the embodiment of the present application, in the presence of an auxiliary meteorological large model, the aforementioned latent space distribution map that is in the latent space and corresponds to the first latent state space vector may specifically include the following steps: obtaining the correction value output by the auxiliary meteorological large model at the corresponding time; encoding the correction value into a latent space correction vector through an encoder; based on the cross-attention mechanism, associating the latent space correction vector to the first latent state space vector, and generating a latent space distribution map.

[0013] According to the first aspect of the embodiment of the present application, the complex terrain three-dimensional wind field prediction method based on graph neural network also includes: based on the wind field distribution prediction information at the next moment, using an autoregressive method to iteratively execute encoding, multi-layer propagation aggregation and decoding processes to obtain the wind field distribution prediction information at any moment.

[0014] In the second aspect, an embodiment of the present application provides a complex terrain three-dimensional wind field prediction system based on a graph neural network, and the complex terrain three-dimensional wind field prediction system includes an acquisition module, a data processing module, a training module, a determination module, a segmentation and exchange module, and an update prediction module; wherein the acquisition module is used to obtain measurement data and terrain data of a complex terrain three-dimensional wind field, and the terrain data includes a digital elevation map and surface roughness; the data processing module is used to perform simulation, sampling and data fusion in accordance with physical constraints based on the terrain data to obtain input information for model training; the training module is used to perform model training based on the input information to obtain a graph diffusion model; the determination module is used to input the measurement data and terrain data into the graph diffusion model to determine a first wind field distribution map at the current moment; wherein the first wind field distribution map is a non-uniform network containing nodes and edges; the segmentation and exchange module is used to segment the first wind field distribution map to construct multiple subgraphs, and add a corresponding central node to each subgraph, and exchange information through an attention mechanism to obtain a second wind field distribution map; the update prediction module is used to update the node data in the second wind field distribution map in the hidden state space through an autoregressive method based on a preset Graphformer model to obtain wind field distribution prediction information at the next moment.

[0015] In a third aspect, an embodiment of the present application provides an electronic device, comprising: a processor, a memory, and a program stored in the memory and executable on the processor, wherein when the program is executed by the processor, the method for predicting three-dimensional wind fields in complex terrain based on graph neural networks in the aforementioned first aspect is implemented.

[0016] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium storing a program or instruction. When the program or instruction is executed by a processor, the method for predicting three-dimensional wind fields in complex terrain based on a graph neural network in the aforementioned first aspect is implemented.

[0017] This application provides a complex terrain three-dimensional wind field prediction method and system based on graph neural network. Compared with the existing technology, it has the following beneficial effects:

[0018] In the first stage of training the graph diffusion model, the present application performs numerical simulation based on terrain data to better analyze and learn the representation of non-uniform wind fields; in the process of predicting wind field distribution based on the trained graph diffusion model, there is no need for simulation processing, and the pure neural network algorithm makes the calculation faster and more efficient, and resource consumption is greatly reduced; wherein, the simulation of the first stage considers the real physical constraints, so that the generated results conform to the real physical laws. In the process of generating the first wind field distribution map, real measurement data is used to provide accuracy, and the graph segmentation algorithm is used to segment the first wind field distribution map to construct multiple subgraphs, establish central nodes, and promote the learning and exchange of long-range information and global information; by updating the node data through the Graphformer model, the wind field distribution prediction information at the next moment can be obtained, realizing low-cost, high-precision and real-time three-dimensional wind field modeling and prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0020] Figure 1 It is a flow chart of a method for predicting three-dimensional wind fields in complex terrain based on a graph neural network provided in an embodiment of the present application;

[0021] Figure 2 yes Figure 1 An exemplary process diagram of S160;

[0022] Figure 3 It is a structural schematic diagram of a complex terrain three-dimensional wind field prediction system based on a graph neural network provided in an embodiment of the present application;

[0023] Figure 4 It is a structural schematic diagram of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0025] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0026] The embodiments of the present application solve the problems of unsatisfactory accuracy in complex three-dimensional wind field modeling and inaccurate wind field forecasting by providing a method and system for predicting three-dimensional wind fields in complex terrain based on a graph neural network.

[0027] The technical solution in the embodiment of the present application is to solve the above technical problems, and the overall idea is as follows:

[0028] Wind farms are often built in hilly, mountainous and coastal areas with varied terrain because these areas can provide ideal wind resources. In these complex terrains, the sampling of wind field spatial data will be affected by the undulations of the terrain, and the distribution of sampling points will change with the fluctuations of the terrain. In order to ensure that the sampled data can meet the Nyquist sampling theorem, that is, the signal information can be fully restored during the sampling process to avoid signal distortion, higher sampling density is required in areas with drastic terrain changes. However, laser wind radar is subject to some technical limitations when acquiring wind field data. For example, the distance attenuation effect will cause the signal strength to decrease at long distances, affecting the accuracy of the data; the coverage of the scanning mode and scanning angle will also limit the sampling of the data. These factors will lead to uneven sampling of laser wind radar data in space.

[0029] In order to better handle these complex terrains and the technical limitations of LiDAR, 3D wind data are usually organized into a non-uniform grid structure. This structure can better adapt to the undulating changes in terrain and conform to the sampling characteristics of LiDAR, thereby improving the accuracy of wind data analysis and processing.

[0030] In the related technologies, most of the existing wind farm modeling technology solutions are based on numerical simulation based on computational fluid dynamics, which consumes a lot of computation. In the face of large-scale wind farm structures, such as large wind farms, the computational consumption is relatively large. In addition, the non-uniform structure of the wind farm will bring great computational complexity to the numerical calculation, and the computational consumption will be further increased. In the aggregation process of graph neural networks, there are problems such as difficulty in capturing long-range dependencies between nodes and insufficient global perception. Nodes that are far away usually find it difficult to capture each other's data relationships, and it is also difficult for nodes to perceive overall environmental changes, which brings challenges to graph neural networks learning complex wind field representations. Existing methods usually focus on establishing the current 3D wind field model and cannot predict future wind fields. Numerical simulation methods cannot give effective long-term predictions due to huge computing resource consumption and are limited by the tiny errors in each step of the simulation; actual measurement methods essentially rely on real-time measurement data and do not have the ability to predict future wind field distribution. Therefore, related technologies have a great disadvantage in long-term wind field forecasting capabilities and are difficult to make accurate forecasts.

[0031] Specifically, wind farm modeling and prediction technologies mainly include: (1) methods based on the combination of sensor measurements and numerical simulation. On the one hand, since numerical simulation requires iterative solution of partial differential equations, it places high demands on computing power. Faced with large-scale wind farm modeling and simulation tasks, computers need to consume a lot of time to solve equations, and real-time and convenience are poor. On the other hand, non-uniform structure grid division is conducive to accurate data calculation and is suitable for non-uniform data structures in practical applications; however, this grid structure is very unfavorable for computer simulation, and complex algorithms are required to calculate numerical solutions, which further increases the computational complexity. (2) Relying only on measured data, the measured data is simply spliced ​​and interpolated to combine modeling. Although this method has good real-time performance, it has poor performance in wind farm accuracy and resolution, and can only judge the general trend but cannot accurately analyze it.

[0032] In terms of long-term wind forecasting, numerical simulation methods have small errors in each simulation step due to computing resource limitations, and cannot accurately predict long-term trends. The method of building a wind field model based solely on actual data can only build a current wind field model and cannot predict future trends.

[0033] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0034] The following first introduces a method for predicting three-dimensional wind fields in complex terrain based on a graph neural network provided in an embodiment of the present application.

[0035] The present application embodiment provides a flow chart of a method for predicting three-dimensional wind fields in complex terrain based on a graph neural network, such as Figure 1 As shown, the complex terrain three-dimensional wind field prediction method may include the following steps S110-S160.

[0036] S110, obtaining measurement data and terrain data of a three-dimensional wind field in a complex terrain, wherein the terrain data includes a digital elevation map and surface roughness.

[0037] S120, performing simulation, sampling and data fusion in accordance with physical constraints based on the terrain data to obtain input information for model training.

[0038] S130: Perform model training based on the input information to obtain a graph diffusion model.

[0039] S140, inputting the measurement data and the terrain data into a graph diffusion model to determine a first wind field distribution graph at a current moment; wherein the first wind field distribution graph is a non-uniform network including nodes and edges.

[0040] S150, segmenting the first wind field distribution map to construct multiple sub-maps, adding a corresponding central node to each sub-map, exchanging information through an attention mechanism, and obtaining a second wind field distribution map.

[0041] S160. Based on a preset Graphformer model, the node data in the second wind field distribution graph is updated in the hidden state space by an autoregressive method to obtain wind field distribution prediction information at the next moment.

[0042] The above is a specific implementation method of the complex terrain three-dimensional wind field prediction method based on graph neural network provided in the embodiment of the present application. The present application uses graph neural network for three-dimensional wind field modeling, and uses graph diffusion model to integrate the real measurement data, digital elevation map and surface roughness data of the complex terrain wind field to obtain the first wind field distribution map that conforms to the laws of physics. Then, the neural network corresponding to the first wind field distribution map is segmented, and multiple subgraphs and corresponding central nodes are constructed to facilitate the exchange of information. Finally, autoregressive prediction is performed based on the Graphformer model, and the output is the wind field distribution prediction information corresponding to the next prediction moment to achieve wind field prediction.

[0043] It can be understood that in the first stage of training the graph diffusion model, the present application performs numerical simulation based on terrain data to better analyze and learn the representation of non-uniform wind fields; in the process of predicting wind field distribution based on the trained graph diffusion model, there is no need for simulation processing, and the calculation based on the pure neural network algorithm makes the calculation faster and more efficient, and resource consumption is greatly reduced; among them, the simulation of the first stage takes into account the real physical constraints, so that the generated results conform to the real physical laws.

[0044] Furthermore, in the process of generating the first wind field distribution map, real measurement data is used to provide accuracy, and a graph segmentation algorithm is used to segment the first wind field distribution map to construct multiple sub-graphs, and a central node is established to promote the learning and exchange of long-range information and global information; the node data is updated through the Graphformer model, and the wind field distribution forecast information at the next moment can be obtained, thereby realizing low-cost, high-precision and real-time three-dimensional wind field modeling and prediction.

[0045] In some embodiments, the above-mentioned simulation, sampling and data fusion in accordance with physical constraints are performed based on terrain data to obtain input information for model training, that is, the above-mentioned S120 may specifically include the following steps:

[0046] S210, using the terrain data as input data for fluid mechanics simulation, randomly selecting boundary conditions and performing numerical simulation based on physical equations to obtain simulation results. It can be understood that the simulation results are obtained based on solving physical equations, strictly comply with physical constraints, and are used to train the model.

[0047] S220: Based on the simulation result, sampling is performed according to the actual sampling position to simulate the sampling of the laser radar in practice and obtain sampling data. It can be understood that the simulation result is sampled according to the actual sampling position to simulate real data.

[0048] S230: Combine the sampling data and the terrain data, and determine the combined data as input information for model training.

[0049] In the embodiment of the present application, in order to provide the map data of the initial wind field distribution, the present application first trains a map diffusion model for combining the measurement data and the terrain data to comprehensively generate the initial first wind field distribution map. In order to ensure that the results given by the map diffusion model conform to the physical laws, that is, the Navier-Stokes equations, the present application uses the simulation results of computational fluid dynamics as a data set in the training of the map diffusion model, and combines the physical loss term to constrain the model.

[0050] In one example, the aforementioned model training is performed based on the input information to obtain a graph diffusion model, that is, the aforementioned S130 may specifically include: comparing the predicted output of the trained intermediate model with the simulation output, calculating the data loss and the physical loss, and gradient back-transferring the training model to obtain the final graph diffusion model. It can be understood that the trained graph diffusion model can take real sampled data as input and learn implicit physical constraints at the same time.

[0051] In some embodiments, the first wind field distribution map is segmented to construct multiple sub-maps, and a corresponding central node is added to each sub-map, and information is exchanged through an attention mechanism to obtain a second wind field distribution map. That is, the aforementioned S150 may specifically include the following steps:

[0052] S310: Group the structure of the first wind field distribution diagram, and add a central node to each group node to promote information exchange between nodes.

[0053] S320, the graph nodes are decomposed into multiple groups of subgraphs by applying the minimum cut algorithm multiple times to represent different local focus areas.

[0054] S330, extracting and integrating the information of each regional node into the corresponding central node through the attention mechanism, and exchanging and updating the information of each central node through the attention mechanism.

[0055] S340. Based on the attention mechanism, the update information of the central node is propagated to each graph node in the same group, so that the graph nodes can learn long-range relationships and global relationships and obtain a second wind field distribution map.

[0056] In the embodiments of the present application, it can be understood that the present application solves the problems of difficulty in capturing long-range dependencies of graph nodes and insufficient global perception capabilities, and uses node grouping and establishment of central nodes to promote information extraction and exchange, thereby improving the information perception capabilities of graph nodes.

[0057] In one example, if Figure 2 As shown, based on the preset Graphformer model, the node data in the second wind field distribution graph is updated in the hidden state space by an autoregressive method to obtain the wind field distribution prediction information at the next moment, that is, the aforementioned S160 may specifically include the following steps:

[0058] S410. Encode the data corresponding to each node in the second wind field distribution diagram into a first latent state space vector through a preset encoder to achieve data mapping.

[0059] S420: Determine a latent space distribution map that is in the latent space and corresponds to the first latent state space vector.

[0060] S430. Perform multi-layer propagation aggregation on the latent space distribution graph through a preset Graphformer model to obtain a second latent state space vector, wherein the Graphformer model is a Transformer that integrates space coding, edge coding, and center coding.

[0061] S440. Decode the second latent state space vector into corresponding wind field data through a decoder corresponding to the encoder to obtain wind field distribution prediction information at the next moment.

[0062] In some embodiments, when there is an auxiliary meteorological large model, the aforementioned determination of the latent space distribution map in the latent space and corresponding to the first latent state space vector, that is, the aforementioned S420, may specifically include the following steps:

[0063] S421. Obtain the correction value output by the auxiliary meteorological large model at the corresponding time.

[0064] S422. Encode the correction value into a latent space correction vector through an encoder.

[0065] S423. Based on the cross attention mechanism, the latent space correction vector is associated with the first latent state space vector, and a latent space distribution map is generated.

[0066] In the embodiment of the present application, it can be understood that the auxiliary meteorological large model can adopt the Fengwu meteorological large model. By adopting the Fengwu meteorological large model as a correction, the error expansion in the autoregressive prediction process is avoided, making the long-term prediction results more reliable.

[0067] In some embodiments, the complex terrain three-dimensional wind field prediction method based on graph neural network also includes: S170, based on the wind field distribution prediction information at the next moment, iteratively executing encoding, multi-layer propagation aggregation and decoding processes in an autoregressive manner to obtain the wind field distribution prediction information at any moment.

[0068] In the embodiment of the present application, it can be understood that after obtaining the wind field distribution prediction information at the next moment, the wind field distribution prediction information at the next moment is used as new input information, and the autoregressive method is adopted to repeat the encoding, multi-layer propagation aggregation and decoding process corresponding to the aforementioned S410-S440. Finally, the wind field distribution prediction information at any moment can be obtained as needed, thereby realizing long-term prediction of complex three-dimensional wind fields.

[0069] In some embodiments, the present application provides a complex terrain three-dimensional wind field prediction system 500 based on graph neural network, such as Figure 3 As shown, the complex terrain three-dimensional wind field prediction system 500 may include the following modules:

[0070] An acquisition module 510 is used to acquire measurement data and terrain data of a three-dimensional wind field in a complex terrain, where the terrain data includes a digital elevation map and surface roughness;

[0071] The data processing module 520 is used to perform simulation, sampling and data fusion in accordance with physical constraints based on terrain data to obtain input information for model training;

[0072] A training module 530 is used to perform model training based on input information to obtain a graph diffusion model;

[0073] A determination module 540 is used to input the measurement data and the terrain data into the graph diffusion model to determine a first wind field distribution graph at a current moment; wherein the first wind field distribution graph is a non-uniform network including nodes and edges;

[0074] A segmentation and exchange module 550 is used to segment the first wind farm distribution map to construct multiple sub-maps, add a corresponding central node to each sub-map, and exchange information through an attention mechanism to obtain a second wind farm distribution map;

[0075] The update prediction module 560 is used to update the node data in the second wind field distribution graph in the hidden state space by autoregression based on the preset Graphformer model to obtain the wind field distribution prediction information at the next moment.

[0076] According to an embodiment of the present application, any multiple modules among the acquisition module 510, the data processing module 520, the training module 530, the determination module 540, the segmentation exchange module 550 and the update prediction module 560 can be combined into one module for implementation, or any one of the modules can be split into multiple modules. Alternatively, at least part of the functions of one or more of these modules can be combined with at least part of the functions of other modules and implemented in one module.

[0077] In some embodiments, the data processing module 520 may be specifically used to:

[0078] The terrain data is used as the input data of the fluid dynamics simulation, the boundary conditions are randomly selected and numerical simulation is performed based on the physical equations to obtain the simulation results;

[0079] Based on the simulation results, sampling is performed according to the actual sampling position to simulate the sampling of the actual laser radar and obtain sampling data;

[0080] The sampling data and the terrain data are combined, and the combined data is determined as input information for model training.

[0081] In some embodiments, the training module 530 can be specifically used to: compare the predicted output of the trained intermediate model with the simulation output, calculate the data loss and physical loss, and gradient backpropagate the training model to obtain the final graph diffusion model.

[0082] In some embodiments, the split exchange module 550 may be specifically used to:

[0083] The structure of the first wind field distribution map is grouped, and a central node is added to each group node to facilitate information exchange between nodes;

[0084] The graph nodes are decomposed into multiple groups of subgraphs by applying the minimum cut algorithm multiple times to represent different local areas of interest;

[0085] The information of each regional node is extracted and integrated into the corresponding central node through the attention mechanism, and the information of each central node is exchanged and updated through the attention mechanism;

[0086] Based on the attention mechanism, the updated information of the central node is propagated to each graph node in the same group, so that the graph nodes can learn long-range and global relationships and obtain the second wind field distribution map.

[0087] In some embodiments, the update prediction module 560 may specifically include:

[0088] The data encoding unit 561 is used to encode the data corresponding to each node in the second wind field distribution diagram into a first hidden state space vector through a preset encoder to achieve data mapping;

[0089] A distribution map determining unit 562, configured to determine a latent space distribution map in the latent space corresponding to the first latent state space vector;

[0090] The propagation aggregation unit 563 is used to perform multi-layer propagation aggregation on the latent space distribution graph through a preset Graphformer model to obtain a second latent state space vector, wherein the Graphformer model is a Transformer that integrates space coding, edge coding and center coding;

[0091] The vector decoding unit 564 is used to decode the second latent state space vector into corresponding wind field data through a decoder corresponding to the encoder to obtain wind field distribution prediction information at the next moment.

[0092] In some embodiments, the distribution map determining unit 562 may be specifically configured to:

[0093] Obtain the correction value output by the auxiliary meteorological model at the corresponding time;

[0094] Encode the correction value into a latent space correction vector through an encoder;

[0095] Based on the cross-attention mechanism, the latent space correction vector is associated with the first latent state space vector, and a latent space distribution map is generated.

[0096] In some embodiments, the complex terrain three-dimensional wind field prediction system 500 based on graph neural network can also include an iterative processing unit 570, which can be specifically used to: based on the wind field distribution prediction information at the next moment, iteratively execute encoding, multi-layer propagation aggregation and decoding processes in an autoregressive manner to obtain the wind field distribution prediction information at any moment.

[0097] Figure 3 Each module in the system shown has the function of implementing each step in the aforementioned complex terrain three-dimensional wind field prediction method based on graph neural network, and can achieve its corresponding technical effect. For the sake of concise description, it will not be repeated here.

[0098] In some embodiments, the present application provides an electronic device, the structure diagram of the electronic device is as follows Figure 4 shown.

[0099] The electronic device may include a processor 610 and a memory 620 storing computer program instructions.

[0100] Specifically, the processor 610 may include a central processing unit (CPU), or an application specific integrated circuit (ASIC), or may be configured to implement one or more integrated circuits of the embodiments of the present application.

[0101] The memory 620 may include a large capacity memory for data or instructions. By way of example and not limitation, the memory 620 may include a hard disk drive (HDD), a floppy disk drive, a flash memory, an optical disk, a magneto-optical disk, a magnetic tape, or a universal serial bus (USB) drive or a combination of two or more of these. In appropriate cases, the memory 620 may include a removable or non-removable (or fixed) medium. In appropriate cases, the memory 620 may be inside or outside the integrated gateway disaster recovery device. In a specific embodiment, the memory 620 is a non-volatile solid-state memory.

[0102] The memory 620 may include a read-only memory (ROM), a random access memory (RAM), a disk storage medium device, an optical storage medium device, a flash memory device, an electrical, optical or other physical / tangible memory storage device. Therefore, typically, the memory 620 includes one or more tangible (non-transitory) computer-readable storage media (e.g., a memory device) encoded with software including computer executable instructions, and when the software is executed (e.g., by one or more processors), it can perform the operations described in any of the complex terrain three-dimensional wind field prediction methods based on graph neural networks in the above-mentioned embodiments.

[0103] The processor 610 reads and executes computer program instructions stored in the memory 620 to implement any one of the complex terrain three-dimensional wind field prediction methods based on graph neural network in the above embodiments.

[0104] In one example, the electronic device may further include a communication interface 630 and a bus 600. Figure 4 As shown, the processor 610, the memory 620, and the communication interface 630 are connected via a bus 600 and communicate with each other.

[0105] The communication interface 630 is mainly used to implement communication between various modules, devices, units and / or equipment in the embodiments of the present application.

[0106] Bus 600 includes hardware, software or both, and the parts of online data flow billing equipment are coupled to each other. For example, but not limitation, bus may include accelerated graphics port (AGP) or other graphics bus, enhanced industrial standard architecture (EISA) bus, front-end bus (FSB), hypertransport (HT) interconnection, industrial standard architecture (ISA) bus, infinite bandwidth interconnection, low pin count (LPC) bus, memory bus, micro channel architecture (MCA) bus, peripheral component interconnection (PCI) bus, PCI-Express (PCI-X) bus, serial advanced technology attachment (SATA) bus, video electronics standard association local (VLB) bus or other suitable bus or two or more of these combinations. In appropriate cases, bus 600 may include one or more buses. Although the present application embodiment describes and shows a specific bus, the present application considers any suitable bus or interconnection.

[0107] In addition, in combination with the complex terrain three-dimensional wind field prediction method based on graph neural network in the above embodiment, the embodiment of the present application can provide a computer storage medium for implementation. The computer storage medium stores computer program instructions; when the computer program instructions are executed by the processor, any one of the complex terrain three-dimensional wind field prediction methods based on graph neural network in the above embodiment is implemented.

[0108] It should be clear that the present application is not limited to the specific configuration and processing described above and shown in the figures. For the sake of simplicity, a detailed description of the known method is omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present application is not limited to the specific steps described and shown, and those skilled in the art can make various changes, modifications and additions, or change the order between the steps after understanding the spirit of the present application.

[0109] The functional blocks shown in the above block diagram can be implemented as hardware, software, firmware or a combination thereof. When implemented in hardware, it can be, for example, an electronic circuit, an application specific integrated circuit (ASIC), appropriate firmware, a plug-in, a function card, etc. When implemented in software, the elements of the present application are programs or code segments that are used to perform the required tasks. The program or code segment can be stored in a machine-readable medium, or transmitted on a transmission medium or a communication link by a data signal carried in a carrier wave. "Machine-readable medium" can include any medium capable of storing or transmitting information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, ROM, flash memory, erasable ROM (EROM), floppy disks, CD-ROMs, optical disks, hard disks, optical fiber media, radio frequency (RF) links, etc. The code segment can be downloaded via a computer network such as the Internet, an intranet, etc.

[0110] It should also be noted that the exemplary embodiments mentioned in this application describe some methods or systems based on a series of steps or devices. However, this application is not limited to the order of the above steps, that is, the steps can be performed in the order mentioned in the embodiment, or in a different order from the embodiment, or several steps can be performed simultaneously.

[0111] Aspects of the present disclosure are described above with reference to the flowchart and / or block diagram of the method, device (system) and computer program product according to the embodiment of the present disclosure. It should be understood that each box in the flowchart and / or block diagram and the combination of each box in the flowchart and / or block diagram can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device to produce a machine so that these instructions executed by the processor of the computer or other programmable data processing device enable the implementation of the function / action specified in one or more boxes of the flowchart and / or block diagram. Such a processor can be, but is not limited to, a general-purpose processor, a special-purpose processor, a special application processor, or a field programmable logic circuit. It can also be understood that each box in the block diagram and / or flowchart and the combination of boxes in the block diagram and / or flowchart can also be implemented by dedicated hardware that performs a specified function or action, or can be implemented by a combination of dedicated hardware and computer instructions.

[0112] In summary, compared with the prior art, this application has the following beneficial effects:

[0113] 1. In the first stage of training the graph diffusion model, this application performs numerical simulation based on terrain data to better analyze and learn the representation of non-uniform wind fields. In the simulation process, real physical constraints are taken into account so that the generated results conform to real physical laws. In the process of predicting wind field distribution based on the trained graph diffusion model, there is no need for simulation processing. The pure neural network algorithm makes the calculation faster and more efficient, and resource consumption is greatly reduced.

[0114] 2. This application uses a graph segmentation algorithm to segment the first wind field distribution map to construct multiple sub-graphs, establishes a central node, promotes the learning and exchange of long-range information and global information, improves the information perception ability of graph nodes, and solves the problem of difficulty in capturing long-range dependency relationships of graph nodes and insufficient global perception ability.

[0115] 3. This application updates the node data through the Graphformer model, and can obtain the wind field distribution prediction information at the next moment, realizing low-cost, high-precision and real-time three-dimensional wind field modeling and prediction; after obtaining the wind field distribution prediction information at the next moment, the autoregressive method is used to iteratively execute the encoding, multi-layer propagation aggregation and decoding process, which can obtain the wind field distribution prediction information at any time, and realize long-term prediction of complex three-dimensional wind fields.

[0116] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A three-dimensional wind field prediction method for complex terrain based on graph neural network, characterized in that: include: Acquire measurement data and terrain data of a three-dimensional wind field in complex terrain, wherein the terrain data includes a digital elevation map and surface roughness; Performing simulation, sampling and data fusion in accordance with physical constraints based on the terrain data to obtain input information for model training; Performing model training based on the input information to obtain a graph diffusion model; Inputting the measurement data and the terrain data into the graph diffusion model to determine a first wind field distribution graph at a current moment; wherein the first wind field distribution graph is a non-uniform network including nodes and edges; The first wind field distribution map is segmented to construct a plurality of sub-maps, and a corresponding central node is added to each of the sub-maps, and information is exchanged through an attention mechanism to obtain a second wind field distribution map; Based on the preset Graphformer model, the node data in the second wind field distribution graph is updated in the hidden state space by an autoregressive method to obtain the wind field distribution prediction information at the next moment.

2. The complex terrain three-dimensional wind field prediction method based on graph neural network according to claim 1 is characterized in that: The simulation, sampling and data fusion in accordance with physical constraints are performed based on the terrain data to obtain input information for model training, including: Using the terrain data as input data for fluid mechanics simulation, randomly selecting boundary conditions and performing numerical simulation based on physical equations to obtain simulation results; Based on the simulation results, sampling is performed according to the actual sampling position to simulate the sampling of the laser radar in practice and obtain sampling data; The sampling data and the terrain data are combined, and the combined data is determined as input information for model training.

3. The complex terrain three-dimensional wind field prediction method based on graph neural network according to claim 2, characterized in that: The performing model training based on the input information to obtain a graph diffusion model includes: The predicted output of the trained intermediate model is compared with the simulation output, the data loss and physical loss are calculated, and the gradient is back-transferred to the training model to obtain the final graph diffusion model.

4. The complex terrain three-dimensional wind field prediction method based on graph neural network according to claim 1, characterized in that: The first wind field distribution map is segmented to construct a plurality of sub-maps, and a corresponding central node is added to each of the sub-maps, and information is exchanged through an attention mechanism to obtain a second wind field distribution map, including: Grouping the structure of the first wind field distribution map, and adding a central node to each group node to facilitate information exchange between nodes; The graph nodes are decomposed into multiple groups of subgraphs by applying the minimum cut algorithm multiple times to represent different local areas of interest; The information of each regional node is extracted and integrated into the corresponding central node through the attention mechanism, and the information of each central node is exchanged and updated through the attention mechanism; Based on the attention mechanism, the update information of the central node is propagated to each graph node in the same group, so that the graph nodes can learn long-range relationships and global relationships and obtain a second wind field distribution map.

5. The complex terrain three-dimensional wind field prediction method based on graph neural network according to claim 1, characterized in that: The method of updating the node data in the second wind field distribution graph in the hidden state space by an autoregressive method based on the preset Graphformer model to obtain the wind field distribution prediction information at the next moment includes: Encoding the data corresponding to each node in the second wind field distribution diagram into a first hidden state space vector through a preset encoder to achieve data mapping; Determine a latent space distribution map in the latent space and corresponding to the first latent state space vector; Perform multi-layer propagation aggregation on the latent space distribution graph through a preset Graphformer model to obtain a second latent state space vector, wherein the Graphformer model is a Transformer that integrates space coding, edge coding and center coding; The second latent state space vector is decoded into corresponding wind field data by a decoder corresponding to the encoder to obtain wind field distribution prediction information at the next moment.

6. The complex terrain three-dimensional wind field prediction method based on graph neural network according to claim 5, characterized in that: In the case where there is an auxiliary meteorological large model, the determining of a latent space distribution map in the latent space and corresponding to the first latent state space vector includes: Obtaining the correction value output by the auxiliary meteorological large model at the corresponding time; encoding the correction value into a latent space correction vector by the encoder; Based on the cross attention mechanism, the latent space correction vector is associated with the first latent state space vector, and a latent space distribution map is generated.

7. The method for predicting three-dimensional wind fields in complex terrain based on graph neural network according to claim 5, characterized in that: Also includes: Based on the wind field distribution prediction information at the next moment, an autoregressive method is used to iteratively perform encoding, multi-layer propagation aggregation and decoding processes to obtain the wind field distribution prediction information at any moment.

8. A complex terrain three-dimensional wind field prediction system based on graph neural network, characterized in that: include: An acquisition module, used to acquire measurement data and terrain data of a three-dimensional wind field in a complex terrain, wherein the terrain data includes a digital elevation map and surface roughness; A data processing module, used for performing simulation, sampling and data fusion in accordance with physical constraints based on the terrain data to obtain input information for model training; A training module, used for performing model training based on the input information to obtain a graph diffusion model; A determination module, used for inputting the measurement data and the terrain data into the graph diffusion model to determine a first wind field distribution graph at a current moment; wherein the first wind field distribution graph is a non-uniform network including nodes and edges; A segmentation and exchange module, used for segmenting the first wind farm distribution map to construct multiple sub-maps, adding a corresponding central node to each of the sub-maps, and exchanging information through an attention mechanism to obtain a second wind farm distribution map; The update prediction module is used to update the node data in the second wind field distribution map in the hidden state space through an autoregressive method based on a preset Graphformer model to obtain the wind field distribution prediction information at the next moment.

9. An electronic device, characterized in that: include: A processor, a memory, and a program stored in the memory and executable on the processor, wherein the program, when executed by the processor, implements a complex terrain three-dimensional wind field prediction method based on a graph neural network as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores programs or instructions, and when the programs or instructions are executed by the processor, the method for predicting three-dimensional wind fields in complex terrain based on graph neural networks as described in any one of claims 1 to 7 is implemented.

Images