A wind farm reconstruction and prediction method and device, electronic equipment and storage medium

Abstract

The application discloses a wind field reconstruction and prediction method and device, electronic equipment and storage medium, and belongs to the technical field of wind field data processing. The method comprises: uniformly encoding historical wind speed data and real-time wind speed data of multiple height layers, so that the model can mine the vertical coupling linkage relationship between wind fields at different height layers and the time evolution law of the wind field; geometric prior information and observation masks are introduced, so that the model can complete wind field flow constraint analysis combined with regional building layout and divide the measured data area and the blank reconstruction area; when the model finally completes global wind field deduction based on multi-dimensional fusion features, not only the accuracy of the wind field reconstruction result is improved, but also multi-step fine wind field prediction is realized, the disadvantages of single wind field deduction dimension, lack of environmental constraints and insufficient prediction accuracy of the traditional technology are solved, and the actual needs of fine wind field reconstruction and short-term accurate wind field prediction in a complex building group scene are met.

Description

Technical Field

[0001] This application relates to the field of wind field data processing technology, and in particular to a wind field reconstruction and prediction method, apparatus, electronic device and storage medium. Background Technology

[0002] In the process of environmental monitoring and wind field simulation, in order to accurately obtain the complete airflow distribution status of different regions across the entire area, so as to meet the practical application needs such as street environment assessment, regional airflow disaster prediction or meteorological travel guidance, it is usually necessary to collect on-site wind speed data based on the deployed wind speed monitoring points, and combine the actual spatial environment of the region to complete the missing wind speed data. At the same time, the wind field development trend in subsequent periods is predicted by combining the past wind field change patterns, so as to obtain dense wind field information with full coverage.

[0003] However, when traditional technologies are used to reconstruct wind fields in various regions and predict wind fields in the short term, they usually only use the scattered and sparse wind speed data collected from a limited number of monitoring points for simple linear interpolation and completion. They do not combine the building layout in the region and the vertical coupling and linkage relationship between wind fields at different heights to carry out constraint analysis, nor can they integrate wind speed change data from multiple consecutive historical moments to explore the evolution trend of wind fields.

[0004] Because traditional techniques rely on only a small amount of scattered wind speed data to make rough calculations, the wind field fitting results in complex areas such as the perimeter of buildings and street passages are prone to deviate significantly from the actual flow state. This not only fails to guarantee the accuracy of the current full-domain wind field reconstruction results, but also makes it difficult to achieve accurate wind field prediction over multiple steps. Consequently, it cannot meet the actual needs of refined wind field reconstruction and short-term accurate wind field prediction in complex building cluster scenarios. Summary of the Invention

[0005] The main purpose of this application is to propose a wind field reconstruction and prediction method, device, electronic device and storage medium. By introducing multi-dimensional features and feature fusion analysis, it effectively avoids the problem that the wind field fitting results in complex areas such as building perimeters and street passages deviate from the actual airflow state, significantly improves the accuracy of the whole-domain wind field reconstruction results, and realizes multi-step fine wind field prediction.

[0006] To achieve the above objectives, one aspect of this application proposes a wind field reconstruction and prediction method, the method comprising: The current sparse wind speed observation data corresponding to different heights in the target area and the historical sparse wind speed observation data corresponding to each height at multiple consecutive historical moments are normalized and mapped to encode multi-channel two-dimensional wind field feature images corresponding to different moments. Obtain the geometric prior information of the target area and the observation mask corresponding to each height; wherein, the geometric prior information is used to characterize the building geometry of the target area; the observation mask is used to mark the observation points with wind speed observation data and the points to be reconstructed without wind speed observation data; The multi-channel two-dimensional wind field feature images, the geometric prior information, and each of the observation masks are input into a trained neural network model, so that the neural network model extracts sparse observation distribution features, multi-height correlation features, building geometric constraint features, and wind field temporal evolution features based on the geometric prior information, the multi-channel two-dimensional wind field feature images, and the observation masks; wherein, the sparse observation distribution features are used to indicate the distribution relationship between the effective observation area and the area to be reconstructed; the multi-height correlation features are used to indicate the vertical coupling relationship of the wind field at different height layers; Based on the sparse observation distribution characteristics, the multi-height correlation characteristics, the building geometric constraint characteristics, and the wind field temporal evolution characteristics, the dense wind field reconstruction result corresponding to the current moment and the dense wind field prediction result for multiple consecutive future moments are generated.

[0007] Furthermore, in some embodiments, the step of extracting sparse observation distribution features, multi-height correlation features, building geometric constraint features, and wind field temporal evolution features based on the geometric prior information, each of the multi-channel two-dimensional wind field feature images, and each of the observation masks includes: The neural network model stitches together the geometric prior information, the multi-channel two-dimensional wind field feature images, and the observation masks along the channel dimension to generate a multi-channel input tensor. Feature extraction is performed on the multi-channel input tensor to output local and global features; wherein, the local features are used to indicate the local variation characteristics between the observation point, building boundary and wind field; the global features are used to indicate the wind field evolution characteristics at different heights within the target area; Cross-scale feature fusion is performed on the local features and the global features to output the sparse observation distribution features, the multi-height correlation features, the building geometric constraint features, and the wind field temporal evolution features.

[0008] Furthermore, in some embodiments, the neural network model includes: an encoder; The process of generating the local features includes: The encoder performs block processing on the multi-channel input tensor to divide it into several local feature blocks, and then the features of each local feature block are fused to generate a joint feature map. The first feature extraction network of the encoder performs feature recognition on each local grid region in the joint feature map to generate observation point distribution features; wherein, the observation point distribution features are used to characterize the wind field arrangement pattern between the observation points and the points to be reconstructed. The building outline in the joint feature map is captured by the first feature extraction network to generate building edge features; wherein, the building edge features are used to characterize the blocking effect of the building edge on the wind field; The first feature extraction network is used to fit the local wind field fluctuations in the joint feature map to generate wind field change features; wherein, the wind field change features are used to characterize the wind field difference between any two grid regions. The distribution characteristics of the observation points, the building edge characteristics, and the wind field variation characteristics are output as the local features.

[0009] Furthermore, in some embodiments, the global features include: vertical correlation features for characterizing the spatial evolution of the upper-level wind field and the lower-level wind field, and temporal correlation features for characterizing the temporal evolution of the wind fields at different altitudes. The process of generating the global features includes: The joint feature map is fully identified by the second feature extraction network of the encoder, and the spatial dependency between the upstream inflow region and the downstream wake region of the wind field is constructed. The long-distance dependency between wind fields at different heights is also constructed. Based on the spatial dependency and the long-distance dependency, the vertical correlation feature is generated. The second feature extraction network identifies the wind field state at different times in the joint feature map, quantifies the evolution trend of the wind field over time, and generates the time-related features based on the evolution trend.

[0010] Furthermore, in some embodiments, the neural network model further includes a decoder; The cross-scale feature fusion of the local and global features, outputting the sparse observation distribution features, the multi-height correlation features, the building geometric constraint features, and the wind field temporal evolution features, includes: The decoder performs upsampling of the observation point distribution features, the building edge features, and the wind field change features at each level to generate the sparse observation distribution features. The decoder performs hierarchical correlation between the vertical correlation features and the wind field change features, quantifies the linkage mapping relationship between wind fields at different altitudes, and generates the multi-altitude correlation features based on the linkage mapping relationship. The decoder fuses the building edge features with the vertical correlation features to quantify the building's constraint strength on the wind field, and generates the building's geometric constraint features based on the constraint strength; wherein, the building's geometric constraint features are used to characterize the degree of influence of building distribution on wind field flow. The wind field temporal evolution features are generated by aggregating the vertical correlation features and the temporal correlation features through the decoder.

[0011] Furthermore, in some embodiments, the generation process of the dense wind field reconstruction results and each of the dense wind field prediction results includes: The neural network model is used to perform cross-correlation calculations between the sparse observation distribution features and the building geometric constraint features to generate a first wind field coupling feature that characterizes the degree of synergistic influence of the observation point layout and the building spatial layout on the wind field distribution. The dense wind field reconstruction result is generated by performing grid interpolation and completion based on the sparse observation distribution characteristics, the building geometric constraint characteristics, and the first wind field coupling characteristics using the neural network model. The neural network model is used to perform a spatiotemporal linkage mapping between the building geometric constraint features and the wind field temporal evolution features to generate a second wind field coupling feature that characterizes the wind field temporal variation characteristics under building constraint conditions. The multi-height correlation features, the wind field temporal evolution features, and the second wind field coupling features are input into the temporal inference network of the neural network model so that the temporal inference network can deduce the prediction results of each of the dense wind fields.

[0012] Furthermore, in some embodiments, the training process of the neural network model includes: Acquire sample data from several regions, and simultaneously acquire labels for the reconstructed real dense wind field and several labels for the predicted real dense wind field corresponding to each sample data from a region; wherein, the sample data from the regions includes: multi-channel two-dimensional wind field feature image samples, geometric prior information samples, and several observation mask samples. Using sample data from each region as model input, and the labels of the reconstructed real dense wind field corresponding to each sample data from each region as well as several labels of the predicted real dense wind field as supervision labels, the neural network model to be trained is iteratively trained until the loss error of the neural network model meets the preset convergence threshold, and the trained neural network model is output.

[0013] To achieve the above objectives, another aspect of this application proposes a wind field reconstruction and prediction device, the device comprising: The feature image generation module is used to normalize and map the current sparse wind speed observation data corresponding to different heights in the target area and the historical sparse wind speed observation data corresponding to each height at multiple consecutive historical moments, and encode the multi-channel two-dimensional wind field feature images corresponding to different moments. The model input data acquisition module is used to acquire the geometric prior information of the target area and the observation mask corresponding to each height; wherein, the geometric prior information is used to characterize the building geometric characteristics of the target area; the observation mask is used to mark the observation points with wind speed observation data and the points to be reconstructed without wind speed observation data; The result output module is used to input the multi-channel two-dimensional wind field feature image, the geometric prior information, and each of the observation masks into a trained neural network model, so that the neural network model extracts sparse observation distribution features, multi-height correlation features, building geometric constraint features, and wind field temporal evolution features based on the geometric prior information, each of the multi-channel two-dimensional wind field feature images, and each of the observation masks; wherein, the sparse observation distribution features are used to indicate the distribution relationship between the effective observation area and the area to be reconstructed; the multi-height correlation features are used to indicate the vertical coupling relationship of the wind field at different height layers; The result output module is also used to generate the dense wind field reconstruction result corresponding to the current moment and predict the dense wind field prediction result for multiple consecutive future moments by using the neural network model based on the sparse observation distribution characteristics, the multi-height correlation characteristics, the building geometric constraint characteristics and the wind field temporal evolution characteristics.

[0014] To achieve the above objectives, another aspect of this application proposes an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the wind field reconstruction and prediction method described above.

[0015] To achieve the above objectives, another aspect of the embodiments of this application proposes a storage medium storing a computer program that, when executed by a processor, implements the wind field reconstruction and prediction method described above.

[0016] The embodiments of this application include at least the following beneficial effects: This application provides a wind field reconstruction and prediction method, device, electronic device and storage medium. The scheme normalizes and maps the current sparse wind speed observation data corresponding to different heights and the historical sparse wind speed observation data of multiple consecutive historical moments at each height, and encodes them into a multi-channel two-dimensional wind field feature image. At the same time, it obtains geometric prior information that can characterize the geometric characteristics of buildings in the target area, as well as an observation mask that can distinguish between observation points and points to be reconstructed. The above-mentioned multi-channel two-dimensional wind field feature image, geometric prior information and observation mask are uniformly input into a trained neural network model. The neural network model is used to fuse multiple types of information to extract sparse observation distribution features, multi-height correlation features, building geometric constraint features and wind field temporal evolution features. Finally, based on the above four types of features, the dense wind field reconstruction result at the current moment and the dense wind field prediction result at multiple consecutive future moments are generated respectively. Compared to traditional techniques that rely solely on scattered and sparse wind speed data for simple linear interpolation, this application utilizes unified encoding of historical and real-time wind speed data across multiple height levels. This allows for the full exploration of vertical coupling relationships between wind fields at different heights and the long-term evolution of wind fields. Furthermore, the introduction of geometric prior information enables the model to incorporate regional building layouts for wind flow constraint analysis. Based on observation masks, the model can effectively learn and differentiate between measured data areas and data gap reconstruction areas. This effectively avoids the problem of wind field fitting results deviating from actual airflow in complex areas such as building perimeters and street passages when the model performs global wind field extrapolation based on multi-dimensional fusion features. This significantly improves the accuracy of global wind field reconstruction results and enables multi-step, refined wind field prediction. It overcomes the shortcomings of traditional techniques, such as single-dimensional wind field extrapolation, lack of environmental constraints, and insufficient prediction accuracy, thus meeting the practical needs for refined wind field reconstruction and short-term accurate wind field prediction in complex building cluster scenarios. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a wind field reconstruction and prediction method provided in an embodiment of this application; Figure 2 This is a schematic diagram illustrating the training and application process of the neural network model provided in the embodiments of this application; Figure 3 This is a schematic diagram of the structure of a wind field reconstruction and prediction device provided in an embodiment of this application; Figure 4 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit it. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this application; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims.

[0019] It is understood that the terms “first,” “second,” etc., used in this application may be used herein to describe various concepts, but unless otherwise stated, these concepts are not limited by these terms. These terms are only used to distinguish one concept from another. For example, without departing from the scope of the embodiments of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the words “if,” “when,” or “in response to a determination” as used herein may be interpreted as “when…” or “when…” or “in response to a determination.”

[0020] As used in this application, the terms "several", "each", etc., "several" include one, two or more, "each" refers to each of the corresponding plurality, and "any" refers to any one of the plurality.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0022] Traditional wind field reconstruction and prediction techniques typically rely on scattered and sparse wind speed data collected from a limited number of monitoring points. They use simple linear interpolation to fill in missing data and then combine this with data from a single moment to predict subsequent wind field trends. Because they fail to consider the geometric layout of buildings in the area, they ignore the obstruction and diversion effects of buildings on the wind field and cannot account for the vertical coupling and linkage between wind fields at different heights. This results in wind field fitting results that deviate significantly from reality in complex areas such as building perimeters and street passages. Furthermore, traditional techniques do not integrate wind speed variation data from multiple consecutive historical moments, making it impossible to accurately uncover the temporal evolution patterns of the wind field and achieve accurate wind field predictions over multiple time steps. Therefore, traditional techniques cannot guarantee the accuracy of current overall wind field reconstruction results, nor can they meet the needs for refined wind field analysis and short-term accurate prediction in complex building cluster scenarios.

[0023] In view of this, the present invention provides a wind field reconstruction and prediction method. This method encodes real-time and historical sparse wind speed data at multiple heights to generate multi-channel two-dimensional wind field feature images corresponding to different times. It then integrates regional geometric prior information and observation masks to extract multi-dimensional wind field features based on a neural network model. This approach goes beyond simple linear interpolation based solely on scattered sparse wind speed data. Instead, it incorporates geometric prior information related to regional building layouts into the entire wind field calculation process, fully considering the obstruction and guidance effects of building structures on airflow formation. This corrects the problem of wind field fitting results deviating from reality in complex areas such as building perimeters and street passages. Furthermore, based on multi-dimensional wind field characteristics, it fully considers the vertical coupling and linkage relationship between wind fields at different heights, fills the gap in the spatial inter-layer wind field correlation logic missing in traditional technologies, and, by leveraging the temporal evolution patterns mined from multiple historical wind speed data, it abandons the crude mode of predicting wind field trends based on data at a single moment, ultimately outputting high-precision dense wind field reconstruction results for the current moment. It can also combine temporal extrapolation logic to achieve multi-step long wind field prediction for multiple consecutive future moments, improving the accuracy, spatiotemporal continuity, and scene adaptability of wind field data. Thus, it can meet the practical application needs of refined wind field analysis and short-term accurate wind field prediction in complex building cluster scenarios.

[0024] Figure 1 This is an optional flowchart of a wind field reconstruction and prediction method provided in an embodiment of this application. Figure 1 The method may include, but is not limited to, steps S1 to S3: Step S1: Normalize and map the current sparse wind speed observation data corresponding to different heights in the target area and the historical sparse wind speed observation data corresponding to each height at multiple consecutive historical moments to encode the multi-channel two-dimensional wind field feature image corresponding to different moments. Indicatively, normalization mapping can transform scattered, heterogeneous, sparse wind speed data into standardized feature images suitable for neural network processing. This eliminates dimensional differences in wind speed data at different altitudes and times, improving the accuracy of subsequent feature extraction. Understandably, normalization mapping typically maps wind speed values ​​to the [0,1] interval, preserving the relative differences in wind speed data while avoiding the interference of extreme values ​​in subsequent calculations.

[0025] Step S2: Obtain the geometric prior information of the target area and the observation mask corresponding to each height; wherein, the geometric prior information is used to characterize the building geometry of the target area; the observation mask is used to mark the observation points with wind speed observation data and the points to be reconstructed without wind speed observation data; Indicatively, geometric prior information is the basic data that reflects the constraining effect of buildings on the wind field in the target area. The observation mask can accurately divide the effective observation area and the blank extrapolation area, providing a basis for positioning for the interpolation completion of subsequent wind field reconstruction.

[0026] Step S3: Input the multi-channel two-dimensional wind field feature image, the geometric prior information, and each of the observation masks into the trained neural network model, so that the neural network model extracts sparse observation distribution features, multi-height correlation features, building geometric constraint features, and wind field temporal evolution features based on the geometric prior information, each of the multi-channel two-dimensional wind field feature images, and each of the observation masks; wherein, the sparse observation distribution features are used to indicate the distribution relationship between the effective observation area and the area to be reconstructed; the multi-height correlation features are used to indicate the vertical coupling relationship of the wind field at different height layers; Based on the sparse observation distribution characteristics, the multi-height correlation characteristics, the building geometric constraint characteristics, and the wind field temporal evolution characteristics, the dense wind field reconstruction result corresponding to the current moment and the dense wind field prediction result for multiple consecutive future moments are generated.

[0027] Indicatively, this step is the core step in achieving accurate wind field reconstruction and prediction. Through a trained neural network model, multi-dimensional input data is fused to extract core features that can characterize wind field distribution, vertical correlation, building constraints, and temporal changes. Based on the extracted features, accurate wind field reconstruction and prediction are performed.

[0028] Steps S1 to S3 of this invention can solve the shortcomings of traditional technologies that rely solely on scattered and sparse wind speed data for simple interpolation and fail to integrate multi-dimensional data. This invention, through normalized encoding, transforms sparse data at multiple heights and times into standardized feature images. Combined with geometric prior information and observation masks, it achieves the synergistic utilization of multi-dimensional data, significantly improving the completeness and effectiveness of wind field data. Furthermore, this invention uses a neural network model to automatically perform feature recognition and extraction, accurately capturing the distribution of observation points, multi-height wind field coupling, building constraints, and temporal evolution patterns. This avoids the wind field fitting deviations caused by traditional technologies ignoring building influences, height correlations, and temporal changes. Simultaneously, it achieves dense wind field reconstruction at the current moment and wind field prediction at multiple future moments, solving the problems of separation between reconstruction and prediction and low prediction accuracy in traditional technologies. This meets the practical needs of refined wind field analysis and short-term prediction in complex building cluster scenarios, enhancing the practicality and application value of wind field data.

[0029] For example, this invention uses a certain urban area as the target area, and sets up wind speed monitoring points at three altitude levels in the urban area: 20m (low altitude), 60m (mid-to-high altitude), and 100m (high altitude), with eight sparse monitoring points at each altitude level. The invention collects sparse wind speed observation data from the eight monitoring points at the three altitude levels at the current moment, and simultaneously collects sparse wind speed observation data for each altitude level for five consecutive historical time periods.

[0030] All collected wind speed data are normalized and mapped, and finally encoded to generate multi-channel two-dimensional wind field feature images corresponding to 6 time points and 3 height levels. The pixels of each feature image correspond to the grid position of the target area, and the pixel value corresponds to the normalized wind speed data.

[0031] The system acquires geometric prior information about the urban area, which may include geometric parameters such as the height, outline boundary, building spacing, and building orientation of all buildings within the area. For example, an office building is 150m high with a rectangular outline, and its spacing from adjacent buildings is 30m. For three height levels (20m, 60m, and 100m), corresponding observation masks are generated. The mask is a two-dimensional matrix with the same size as the multi-channel two-dimensional wind field feature image. When marked as 1, it represents a monitoring point with wind speed observation data; when marked as 0, it represents a point to be reconstructed without wind speed observation data. This achieves accurate division between the effective observation area and the area to be reconstructed.

[0032] The multi-channel two-dimensional wind field feature images generated at the above six time points, along with the acquired geometric prior information of the urban area and the observation masks corresponding to the three height layers, are input into the trained neural network model. The neural network model extracts four core features through its built-in feature extraction mechanism: sparse observation distribution features, which can clarify the spatial distribution relationship between the eight monitoring points and the remaining points to be reconstructed; multi-height correlation features, which can represent the coupling relationship between wind speeds at 20m and 60m heights, as well as the influence of high-altitude wind speeds on low-altitude wind speeds; building geometric constraint features, which can represent the intensity of the blocking and guiding effect of a 150m-high office building on the surrounding wind field; and wind field temporal evolution features, which can represent the wind speed change trend from five historical moments to the current moment, such as the pattern of gradually increasing wind speed or clockwise wind direction deflection.

[0033] Based on the four core features extracted above, the neural network model can generate a full-area dense wind field reconstruction result for the urban area at the current moment. It can cover all grid locations, including the wind speed data of the original points to be reconstructed, and present the actual wind field distribution in complex areas such as building perimeters and street passages. For example, the distribution shows that the wind speed decreases on the leeward side of office buildings and increases on the windward side. At the same time, through the neural network model, the dense wind field prediction results for each of the next three moments can also be predicted, which can clarify the subsequent wind direction change trend and provide data support for pedestrian travel or building ventilation control in the target area.

[0034] For step S3, in some embodiments, the following training process may be included before applying the neural network model: Acquire sample data from several regions, and simultaneously acquire labels for the reconstructed real dense wind field and several labels for the predicted real dense wind field corresponding to each sample data from a region; wherein, the sample data from the regions includes: multi-channel two-dimensional wind field feature image samples, geometric prior information samples, and several observation mask samples. To illustrate, 100 different types of regions can be selected as sample regions, such as urban areas, residential communities, and industrial parks. Multiple sets of data are collected for each sample region to form regional sample data: multi-channel two-dimensional wind field feature image samples, geometric prior information samples, and observation mask samples. Simultaneously, using high-precision wind field monitoring equipment, the real dense wind field reconstruction results for each sample region at the corresponding time (serving as the real label of the reconstruction results), as well as real dense wind field data for multiple future timeframes (serving as the real label of the prediction results), form a complete sample dataset.

[0035] Using sample data from each region as model input, and using the labels of the reconstructed real dense wind field corresponding to each sample data from each region and several labels of the predicted real dense wind field as supervision labels, the neural network model to be trained is iteratively trained until the loss error of the neural network model meets the preset convergence threshold, and the trained neural network model is output. Indicatively, the present invention can reduce the loss error between the model output and the true label by continuously adjusting the model parameters, such as the parameters of the encoder, decoder and temporal inference network of the model, until the error reaches a preset convergence threshold, so as to ensure that the model can accurately extract multi-dimensional wind field features and generate reliable wind field reconstruction and prediction results.

[0036] For example, the regional sample data of the aforementioned 100 sample areas are input one by one into the neural network model to be trained. The corresponding real dense wind field reconstruction labels and prediction labels are used as supervision labels. During training, mean squared error is used as the basic data loss calculation method, and a joint loss function is constructed by combining wind field physical consistency constraint loss to complete the iterative convergence training of the model. The preset convergence threshold can be set to 0.001. After each iteration of training, the parameters of the model's encoder, decoder, and temporal inference network are adjusted to reduce the loss error. The iterative training process is repeated until the model's loss error drops below 0.001, at which point training stops, and the trained neural network model is output. Indicatively, the neural network model of this invention can adapt to the wind field reconstruction and prediction needs of different types of regions and has strong generalization ability.

[0037] In some embodiments, when the neural network model extracts sparse observation distribution features, multi-height correlation features, building geometric constraint features, and wind field temporal evolution features based on the geometric prior information, the multi-channel two-dimensional wind field feature images, and the observation masks, the specific steps include: Step S301: The geometric prior information, each of the multi-channel two-dimensional wind field feature images, and each of the observation masks are stitched together along the channel dimension using the neural network model to generate a multi-channel input tensor; Schematic illustration: To address the problem of fragmented processing and inability to establish correlations among multi-dimensional data, this embodiment of the invention standardizes and concatenates input data of different types and dimensions to form a unified multi-channel input tensor. This ensures that the neural network model can simultaneously process multi-dimensional data and accurately uncover the intrinsic correlations between data. For example, geometric prior information is used as one channel, multi-channel two-dimensional wind field feature images at six time points are used as six channels, and observation masks at three height levels are used as three channels. By concatenating along the channel dimensions, a multi-channel input tensor with the dimension of [grid size × grid size × (1 + 6 + 3)] is finally generated, where 1 represents the geometric prior information channel, 6 represents the wind field feature image channels at six time points, and 3 represents the observation mask channels at three height levels. This achieves unified integration of all input data.

[0038] Step S302: Extract features from the multi-channel input tensor and output local features and global features; wherein, the local features are used to indicate the local variation characteristics between the observation point, the building boundary and the wind field; the global features are used to indicate the wind field evolution characteristics at different heights within the target area; Indicatively, local and global features complement each other. Local features focus on local details of the target area, such as wind field changes around a single observation point or a single building, to capture fine-scale wind field patterns. Global features, on the other hand, focus on the overall patterns of the entire target area, such as the overall correlation of wind fields at different altitudes and the evolution trend of wind fields across the entire area, to capture coarse-scale wind field patterns. The combination of the two achieves comprehensive feature coverage of both local details and overall patterns, avoiding the one-sidedness of feature extraction.

[0039] In this embodiment of the invention, feature extraction can be performed on the generated multi-channel input tensor to output local and global features. Local features may include: wind speed gradients around the observation point and abrupt wind field changes at building boundaries; global features may include: the overall vertical correlation between wind fields at three altitude levels, such as the pattern that low-altitude wind speed increases synchronously with high-altitude wind speed; global features may also include: the overall flow trend of the wind field across the entire urban area, such as an overall southerly wind direction and a trend of wind speed gradually decreasing from south to north.

[0040] Step S303: Perform cross-scale feature fusion on the local features and the global features to output the sparse observation distribution features, the multi-height correlation features, the building geometric constraint features, and the wind field temporal evolution features.

[0041] To illustrate, local features belong to fine-scale features, while global features belong to coarse-scale features. By fusing across scales, the complementary optimization of the two scales can be achieved, thereby combining local detailed patterns with global overall patterns and accurately separating four types of core wind field features. This provides more accurate and comprehensive feature support for subsequent wind field reconstruction and prediction.

[0042] Specifically, when performing cross-scale fusion of extracted local and global features, the details of local observation points and the wind field characteristics at building boundaries can be combined with the global multi-height wind field correlation and the global temporal evolution pattern to ultimately separate four types of core features. For example, by combining the distribution of local observation points with the global regional division, sparse observation distribution features can be obtained; by combining the details of local height-layer wind fields with the global vertical correlation pattern, multi-height correlation features can be obtained; by combining the abrupt changes in local building boundary wind fields with the influence of global building distribution, building geometric constraint features can be obtained; and by combining local moment-to-moment wind field changes with global temporal trends, wind field temporal evolution features can be obtained.

[0043] Therefore, in this embodiment of the invention, by extracting local and global features step by step, dual coverage of local details and global patterns is achieved, avoiding the problem of incomplete feature extraction caused by traditional techniques that only focus on local or global features. Moreover, through cross-scale feature fusion, complementary optimization of local and global features can be achieved, accurately separating four types of core wind field features, providing more accurate feature support for subsequent wind field reconstruction and prediction, and further improving the accuracy of reconstruction and prediction.

[0044] In some embodiments, the neural network model includes: an encoder; the encoder is used to perform block processing, feature fusion and local feature extraction on multi-channel input tensors, and its core component includes a first feature extraction network, which can be used to extract local detail features.

[0045] The process of generating local features includes: Step S304: The multi-channel input tensor is divided into blocks by the encoder to form several local feature blocks. Then, the features of each local feature block are fused to generate a joint feature map. Indicatively, block processing can decompose a large multi-channel input tensor into multiple small-sized local feature blocks, reducing the computational cost of feature extraction. At the same time, through feature fusion, the information of each local feature block can be integrated to generate a joint feature map that reflects local correlations, laying the foundation for subsequent local feature extraction and ensuring that local features can reflect the intrinsic correlation between wind fields, buildings and observation points within the region.

[0046] For example, the encoder divides the resulting multi-channel input tensor into blocks, creating multiple local feature blocks. Each local feature block covers a certain range of grid areas, such as a small block in an urban area. By fusing features from multiple local feature blocks, the wind field data, geometric prior data, and observation mask data from each feature block can be integrated to generate a joint feature map. This joint feature map can clearly reflect the relationship between the wind field and buildings and observation points in each local area.

[0047] Step S305: The first feature extraction network of the encoder is used to perform feature recognition on each local grid region in the joint feature map to generate observation point distribution features; wherein, the observation point distribution features are used to characterize the wind field arrangement pattern between the observation points and the points to be reconstructed. Schematic, the first feature extraction network focuses on the observation point information in the joint feature map. By identifying the spatial location of the observation point and the surrounding wind field data, it can generate features that reflect the wind field correlation between the observation point and the point to be reconstructed, providing an accurate basis for subsequent wind field interpolation and avoiding the deviation caused by traditional interpolation ignoring the distribution pattern of the points.

[0048] For example, the first feature extraction network can identify multiple local grid regions in the joint feature map one by one, focusing on capturing the spatial location of the eight monitoring points (the number of points corresponding to each height layer) and the wind speed difference between the monitoring points and the surrounding points to be reconstructed, thereby generating the distribution features of the observation points.

[0049] Step S306: Capture the building outline in the joint feature map through the first feature extraction network to generate building edge features; wherein, the building edge features are used to characterize the blocking effect of the building edge on the wind field; Schematic, building edges are a key factor affecting wind flow. The first feature extraction network captures the building outline boundaries in the joint feature map and extracts the wind field variation characteristics at the building edges, thereby generating building edge features to clarify the local influence of buildings on the wind field and make up for the shortcomings of traditional techniques that ignore the influence of building edges.

[0050] For example, the first feature extraction network captures the outline boundaries of all buildings in the urban area in the joint feature map, such as the rectangular outline of a 150m office building and the outlines of surrounding multi-story buildings. It can focus on identifying characteristics such as sudden changes in wind speed and wind direction deflection at the building edges. For example, the wind speed increases on the windward side edge of the office building and decreases on the leeward side edge, which will form vortices. This makes the final building edge features accurately characterize the blocking strength and influence range of each building edge on the wind field.

[0051] Step S307: Fit the local wind field fluctuations in the joint feature map using the first feature extraction network to generate wind field change features; wherein, the wind field change features are used to characterize the wind field difference between any two grid regions; Schematic, the wind field variation features focus on the detailed differences in the wind field within a local area. By fitting the wind speed and direction differences between adjacent grids or grids in different regions in the joint feature map, the model can generate features that can reflect the local wind field fluctuation patterns, providing detailed support for the interpolation completion of subsequent wind field reconstruction, and ensuring that the wind speed data of the points to be reconstructed conforms to the actual wind field fluctuation patterns.

[0052] For example, the first feature extraction network fits the wind field data of each grid region and its adjacent grid regions in the joint feature map, calculates the wind speed difference and wind direction angle between any two grid regions, and thus generates wind field variation features. These wind field variation features can clearly reflect the fluctuation pattern of the local wind field. For instance, these features show that the wind speed fluctuation is smaller in the street passage and larger around the buildings, providing accurate fluctuation reference for subsequent wind speed completion at the points to be reconstructed.

[0053] Step S308: Output the distribution characteristics of the observation points, the building edge characteristics, and the wind field change characteristics as the local features; Indicatively, the three types of detailed features extracted above are integrated to form a complete local feature, which can cover three core local dimensions: observation point, building edge, and local wind field fluctuation. This provides high-quality fine-scale feature support for subsequent cross-scale feature fusion, ensuring the effectiveness of cross-scale fusion.

[0054] Therefore, in this embodiment of the invention, by using encoder block processing and feature fusion to generate a joint feature map, the computational load of feature extraction is reduced, and the association and integration of multi-dimensional data in local areas are realized, solving the problems of scattered local data and inability to establish associations in traditional technologies. Based on the first feature extraction network in the model, this invention can accurately extract three types of local features: the distribution of observation points, building edges, and local wind field fluctuations. This achieves comprehensive coverage of the core details of the local wind field and avoids the shortcomings of traditional technologies in extracting local features incompletely, such as missing the influence of wind fields at building edges. This invention also clarifies the specific composition and generation process of local features, making the extraction of local features more targeted and standardized, providing high-quality fine-scale feature support for subsequent cross-scale feature fusion, and further improving the accuracy of wind field reconstruction and prediction.

[0055] In some embodiments, the global features include: vertical correlation features for characterizing the spatial evolution of wind fields between upper and lower altitudes, and temporal correlation features for characterizing the temporal evolution of wind fields at different altitudes. Schematic, the two main components of the global features cover both spatial and temporal dimensions. Vertical correlation features address the coupling problem of wind fields at different altitudes, while temporal correlation features address the temporal evolution of wind fields. The combination of the two achieves comprehensive coverage of the wind field patterns across the entire domain, providing coarse-scale support for subsequent cross-scale fusion and result generation.

[0056] The process of generating the global features includes: Step S309: The joint feature map is fully identified through the second feature extraction network of the encoder, the spatial dependency between the upstream inflow region and the downstream wake region of the wind field is constructed, and the long-distance dependency between wind fields at different heights is constructed. Based on the spatial dependency and the long-distance dependency, the vertical correlation feature is generated. Schematic, the second feature extraction network can identify the spatial correlation patterns of the wind field across the entire region. It establishes spatial dependencies by identifying the upstream inflow region and the downstream wake region in the joint feature map, while capturing long-distance correlations of wind fields at different altitudes. Based on the two different correlations mentioned above, it generates the final vertical correlation features, which can clarify the spatial evolution patterns of wind fields at high and low altitudes and make up for the shortcomings of traditional technologies that ignore the vertical coupling of wind fields at multiple altitudes.

[0057] For example, based on the above-mentioned urban area scenario, the second feature extraction network of the present invention can determine the upstream inflow area (e.g., the southern area) and the downstream wake area (e.g., the northern area) of the wind field in the urban area when performing global recognition on the joint feature map. At the same time, it can construct the spatial dependency relationship between the two, such as when the wind speed in the upstream inflow area increases, the wind speed in the downstream wake area increases synchronously, and the wind direction remains consistent. It can also construct the long-distance dependency relationship between the wind fields at three height levels (20m, 60m, and 100m). For example, after the wind speed at 100m altitude changes, the wind speeds at 60m and 20m altitudes will change synchronously after 10 minutes, and the change amplitude is proportional. Based on the above two dependency relationships, the vertical correlation feature is finally generated to characterize the spatial coupling law of the wind fields at different height levels.

[0058] Step S3010: Identify the wind field state at different times in the joint feature map through the second feature extraction network, quantify the evolution trend of the wind field over time, and generate the time-related features based on the evolution trend; Indicatively, temporal correlation features can quantify the temporal evolution of the wind field across the entire region. By identifying the wind field state at different times in the joint feature map, the temporal change trends of wind speed and wind direction can be quantified, thereby deriving temporal correlation features. This provides reliable temporal support for subsequent wind field prediction and solves the problem that traditional technologies cannot accurately capture temporal changes in the wind field.

[0059] Therefore, the embodiments of the present invention can achieve full-domain wind field pattern coverage in both spatial and temporal dimensions, solving the problem that traditional technologies only focus on local areas and ignore the spatiotemporal patterns of the entire domain; and by constructing spatial and long-distance dependencies through a second feature extraction network, it accurately captures the vertical coupling patterns of high-altitude and low-altitude wind fields, avoiding the wind field fitting deviation caused by the inability of traditional technologies to reflect the correlation of wind fields at multiple altitudes; moreover, it can quantify the temporal evolution trend of the wind field and generate temporal correlation features, providing reliable temporal support for subsequent short-term multi-step wind field prediction, solving the problem that traditional technologies cannot accurately capture the temporal changes of the wind field, resulting in low prediction accuracy; In some embodiments, the neural network model further includes a decoder; the decoder is used to receive local features and global features output by the encoder, perform cross-scale fusion, and separate and output four types of core wind field features. Indicatively, the decoder can achieve mutual fusion of local features and global features, while accurately separating the fused features into four types of core wind field features, ensuring that the features can directly support subsequent wind field reconstruction and prediction.

[0060] Therefore, when performing cross-scale feature fusion on the local features and the global features to output the sparse observation distribution features, the multi-height correlation features, the building geometric constraint features, and the wind field temporal evolution features, the specific steps include: Step S3011: The sparse observation distribution features are generated by upsampling the observation point distribution features, the building edge features, and the wind field change features step by step through the decoder. Indicatively, the purpose of stepwise upsampling is to amplify the size of local features, transform fine-scale local features into features with the same size as global features, and integrate information from the three types of local features to generate sparse observation distribution features that can accurately indicate the distribution relationship between the effective observation area and the area to be reconstructed, providing accurate positioning for subsequent wind field interpolation and completion.

[0061] For example, the decoder upsamples the extracted observation point distribution features, building edge features, and wind field variation features step by step. During the amplification process, it integrates the information from the three types of features to finally generate sparse observation distribution features. It can be understood that these sparse observation distribution features can accurately mark the spatial location of all valid observation points within the area and the distribution range of the points to be reconstructed. Simultaneously, combined with building edge features and wind field variation features, it clarifies the spatial relationships between the points to be reconstructed, the observation points, and the buildings, providing accurate positioning and reference for subsequent wind field interpolation and completion.

[0062] Step S3012: The vertical correlation feature and the wind field change feature are hierarchically correlated by the decoder to quantify the linkage mapping relationship between wind fields at different heights, and the multi-height correlation feature is generated based on the linkage mapping relationship. Indicatively, this step combines global vertical correlation features with local wind field change features through hierarchical association, quantifies the linkage relationship of wind fields at different altitudes, and generates multi-altitude correlation features so that the model can learn the mutual influence rules of wind fields at different altitudes, providing vertical correlation support for wind field reconstruction and prediction.

[0063] For example, the decoder hierarchically correlates vertical correlation features with local wind field change features, quantifying the linkage mapping relationship of wind fields at multiple altitude levels. For instance, when the wind speed at 100m altitude increases by 1m / s, the wind speed at 60m altitude increases by 0.8m / s, and the wind speed at 20m altitude increases by 0.5m / s, with the wind direction remaining consistent. Based on this linkage mapping relationship, multi-altitude correlation features can be derived to accurately characterize the coupling and linkage patterns of wind fields at different altitude levels.

[0064] Step S3013: The building edge features and the vertical correlation features are fused by the decoder to quantify the constraint strength of the building on the wind field, and the building geometric constraint features are generated based on the constraint strength; wherein, the building geometric constraint features are used to characterize the degree of influence of building distribution on wind field flow. Indicatively, the constraint effect of buildings on wind fields is reflected both in the abrupt changes in wind fields at local building edges and in the overall impact of building distribution across the entire region on wind fields at multiple heights. By integrating building edge features and vertical correlation features, the constraint strength of buildings can be quantified, generating building geometric constraint features. This allows the model to learn the impact of building distribution on wind flow, ensuring that subsequent wind field reconstruction results are more closely aligned with the actual building environment.

[0065] Step S3014: Aggregate the vertical correlation features and the temporal correlation features through the decoder to generate the wind field temporal evolution features; To illustrate, the temporal evolution of wind fields not only includes the changing trend in the time dimension, but is also affected by the multi-altitude wind field correlation in the spatial dimension. This embodiment of the invention obtains the temporal evolution features of wind fields by aggregating the above two types of features, so that the extracted temporal evolution features of wind fields can fully reflect the spatiotemporal evolution law of wind fields and provide comprehensive spatiotemporal support for wind field prediction.

[0066] For example, the temporal evolution characteristics of the wind field can clearly reflect the changes in the wind field in the urban area over time. Moreover, not only do the wind speed and direction change, but the coupling relationship of the wind field at different altitudes also evolves synchronously. For instance, at 13:00, the coupling ratio of upper-level and lower-level wind speeds is 1:0.5, and at 14:00, the coupling ratio becomes 1:0.6. Therefore, based on these temporal evolution characteristics of the wind field, a comprehensive spatiotemporal evolution basis can be provided for subsequent short-term multi-step prediction of the wind field.

[0067] Therefore, the embodiments of the present invention realize cross-scale feature fusion based on the decoder, which can combine local features with global features, solve the problem of separation of local and global features in traditional technology, thus making it impossible to achieve complementary optimization, and improve the integrity and accuracy of features; In the processing of the decoder of this invention, the generation of each feature combines local and global information, making the features more targeted and practical, and avoiding the shortcomings of traditional technology in feature extraction being one-sided and unable to support accurate reconstruction and prediction.

[0068] Please see Figure 2 , Figure 2 The flowchart illustrating the overall implementation of a wind field reconstruction and prediction method provided in this application demonstrates the entire process of the invention from data preparation and model training to wind field reconstruction and prediction output. The specific implementation process is as follows: The first stage is the model training phase, which involves iterative optimization of the neural network based on a pre-built wind field sample dataset. During training, this invention can input a wind field sample dataset containing wind field feature image samples at multiple heights and times, as well as building geometry prior information samples, observation mask samples, and corresponding real dense wind field labels, into the neural network to be trained for iterative training. After each training iteration, the model updates the parameters of the encoder, decoder, and time-series inference network through the backpropagation algorithm, continuously reducing the loss error between the model output and the real labels until the model's loss error meets a preset convergence threshold. At this point, training stops, and the trained neural network model is output. Next, in the reasoning stage of actual wind field reconstruction and prediction, real-time sparse wind speed data of the target area can be collected, and historical sparse wind speed data of multiple consecutive historical moments at various heights in the area can be obtained. Both types of data are sent to the normalization mapping and encoding module. Through linear brightness mapping, the wind speed data at different heights and times are normalized to the [0,1] interval. After eliminating the difference in dimensions, the multi-channel two-dimensional wind field feature images corresponding to different times are encoded, thereby turning the scattered wind speed data into a standardized format that the model can directly process. Further acquire prior geometric information about the building, such as the building outline, height, and signed distance function of the target area, and generate observation masks corresponding to each height layer to clarify the effective and blank areas of the data, thus avoiding unfounded and blind data filling and extrapolation in the model.

[0069] The multi-channel two-dimensional wind field feature image generated by encoding, the prior information of building geometry, and the observation masks of each height layer are stitched together along the channel dimension to generate a unified multi-channel input tensor. This integrates wind field data, building information, and observation point information, allowing the model to process multi-source data simultaneously and explore the correlations between them.

[0070] A multi-channel input tensor is fed into the encoder, which first divides the input tensor into blocks, then fuses them to generate a joint feature map. Subsequently, the encoder extracts local and global features in parallel through a shallow network (i.e., the first feature extraction network) and a deep network (i.e., the second feature extraction network). The shallow network focuses on details, extracting three types of local features: the distribution features of observation points, the constraint features of building edges, and the features of wind field changes. This allows for accurate capture of the distribution patterns of observation points and the area to be reconstructed, the obstruction effect of building edges on the wind field, and the differences in wind field fluctuations in local areas. The deep network focuses on global patterns, extracting two types of global features: vertical correlation features and temporal correlation features. These represent the spatial coupling relationship of wind fields at different altitudes and the temporal trend of wind field evolution over time, respectively. This ensures that local details are not lost while the overall wind field trend is controlled.

[0071] The local and global features output by the encoder are fed into the decoder, which then uses a cross-scale feature fusion mechanism to fuse and optimize shallow local details and deep global patterns. Subsequently, four types of core wind field features are separated: sparse observation distribution features, multi-height correlation features, building geometric constraint features, and wind field temporal evolution features. These features can comprehensively cover the spatial distribution, vertical coupling, building constraints, and temporal evolution of the wind field, thus providing accurate support for the generation of subsequent results.

[0072] Based on the four core features output by the decoder, the model can reconstruct the dense wind field at the current moment and predict the dense wind field at multiple future moments. On the one hand, it can generate the dense wind field reconstruction result at the current moment through grid interpolation, fully covering all grid points in the target area and accurately presenting the wind field distribution in complex areas such as building perimeters and street passages. On the other hand, it can obtain the dense wind field prediction results for multiple consecutive future moments through deduction, realizing short-term multi-step wind field prediction and providing data support for applications such as urban wind environment assessment and building ventilation design.

[0073] Through the above process, this invention achieves accurate reconstruction and short-term prediction of dense wind fields across the entire region from sparse observation data. Based on the massive samples learned during the training phase, the model learns the physical laws of wind fields, enabling it to obtain accurate wind field reconstruction results and multi-step prediction results of wind fields through multi-scale feature extraction and fusion during the inference phase. This effectively solves the shortcomings of traditional wind field reconstruction and prediction methods, which have low accuracy and cannot adapt to complex urban scenarios.

[0074] In some embodiments, the specific generation process of the dense wind field reconstruction results and wind field prediction results of the present invention includes: Step S3015: The sparse observation distribution features and the building geometric constraint features are cross-correlated calculated using the neural network model to generate a first wind field coupling feature that characterizes the degree of synergistic influence of the observation point layout and the building spatial layout on the wind field distribution. Indicatively, through cross-correlation calculations, the influence of the layout of observation points and the spatial layout of buildings can be integrated to clarify the wind field distribution pattern under the combined effect of the two, providing accurate spatial coordination support for wind field reconstruction at the current moment and avoiding the deviation caused by traditional interpolation ignoring the synergistic influence of multiple factors.

[0075] Step S3016: The dense wind field reconstruction result is generated by performing grid interpolation and completion using the neural network model based on the sparse observation distribution characteristics, the building geometric constraint characteristics, and the first wind field coupling characteristics. To illustrate, if a point to be reconstructed is located on the leeward side of an office building, then this invention can combine the wind speed data of two surrounding observation points, the building constraint strength, and the first wind field coupling feature (which indicates that the wind speed on the leeward side is lower) to complete the wind speed data of that point, and finally generate a dense wind field reconstruction result of the global grid at the current moment, which can accurately present the actual wind field distribution in complex areas such as the area around the building and street passages.

[0076] Step S3017: The building geometric constraint features and the wind field temporal evolution features are spatiotemporally linked and mapped through the neural network model to generate a second wind field coupling feature for characterizing the wind field temporal change characteristics under building constraint conditions. Schematic, the second wind field coupling feature can be reflected in the change of urban wind field over time under the constraint of the office building. For example, when the wind speed increases, the wind speed on the leeward side of the office building increases more slowly than that on the windward side, and the wind direction deflection angle is smaller. Based on this second wind field coupling feature, a spatiotemporal coordination constraint basis can be provided for subsequent wind field prediction.

[0077] Step S3018: Input the multi-height correlation features, the wind field temporal evolution features, and the second wind field coupling features into the temporal inference network of the neural network model, so that the temporal inference network can deduce the prediction results of each of the dense wind fields; Indicatively, the temporal extrapolation network can accurately extrapolate the wind field state at multiple future moments based on three types of features: multi-height correlation of the input, temporal evolution of the wind field, and coupling features of the second wind field, generating dense wind field prediction results and solving the shortcomings of traditional technology in low prediction accuracy and inability to achieve multi-step prediction.

[0078] For example, the multi-height correlation features indicating the linkage between wind fields at different heights, the temporal evolution features indicating the spatiotemporal evolution of wind fields, and the second wind field coupling features indicating the temporal changes under building constraints are input into the time series extrapolation network to extrapolate the dense wind field prediction results for the next three time points. For example, it is predicted that the overall wind speed in the urban area will increase to 5.5 m / s at 14:15, and the wind direction will continue to deflect clockwise by 5°.

[0079] In this embodiment of the invention, by generating first and second wind field coupling features, the influence quantification of spatial coordination and spatiotemporal coordination is realized respectively, which solves the problems of traditional technology ignoring the synergistic influence of multiple factors and large deviations in wind field fitting and prediction; and by combining the three types of features for grid interpolation completion, compared with traditional simple linear interpolation, the completion accuracy of wind speed data at the point to be reconstructed is greatly improved, ensuring that the dense wind field reconstruction result fits the actual wind field distribution, which is especially suitable for complex areas such as the area around buildings; This invention can also be based on a time-series inference network and combined with multi-dimensional spatiotemporal features to achieve dense wind field prediction for multiple consecutive future moments, which solves the shortcomings of traditional technology in low prediction accuracy and inability to achieve multi-step prediction, and improves the practicality of wind field prediction.

[0080] In some embodiments, to more clearly illustrate the actual processing procedure of the neural network model of the present invention, the specific implementation process is as follows: The neural network model used in this invention is built using a collaborative architecture of encoding and decoding. The encoding end of the model introduces a hierarchical structure and combines a shift window self-attention mechanism, which can effectively model the long-range spatial correlation of urban wind fields in a multi-scale feature space and accurately capture the global flow patterns such as the direction of large-scale urban airflow and the interconnection of street airflows. The decoding end of the model follows the core structure of the U-Net network and adopts a method of progressive upsampling combined with skip connection feature fusion to achieve super-resolution restoration and reconstruction from low-dimensional sparse observation features to high-resolution global dense wind field features, accurately restoring the refined local wind field structure such as street valleys and building bypasses.

[0081] To achieve historical time-series information fusion and accurate prediction of future multi-step wind fields, this invention also employs a sliding time window approach to unify sparse observation data from multiple consecutive historical moments as network input. A pluggable time fusion module (TFM) is embedded in the front end of the encoder to complete the fusion of historical time-series features. This time-series fusion module can flexibly select various structures such as time-series shifting modules, causal time-series convolutional networks, recurrent neural networks, and time-series self-attention Transformers to adapt to different real-time and accuracy requirements. Based on the fused time-series features, the reconstruction results of the dense wind field at the current moment and the prediction results of the dense wind field in the future multi-step process can be output simultaneously in the shared reconstruction task branch and prediction task branch, realizing the integrated output of wind field reconstruction and time-series prediction.

[0082] During the model training phase, this invention can also introduce a joint training mechanism by physical consistency constraints. The overall loss function of the model is composed of the wind field reconstruction data error term, the multi-step wind field prediction error term, and the physical consistency constraint loss term. The proportion of each loss term is balanced by setting different weight coefficients.

[0083] Schematic, the model training of the present invention employs a joint loss function that co-optimizes data-driven error and physical constraint error. Its specific expression is: ; in, This represents the wind field reconstruction error for the current window, used to measure the deviation between the current wind field output by the model and the actual wind field. Mean squared error can be used as the calculation method. The wind field prediction error over multiple future steps is used to measure the deviation between the model's output time-series prediction and the actual future wind field. It can also be calculated using the mean squared error. The physical consistency constraint loss is used to ensure that the wind field output by the model conforms to the basic laws of fluid dynamics. , and These are the weighting coefficients for each loss term, which can be adjusted according to the actual application scenario; Furthermore, the physical consistency constraint of this invention can be established based on the continuity equation of three-dimensional incompressible flow, which is the fundamental law of mass conservation in fluid mechanics, and its form is: ; Among them, among them, , , The wind field is respectively , , The velocity components are distributed in three directions. To transform this physical constraint into a loss term that can be used for model training, this invention employs the finite difference method to calculate the velocity field components output by the neural network in three directions. , , The gradient in the direction is used to construct the divergence loss of the velocity field, which is the physical consistency constraint loss. During training, the divergence loss is minimized through backpropagation, forcing the velocity field divergence output by the model to be as close to zero as possible, thereby enabling the model to learn the wind field distribution pattern that conforms to the law of conservation of mass.

[0084] Through the aforementioned physical consistency constraint training, this invention can effectively avoid abnormal outputs that do not conform to the laws of fluid physics, which are generated by the model relying solely on data fitting. This significantly improves the quality conservation, physical consistency, and temporal stability of wind field reconstruction and prediction results. For example, in complex flow scenarios such as building bypass and street valley eddies, this invention can greatly reduce non-physical errors in the model output, making the final wind field results more consistent with the actual flow characteristics of urban airflow.

[0085] Meanwhile, this invention can further realize the deep integration and application of architectural geometric prior information. The symbolic distance function (SDF) representing the spatial distribution of buildings is set as an independent input channel and input into the neural network together with sparse observation data and observation mask data. The symbolic distance function is used to accurately represent the spatial distance relationship between each grid point and the boundary of the building entity, which enhances the network's ability to learn the features of airflow in the near-wall area of ​​the building and the wind field structure around the building, and further reduces the wind field inference error in complex urban scenarios.

[0086] Finally, the multi-channel pseudo-color wind field feature image output by the neural network is split according to the preset channel rules, and the previously preset inverse mapping relationship between brightness and wind speed values ​​is called to restore the complete wind speed component field in each height plane one by one. The final output is multi-height three-dimensional refined wind field data that can be directly used for engineering applications such as urban wind environment assessment, building ventilation design, and urban airflow disaster prediction.

[0087] This invention significantly improves the real-time reconstruction accuracy and short-time multi-step wind field prediction accuracy of sparse wind field observation in complex urban building cluster scenarios through improved model architecture, physical constraint training, and geometric information constraint optimization methods. It effectively compensates for the result deviation caused by the traditional interpolation and extrapolation methods ignoring building flow around and temporal evolution laws, and further enhances the practicality and engineering adaptability of the method in actual urban meteorological monitoring, urban planning wind environment simulation and other scenarios.

[0088] Please see Figure 3This application also provides a wind field reconstruction and prediction device, which can implement the above-mentioned wind field reconstruction and prediction method. The device includes: The feature image generation module is used to normalize and map the current sparse wind speed observation data corresponding to different heights in the target area and the historical sparse wind speed observation data corresponding to each height at multiple consecutive historical moments, and encode the multi-channel two-dimensional wind field feature images corresponding to different moments. The model input data acquisition module is used to acquire the geometric prior information of the target area and the observation mask corresponding to each height; wherein, the geometric prior information is used to characterize the building geometric characteristics of the target area; the observation mask is used to mark the observation points with wind speed observation data and the points to be reconstructed without wind speed observation data; The result output module is used to input the multi-channel two-dimensional wind field feature image, the geometric prior information, and each of the observation masks into a trained neural network model, so that the neural network model extracts sparse observation distribution features, multi-height correlation features, building geometric constraint features, and wind field temporal evolution features based on the geometric prior information, each of the multi-channel two-dimensional wind field feature images, and each of the observation masks; wherein, the sparse observation distribution features are used to indicate the distribution relationship between the effective observation area and the area to be reconstructed; the multi-height correlation features are used to indicate the vertical coupling relationship of the wind field at different height layers; The result output module is also used to generate the dense wind field reconstruction result corresponding to the current moment and predict the dense wind field prediction result for multiple consecutive future moments by using the neural network model based on the sparse observation distribution characteristics, the multi-height correlation characteristics, the building geometric constraint characteristics and the wind field temporal evolution characteristics.

[0089] It is understood that the content of the above method embodiments is applicable to the present device embodiments. The specific functions implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0090] It should be noted that the device embodiments described above are merely illustrative. The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0091] Those skilled in the art will clearly understand that, for convenience and simplicity, the specific working process of the device described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0092] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the aforementioned wind field reconstruction and prediction method. This electronic device can include any smart terminal such as a tablet computer or an in-vehicle computer.

[0093] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0094] Please see Figure 4 , Figure 4 This illustrates the hardware structure of an electronic device according to another embodiment, the electronic device comprising: The processor can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to achieve the technical solutions provided in the embodiments of this application. The memory can be implemented in the form of read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory can store the operating system and other applications. When the technical solutions provided in the embodiments of this application are implemented through software or firmware, the relevant program code is stored in the memory and called and executed by the processor. Input / output interfaces are used to implement information input and output; The communication interface is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, Wi-Fi, Bluetooth, etc.). A bus is used to transfer information between various components of a device, such as processors, memory, input / output interfaces, and communication interfaces. The processor, memory, input / output interface, and communication interface are interconnected within the device via a bus.

[0095] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the terminal device, connecting all parts of the terminal device via various interfaces and lines.

[0096] The memory can be used to store the computer program. The processor implements various functions of the terminal device by running or executing the computer program stored in the memory and calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function, etc.; the data storage area may store data created based on the use of the mobile phone, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, RAM, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device or other volatile solid-state storage device.

[0097] This application embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the above-described wind field reconstruction and prediction method.

[0098] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0099] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described wind field reconstruction and prediction method.

[0100] It is understood that the content of the above method embodiments is applicable to the embodiments of this computer program product. The specific functions implemented by the embodiments of this computer program product are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0101] Those skilled in the art will understand that all or some of the steps, apparatuses, or functional modules / units in the methods disclosed above can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0102] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A wind field reconstruction and prediction method, characterized in that, The method includes: The current sparse wind speed observation data corresponding to different heights in the target area and the historical sparse wind speed observation data corresponding to each height at multiple consecutive historical moments are normalized and mapped to encode multi-channel two-dimensional wind field feature images corresponding to different moments. Obtain the geometric prior information of the target area and the observation mask corresponding to each height; wherein, the geometric prior information is used to characterize the building geometry of the target area; the observation mask is used to mark the observation points with wind speed observation data and the points to be reconstructed without wind speed observation data; The multi-channel two-dimensional wind field feature image, the geometric prior information, and each of the observation masks are input into a trained neural network model, so that the neural network model splices the geometric prior information, each of the multi-channel two-dimensional wind field feature images, and each of the observation masks along the channel dimension to generate a multi-channel input tensor. Feature extraction is performed on the multi-channel input tensor to output local and global features; wherein, the local features are used to indicate the local variation characteristics between the observation point, building boundary and wind field; the global features are used to indicate the wind field evolution characteristics at different heights within the target area; Cross-scale feature fusion is performed on the local features and the global features to output the sparse observation distribution features, the multi-height correlation features, the building geometric constraint features, and the wind field temporal evolution features; wherein, the sparse observation distribution features are used to indicate the distribution relationship between the effective observation area and the area to be reconstructed; the multi-height correlation features are used to indicate the vertical coupling relationship of the wind field at different height layers; Based on the sparse observation distribution characteristics, the multi-height correlation characteristics, the building geometric constraint characteristics, and the wind field temporal evolution characteristics, the dense wind field reconstruction result corresponding to the current moment and the dense wind field prediction result for multiple consecutive future moments are generated.

2. The wind field reconstruction and prediction method according to claim 1, characterized in that, The neural network model includes: an encoder; The process of generating the local features includes: The encoder performs block processing on the multi-channel input tensor to divide it into several local feature blocks, and then the features of each local feature block are fused to generate a joint feature map. The first feature extraction network of the encoder performs feature recognition on each local grid region in the joint feature map to generate observation point distribution features; wherein, the observation point distribution features are used to characterize the wind field arrangement pattern between the observation points and the points to be reconstructed. The building outline in the joint feature map is captured by the first feature extraction network to generate building edge features; wherein, the building edge features are used to characterize the blocking effect of the building edge on the wind field; The first feature extraction network is used to fit the local wind field fluctuations in the joint feature map to generate wind field change features; wherein, the wind field change features are used to characterize the wind field difference between any two grid regions. The distribution characteristics of the observation points, the building edge characteristics, and the wind field variation characteristics are output as the local features.

3. The wind field reconstruction and prediction method according to claim 2, characterized in that, The global features include: vertical correlation features used to characterize the spatial evolution of the upper-level wind field and the lower-level wind field, and temporal correlation features used to characterize the temporal evolution of the wind field at different altitudes. The process of generating the global features includes: The joint feature map is fully identified by the second feature extraction network of the encoder, and the spatial dependency between the upstream inflow region and the downstream wake region of the wind field is constructed. The long-distance dependency between wind fields at different heights is also constructed. Based on the spatial dependency and the long-distance dependency, the vertical correlation feature is generated. The second feature extraction network identifies the wind field state at different times in the joint feature map, quantifies the evolution trend of the wind field over time, and generates the time-related features based on the evolution trend.

4. The wind field reconstruction and prediction method according to claim 3, characterized in that, The neural network model further includes: a decoder; The cross-scale feature fusion of the local and global features, outputting the sparse observation distribution features, the multi-height correlation features, the building geometric constraint features, and the wind field temporal evolution features, includes: The decoder performs upsampling of the observation point distribution features, the building edge features, and the wind field change features at each level to generate the sparse observation distribution features. The decoder performs hierarchical correlation between the vertical correlation features and the wind field change features, quantifies the linkage mapping relationship between wind fields at different altitudes, and generates the multi-altitude correlation features based on the linkage mapping relationship. The decoder fuses the building edge features with the vertical correlation features to quantify the building's constraint strength on the wind field, and generates the building's geometric constraint features based on the constraint strength; wherein, the building's geometric constraint features are used to characterize the degree of influence of building distribution on wind field flow. The wind field temporal evolution features are generated by aggregating the vertical correlation features and the temporal correlation features through the decoder.

5. The wind field reconstruction and prediction method according to claim 4, characterized in that, The generation process of the dense wind field reconstruction results and each of the dense wind field prediction results includes: The neural network model is used to perform cross-correlation calculations between the sparse observation distribution features and the building geometric constraint features to generate a first wind field coupling feature that characterizes the degree of synergistic influence of the observation point layout and the building spatial layout on the wind field distribution. The dense wind field reconstruction result is generated by performing grid interpolation and completion based on the sparse observation distribution characteristics, the building geometric constraint characteristics, and the first wind field coupling characteristics using the neural network model. The neural network model is used to perform a spatiotemporal linkage mapping between the building geometric constraint features and the wind field temporal evolution features to generate a second wind field coupling feature that characterizes the wind field temporal variation characteristics under building constraint conditions. The multi-height correlation features, the wind field temporal evolution features, and the second wind field coupling features are input into the temporal inference network of the neural network model so that the temporal inference network can deduce the prediction results of each of the dense wind fields.

6. The wind field reconstruction and prediction method according to claim 1, characterized in that, The training process of the neural network model includes: Acquire sample data from several regions, and simultaneously acquire labels for the reconstructed real dense wind field and several labels for the predicted real dense wind field corresponding to each sample data from a region; wherein, the sample data from the regions includes: multi-channel two-dimensional wind field feature image samples, geometric prior information samples, and several observation mask samples. Using sample data from each region as model input, and the labels of the reconstructed real dense wind field corresponding to each sample data from each region as well as several labels of the predicted real dense wind field as supervision labels, the neural network model to be trained is iteratively trained until the loss error of the neural network model meets the preset convergence threshold, and the trained neural network model is output.

7. A wind field reconstruction and prediction device, characterized in that, The device includes: The feature image generation module is used to normalize and map the current sparse wind speed observation data corresponding to different heights in the target area and the historical sparse wind speed observation data corresponding to each height at multiple consecutive historical moments, and encode the multi-channel two-dimensional wind field feature images corresponding to different moments. The model input data acquisition module is used to acquire the geometric prior information of the target area and the observation mask corresponding to each height; wherein, the geometric prior information is used to characterize the building geometric characteristics of the target area; the observation mask is used to mark the observation points with wind speed observation data and the points to be reconstructed without wind speed observation data; The output module is used to input the multi-channel two-dimensional wind field feature images, the geometric prior information, and each of the observation masks into a trained neural network model, so that the neural network model splices the geometric prior information, each of the multi-channel two-dimensional wind field feature images, and each of the observation masks along the channel dimension to generate a multi-channel input tensor; extract features from the multi-channel input tensor to output local features and global features; perform cross-scale feature fusion on the local features and global features to output the sparse observation distribution features, the multi-height correlation features, the building geometric constraint features, and the wind field temporal evolution features; wherein, the local features are used to indicate the local variation characteristics between the observation point, the building boundary, and the wind field; the global features are used to indicate the wind field evolution characteristics at different heights within the target area; the sparse observation distribution features are used to indicate the distribution relationship between the effective observation area and the area to be reconstructed; and the multi-height correlation features are used to indicate the vertical coupling relationship of the wind field at different height layers; The result output module is also used to generate the dense wind field reconstruction result corresponding to the current moment and predict the dense wind field prediction result for multiple consecutive future moments by using the neural network model based on the sparse observation distribution characteristics, the multi-height correlation characteristics, the building geometric constraint characteristics and the wind field temporal evolution characteristics.

8. An electronic device, characterized in that, The electronic device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the wind field reconstruction and prediction method according to any one of claims 1 to 6.

9. A storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements a wind field reconstruction and prediction method according to any one of claims 1 to 6.

Images