Urban low-level turbulence probabilistic forecasting method based on turbulence observation constrained diffusion model

Abstract

The application discloses a kind of urban low-altitude turbulence probability prediction methods based on eddy observation constraint diffusion model, belong to low-altitude meteorological monitoring and prediction technical field.The application assimilates and fuses building scale CFD background state field and eddy correlation observation, uses three-dimensional variation assimilation to construct historical turbulence state analysis field in offline stage, generates current turbulence state analysis field using optimal interpolation method of pre-computed gain matrix in online stage;According to current inflow condition, CFD background state field is matched and interpolated from CFD sample library to obtain current;Turbulence state analysis field, city form information and forecast time limit are jointly used as generation condition of diffusion model, generate multiple possible future low-altitude turbulence state field, and calculate to obtain probability risk field.The application can form urban low-altitude turbulence probability prediction field under different forecast time limit, provide spatial continuous, time definite prediction information for low-altitude take-off and landing area, flight route and key area turbulence risk judgment.

Description

Technical Field

[0001] This invention relates to low-altitude meteorological monitoring and forecasting technology, specifically to a probabilistic forecasting method for urban low-altitude turbulence based on an eddy observation-constrained diffusion model. Background Technology

[0002] With the development of the low-altitude economy, low-altitude flight activities are increasing, placing higher demands on the refined forecasting of near-surface wind fields and turbulence in cities. Dense urban buildings, complex street-to-valley passages, and building wakes can easily cause sudden increases in low-altitude turbulence, posing a serious threat to the operational safety of low-altitude aircraft. Existing meteorological support is mostly geared towards conventional weather scales or single-point risk assessment, which is insufficient to meet the needs of urban low-altitude flights for regionalized, minute-level, and probabilistic turbulence forecasting. Eddy covariance observations can obtain high-frequency three-dimensional wind speed information, but a limited number of discrete point observations can only reflect local turbulence states, making it difficult to form a continuous three-dimensional low-altitude turbulence state field within the target area. Microscale computational fluid dynamics simulations can provide the spatial structure of wind fields and turbulence states in complex urban building environments, but the computational cost is high and real-time observation corrections are lacking, making real-time rolling operation difficult and prone to deviations from actual atmospheric conditions. Numerical weather forecast products can provide large-scale meteorological backgrounds, but their spatiotemporal resolution is insufficient to directly meet the needs of urban building-scale low-altitude turbulence forecasting. Existing artificial intelligence prediction methods can improve flow field prediction efficiency, but they are mostly based on deterministic outputs, making it difficult to express the uncertainty of urban low-altitude turbulence evolution. In recent years, probabilistic generative models, represented by diffusion models, have been able to generate multiple possible future flow fields through random sampling, providing a new technical approach for expressing the uncertainties of complex flow fields and reconstructing sparse observation constraints. However, if real-time meteorological observations are not explicitly assimilated and fused with the physical background field, the generated future flow field samples may deviate from the current true turbulence state, making them difficult to directly use for probabilistic forecasting of urban low-level turbulence. Therefore, to address the problems that sparse eddy covariance observations are unable to form a continuous three-dimensional turbulence state field and that single deterministic predictions are insufficient to express the uncertainty of future turbulence risks, it is necessary to propose a probabilistic forecasting method for urban low-level turbulence that combines the assimilation analysis field of eddy covariance observations with a probabilistic generative model. Summary of the Invention

[0003] To address the problems of existing technologies, this invention proposes a probabilistic forecasting method for urban low-altitude turbulence based on an eddy covariance observation-constrained diffusion model. This method acquires eddy covariance observation data in real time and constructs probabilistic generation condition information including a background state field and an urban morphological condition field. It then uses the eddy covariance observation-constrained diffusion model to generate multiple possible low-altitude turbulence state field samples for the next 10-15 minutes. Finally, based on the proportion of strong turbulence risk indicators exceeding a preset turbulent kinetic energy threshold in the generated samples, the probability of strong turbulence risk is calculated, forming an urban low-altitude turbulence probabilistic risk field and outputting probabilistic forecast information.

[0004] The method for probabilistic prediction of urban low-altitude turbulence based on an eddy observation-constrained diffusion model of the present invention includes the following steps:

[0005] 1) Construct an eddy covariance observation network:

[0006] The target prediction area was delineated and divided into a three-dimensional grid; an eddy covariance observation network including observation points at multiple different locations was set up.

[0007] 2) Calculate the characteristics of the eddy covariance observation state:

[0008] High-frequency observation data were collected at each observation point, and the high-frequency observation data were preprocessed to obtain the observation time series under a unified time reference.

[0009] Within the set sliding time window, the eddy covariance observation state characteristics of each observation point are calculated based on the observation time series;

[0010] 3) Obtain the CFD background state field and urban morphological conditions;

[0011] According to the preset incoming flow conditions, the computational fluid dynamics (CFD) model is run to simulate and obtain the CFD background state field within the target prediction area, and a building-scale CFD sample library is established.

[0012] Obtain urban morphological conditions;

[0013] 4) Constructing training samples for the diffusion model:

[0014] The background state field of CFD is subjected to three-dimensional variational assimilation correction using eddy covariance observation state characteristics, and the assimilation correction result is further corrected by physical constraints to obtain the historical turbulence state analysis field. The historical turbulence state analysis field at any time t and the urban morphological conditions are used as input conditions for the diffusion model, with the corresponding future time... The historical turbulence state analysis field is used as the training target. The input conditions and the training target together constitute the training samples of the diffusion model, and τ is the prediction time.

[0015] 5) Offline training of the eddy covariance observation-constrained diffusion model:

[0016] An eddy covariance observation-constrained diffusion model is established, which includes a forward noise addition process, a reverse noise reduction process, and a noise reduction network for noise estimation.

[0017] Offline training is performed on the eddy covariance observation constraint diffusion model. Standard Gaussian random noise is gradually added to the training target through a forward noise addition process according to multiple preset noise intensity levels to obtain the noisy state field under different noise intensity levels. Then, the noisy state field, the corresponding noise intensity level, and the input conditions are fed into the denoising network, and the denoising network outputs the estimated value of the added noise.

[0018] The model parameters are updated by reducing the difference between the actual added random noise and the estimated output noise; after repeated training with multiple training samples, the trained eddy covariance observation constraint diffusion model is obtained.

[0019] 6) Online Forecast:

[0020] Based on the current incoming flow conditions, determine the two adjacent incoming flow wind speed levels and the two adjacent incoming flow wind direction zones in the CFD sample library obtained in step 3), and interpolate to obtain the CFD background state field corresponding to the current incoming flow conditions.

[0021] Based on the current eddy covariance observation state characteristics of each observation point, the CFD background state field is optimally interpolated and assimilated to obtain the current turbulence state analysis field.

[0022] N standard Gaussian random noise fields, along with the current turbulence state analysis field, urban morphology conditions, and prediction time, are input into the trained eddy covariance observation constraint diffusion model. Through the stepwise reverse denoising process, the denoising network is invoked to generate N future low-altitude turbulence state field samples online, where N is a natural number ≥ 2.

[0023] 7) Calculate the probability risk field:

[0024] Based on N future low-altitude turbulence state field samples, the probability of strong turbulence risk occurring at each three-dimensional grid node, flight path, or low-altitude take-off and landing point within the target prediction area is calculated, and the probability risk field is obtained.

[0025] In step 1), the target prediction area is defined based on the low-altitude take-off and landing points, flight routes, building distribution, and prevailing wind direction. The target prediction area is a three-dimensional spatial region where low-altitude turbulence risk prediction results need to be output, covering the low-altitude take-off and landing points, flight routes, and their surrounding influence areas. The horizontal scale is set to several hundred meters to several kilometers depending on the application scenario, and the vertical range is the altitude range involved in low-altitude flight activities, preferably from the ground to below 1000 m. The target prediction area is divided into a three-dimensional grid according to the set horizontal and vertical resolution. The horizontal size of the grid is set to 5 m to 100 m, and the vertical size of the grid is set to 2 m to 50 m. Preferably, the horizontal size of the grid is 10 m to 50 m, and the vertical size of the grid is 5 m to 20 m. Based on the actual application scenario, an eddy covariance observation network is deployed within the target prediction area, including observation points at multiple different locations. If the flight path is located at the edge of the target prediction area, the eddy covariance observation network also includes observation points deployed upstream of the prevailing wind direction within the target prediction area. These observation points include an upstream sentinel station, a central master station, one or more lateral auxiliary stations, and one or more altitude monitoring stations. The upstream sentinel station is located upstream or outside the prevailing wind direction within the target prediction area to detect incoming flow disturbance signals in advance. A central master station is set at the center of the target prediction area, one or more lateral auxiliary stations are set at the edge of the target prediction area, and one or more altitude monitoring stations are set at different altitudes along the vertical direction within the target prediction area to characterize the changes in low-altitude turbulence at different locations and altitudes within the target prediction area. The location identification information of each observation point is obtained to determine the spatial correspondence between the observation point and the three-dimensional grid nodes, and is used for subsequent construction of the interpolation matrix, calculation of observation increments, and assimilation correction. Eddy coherence systems are set up at each observation point to collect high-frequency three-dimensional wind speed, temperature, and pressure pulsation data in real time. Each eddy coherence system is connected to the data processing platform via wired or wireless communication. The data processing platform includes one or more of an edge computing unit, a central server, or a cloud server, used to receive, process, store, and predict low-altitude turbulence data. Each eddy coherence system includes at least a three-dimensional ultrasonic anemometer and a pressure pulsation sensor, a data acquisition unit, a time synchronization module, and an edge computing module. The three-dimensional ultrasonic anemometer and the pressure pulsation sensor are connected to the data acquisition unit, the three-dimensional ultrasonic anemometer and the pressure pulsation sensor and the data acquisition unit are connected to the time synchronization module, and the data acquisition unit and the time synchronization module are connected to the edge computing module. The edge computing module is connected to the data processing platform via wired or wireless communication. The sampling frequency of the high-frequency observation data is preferably 1–20 Hz. When deploying the eddy coherence observation network, the location identification information of each observation point is obtained; the location identification information includes the spatial coordinates, observation height, observation point type, and the position of the corresponding three-dimensional grid node of the observation point.

[0026] In step 2), the preprocessing of high-frequency observation data includes: time synchronization, coordinate rotation, outlier removal, trend term processing, and quality control, to obtain the observation time series under a unified time reference. The high-frequency observation data obtained from the eddy covariance system measurements at each observation point includes average wind speed. Weather and wind direction And turbulent kinetic energy (TKE). Because wind direction is an angular variable, it exhibits abrupt changes around 0° / 360°, making direct linear interpolation, assimilation, and neural network training unsuitable. It is converted to an easterly wind component. and northerly wind component Afterwards, the wind field becomes a continuous variable, making it more suitable as input for matrix assimilation and diffusion models. The average wind speed... and weather wind direction Converted to easterly wind component and northerly wind component :

[0027]

[0028] Eddy covariance observation characteristics include easterly wind component Northward wind component And turbulent kinetic energy (TKE). Eddy-related observational state characteristics are measured turbulent properties that exist only at discrete observation points, while the turbulent state analysis field is a continuous description that extends these discrete characteristics to the entire three-dimensional grid space through assimilation.

[0029] In step 3), based on the three-dimensional geometry of the buildings and the ground boundaries of the target prediction area, a building-scale CFD model is established using the steady-state Reynolds-averaged Navier-Stokes method. A turbulent closure scheme capable of directly outputting turbulent kinetic energy is employed. According to preset inflow conditions, the CFD model is used to simulate and generate the CFD background state field within the target prediction area. Inflow conditions include inflow wind speed and direction. These conditions are pre-classified, with multiple inflow wind speed levels and inflow wind direction zones set. The CFD model is run for each level of inflow wind speed and each zone's inflow wind direction combination, forming a building-scale CFD sample library. The CFD background state field includes the eastward and northward wind components and turbulent kinetic energy of each three-dimensional grid node in the target prediction area.

[0030] Based on the average incoming wind speed and average incoming wind direction observed from the upstream sentinel station, the wind speed levels and wind direction zones of two adjacent incoming winds are determined in the CFD sample library. The CFD background state field corresponding to the current incoming wind conditions is obtained by bilinear interpolation of adjacent CFD samples with wind speed and wind direction weights; the wind direction weights are calculated according to the shortest circumference direction to avoid wind direction jumps near 0° / 360°.

[0031] Urban morphological conditions include building height field Building planar density field and the direction of the town. Urban morphological conditions remain unchanged within the same target prediction area.

[0032] In step 4), the lead time is predicted. The time interval is set to 10-15 minutes. For any historical moment, based on the incoming wind speed and direction observed by the upstream sentry station at that moment, the historical CFD background state field is obtained according to step 3). The three-dimensional variational assimilation correction of the CFD background state field using eddy covariance observation state characteristics includes the following steps: [The text then abruptly shifts to a different topic:] ...the historical CFD background state field... All eddy covariance observation state features in the data are arranged into a background state vector. :

[0033]

[0034] Where J is the number of 3D mesh nodes. , and These represent the background easterly wind component, background northerly wind component, and background turbulent kinetic energy at the j-th 3D grid node, respectively, j=1,…J;

[0035] The measured eddy covariance state characteristics of each eddy covariance observation point at the same time are arranged into an observation vector. :

[0036]

[0037] in, The number of observation points, , and Let i represent the measured easterly wind component, northerly wind component, and turbulent kinetic energy at the i-th observation point, i=1,…I;

[0038] Historical turbulent state analysis field is achieved by minimizing the following three-dimensional variational cost function. get:

[0039]

[0040] in, For the three-dimensional variational cost function, the first term measures the deviation between the state field to be determined and the CFD background state field, and the second term measures the deviation between the state field to be determined at the observation point location and the measured eddy current value. Let be the eddy covariance observation state vector to be determined. This is the interpolation matrix from 3D mesh nodes to observation points. The spatial correspondence between observation points and 3D mesh nodes is determined by the positional identifiers of each observation point, thus constructing the interpolation matrix from 3D mesh nodes to observation points. The background error covariance matrix, Let be the observation error covariance matrix. The smaller the background error or observation error, the stronger the constraint effect of the corresponding information on assimilation.

[0041] The three-dimensional variational cost function is obtained through constrained optimization. The minimum state vector is used as the state vector for historical turbulence state analysis. And restore it to the historical turbulence state analysis field. Physical constraint corrections are performed during the assimilation correction process and after the solution is completed: eastward wind component of the 3D mesh nodes inside the building. Northward wind component The turbulent kinetic energy (TKE) is set to zero; TKE values ​​less than zero within the fluid region are corrected to zero or a non-negative lower limit, thereby ensuring that the analysis field satisfies the building boundary constraints and the non-negative turbulent kinetic energy constraints.

[0042] For any time t, the historical turbulence state analysis field and urban morphology conditions are combined into the input conditions of the diffusion model. :

[0043]

[0044] in, Let t be the easterly wind component, the northerly wind component, and the turbulent kinetic energy at time t, i.e., the historical turbulent state analysis field.

[0045] The training objective is to find future moments in the same historical sample. Historical turbulent state analysis field.

[0046] Each training sample consists of clearly defined input conditions and training objectives.

[0047] In step 5), an eddy covariance observation constraint diffusion model is established. The denoising network uses a U-shaped convolutional neural network (U-Net), but a 3D convolutional network, Transformer network, or other neural network structures capable of handling 3D mesh fields can also be used. This network is used to extract spatial features from the noisy 3D state field and estimate the added noise under input condition constraints.

[0048] The training target is denoted as the noise-free target state field, which is the historical turbulence state analysis field for future time steps. Standard Gaussian random noise is added to the training target according to several preset noise intensity levels, resulting in noisy state fields at different noise intensity levels:

[0049]

[0050] in, For the first Noisy state field at each noise intensity level For a noise-free target state field, Standard Gaussian random noise, As of the date The cumulative noise retention factor at each noise intensity level gradually decreases as the noise addition steps increase. =1,…S, where S is the number of noise intensity levels. The noise intensity levels characterize the proportion of random noise in the target state field. Low noise intensity levels correspond to slight noise perturbations, in which case the noisy state field mainly retains information from the original target state field; high noise intensity levels correspond to strong noise perturbations, in which case the noisy state field is mainly composed of random noise components. The number of noise intensity levels is preset based on model training stability, generation accuracy, and computational resources, preferably between 50 and 1000.

[0051] Output of the denoising network Represented as:

[0052]

[0053] in, For the denoising network to be trained, These are the parameters for the denoising network. During training, the error between the actual noise and the predicted noise is used as the training loss to update the denoising network parameters. After training, the eddy observation-constrained diffusion model is obtained.

[0054] The eddy covariance observation-constrained diffusion model is implemented in the form of a computer program; offline training is completed on a central server, cloud server, or local computing server; after training, the model parameter file is obtained and deployed to edge computing units, central servers, or cloud servers for online probabilistic forecasting.

[0055] In step 6), during online operation, based on the current incoming flow conditions (incoming wind speed and direction) of the upstream sentinel station, the current CFD background state vector is obtained according to step 3). Based on the current easterly wind component at each observation point. Northward wind component Constructing the current observation vector with turbulent kinetic energy TKE The current turbulence state analysis vector is obtained using the optimal interpolation method. :

[0056]

[0057] in,

[0058] This is the difference between the measured value and the CFD background value at the observation point location, i.e., the observation increment; The optimal interpolation gain matrix is ​​used; the current turbulence state analysis vector is used. The current turbulent state analysis field is reconstructed using a three-dimensional mesh. .

[0059] Perform the same physical constraint correction as in step 4), adjusting the easterly wind component of the grid nodes inside the building. Northward wind component And set the turbulent kinetic energy TKE to zero, and set the less than zero kinetic energy in the fluid region to zero. The value is corrected to zero, thus obtaining the current turbulent state analysis field;

[0060] The current input conditions are constructed by combining the current turbulence state analysis field and urban morphology conditions. :

[0061]

[0062] Generate separately Each independent standard Gaussian random noise field is fed into a trained diffusion model along with the current input conditions. Through repeated calls to the denoising network during the reverse denoising process, a new denoising model is generated. Sample of future low-altitude turbulent state field:

[0063]

[0064] This represents the nth future low-altitude turbulent state field. Different random noise fields correspond to different future state field samples, and all samples are constrained by the current turbulent state analysis field and urban morphology conditions. The number N of future low-altitude turbulent state field samples is determined based on the probability estimation accuracy requirements and online computation time, preferably between 20 and 100. Each generated sample undergoes the same physical constraint correction as in step 4), ensuring that the eddy covariance observation state characteristics of the three-dimensional mesh inside the building are zero, and guaranteeing that the turbulent kinetic energy of the fluid region is non-negative.

[0065] In step 7), the turbulent kinetic energy TKE is used as the criterion for strong turbulence. For the first... 3D grid nodes and prediction time Probability of strong turbulence Represented as:

[0066]

[0067] in, For the first The future low-altitude turbulent state field at the th 3D grid nodes and prediction time Turbulent kinetic energy below, The threshold for turbulent kinetic energy. This is an indicator function, taking the value 1 if the condition in parentheses is true, and 0 otherwise. For small, low-altitude aircraft with weak anti-interference capabilities, A value of 0.5~1.5 m² / s² can be used; for low-altitude aircraft with strong anti-interference capabilities. A value of 2.0~5.0 m² / s² is acceptable.

[0068] Based on the probability of strong turbulence occurrence at each three-dimensional grid node, a three-dimensional urban low-altitude turbulence probability forecast field is formed for the target prediction area under different prediction timeframes.

[0069] Furthermore, risk warning information can be output based on probabilistic forecast fields. Multiple different risk levels can be set according to the probability of strong turbulence occurring.

[0070] The 10-15 minute range is set according to a fixed prediction interval. Divided into multiple prediction timeframes. For the same 3D grid risk area, if its continuity... If a predicted risk reaches the warning threshold, then the duration of the risk is... Represented as:

[0071]

[0072] in, A fixed forecast interval between adjacent forecast leads. The duration of risk in the three-dimensional mesh risk area. This determines the number of consecutive risk warnings that can be predicted within a given timeframe. Based on the spatial relationship between the risk area and low-altitude takeoff and landing points and flight routes, recommended takeoff and landing windows and route risk alerts can also be output.

[0073] Advantages of this invention:

[0074] (1) This invention integrates the background state field of computational fluid dynamics at the building scale with eddy covariance observations, and converts discrete point observations into a continuous three-dimensional turbulence state analysis field within the target prediction area, thus making up for the insufficiency of single point or a small number of observations in reflecting the spatial distribution of urban low-altitude turbulence.

[0075] (2) In the offline stage, the present invention uses three-dimensional variational assimilation to construct the historical turbulence state analysis field, and in the online stage, it uses the optimal interpolation method of pre-calculated gain matrix to generate the current turbulence state analysis field, so that the diffusion model training data is consistent with the online input, while reducing the amount of online computation.

[0076] (3) Based on the current incoming wind speed and direction, the present invention matches and interpolates the background state field from a pre-established computational fluid dynamics sample library, eliminating the need to rerun the computational fluid dynamics model for each online forecast, thus improving the operational efficiency of urban low-altitude turbulence probability forecasting.

[0077] (4) The present invention uses the current turbulence state analysis field, urban morphology information and prediction time as the generation conditions of the diffusion model, so that the generation result is simultaneously constrained by the current actual turbulence state and urban spatial characteristics such as building height, building density and street valley direction, thereby reducing the problem of inconsistency between the future state field and the current actual state.

[0078] (5) This invention uses a diffusion model to generate multiple possible future low-altitude turbulence state fields, instead of outputting only a single deterministic result, which can reflect the multiple possibilities and uncertainties of the future urban low-altitude turbulence evolution.

[0079] (6) Based on the proportion of turbulent kinetic energy exceeding the preset turbulent kinetic energy threshold in multiple future state field samples, this invention directly calculates the probability of strong turbulence occurring at each three-dimensional grid node and under the prediction time. The probability meaning is clear, and there is no need to set multiple index weights or construct a complex comprehensive risk index.

[0080] (7) The present invention can form a three-dimensional urban low-altitude turbulence probability forecast field under different prediction time, providing spatially continuous and temporally clear forecast information for the assessment of turbulence risks in low-altitude take-off and landing areas, flight routes and other key areas. Attached Figure Description

[0081] Figure 1 This is a flowchart of the urban low-altitude turbulence probability prediction method based on the eddy observation constrained diffusion model of the present invention;

[0082] Figure 2 This is a schematic diagram of the eddy covariance observation network and target prediction area layout, representing an embodiment of the urban low-altitude turbulence probability prediction method based on the eddy observation constrained diffusion model of the present invention. Detailed Implementation

[0083] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0084] like Figure 1 As shown in this embodiment, the urban low-altitude turbulence probability prediction method based on an eddy observation-constrained diffusion model includes the following steps:

[0085] 1) Construct an eddy covariance observation network:

[0086] Based on low-altitude take-off and landing points, flight paths, urban building distribution, and prevailing wind direction, a target prediction area is delineated, covering low-altitude take-off and landing points, flight paths, and their surrounding influence zones. The horizontal scale is set to several hundred meters to several kilometers, and the vertical range is from the ground to below 1000 m, depending on the application scenario. The target prediction area is divided into a three-dimensional grid according to the set horizontal and vertical resolutions, with the horizontal grid size set to 10 m to 50 m and the vertical grid size to 5 m to 20 m. According to the actual application scenario, an eddy covariance observation network including observation points at multiple different locations is deployed within the target prediction area and upstream of the prevailing wind direction, such as... Figure 2 As shown, the observation points include an upstream sentinel station, a central master station, one or more lateral auxiliary stations, and one or more altitude monitoring stations. The upstream sentinel station is located upstream or outside the prevailing wind direction of the target prediction area to detect incoming flow disturbances in advance. A central master station is set at the center of the target prediction area, and one or more lateral auxiliary stations are set at the edges of the target prediction area. Multiple altitude monitoring stations are set at different vertical heights (h1, h2, h3…) within the target prediction area to characterize low-level turbulence changes at different locations and altitudes within the target prediction area. The location identification information of each observation point is acquired to determine the spatial correspondence between the observation point and the 3D grid nodes, and this information is used for subsequent construction of the interpolation matrix, calculation of observation increments, and assimilation correction. Eddy covariance systems are set up at each observation point to collect high-frequency 3D wind speed, temperature, and pressure pulsation data in real time. Each eddy covariance system is connected to the data processing platform via wired or wireless communication. The data processing platform includes one or more of edge computing units, a central server, or a cloud server to receive, process, store, and predict low-level turbulence data. The eddy covariance system includes at least a 3D ultrasonic anemometer, a barometric pressure pulsation sensor, a data acquisition unit, a time synchronization module, and an edge computing module. The 3D ultrasonic anemometer and barometric pressure pulsation sensor are connected to the data acquisition unit. The 3D ultrasonic anemometer, barometric pressure pulsation sensor, and data acquisition unit are connected to the time synchronization module. The data acquisition unit and time synchronization module are connected to the edge computing module. The edge computing module is connected to the data processing platform via wired or wireless communication. The sampling frequency of the high-frequency observation data is 10 Hz. When deploying the eddy covariance observation network, the location identification information of each observation point is acquired; the location identification information includes the spatial coordinates, observation height, observation point type, and the position of its corresponding 3D grid node.

[0087] 2) Calculate the characteristics of the eddy covariance observation state:

[0088] High-frequency observation data were collected at each observation point. The high-frequency observation data were preprocessed, including time synchronization, coordinate rotation, outlier removal, trend term processing, and quality control, to obtain the observation time series under a unified time reference.

[0089] Within a set 5-minute sliding time window, the eddy covariance observation state characteristics of each observation point are calculated based on the observation time series; high-frequency observation data, including average wind speed, are collected at each observation point. Weather and wind direction And turbulent kinetic energy (TKE), average wind speed and weather wind direction Converted to easterly wind component and northerly wind component :

[0090]

[0091] Eddy covariance observation characteristics include easterly wind component Northward wind component And turbulent kinetic energy (TKE). Eddy-related observational state characteristics are measured turbulent properties that exist only at discrete observation points, while the turbulent state analysis field is a continuous description that extends these discrete characteristics to the entire three-dimensional grid space through assimilation.

[0092] 3) Obtain the CFD background state field and urban morphological conditions;

[0093] Based on the three-dimensional geometry of buildings and ground boundaries of the target prediction area, a building-scale CFD model is established using the steady-state Reynolds-averaged Navier-Stokes method. A turbulent closure scheme capable of directly outputting turbulent kinetic energy is employed. The CFD model is used to simulate and generate the CFD background state field within the target prediction area according to preset inflow conditions. Inflow conditions include inflow wind speed and direction. These conditions are pre-classified, with multiple inflow wind speed levels and inflow wind direction zones established. The CFD model is run separately for each level of inflow wind speed and zone inflow wind direction combination, forming a building-scale CFD sample library. The CFD background state field includes the easterly wind component, northerly wind component, and turbulent kinetic energy of each three-dimensional grid node in the target prediction area. Urban morphological conditions include the building height field. Building planar density field and the direction of the town. Urban morphological conditions remain unchanged within the same target prediction area.

[0094] 4) Constructing training samples for the diffusion model:

[0095] The historical turbulence state analysis field is obtained by performing three-dimensional variational assimilation correction:

[0096] For any given historical moment, based on the incoming wind speed and direction observed by the upstream sentry station at that moment, the historical CFD background state field is obtained according to step 3). The historical CFD background state field All eddy covariance observation state features in the data are arranged into a background state vector. :

[0097]

[0098] Where J is the number of 3D mesh nodes. , and These represent the background easterly wind component, background northerly wind component, and background turbulent kinetic energy at the j-th 3D grid node, respectively, j=1,…J;

[0099] The measured eddy covariance state characteristics of each eddy covariance observation point at the same time are arranged into an observation vector. :

[0100]

[0101] in, The number of observation points, , and Let i represent the measured easterly wind component, northerly wind component, and turbulent kinetic energy at the i-th observation point, i=1,…I;

[0102] Historical turbulent state analysis field is achieved by minimizing the following three-dimensional variational cost function. get:

[0103]

[0104] in, For the three-dimensional variational cost function, the first term measures the deviation between the state field to be determined and the CFD background state field, and the second term measures the deviation between the state field to be determined at the observation point location and the measured eddy current value. Let be the eddy covariance observation state vector to be determined. This is the interpolation matrix from 3D mesh nodes to observation points. The spatial correspondence between observation points and 3D mesh nodes is determined by the positional identifiers of each observation point, thus constructing the interpolation matrix from 3D mesh nodes to observation points. The background error covariance matrix, This is the observation error covariance matrix; the smaller the background error or observation error, the stronger the constraint effect of the corresponding information in assimilation.

[0105] interpolation matrix The method employs trilinear interpolation, which involves linearly weighting the state values ​​of each grid node based on the spatial positions of the eight adjacent 3D grid nodes around the observation point, thereby obtaining the interpolation result corresponding to the position of the observation point. That is, the eddy covariance observed state vector to be determined is obtained through the interpolation matrix. Interpolation mapped to the locations of each observation point;

[0106] Background error covariance matrix For the easterly wind component respectively Northward wind component The turbulent kinetic energy TKE model is established, and no cross-covariance is set between different variables; for variables... , No. The and the first Background error covariance between 3D mesh nodes Defined as:

[0107]

[0108] in, For the first The and the first Horizontal distance between three-dimensional mesh nodes Let J be the height difference between the j-th and l-th grid nodes, where j, l = 1, ..., J; For variables The background error variance is calculated by interpolating the CFD background state field to each observation point, calculating the difference between it and the eddy covariance observation state characteristics obtained at the same time and location, and using the statistical variance of all effective differences of this variable as the background error variance. ; and Variables The horizontal and vertical correlation scales are determined by the spatial correlation function of the historical residuals from multiple observation points, with the residual correlation coefficients reduced to their initial values. The corresponding horizontal and vertical distances; observation error covariance matrix Set as a diagonal matrix, determined based on the measurement error variance obtained from the instrument calibration test;

[0109] The three-dimensional variational cost function is obtained through constrained optimization. The minimum state vector is used as the state vector for historical turbulence state analysis. And restore it to the historical turbulence state analysis field. .

[0110] Physical constraint corrections are performed during the assimilation correction process and after the solution is completed: easterly wind component of the 3D mesh nodes inside the building. Northward wind component The turbulent kinetic energy (TKE) is set to zero; TKE values ​​less than zero within the fluid region are corrected to zero or a non-negative lower limit, thus ensuring the analysis field satisfies the building boundary constraints and the non-negativity constraint of the turbulent kinetic energy. To avoid 0° / 360° angular jumps in the street valley direction, the field in the street valley direction is... Convert to: enter.

[0111] For any time t, the historical turbulence state analysis field and urban morphology conditions are combined into the input conditions of the diffusion model. :

[0112]

[0113] in, Let t be the easterly wind component, the northerly wind component, and the turbulent kinetic energy at time t, i.e., the historical turbulent state analysis field.

[0114] The historical turbulence state analysis field and urban morphological conditions at any given time t are used as input conditions, with the corresponding future time as the input. The historical turbulence state analysis field is used as the training objective. The input conditions and the training objective together constitute the training samples for the diffusion model. The predicted lead time is set to 15 minutes.

[0115] 5) Offline training of the eddy covariance observation-constrained diffusion model:

[0116] An eddy covariance observation constraint diffusion model is established, which includes a forward noise addition process, a reverse noise reduction process, and a noise reduction network for noise estimation. The noise reduction network adopts a U-shaped convolutional neural network (U-Net) to extract spatial features from the noisy three-dimensional state field and estimate the added noise under the input condition constraints.

[0117] Offline training was performed on the eddy covariance observation-constrained diffusion model. The training target was denoted as a noise-free target state field, which is the historical turbulence state analysis field for future time moments. Standard Gaussian random noise was added to the training target according to several preset noise intensity levels to obtain noisy state fields at different noise intensity levels:

[0118]

[0119] in, For the first Noisy state field at each noise intensity level For a noise-free target state field, Standard Gaussian random noise, As of the date The cumulative noise retention coefficient is calculated at each noise intensity level, and gradually decreases with increasing noise levels. The noise intensity level characterizes the proportion of random noise in the target state field. A low noise intensity level corresponds to slight noise perturbation, in which case the noisy state field mainly retains information from the original target state field; a high noise intensity level corresponds to strong noise perturbation, in which case the noisy state field is mainly composed of random noise components. =1,…S, where S is the number of noise intensity levels. The number of noise intensity levels is preset to 200 based on model training stability, generation accuracy, and computational resources.

[0120] The noisy state field, the corresponding noise intensity level, and the input conditions are then fed into the denoising network, which outputs an estimate of the added noise. Represented as:

[0121]

[0122] in, For the denoising network to be trained, The parameters are used for the denoising network. During training, the error between the real noise and the predicted noise is used as the training loss to update the denoising network parameters. After training, the eddy observation-constrained diffusion model is obtained.

[0123] The eddy covariance observation-constrained diffusion model is implemented in the form of a computer program; offline training is completed on a central server, cloud server, or local computing server; after training, the model parameter file is obtained and deployed to edge computing units, central servers, or cloud servers for online probabilistic forecasting.

[0124] 6) Online Forecast:

[0125] During online operation, based on the current incoming flow conditions (incoming wind speed and direction) at the upstream sentinel station, the system determines two adjacent incoming wind speed levels and two adjacent incoming wind direction zones from the CFD sample library obtained in step 3), interpolates to obtain the CFD background state field corresponding to the current incoming flow conditions, and obtains the current CFD background state vector. Based on the current easterly wind component at each observation point Northward wind component Constructing the current observation vector with turbulent kinetic energy TKE The current turbulence state analysis vector is obtained using the optimal interpolation method. :

[0126]

[0127] in, This is the difference between the measured value and the CFD background value at the observation point location, i.e., the observation increment; For the optimal interpolation gain matrix The current turbulence state analysis vector is then reconstructed into the current turbulence state analysis field using a three-dimensional mesh. The optimal interpolation gain matrix is:

[0128]

[0129] The interpolation matrix, background error covariance matrix, and observation error covariance matrix used for online optimal interpolation are determined using the same methods as in step 4). Since the target region grid and observation point locations are fixed, the optimal interpolation gain matrix... It can be pre-calculated and stored during the offline phase; during online operation, only the observation increment needs to be calculated and matrix multiplication needs to be completed.

[0130] Perform the same physical constraint correction as in step 4), adjusting the easterly wind component of the grid nodes inside the building. Northward wind component And set the turbulent kinetic energy TKE to zero, and set the less than zero kinetic energy in the fluid region to zero. The value is corrected to zero, thus obtaining the current turbulent state analysis field:

[0131] The current input conditions are constructed by combining the current turbulence state analysis field and urban morphology conditions. :

[0132]

[0133] Generate separately Each independent standard Gaussian random noise field is fed into a trained diffusion model along with the current input conditions. Through repeated calls to the denoising network during the reverse denoising process, a new denoising model is generated. Sample of future low-altitude turbulent state field:

[0134]

[0135] This represents the nth future low-altitude turbulent state field. Different random noise fields correspond to different future state field samples, and all samples are constrained by the current turbulent state analysis field and urban morphology conditions. The number N of future low-altitude turbulent state field samples is determined to be 50 based on the probability estimation accuracy requirements and online computation time. Each generated sample undergoes the same physical constraint correction as in step 4), ensuring that the eddy covariance observation state characteristics of the 3D mesh inside the building are zero and that the turbulent kinetic energy of the fluid region is non-negative.

[0136] 7) Calculate the probability risk field:

[0137] Based on N future low-altitude turbulent state field samples, the turbulent kinetic energy TKE is used as the criterion for strong turbulence. For the ... 3D grid nodes and prediction time Probability of strong turbulence Represented as:

[0138]

[0139] in, For the first The future low-altitude turbulent state field at the th 3D grid nodes and prediction time Turbulent kinetic energy below, The turbulent kinetic energy threshold is predetermined for the target aircraft. This is an indicative function. It takes the value 1 when the condition in parentheses is true (turbulent kinetic energy is greater than or equal to the turbulent kinetic energy threshold), and 0 otherwise (turbulent kinetic energy is less than the turbulent kinetic energy threshold). This yields the proportion of samples whose turbulent kinetic energy reaches the turbulent kinetic energy threshold out of the total sample. A strong turbulence risk level is set based on the turbulent kinetic energy threshold. For example, when TKE < 0.5 m² / s², it is classified as a low-turbulence state; when 0.5 m² / s² ≤ TKE < 2.0 m² / s², it is classified as a moderate-turbulence state; and when TKE ≥ 2.0 m² / s², it is classified as a strong-turbulence state.

[0140] The turbulent kinetic energy threshold of the target aircraft is pre-calibrated based on the allowable turbulence intensity provided by the aircraft manufacturer, flight test results, or historical flight attitude disturbance records, and remains fixed in the probabilistic prediction of the same type of aircraft. When the above data is unavailable, the turbulent kinetic energy threshold can be determined based on the statistical distribution of historical strong turbulence events in the target area.

[0141] Based on the probability of strong turbulence occurrence at each three-dimensional grid node, a three-dimensional urban low-altitude turbulence probability forecast field is formed for the target prediction area under different prediction timeframes.

[0142] Furthermore, risk warning information can be output based on the probability forecast field. Low-risk, medium-risk, and high-risk levels are set according to the probability of strong turbulence occurrence: a probability less than 0.3 is judged as low risk; a probability of not less than 0.3 and less than 0.6 is judged as medium risk; and a probability of not less than 0.6 is judged as high risk. The medium-risk and high-risk levels are set as early warning risks.

[0143] The 12-15 minute range is set according to a fixed prediction interval. =1 min is divided into four prediction timeframes. For the same 3D grid risk area, if its continuity... =If the risk warning is triggered within 2 prediction timeframes, then the duration of the risk is...

[0144] Represented as:

[0145]

[0146] Wherein, represents the fixed prediction interval between adjacent prediction timeframes. The duration of risk in the three-dimensional mesh risk area. This refers to the number of consecutive prediction periods that reach the warning risk level. Specifically, if two consecutive prediction periods reach the medium or high risk level, it is considered a persistent risk, with a risk duration of at least 2 minutes. Based on the spatial relationship between the risk area and low-altitude takeoff and landing points and flight routes, recommended takeoff and landing windows and route risk alerts can also be output.

[0147] Finally, it should be noted that the purpose of disclosing the embodiments is to help further understand the present invention. However, those skilled in the art will understand that various substitutions and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the content disclosed in the embodiments, and the scope of protection of the present invention is defined by the claims.

Claims

1. A method for probabilistic prediction of urban low-altitude turbulence based on an eddy observation-constrained diffusion model, characterized in that, The probability prediction method includes the following steps: 1) Construct an eddy covariance observation network: The target prediction area was delineated and divided into a three-dimensional grid; an eddy covariance observation network including observation points at multiple different locations was set up. 2) Calculate the characteristics of the eddy covariance observation state: High-frequency observation data were collected at each observation point, preprocessed, and the eddy covariance observation state characteristics of each observation point were calculated. 3) Obtain the CFD background state field and urban morphological conditions; According to the preset incoming flow conditions, the computational fluid dynamics (CFD) model is run to simulate and obtain the CFD background state field within the target prediction area, and a building-scale CFD sample library is established. Obtain urban morphological conditions; 4) Constructing training samples for the diffusion model: The background state field of CFD is subjected to three-dimensional variational assimilation correction using eddy covariance observation state characteristics, and the assimilation correction result is further corrected by physical constraints to obtain the historical turbulence state analysis field. The historical turbulence state analysis field at any time t and the urban morphological conditions are used as input conditions, with the corresponding future time... The historical turbulence state analysis field is used as the training target to form the training sample, and τ is the prediction time. 5) Offline training of the eddy covariance observation-constrained diffusion model: A constrained diffusion model based on eddy covariance observations is established. The model includes a forward noise addition process, a reverse noise reduction process, and a noise reduction network. The eddy covariance observation constraint diffusion model was trained offline using training samples to obtain the trained eddy covariance observation constraint diffusion model. 6) Online Forecast: Based on the current incoming flow conditions, the corresponding CFD background state field is obtained from the CFD sample library; Based on the current eddy covariance observation state characteristics of each observation point, the CFD background state field is optimally interpolated and assimilated to obtain the current turbulence state analysis field. N standard Gaussian random noise fields, along with the current turbulence state analysis field, urban morphology conditions, and prediction time, are input into the trained eddy covariance observation constraint diffusion model. Through a stepwise reverse denoising process, N future low-altitude turbulence state field samples are generated online, where N is a natural number ≥ 2. 7) Calculate the probability risk field: Based on N future low-altitude turbulence state field samples, the probability of strong turbulence risk occurring in the target prediction area is calculated, and the probability risk field is obtained.

2. The probability prediction method according to claim 1, characterized in that, In step 1), the observation points include a central main station, one or more lateral auxiliary stations, and one or more altitude monitoring stations. Eddy covariance systems are set up at each observation point, and each eddy covariance system is connected to the data processing platform via wired or wireless communication. Each eddy covariance system includes at least a three-dimensional ultrasonic anemometer, a pressure pulsation sensor, a data acquisition unit, a time synchronization module, and an edge computing module. The three-dimensional ultrasonic anemometer and the pressure pulsation sensor are connected to the data acquisition unit, the three-dimensional ultrasonic anemometer, the pressure pulsation sensor, and the data acquisition unit are connected to the time synchronization module, and the data acquisition unit and the time synchronization module are connected to the edge computing module. The edge computing module is connected to the data processing platform via wired or wireless communication.

3. The probability prediction method according to claim 1, characterized in that, In step 2), the preprocessing of high-frequency observation data includes: time synchronization, coordinate rotation, outlier removal, trend term processing, and quality control, resulting in an observation time series under a unified time reference. The high-frequency observation data obtained from the eddy covariance system measurements at each observation point includes average wind speed. Weather and wind direction And turbulent kinetic energy (TKE), average wind speed and weather wind direction Converted to easterly wind component and northerly wind component : Within a set sliding time window, the eddy covariance observation state characteristics of each observation point are calculated based on the observation time series; the eddy covariance observation state characteristics include the easterly wind component. Northward wind component And turbulent kinetic energy TKE.

4. The probability prediction method according to claim 1, characterized in that, In step 3), based on the three-dimensional geometry of the buildings and the ground boundary of the target prediction area, a building-scale CFD model is established, and a turbulent closure scheme that can directly output turbulent kinetic energy is adopted. According to the preset inflow conditions, the CFD model is used to simulate and generate the CFD background state field within the target prediction area.

5. The probability prediction method according to claim 4, characterized in that, In step 3), the incoming flow conditions include incoming wind speed and incoming wind direction. The incoming flow conditions are pre-classified, and multiple incoming wind speed levels and incoming wind direction zones are set. For each level of incoming wind speed and zone of incoming wind direction combination, a CFD model is run to form a building-scale CFD sample library.

6. The probability prediction method according to claim 1, characterized in that, In step 4), the three-dimensional variational assimilation correction of the CFD background state field using eddy covariance observation state characteristics includes the following steps: All eddy covariance observation state features in the historical CFD background state field are arranged into a background state vector. : Where J is the number of 3D mesh nodes. , and These represent the background easterly wind component, background northerly wind component, and background turbulent kinetic energy at the j-th 3D grid node, respectively, j=1,…J; The eddy covariance observation state characteristics of each eddy covariance observation point at the same time are arranged into an observation vector. : in, The number of observation points, , and Let i represent the measured easterly wind component, northerly wind component, and turbulent kinetic energy at the i-th observation point, i=1,…I; Historical turbulent state analysis field is achieved by minimizing the following three-dimensional variational cost function. get: in, It is a three-dimensional variational cost function; Let be the eddy covariance observation state vector to be determined. This is the interpolation matrix from the 3D mesh to the observation point. The background error covariance matrix, The observation error covariance matrix; The three-dimensional variational cost function is obtained through constrained optimization. The minimum state vector is used as the state vector for historical turbulence state analysis. And restore it to the historical turbulence state analysis field. .

7. The probability prediction method according to claim 6, characterized in that, In step 5), the eddy covariance observation constraint diffusion model is trained offline. Standard Gaussian random noise is gradually added to the training target through a forward noise addition process according to multiple preset noise intensity levels to obtain the noisy state field under different noise intensity levels. Then, the noisy state field, the corresponding noise intensity level and the input conditions are sent to the denoising network, and the denoising network outputs the estimated value of the added noise. The model parameters are updated by reducing the difference between the actual added random noise and the estimated output noise; after repeated training with multiple training samples, the trained eddy covariance observation constraint diffusion model is obtained.

8. The probability prediction method according to claim 7, characterized in that, The training target is denoted as a noise-free target state field, which is the historical turbulence state analysis field at future moments. Standard Gaussian random noise is added to the training target according to multiple preset noise intensity levels to obtain noisy state fields at different noise intensity levels: in, For the first Noisy state field at each noise intensity level For a noise-free target state field, Standard Gaussian random noise, As of the date The cumulative noise retention factor at each noise intensity level gradually decreases as the noise addition steps increase. =1,…S, where S is the number of noise intensity levels; Output of the denoising network Represented as: in, For the denoising network to be trained, For the parameters of the denoising network, The input conditions are used as input conditions; during training, the error between the real noise and the predicted noise is used as the training loss to update the parameters of the denoising network. After training, the eddy observation constrained diffusion model is obtained.

9. The probability prediction method according to claim 8, characterized in that, In step 6), during online operation, based on the current incoming flow conditions from the upstream sentinel station, the CFD background state field corresponding to the current incoming flow conditions is interpolated from the CFD sample library obtained in step 3), and the current CFD background state vector is obtained. The current observation vector is constructed based on the current easterly wind component, northerly wind component, and turbulent kinetic energy at each observation point. The current turbulence state analysis vector is obtained using the optimal interpolation method. : in, The optimal interpolation gain matrix is ​​used; the current turbulence state analysis vector is used. The current turbulent state analysis field is reconstructed using a three-dimensional mesh. .

10. The probability prediction method according to claim 9, characterized in that, In step 7), based on multiple future low-altitude turbulent state field samples, the turbulent kinetic energy TKE is used as the criterion for strong turbulence; for the first... 3D grid nodes and prediction time Probability of strong turbulence Represented as: in, For the first The future low-altitude turbulent state field at the th 3D grid nodes and prediction time Turbulent kinetic energy below, The threshold for turbulent kinetic energy. Let N be the indicative function, and N be the number of future low-altitude turbulent state field samples.

Images