Extreme gale weather prediction method based on a large meteorological prediction model
By using multi-source assimilation and micro-downburst calculation based on a large meteorological forecast model, combined with boundary layer similarity mapping and hysteresis kernel, high-precision gust distribution and over-threshold probability are generated, solving the problem of insufficient reliability in the prediction of extreme wind weather in existing technologies, and realizing accurate forecasts at the minute and kilometer levels.
Patent Information
- Application Number
- CN202511612798.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-06
- Publication Date
- 2026-02-24
- Estimated Expiration
- 2045-11-06
AI Technical Summary
Existing technologies cannot achieve minute-level or kilometer-level landing area boundaries, pixel-level over-threshold probability, and calibrable alarm benchmarks, resulting in insufficient reliability in the location, timing, and probability of extreme wind weather forecasts, affecting power transmission and distribution, airport operations, port operations, wind turbine units, and forest area safety.
Based on a large meteorological forecast model, a consistent three-dimensional physical field is formed through multi-source assimilation. By combining micro-downburst calculation, boundary layer similarity mapping and hysteresis kernel, gust distribution and over-threshold probability are generated. An optimal transport and conditional diffusion probability generation model with advection constraints is used to perform parameter calibration and event consistency verification.
It achieves high-precision forecasting of extreme wind weather, improves the reliability of positioning, timing and probability, and supports decisions such as scheduling, flow control, shutdown and emergency deployment.
Smart Images

Figure CN121069530B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of meteorological disaster monitoring, early warning and numerical forecasting fusion technology, and in particular to a method for predicting extreme wind weather based on a large meteorological forecasting model. Background Technology
[0002] Extreme winds impact power transmission and distribution, as well as the safety of transmission corridors, causing line trips and localized power outages; they affect airport operations, low-altitude air routes, and ground operations; pose risks to port loading and unloading, offshore operations, and cross-sea passage; damage urban construction sites, high-rise building facades, and temporary structures; cause wind turbines to face shutdowns and fatigue accumulation due to exceeding cutoff wind speeds; trigger fallen trees and secondary blockages in forest areas; and increase pricing and compensation uncertainty in insurance and reinsurance. The industry needs minute-level and kilometer-level coverage boundaries, pixel-level over-threshold probabilities, and calibrable alarm benchmarks to support decisions such as dispatching, flow control, shutdown, and emergency deployment. Summary of the Invention
[0003] To address the numerous problems existing in the prior art, this invention provides a method for predicting extreme wind weather based on a large meteorological forecasting model. Starting with a consistent three-dimensional physical field obtained through multi-source assimilation, this invention uses a time-progressive micro-downburst calculation program to generate downsinking potential energy, cold pool depth, and dummy temperature difference. A gust baseline is obtained through boundary layer similarity mapping and a hysteresis kernel. Combined with an optimal transport and conditional diffusion probability generation model constrained by advection, the pixel gust distribution and over-threshold probability are output. Finally, closed-loop parameter updates are performed using monotonically normalized flow and event consistency verification to improve the reliability of positioning, timing, and probability.
[0004] A method for predicting extreme wind weather based on a large-scale meteorological forecasting model includes:
[0005] The system acquires three-dimensional physical fields for numerical weather prediction, weather radar, meteorological satellites and ground meteorological observations, unifies the grid and assimilates the data through observation operators to form a consistent three-dimensional physical field and atmospheric stability and low-level wind shear.
[0006] Time progression is performed and vertical acceleration is calculated based on a micro-downburst calculation program. The sinking potential energy is obtained by integration, the depth of the cold pool and the dummy temperature difference are diagnosed, and the location and probability of occurrence of the cold pool front are constructed.
[0007] Based on the sinking potential energy, cold pool depth and illusory temperature difference, coarse resolution gusts are obtained through boundary layer similarity mapping and hysteresis kernel. Optimal transmission downsampling with advection constraints is adopted and the cold pool front is used as an insurmountable boundary. Combined with the conditional diffusion probability generation model, the pixel gust distribution and over-threshold probability are output and the wind damage polygon is extracted.
[0008] Aligning radar lower-level divergence, satellite brightness temperature cooling rate, lightning density variation, and ground gusts, applying monotonically normalized flow for probabilistic calibration, and using crossover ratio and conformal coverage for event consistency verification, the boundary layer similarity mapping parameters, hysteresis kernel, and generation variance are updated under amplitude-limited conditions.
[0009] Preferably, when acquiring and assimilating the three-dimensional physical field of numerical weather prediction, weather radar, meteorological satellite and ground meteorological observations, the observation operator generates radar reflectivity simulation and radial velocity simulation, satellite brightness temperature simulation and cloud top height simulation, and ground wind speed simulation and ground pressure simulation, respectively. In a unified grid and terrain-following vertical coordinate, variational assimilation weighted by observation error is used to generate a consistent three-dimensional physical field and atmospheric stability and low-level wind shear.
[0010] Preferably, when advancing time, a neural operator is used to update the three-dimensional physical field, and the update result is corrected by using the residuals of the continuity equation and the total water conservation residuals as regularization terms.
[0011] Preferably, when calculating the vertical acceleration based on the micro-downburst calculation program, the vertical acceleration is obtained by adding the evaporative cooling buoyancy and the precipitation load. The sinking potential energy is obtained by integrating the negative value of the vertical acceleration along the height. The depth of the cold pool and the virtual temperature difference are determined based on the vertical distribution of the virtual temperature difference. The horizontal set function of the location and occurrence probability of the cold pool front is constructed by the virtual temperature difference gradient and the low-level divergence.
[0012] Preferably, when obtaining coarse-resolution gust winds based on sinking potential energy, cold pool depth, and virtual temperature difference through boundary layer similarity mapping, the boundary layer similarity mapping takes frictional velocity, stability parameters, surface roughness, turbulent kinetic energy, and near-surface wind shear as inputs, and applies monotonicity constraints to surface roughness and stability parameters.
[0013] Preferably, when obtaining the coarse-resolution gust based on the sinking potential energy, cold pool depth, and dummy temperature difference through a hysteresis kernel, the hysteresis kernel is composed of a time convolution kernel formed by adding two exponential kernels, which is used to align the time series of the coarse-resolution gust with the time scale of the cold pool evolution.
[0014] Preferably, when using advection-constrained optimal transport downsampling and taking the cold pool front as an insurmountable boundary, the near-surface average wind field is used as the advection constraint, and the boundary conditions of zero normal flux and tangential energy redistribution are applied at the location of the cold pool front to obtain the fine-scale field of gusts.
[0015] Preferably, when combining the conditional diffusion probability generation model to output the pixel gust distribution and the over-threshold probability, the fine-scale field of gusts, surface roughness, terrain data, building height, low-level wind shear, turbulent kinetic energy, cold pool front occurrence probability, low-level divergence and ensemble confidence weight are used as the condition set, and energy consistency regularization, topological persistence screening and extreme value distribution embedding are applied to the pixel gust distribution.
[0016] Preferably, after aligning the radar lower-level divergence, satellite brightness temperature cooling rate, lightning density changes, and ground gusts, a monotonically normalized flow is applied to perform probability calibration on the over-threshold probability. The monotonically normalized flow adopts a piecewise affine transformation while maintaining monotonicity and endpoint invariance, and the consistency between the Brill score and the reliability curve within the sliding time window is used as the optimization objective.
[0017] Preferably, when performing event consistency verification using cross-union ratio and conformal coverage, a synthetic loss is formed based on cross-union ratio, peak occurrence time difference, and cold pool front geometric deviation, and the boundary layer similarity mapping parameters, hysteresis kernel, and generation variance are updated under amplitude-limited conditions.
[0018] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows:
[0019] By integrating numerical weather prediction's three-dimensional physical field, weather radar, meteorological satellites, and ground meteorological observations through "observation operator + variational assimilation," a unified output of consistent three-dimensional physical field, stability, and low-level wind shear was achieved.
[0020] By connecting microphysical sinking with near-ground gusts through a "micro-downburst calculation program + sinking potential energy", a triggering forcing quantity consistent with cold pool dynamics was achieved.
[0021] By characterizing the surface roughness and stratification memory effect through "boundary layer similarity mapping + hysteresis kernel", physical constraints on the amplitude and duration of gusts were achieved.
[0022] By using "optimal transport downsampling under advection constraints" and taking the cold pool front as an insurmountable boundary, the geometric consistency and energy conservation of fine-scale gusts on both sides of the front were achieved.
[0023] By generating pixel gust distributions on a fine-scale base map using a "conditional diffusion probability generation model", the same source output of the over-threshold probability and the wind damage polygon is achieved.
[0024] By calibrating the over-threshold probability using a "monotonically normalized flow", consistency between probability and observation frequency and interpretable threshold selection are achieved.
[0025] By verifying the consistency of events through "intersection-union ratio and conformal coverage", joint constraints on spatial overlap, occurrence time and geometric deviation are achieved;
[0026] By applying "amplitude-limited online parameter updates" to the boundary layer similarity mapping parameters, hysteresis kernel, and generation variance, stable adaptation under incomplete observation conditions is achieved. Attached Figure Description
[0027] Figure 1 This is a flowchart illustrating the method of the present invention. Detailed Implementation
[0028] The embodiments of the present disclosure will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the disclosure. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of the present disclosure for ease of explanation. However, it will be apparent that one or more embodiments may be practiced without these specific details. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concepts of the present disclosure.
[0029] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit this disclosure. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.
[0030] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.
[0031] like Figure 1 As shown, a method for predicting extreme wind weather based on a large-scale meteorological forecasting model includes:
[0032] The system acquires three-dimensional physical fields for numerical weather prediction, weather radar, meteorological satellites and ground meteorological observations, unifies the grid and assimilates the data through observation operators to form a consistent three-dimensional physical field and atmospheric stability and low-level wind shear.
[0033] Preferably, when acquiring and assimilating the three-dimensional physical field of numerical weather prediction, weather radar, meteorological satellite and ground meteorological observations, the observation operator generates radar reflectivity simulation and radial velocity simulation, satellite brightness temperature simulation and cloud top height simulation, and ground wind speed simulation and ground pressure simulation, respectively. In a unified grid and terrain-following vertical coordinate, variational assimilation weighted by observation error is used to generate a consistent three-dimensional physical field and atmospheric stability and low-level wind shear.
[0034] This invention aligns the three-dimensional physical field of numerical weather prediction, weather radar volume scan data, meteorological satellite brightness temperature and cloud top height products, and ground meteorological observations into a unified grid and terrain-following vertical coordinate system according to time windows. The time windows employ a symmetrical sliding method to ensure controllable errors in the observation and reporting times involved in the assimilation. The unified grid uses the same projection and horizontal resolution as the operational area, and a terrain-following hierarchical division is adopted in the vertical direction, maintaining consistency between the grid and hierarchy throughout the entire process. To reduce false signals, ground feature and biological echoes are removed from the weather radar, and the consistency of scanning angle and cloud detection is checked for meteorological satellites. Outliers and physical boundaries are checked for ground meteorological observations; the results of these checks are used as weights input into the assimilation stage.
[0035] Secondly, observation operators matching the three-dimensional physical field are established for various sensors. The weather radar observation operator maps the distribution of condensate and wind vectors in the three-dimensional physical field into simulated radar reflectivity and radial velocity; the meteorological satellite observation operator generates simulated brightness temperature and cloud top height based on the approximate relationship between cloud top temperature, cloud top height, and radiation transfer; the ground meteorological observation operator extracts near-surface wind and surface air pressure from the three-dimensional physical field into simulated surface wind speed and surface air pressure. The above observation operators are strictly aligned with a unified grid, ensuring that each observation corresponds to a unique grid position and vertical level in space, and includes observation error variance for use in assimilation weights.
[0036] The core fusion is achieved through variational assimilation, aiming to preserve the background field structure of numerical weather prediction while making the simulated observations as close as possible to real observations. The following cost function is used as the minimization objective of the assimilation:
[0037]
[0038] in, This represents the state vector of a three-dimensional physical field. The background field state vector, The background error covariance matrix, For the first Observational operators, For the first Class observation vector, For the first Observation error covariance matrix For the cost function, superscript This is a transpose. Numerical optimization is used to iteratively solve the problem within a time window until the cost function converges; the process stops when the cost function decreases below a threshold or reaches the maximum number of iterations. To ensure stability, a batch insertion of observations and quality weights is adopted: observations that pass quality checks are assigned higher weights, while observations from marginal perspectives or under conditions of heavy precipitation attenuation are assigned lower weights.
[0039] After assimilation, a consistent three-dimensional physical field is obtained, and atmospheric stability and low-level wind shear are calculated based on this field. Stability is based on the vertical profiles of temperature and water vapor, outputting a set of indices for convective instability and stratification characteristics; low-level wind shear is defined using the wind vector difference between two fixed-height layers.
[0040]
[0041] in, The modulus of low-level wind shear. The horizontal wind vector at the upper level. The horizontal wind vector at the lower level. The height of the upper level The lower layer height is used. The shear calculation relies solely on the assimilated three-dimensional physical field, avoiding spatial voids caused by sparse observations and ensuring consistency with variables in subsequent gust mappings.
[0042] To facilitate data tracing in subsequent processes, this invention records the time window, observation coverage ratio, quality weights of various observations, and optimized convergence indexes while outputting consistent three-dimensional physical fields, atmospheric stability, and low-level wind shear. If the observation coverage of a certain area is insufficient, a lower confidence weight is recorded in the corresponding grid. Subsequent time series advancement and probability generation can adopt a conservative strategy accordingly to reduce the risk of error propagation.
[0043] Example: A unified grid is established in areas with mixed urban clusters and complex terrain, with horizontal resolution and vertical hierarchy configured according to operational specifications. A time window of 30 minutes before and after the reporting start time is selected, aggregating the three-dimensional physical field of numerical weather prediction, weather radar volume scan, meteorological satellite brightness temperature and cloud top height products, and automatic ground station wind and air pressure data. After quality control, observation operators are used to generate simulated quantities of radar reflectivity, radial velocity, brightness temperature, cloud top height, surface wind speed, and surface air pressure. A cost function is constructed and minimized to obtain a consistent three-dimensional physical field. Based on this three-dimensional physical field, the set of stability indices and wind shear at both altitude levels are calculated. The data, along with the reliable weights, is output to provide input for subsequent time progression, sinking potential energy, cold pool depth, and dummy temperature difference diagnosis. This embodiment can automatically reduce the weights of the corresponding grids during periods of sparse observation, ensuring the spatial continuity and physical consistency of the assimilation results, and laying a reliable foundation for the probabilistic generation of gusts in subsequent steps.
[0044] Time progression is performed and vertical acceleration is calculated based on a micro-downburst calculation program. The sinking potential energy is obtained by integration, the depth of the cold pool and the dummy temperature difference are diagnosed, and the location and probability of occurrence of the cold pool front are constructed.
[0045] The time progression employs a fixed-step rolling method. The propulsion engine reads wind, temperature, humidity, and condensate data from a consistent 3D physics field and outputs the 3D physics field for the next time step. The side boundaries are consistent with numerical weather prediction, a sponge layer is set at the top to absorb disturbances, and actual terrain and surface roughness are used at the bottom for near-surface layer diagnostics. The propulsion output is directly fed into the micro-downburst calculation program, avoiding information loss caused by cross-coordinate interpolation.
[0046] The micro-downburst calculation program calculates vertical acceleration at each vertical layer. Vertical acceleration consists of two parts: a buoyancy term characterized by illusory temperature difference, and a downward pull term generated by the mass carried by precipitation on the air column. The buoyancy term is obtained by the difference between the environmental profile and the grid profile, while the precipitation downward pull term is obtained by parameterizing the rainwater and hail content. Both terms are derived from the time-progressed three-dimensional physics field, without relying on additional sensors. The program performs amplitude limiting and spatial smoothing for outliers to ensure numerical stability.
[0047] This invention uses subsidence potential energy as the core variable connecting microphysical subsidence and near-ground gusts, and the core calculation expression is as follows:
[0048]
[0049] in, For sinking potential energy, The coordinates of the horizontal position are the meridional coordinates. The coordinates are the latitudinal coordinates of the horizontal position. This is the height near the top of the stratum. The upper bound of the integral is the height. This represents vertical acceleration. The above expression only accumulates layers that accelerate downwards, avoiding interference from upward motion.
[0050] Cold pool diagnosis is based on subsidence potential energy and illusory temperature difference field. First, the illusory temperature difference field is calculated, defined as the ambient illusory temperature minus the grid illusory temperature. For each horizontal grid, the layer thickness continuously exceeding a threshold is searched from bottom to top as the cold pool depth, and the illusory temperature difference within that layer thickness is averaged to form a near-surface illusory temperature difference index. The threshold is set regionally and written into metadata for easy auditing and reproduction.
[0051] The location of the cold pool front was determined in three steps. First, candidate region selection: candidate zones were marked in a grid with significant virtual temperature differences and positive near-surface divergence. Second, boundary refinement: within the candidate zones, continuous maximal gradient connected zones were searched along the horizontal gradient direction of the virtual temperature difference to obtain the initial frontal polygon. Third, geometric smoothing and fracture repair: the polygon was smoothed with length and curvature constraints, isolated short segments were removed, and adjacent fracture segments were connected to obtain a coherent front. The probability of frontal occurrence was obtained by weighting the virtual temperature difference intensity, near-surface divergence intensity, and virtual temperature difference gradient intensity, followed by a monotonically bounded compression mapping. The weights were calibrated regionally based on historical cases.
[0052] In engineering applications, the time-tracking and calculation programs share the same grid and vertical layer, reducing interpolation errors. Uncertain regions caused by observation occlusion or echo attenuation are recorded with lower confidence weights in the fields, and subsequent probabilistic steps employ conservative strategies accordingly. All intermediate results are output as grid products, including timestamps, coordinate information, units, threshold source, and version number. To avoid non-physical energy bursts, an upper limit is set for the sinking potential energy in extremely dry and cold environments; the clamping reference source and threshold are also recorded in the metadata.
[0053] Example: During a severe summer convection event, a unified grid was established in the area where the urban cluster and the piedmont meet. After time-progression, the 3D physics field was input into a micro-downburst calculation program to obtain the vertical acceleration field, and the subsidence potential energy grid was calculated according to the core expression. A continuous cold pool depth zone with positive near-surface divergence was formed in the leeward region of the city, and candidate zones were marked. A high-gradient connected zone was formed along the horizontal gradient direction of the illusory temperature difference, and the frontal polygon was obtained after smoothing and fracture repair, while a frontal occurrence probability grid was generated. Finally, the subsidence potential energy, cold pool depth, illusory temperature difference, frontal polygon, and frontal occurrence probability were output, providing direct input for subsequent gust mapping and probability generation. This example automatically reduces the confidence weight in the radar-blocked area, but the algorithm flow remains unchanged, ensuring the spatial consistency and temporal continuity of diagnostic variables.
[0054] Preferably, when advancing time, a neural operator is used to update the three-dimensional physical field, and the update result is corrected by using the residuals of the continuity equation and the total water conservation residuals as regularization terms.
[0055] The input for time-progression is a consistent three-dimensional physical field, including horizontal wind, vertical wind, temperature, virtual potential temperature, pressure, air density, and the mixing ratio of various condensates. Before entering the model, the input is mean-reduced and scaled by channel, and missing measurements are masked. The boundary conditions are consistent with numerical weather prediction, with open relaxation for the lateral boundaries, a sponge layer at the top, and the bottom retaining topography and surface roughness for near-surface diagnosis.
[0056] The neural operator employs a structure combining frequency-domain convolution and spatial local convolution. The input is the current 3D physics tensor, and the output is an increment tensor of the same size. The propagation rule is:
[0057]
[0058] in, This is a three-dimensional physical field state vector. To advance the time variable of time, the accumulated time is taken from the starting time of reporting as zero. For time step, This is a neural operator. The update does not modify the grid or coordinates; it only numerically provides the field for the next time step.
[0059] To suppress long-term drift and ensure consistency of key conserved quantities, a two-part cost function is constructed to perform small-step corrections on the neural operator output:
[0060]
[0061] in, air density, It is a three-dimensional wind vector. The sum of the mixing ratios of the total water content. For the main water source confluence, These are the weighting coefficients. This is the L2 norm of the sum of squares at each grid point. The first term corresponds to the continuity residual of air quality, and the second term corresponds to the total water conservation residual. The source and sink terms are diagnostic quantities from microphysical processes, ensuring that the contribution of water phase transition is correctly subtracted or added.
[0062] The correction process involves one to three small iterations at each time step. Specifically, the residuals are calculated based on the updated 3D physics field. Small-scale corrections are then made to the wind and water vapor fields according to the gradient direction of the cost function. The cost function is re-evaluated after correction. If the cost function decreases below a preset threshold, the iteration stops. The correction amount is limited by the upper bound of the amplitude of each channel to prevent excessive interference with the master solution of the neural operator. The numerical differencing used is central differencing, with the vertical direction consistent with the terrain-following coordinates. The divergence of the boundary cells is calculated using first-order extrapolation to ensure the continuity of the cost function at the boundaries.
[0063] To ensure seamless integration with subsequent modules, the output of this step at each time step includes: the corrected 3D physics field, the air quality continuity residual raster, the total water conservation residual raster, and the cost function value. If the continuity residual or the total water conservation residual is consistently large in a certain region, the confidence weight for that region is recorded in the metadata, allowing for the use of more conservative parameters when diagnosing sinking potential energy and cold pools. All the above data includes timestamps and version numbers for easy auditing and playback.
[0064] In terms of training and deployment, the neural operator is trained offline using historical reanalysis and high-resolution numerical simulations under supervision, while incorporating a physical regularization term consistent with the cost function to enable the model to learn an incremental mapping that is close to conservation. During online inference, small-step corrections are still performed to adapt to background errors caused by real-time observations. The effect of this design is that during the rapid development of strong convection, the wind field and water vapor field will not exhibit non-physical divergence or total water level jumps, the subsequent calculations of sinking potential energy, cold pool depth, and illusory temperature difference are more stable, and the spatial geometry of the cold pool front is more consistent.
[0065] Example: In a business area covering 200×200 km and 40 vertical layers, rolling prediction is performed with a time step of 3 minutes. Each step first generates and updates the 3D physical field increment using a neural operator, then calculates the continuity residual and the total water conservation residual. If the cost function exceeds a threshold, two small-step corrections are performed, with the channel amplitude of a single correction not exceeding 10% of the channel range. After processing, the corrected 3D physical field and residual grid are output and entered into the subsidence potential and cold pool diagnostic process. Field tests show that this process can significantly reduce the anomalous drift of near-surface divergence during periods of heavy precipitation echo propagation, ensuring the robustness of downstream modules.
[0066] Preferably, when calculating the vertical acceleration based on the micro-downburst calculation program, the vertical acceleration is obtained by adding the evaporative cooling buoyancy and the precipitation load. The sinking potential energy is obtained by integrating the negative value of the vertical acceleration along the height. The depth of the cold pool and the virtual temperature difference are determined based on the vertical distribution of the virtual temperature difference. The horizontal set function of the location and occurrence probability of the cold pool front is constructed by the virtual temperature difference gradient and the low-level divergence.
[0067] Within a unified grid and terrain-following vertical coordinate system, the system processes the 3D physics field in a fixed-time-step rolling manner. First, it obtains the current wind, temperature, humidity, and condensate field through time progression. Then, it calls a micro-downburst calculation program to diagnose vertical acceleration at each vertical layer. Vertical acceleration consists of buoyancy and precipitation load. Buoyancy is characterized by the illusory temperature difference, and precipitation load is parameterized from rainwater and hail content. Outliers are amplitude-limited and smoothed on a small scale along the vertical direction to suppress numerical oscillations. The core expression is:
[0068]
[0069] It is vertical acceleration; This is the buoyancy term; This refers to the precipitation load item; For height.
[0070] The sinking potential energy is calculated based on vertical acceleration. In the implementation, a trapezoidal method is used in the vertically discrete layers, accumulating contribution only to layers accelerating downwards. An upper limit is set in a dry, cold environment to prevent abnormal energy accumulation, and the threshold and version information are written into the metadata. The core expression is: .
[0071] Cold pool diagnosis is based on the vertical distribution of illusory temperature difference. The system first calculates the illusory temperature difference field (ambient illusory temperature minus grid illusory temperature), and then searches for layers exceeding a threshold continuously from bottom to top within each horizontal grid, which are taken as the cold pool depth. Within this layer thickness, the illusory temperature difference is averaged over a height to obtain the near-surface illusory temperature difference index. The threshold is calibrated by region, and its source and number are recorded in the metadata. The output cold pool depth raster, illusory temperature difference raster, and subsidence potential energy raster are consistent in grid, timestamp, and units, facilitating direct reading by subsequent modules.
[0072] Cold pool front extraction consists of three steps: candidate region selection, boundary refinement, and geometric repair. The candidate region comprises a grid with significant virtual temperature difference and positive low-level divergence. Boundary refinement extracts continuous high-gradient connected zones along the horizontal gradient direction of the virtual temperature difference within the candidate region, forming an initial polygonal line. Geometric repair smooths the polygonal line using length and curvature constraints, eliminating isolated short segments and connecting adjacent fault segments to obtain a coherent frontal polygonal line. The probability of frontal occurrence is determined by weighting the intensity of virtual temperature difference, low-level divergence, and virtual temperature difference gradient, and then normalizing it to the 0-1 interval using a monotonically bounded compression mapping. The weights are fitted to historical samples of the region and can be set according to terrain type.
[0073] In terms of data organization, each time step outputs a sinking potential energy grid, a cold pool depth grid, a dummy temperature difference grid, a cold pool front polygonal line, and a frontal occurrence probability grid, along with confidence weights and quality indicators. All products and time-progression phases use the same grid and vertical definition to avoid error propagation caused by secondary interpolation; confidence weights are reduced for observed occlusions or inferred unstable regions, allowing subsequent probabilistic steps to adopt conservative strategies.
[0074] Example: During a severe summer convection event, the operational domain was 200×200 km, with 40 vertical layers and a time step of 3 minutes. After time progression, the 3D physics field was input into the micro-downburst calculation program. Vertical acceleration and downdraft potential energy were calculated according to the core expressions described above, and cold pool depth and dummy temperature difference diagnosis were completed. The system identified a continuous cold pool zone downwind of the city, extracted high-gradient connectivity lines along the horizontal gradient of the dummy temperature difference, obtained frontal polygons through geometric repair, and simultaneously generated an occurrence probability grid. This example demonstrates that this process can stably depict the cold pool advancement path and the potential downburst impact range, and reduce the risk of false alarms in radar-blocked areas through trusted weight downgrading.
[0075] Based on the sinking potential energy, cold pool depth and illusory temperature difference, coarse resolution gusts are obtained through boundary layer similarity mapping and hysteresis kernel. Optimal transmission downsampling with advection constraints is adopted and the cold pool front is used as an insurmountable boundary. Combined with the conditional diffusion probability generation model, the pixel gust distribution and over-threshold probability are output and the wind damage polygon is extracted.
[0076] Within a unified grid and terrain-following vertical coordinate system, a coarse-resolution gust map is first obtained based on subsidence potential energy, cold pool depth, and illusory temperature difference. Then, a fine-scale gust base map is generated through optimal transport downsampling under advection constraints. Finally, a conditional diffusion probability generation model provides the pixel gust distribution and over-threshold probability, and wind damage polygons are extracted. The process and data interface remain consistent with the preceding steps to facilitate engineering integration and auditing.
[0077] First, the gust amplitude principal term is constructed using subsidence potential energy, cold pool depth, and illusory temperature difference. Then, a boundary layer similarity mapping is superimposed to obtain a coarse-resolution gust. The core calculation expression is:
[0078]
[0079] The coarse resolution is 10-meter gusts; This is the potential energy for sinking; This refers to the depth of the cold pool; This is the virtual temperature difference; This is the scaling factor; This is the boundary layer similarity mapping factor.
[0080] To characterize cold pool propulsion and wake effects, hysteresis convolution is applied to the time series:
[0081]
[0082]
[0083] To predict the timeliness; It is a delayed kernel; and As weight; and Time scale; This is the convolution time variable.
[0084] The boundary layer similarity mapping factor is obtained by mapping friction velocity, stability parameter, surface roughness, turbulent kinetic energy and near-ground wind shear. It is trained on historical samples and fine-tuned in small steps with amplitude limit during inference.
[0085] Subsequently, optimal transport downsampling with advection constraints is performed. Using the square of the coarse-resolution gust as the energy density input, semi-Lagrange advection is first performed based on the near-surface mean wind. Then, regularized optimal transport redistributes the energy across the fine grid, with the cold pool front as the insurmountable boundary: normal flux is zero, and energy slippage is allowed along the tangential direction. To avoid non-physical deformation of the fine grid, a small-amplitude deformation smoothing projection is introduced. The output is a fine-scale gust base map, consistent with the input grid coordinates.
[0086] A conditional diffusion probability generation model is constructed under the condition of a fine-scale gust base map. The condition set includes the fine-scale gust base map, surface roughness, topographic data, building height, low-level wind shear, turbulent kinetic energy, cold pool front occurrence probability, and near-surface divergence. The model generates several gust samples using sliding window tiles and applies energy consistency constraints and topological connectivity filtering to suppress isolated spikes and discontinuous bands. For extreme tails, statistical tail fitting is used to improve the amplitude and probability characterization accuracy of strong wind pixels. The over-threshold probability is obtained from the sample frequency.
[0087]
[0088] This represents the probability of a pixel exceeding a threshold. The number of samples; For the first One sample of gusts; The threshold value is used.
[0089] Connected components are extracted in the probability field through continuous thresholding and morphological refinement to generate wind damage polygons, and the peak time, peak gust and quantile gust are saved as polygon attributes.
[0090] In engineering applications: coarse-resolution and fine-scale gust base maps are processed in the same coordinate system to avoid interpolation errors; hysteresis kernel parameters and boundary layer similarity mapping factors are updated with amplitude limits during runtime to ensure stability; the optimal transmission step uses a fixed number of iterations and temperature parameters to ensure timeliness; in the probability generation stage, the sample quantity and tile overlap are adaptively set according to the regional sample size, and random seeds and version information are recorded uniformly. All outputs include metadata for raster and vector products, such as timestamps, units, parameter versions, and quality flags, for direct consumption and traceability by downstream modules.
[0091] Example: In a severe convection forecast, the subsidence potential energy, cold pool depth, and dummy temperature difference from the preceding diagnostics are input to calculate coarse-resolution gusts and apply hysteresis convolution. At the junction of urban clusters and plains, advection is performed based on near-surface average wind, and optimal transport downsampling is completed to obtain a fine-scale gust base map that is more sensitive to road and building scales. Conditional diffusion sampling is performed on a condition set including building height and surface roughness to obtain a sample set and a probabilistic grid exceeding the threshold. Two wind damage polygon bands are obtained through connected component extraction, which coincide with the radar's lower-level divergence enhancement area. This process maintains that the normal does not cross at the cold pool front, avoiding the diffusion of high-value gusts across the front and improving the spatial positioning accuracy of the extreme wind area.
[0092] Preferably, when obtaining coarse-resolution gust winds based on sinking potential energy, cold pool depth, and virtual temperature difference through boundary layer similarity mapping, the boundary layer similarity mapping takes frictional velocity, stability parameters, surface roughness, turbulent kinetic energy, and near-surface wind shear as inputs, and applies monotonicity constraints to surface roughness and stability parameters.
[0093] Within a unified grid and terrain-following vertical coordinate system, the subsidence potential energy grid, cold pool depth grid, and dummy temperature difference grid, output from previous steps, are first read. Friction velocity, stability parameters, surface roughness, turbulent kinetic energy, and near-surface wind shear are then derived from the 3D physics field. Friction velocity is calculated from surface stress and air density; stability parameters are taken from the dimensionless stability of the near-surface scheme; surface roughness is derived from underlying surface data; and turbulent kinetic energy and near-surface wind shear are directly derived from the 3D physics field or its diagnostic quantities. All variables are spatially bilinearly aligned, and missing measurements are masked and weighted in subsequent mappings.
[0094] The coarse-resolution gust is obtained by multiplying the "gust baseline amplitude" and the "boundary layer similarity mapping factor". The gust baseline amplitude is jointly determined by the subsidence potential energy, cold pool depth, and illusory temperature difference, serving as a ensemble of forcing intensity. The boundary layer similarity mapping factor is used to correct the forcing intensity to the gust response at a height of 10 meters, explicitly incorporating near-surface dynamic and thermal information. The core calculation expression is as follows:
[0095]
[0096] The coarse resolution is 10-meter gusts; The baseline amplitude of the gust is composed of the sinking potential energy, the depth of the cold pool, and the virtual temperature difference. This is the boundary layer similarity mapping factor.
[0097] The boundary layer similarity mapping factor is in log-linear form to facilitate monotonic constraints and partitioning.
[0098]
[0099] The frictional speed; For stability parameters; It is a monotonic transformation function of the stability parameter and decreases monotonically as stability increases; For surface roughness; It is turbulent kinetic energy; This is near-surface wind shear; to These are the coefficients that need to be calibrated. To satisfy physical rationality, a monotonicity constraint is implemented: an increase in surface roughness should reduce the gust response by 10 meters, and the constraint is... Stability enhancement should suppress gust response, with constraints as follows: As stability monotonically decreases and To avoid extreme responses, During runtime, it is restricted to a preset upper and lower bound, which are set and recorded in the metadata according to climate zone and underlying surface type.
[0100] The calibration process is conducted separately. A sample set is constructed using historical data on station gusts and grid assimilation fields, and quantile regression coefficients are used to fit the data. to The regression loss incorporates a penalty or projection operation based on coefficient sign constraints, thereby obtaining a mapping more sensitive to strong wind quantiles while maintaining monotonicity. To reduce sampling bias, samples are stratified by terrain type, season, and stability; the stability transformation function... A single-question table function is used to ensure that it is monotonic and differentiable over the entire stability range.
[0101] The process first calculates the baseline amplitude and mapping factor of the gust, outputting the coarse-resolution gust. Simultaneously, it records the mapping factor, the quality label of the input features, and their weights. If frictional velocity or turbulent kinetic energy is missing, empirical substitution relationships for near-surface wind speed, stability, and wind shear are used for estimation, and the source is noted in the quality label. For high-roughness areas such as forests and high-rise buildings, surface roughness is derived from a high-resolution underlying surface database. If necessary, a building height correction term is added within urban functional areas, without altering the monotonicity constraints and mapping structure described above. All outputs maintain the same grid and timestamps as the preceding steps, facilitating direct use in subsequent advection and optimal transport downsampling.
[0102] Example: During a strong cold pool outreach process, the operational domain is 200×200 km, and the time step is 3 minutes. The system reads the sinking potential energy, cold pool depth, and illusory temperature difference, and combines this with the surface roughness grid, turbulent kinetic energy, and near-surface wind shear to obtain the coarse-resolution gust wind using the above calculation expression. In urban agglomeration areas, the surface roughness is relatively large, and the monotonicity constraint is... Automatically reduce the mapping factor; in the afternoon unstable stratification region, the stability parameter decreases. As the scale increases, the mapping factor also increases. The resulting coarse-resolution gust field exhibits a banded enhancement at the cold pool front, showing good consistency with the gust distribution at ground stations, and provides stable input for subsequent fine-scale generation and wind damage polygon extraction.
[0103] Preferably, when obtaining the coarse-resolution gust based on the sinking potential energy, cold pool depth, and dummy temperature difference through a hysteresis kernel, the hysteresis kernel is composed of a time convolution kernel formed by adding two exponential kernels, which is used to align the time series of the coarse-resolution gust with the time scale of the cold pool evolution.
[0104] Within a unified grid and terrain-following vertical coordinate system, the baseline amplitude of the gust and related diagnostic parameters are first obtained from the preceding steps. Then, the coarse-resolution 10-meter gust is corrected using a time lag method to align the gust time series with the time scale of cold pool development and attenuation. The lag employs causal convolution, which avoids "future information leakage" and reflects the physical process of "rapid rise and slow fall".
[0105] The core calculation expression is as follows:
[0106]
[0107]
[0108] The coarse resolution is 10-meter gusts; The baseline amplitude of the gust is obtained by combining the subsidence potential energy, the depth of the cold pool, and the virtual temperature difference. It is a delayed kernel; and As weight; and Time scale; For convolution integral variables; For time; This is the convolution truncation length.
[0109] The implementation steps are as follows: First, construct the time series of gust baseline amplitude for each grid point over a past period, keeping the time step consistent with the previous progression. Second, set the hysteresis kernel parameters based on physical diagnostics: map the growth rate of the cold pool depth and the cold pool front propagation speed to the rise time scale to obtain... By mapping the rate of increase in virtual temperature difference to the rate of decrease in precipitation to a decay time scale, we obtain... The weights are used to control the proportion of "fast branches" and "slow branches," and can be set according to the precipitation intensity and near-surface divergence intensity, so that the stronger the precipitation, the greater the weight of the slow branch. All parameters are subject to engineering limits, and their source and version number are recorded in the metadata. Subsequently, the hysteresis kernel is normalized so that the discrete summation of the kernel within the cutoff length is 1, ensuring the conservation of amplitude energy.
[0110] The numerical implementation employs a causal finite impulse response (FIR) approach: discrete convolutions are performed with time steps as the sampling interval, and the upper bound is empirically chosen to be several times the size of the time scale. To improve efficiency, only the sum of the most recent convolutional segment is incrementally updated in each step; a "warm start" is used at the boundaries, bringing in the internal state of the previous step to avoid the cold start bias of the first convolution. For input sequences with extreme transitions, amplitude limiting and first-order temporal smoothing are performed first to prevent abnormal spikes from forming non-physical tails after convolution.
[0111] The output includes a coarse-resolution 10-meter gust time series, a hysteresis kernel parameter grid for tracing, and a quality flag. The quality flag is downgraded when the cold pool diagnosis is unstable or the observation coverage is insufficient, facilitating a conservative strategy in the subsequent fine-scale generation stage. To coordinate with subsequent advection and optimal transport downsampling, the output grid, timestamps, and units are consistent with those of the previous steps.
[0112] Example: During a strong convection event, the above process was applied to several grids downwind of an urban cluster. The cold pool depth rapidly increased, and the cold pool front propagation speed was high, resulting in a smaller rise timescale and a larger decay timescale provided by the system. High precipitation intensity increased the weight of the slow branch. After convolution, the coarse-resolution 10-meter gust rapidly increased as the front passed and slowly subsided after precipitation and cold pool weakening, matching the temporal pattern of the gust records from ground-based automatic weather stations. This hysteresis processing eliminated short-period oscillations in the baseline amplitude, making the gust peak value and duration more consistent with the cold pool evolution, providing stable input for subsequent advection-constrained downsampling and probability generation.
[0113] Preferably, when using advection-constrained optimal transport downsampling and taking the cold pool front as an insurmountable boundary, the near-surface average wind field is used as the advection constraint, and the boundary conditions of zero normal flux and tangential energy redistribution are applied at the location of the cold pool front to obtain the fine-scale field of gusts.
[0114] Within a unified grid and terrain-following vertical coordinate system, the inputs are coarse-resolution 10-meter gust winds, near-surface average wind fields, cold pool front polygons, and their normal and tangential information. To obtain physically constrained fine-scale gust fields on a fine grid, the process sequentially performs advection extrapolation, optimal transport redistribution, and frontal boundary constraint superposition correction, while maintaining energy conservation and geometric consistency without crossing fronts.
[0115] First, the coarse-resolution 10-meter gust wind is converted into an energy density grid, i.e., the energy density is obtained by squaring the gust amplitude for each grid. Then, at the center of each fine grid, the energy density field after advection is obtained by backtracking along the reverse direction of the near-surface mean wind to the position of the previous time step and using bilinear interpolation. This step is implemented using a semi-Lagrange method to avoid time step limitations and reduce numerical diffusion. The core expression is as follows:
[0116]
[0117] The energy density after advection; The energy density before advection; It is a spatial position vector; This represents the near-surface average wind vector. For time step.
[0118] To redistribute the energy density after advection on a fine-mesh grid and restore details, a discrete optimal transport problem is constructed. Using coarse-mesh elements as the source and fine-mesh elements as the sink, the goal is to achieve energy conservation through redistribution with minimal spatial displacement while adhering to the impassable condition of the frontal boundary. The core expression is as follows:
[0119] in , For source unit sink unit The amount of data transmitted; The cost of spatial displacement is generally taken as the square of the distance between the source and sink centers; The energy density of the source unit; The target energy density or capacity of the sink cell (set according to grid area and total conservation). To determine the crossing indicator, if the line connecting the source and sink crosses the front in the normal direction, then... Forced In implementation, crossing is determined by intersecting the source-sink line with the frontal polygonal line and comparing the normal signs; for transmissions located within the frontal tangential zone, energy redistribution in the tangential direction is allowed.
[0120] An entropy-regularized approximation algorithm is used for the solution, with a fixed number of iterations to meet timeliness requirements. To avoid non-physical ringing or excessive clustering on the fine mesh, a weak smoothing term is introduced into the transport quantity during iteration, and an upper bound is applied to the single-step displacement to prevent long-distance jumps. After redistribution, the energy density field of the fine mesh is obtained, and the square root is taken to recover the gust amplitude, forming a fine-scale gust field; simultaneously, the total transport quantity error and energy conservation deviation are output as quality indicators.
[0121] Frontal boundary constraints are handled in two levels. The first level is a hard constraint, which eliminates normal-crossing transmission channels in the optimal transmission problem based on the crossing judgment. The second level is a soft constraint, which sets a narrow band on each side of the front, increases the cost weight for tangential transmission from fine mesh to fine mesh, so that energy is preferentially redistributed along the front direction rather than crossing. For regions with uncertain frontal geometry, the narrow band width and cost weight are weighted using the front occurrence probability. The higher the probability, the stricter the normal prohibition and the higher the priority of tangential channels.
[0122] In engineering applications: all raster products use the same coordinates and timestamps as previous steps to avoid redundant interpolation; the construction of cross-jurisdictional and tangential zones is completed on the vector layer and then projected onto the grid; when near-surface average wind is missing, it is filled in with the median of the neighborhood and the weight of that area is reduced; energy conservation is corrected at each step, and if the conservation error exceeds the threshold, the fine grid is normalized proportionally. In addition to the fine-scale gust field, the output also includes transmission cost statistics, conservation correction coefficients, and boundary constraint parameter versions for easy traceability and reproduction.
[0123] Example: During a rapid outward expansion of a cold pool, the coarse grid resolution was 2 km, the fine grid resolution was 500 m, and the time step was 3 minutes. The system first backtracked to obtain the energy density after advection based on the near-surface average wind field, and then redistributed it to the fine grid using optimal transmission. At the front, the normal channel was blocked, and the energy slid tangentially along the front and accumulated on the windward side. After recovering to gusts, the fine-scale field formed a banded high value at the frontal front that did not cross the front, which coincided with the location of the gust enhancement zone of the ground automatic weather station, with a conservation error of less than 1%, meeting the input requirements of the subsequent probability generation module.
[0124] Preferably, when combining the conditional diffusion probability generation model to output the pixel gust distribution and the over-threshold probability, the fine-scale field of gusts, surface roughness, terrain data, building height, low-level wind shear, turbulent kinetic energy, cold pool front occurrence probability, low-level divergence and ensemble confidence weight are used as the condition set, and energy consistency regularization, topological persistence screening and extreme value distribution embedding are applied to the pixel gust distribution.
[0125] Within a unified grid and terrain-following vertical coordinate system, inputs include fine-scale gust fields, surface roughness, terrain data, building height, low-level wind shear, turbulent kinetic energy, cold pool front occurrence probability, low-level divergence, and ensemble confidence weights. These channels are pixel-aligned and scaled; dynamic channels use timestamps consistent with time progression, while static channels use a unified regional projection. These channels are then grouped into a conditional set and fed into a conditional diffusion probability generation model to generate a gust sample set across the entire domain using a tile-based approach.
[0126] The conditional diffusion probabilistic generation model employs a back-and-forth process of noise injection and progressive denoising. The backbone network is an encoder-decoder structure, and the condition set is injected into the network through channel splicing and self-attention injection. During sampling, a random seed and step number are recorded. Tiles are overlapped and weighted fusion is performed at the edges to ensure spatial continuity. The set confidence weight is superimposed as a pixel weight on the noise schedule and conditional scaling, making the output more conservative in areas with weak observation support.
[0127] To ensure consistency between the samples and physical forcing, a combination of energy consistency regularization and projection is used: after each tile sampling, the average energy of the sample set is compared with the reference energy of the fine-scale field of the gust, and a proportional adjustment is made within the allowable range before proceeding to the next batch. The core expression is as follows:
[0128]
[0129] For energy uniformity regularization, For the sample size, For the first A 10-meter gust of wind in one sample, The reference energy density is the squared amplitude of the fine-scale field of the gust. This regularization can be used for offline training or as a basis for amplitude fine-tuning after online sampling.
[0130] To suppress isolated noise and discontinuous bands, topological persistence screening is performed on the pixel extrema and connectivity patterns of the sample set: First, local peaks and saddle points are identified on the bitmap of the samples, and the peak-saddle difference is calculated as persistence. Patches below the threshold are deleted or merged with neighboring high-persistence patches. The filtered sample set is then embedded with extremum distribution: quantile sequences are extracted from the samples by pixel, and high quantiles are quantized and mapped using regionalized extremum tail parameters, so that the probability tail of strong wind areas is consistent with historical statistics, while maintaining the energy consistency regularity.
[0131] The probability of exceeding the threshold is calculated from the sample frequency:
[0132]
[0133] This represents the probability of a pixel exceeding a threshold. The number of samples; For the first One sample of gusts; The threshold is used. The source and version of the threshold are recorded in the metadata. The probability map is used to extract wind damage polygons from the connected components constrained by size and shape. At the same time, the median gust, quantile gust, probability of reaching the threshold, and peak occurrence time are summarized for each polygon.
[0134] In engineering applications: missing measurements in all channels are masked and weighted less during loss and sampling; static channels in urban areas are cropped with additional building height to reflect abrupt changes in underlying surface roughness; model inference uses fixed steps and temperature parameters to ensure timeliness; each run outputs raster and vector products with timestamps, units, parameter versions, and quality marks to meet traceability requirements.
[0135] Example: During a cold pool advance in an urban cluster, a condition set was created by combining fine-scale gust fields with factors such as building height, surface roughness, and low-level wind shear. This generated 200 gust samples using a 256×256 tile configuration with 64-pixel overlap. Amplitude fine-tuning with energy consistency regularization maintained consistency between the sample average energy and the reference energy. Topological persistence filtering removed scattered small patches, and extreme value distribution embedding enhanced the amplitude characterization of high quantiles. The resulting over-threshold probability bands are distributed along the cold pool front, coinciding with the location of gust enhancement areas from ground-based automatic weather stations. Furthermore, they exhibit a finer spatial structure in densely built-up areas, providing refined input for alarm and emergency response plans.
[0136] Aligning radar lower-level divergence, satellite brightness temperature cooling rate, lightning density variation, and ground gusts, applying monotonically normalized flow for probabilistic calibration, and using crossover ratio and conformal coverage for event consistency verification, the boundary layer similarity mapping parameters, hysteresis kernel, and generation variance are updated under amplitude-limited conditions.
[0137] Within a unified grid and terrain-following vertical coordinate system, radar lower-level divergence, satellite brightness temperature cooling rate, lightning density changes, and ground gust observations are first aligned to a fine-scale grid according to time windows and spatial indices, forming pixel-level labels and event polygon ground truth values. Pixel-level labels are determined by whether ground gusts exceed a threshold, and event polygon ground truth values are obtained by intersecting the observed connected regions exceeding the threshold with the radar lower-level divergence enhancement area. To avoid error propagation, lower confidence weights are applied to obstructed and sparse areas, and their source and timestamp are recorded in the metadata.
[0138] Probabilistic calibration employs a monotonically normalized flow to map the over-threshold probability of uncalibrated pixels to calibration probabilities. The mapping is a piecewise affine monotonic function, with endpoints maintained at zero and one, and node positions and slopes calibrated based on observations within a sliding time window. The core calculation expression is as follows:
[0139]
[0140] The probability of an uncalibrated pixel exceeding the threshold. To calibrate the probability, For the first Each node position The slope is non-negative. For the number of segments, To obtain the positive operator, This represents all nodes and slope parameters to be calibrated. A reliable monotonic mapping is achieved by minimizing the deviation between the Brill score and the reliability curve, and by using non-negativity of the slope between nodes and endpoint constraints as regularization.
[0141] Event consistency verification employs two metrics: Cross-Union Ratio (CIRR) and Conformal Coverage. CIRR is used to obtain predicted event polygons by thresholding the calibration probability. These polygons are then matched one-to-one with the ground truth polygons to calculate the overlap ratio and identify the locations of any missing components. Conformal Coverage uses pixel-level inconsistency scores to construct a threshold set that guarantees coverage, forming a globally uniform confidence threshold to ensure that the coverage rate of positive examples reaches the target at a given significance level.
[0142]
[0143]
[0144] These are the quantile values for non-consistent scores. At the significance level, For pixel tags, The set of accepted predictions. Through... The event set of connected component extraction and minimum polygon envelope generation is obtained, and the index of the under-covered region is written back as the positioning basis for subsequent parameter fine-tuning.
[0145] To reduce systematic bias, key module parameters are updated in small steps under amplitude constraints. The updated parameters include the coefficients of the boundary layer similarity mapping, the weights and time scale of the hysteresis kernel, and the variance generated by diffusion. The core calculation expression is as follows:
[0146]
[0147] For the parameter vector to be updated, The parameter adjustment is obtained by combining calibration loss and event consistency loss. This is the upper bound of the single-step amplitude. To ensure stability, first adjust the normalized scale within the module. Perform channel-level scaling, then amplitude clipping, and record the version number and affected area after the update.
[0148] In terms of the engineering workflow, probability calibration and event consistency verification are performed using a rolling time window. Each time step outputs a calibration probability raster, pixel quantile intervals, event polygon sets, and their attribute fields, including peak gust, median gust, threshold probability, peak time, and quality indicators. If the cross-union ratio is consistently low and conformal coverage is insufficient, only the parameters of the affected sub-regions are updated with limited amplitude to avoid global drift. All outputs maintain the same coordinates and timestamps as the preceding modules for easy auditing and playback.
[0149] Example: During a strong convection outreach process, the radar's lower-level divergence enhancement zone, the satellite brightness temperature continuously decreasing zone, the lightning density surge zone, and the ground gust zone were aligned to a fine-scale grid. After applying monotonically normalized flow to the uncalibrated probabilities, the reliability curve approached the diagonal, and the Brill score decreased. A conformal coverage threshold was calculated at a significance level of 0.1. The intersection-union ratio (IU) of the predicted set and the ground truth polygon significantly improved near the frontal principal axis, indicating that insufficient coverage was mainly concentrated on the radar obstruction side. Based on this, the system made amplitude-limited fine-tuning of the roughness coefficient and the slow branch weight of the hysteresis kernel in the boundary layer similarity mapping, increasing the diffusion generation variance in the obstructed area. The final output event polygon was more consistent with the observed area, and the pixel-level quantile interval coverage reached the target level.
[0150] Preferably, after aligning the radar lower-level divergence, satellite brightness temperature cooling rate, lightning density changes, and ground gusts, a monotonically normalized flow is applied to perform probability calibration on the over-threshold probability. The monotonically normalized flow adopts a piecewise affine transformation while maintaining monotonicity and endpoint invariance, and the consistency between the Brill score and the reliability curve within the sliding time window is used as the optimization objective.
[0151] Within a unified grid and terrain-following vertical coordinate system, radar lower-level divergence, satellite brightness temperature cooling rate, lightning density variation, and ground gust observations are aligned using the same time window and spatial index. Radar lower-level divergence uses the horizontal divergence field of the lowest available volume scan layer; the satellite brightness temperature cooling rate is the backward difference of the brightness temperature time series; lightning density variation is the increment of lightning strike counts per unit time; and ground gust observations are from automatic weather stations. After alignment, pixel-level samples are formed: each pixel contains an uncalibrated over-threshold probability, evidence channel, and a label indicating whether it exceeds the threshold. Missing measurements are masked and downweighted in subsequent optimization.
[0152] Probabilistic calibration employs a monotonically normalized flow to map uncalibrated probabilities to calibrated probabilities. The mapping is a piecewise affine curve with overall monotonicity, and endpoints are fixed at zero and one, facilitating reuse across different regions and seasons. This is achieved by imposing a non-negativity constraint on the slope and fixing it. This ensures that the mapping is monotonic and the endpoints remain unchanged.
[0153] The optimization objective combines the Brill score within a sliding time window with the consistency of the reliability curve. The Brill score measures the consistency between probability and observation, while the consistency of the reliability curve requires that the average observation frequency of each probability bin be close to the bin center. The core expression of the Brill score is as follows:
[0154]
[0155] Rate Brill The number of samples within the window. For the first The calibration probability of each pixel. For the first The binary labels for each pixel are used. In the actual implementation, the binning bias between the Brill score and the reliability curve is minimized, and a slope smoothing term is added to prevent overfitting and broken line oscillations in sparse sample regions.
[0156] The training and application process is as follows. First, a sliding time window (e.g., from the last few tens of minutes to several hours) is constructed to collect the assimilated uncalibrated probabilities and pixel labels, and the evidence strength and coverage of each pixel are statistically analyzed. Second, node positions are initialized as uniform quantiles with a constant slope; smaller sample weights are assigned to regions with missing data and weak evidence. Subsequently, constrained optimization is used to solve the parameter set. After each iteration, the slope is projected to be non-negative, and recalibrated to keep the endpoints unchanged. After optimization convergence, the parameters are frozen, and the global uncalibrated probability at the current time is mapped to obtain the calibrated probability raster. To ensure timeliness, the output is performed immediately after the window ends, and the optimal parameters of the previous window are used as the initial values for the new window.
[0157] The integration with the evidence channel follows the principle of "calibration before fusion, weights used for constraints": the evidence channel only affects the stabilization of sample weights and reliability bins, and does not directly enter the mapping function. When radar obstruction, satellite shadows, or lightning sensors are offline, samples in that area are automatically downweighted, and the mapping curve approaches the historical steady state, avoiding overcalibration caused by temporary lack of evidence. For significantly different underlying surfaces such as urban and mountainous areas, a partitioned curve is used: node locations are shared, while slope parameters are independent for each partition, maintaining overall consistency while reflecting local differences.
[0158] In engineering applications: probabilities and labels are strictly aligned at the pixel level; binning uses equal division to balance the sample size of each bin; when a bin has insufficient samples, adjacent bins are merged; in addition to calibrating the probability raster, the output includes the number of samples per bin, observation frequency, binning bias, and the node coordinates and slope of the final curve for easy auditing. All results are written to the same grid and timestamp as the upstream data, along with parameter version and training window information.
[0159] Example: During a strong convection event, a sliding window of approximately 2 hours was selected to aggregate the overall uncalibrated probability and pixel labels. Ten nodes were set at equal quantiles, with an initial slope of constant. Constrained iterative optimization was used to minimize the Brillouin score and reliability deviation. After calibration, the reliability curve converged from an upwardly convex shape deviating from the diagonal to near the diagonal, the Brillouin score decreased, and the false alarm rate in the low-probability segment decreased. In the radar-blocked foreland zone, due to the smaller sample weights, the curve remained flat, avoiding excessive bending. The final output calibration probability served as input for subsequent event consistency checks and parameter amplitude-constrained updates, improving the interpretability and stability of alarm threshold selection.
[0160] Preferably, when performing event consistency verification using cross-union ratio and conformal coverage, a synthetic loss is formed based on cross-union ratio, peak occurrence time difference, and cold pool front geometric deviation, and the boundary layer similarity mapping parameters, hysteresis kernel, and generation variance are updated under amplitude-limited conditions.
[0161] Within a unified grid, radar lower-level divergence, satellite brightness temperature cooling rate, lightning density changes, and ground gust observations are first aligned using the same time window and spatial index to obtain pixel labels and event polygon ground truth values. Probabilistic products utilize the calibration probabilities and wind damage polygons output from previous steps. For each time step, predicted event polygons are generated based on the calibration probabilities and matched one-to-one with the ground truth set; simultaneously, peak occurrence times are extracted from the ground gust time series, and observed cold pool front polygons are constructed using radar lower-level divergence and cold pool diagnostics. Event consistency verification focuses on the intersection-union ratio (IUGR) and conformal coverage: IUGR measures spatial overlap, and conformal coverage uses the previously defined inconsistency threshold to calculate the positive example coverage rate, constraining overall reliability.
[0162] To unify the measurement of spatial, temporal, and geometric biases, a composite loss is constructed with minimization as the objective. The core calculation expression is as follows:
[0163]
[0164]
[0165] It is the combined loss of the basic space, time, and geometric biases; To predict the intersection-union ratio of the event polygon and the ground truth polygon; The weighted median for predicting the peak gust time within an event; This is the weighted median of the peak gust occurrence time corresponding to the true value region; To predict the average normal deviation between the cold pool front breakline and the observed front breakline (by calculating the average distance from the point to the line along the observed normal); The total loss is synthesized for event consistency; This is the conformal coverage rate (the proportion of positive examples covered by the preceding conformal threshold). For target coverage level These are non-negative weighting coefficients. Spatial matching uses nearest neighbor pairing and retains pairs with the maximum intersection-union ratio when one-to-many relationships exist; temporal statistics downgrade abnormal stations with weights; geometric deviation is projected along the normal of the observed front to avoid tangential drift affecting the metric.
[0166] Parameter updates are performed under "range-limited" conditions, with only minor adjustments made to three types of modules: coefficients of the boundary layer similarity mapping, weights and time scales of the hysteresis kernel, and variance of the probability generation stage. The core update rules are as follows:
[0167]
[0168] The parameter vector to be updated (including similarity mapping coefficients, hysteresis kernel weights and time scales, and generation variance); This is the step size coefficient; This is the upper bound of the single-step amplitude; The gradient of the synthesis loss with respect to the parameters; This is a component-wise pruning operator. To ensure physical consistency, after the update, a monotonicity check is applied to the surface roughness coefficient and stability coefficient in the similarity mapping, lower and upper bounds are applied to the time scale of the hysteresis kernel, and non-negativity and upper bound constraints are applied to the generation variance; if the check fails, it reverts to the previous version.
[0169] The engineering workflow is a rolling "evaluation-update-writeback" process. In the evaluation phase, the intersection-union ratio, peak time difference, and geometric bias are calculated to obtain the synthesis loss and coverage difference. In the update phase, gradients are calculated and amplitude clipping is performed by region to avoid overfitting in sparsely observed or occluded areas. In the writeback phase, parameter writeback is only enabled for sub-regions with persistently high losses; other regions maintain the previous version. All outputs include validation statistics, parameter version numbers, affected region masks, and quality flags to ensure traceability.
[0170] Example: In a case study of rapid outward expansion in a cold pool, the sliding time window was set to 120 minutes. Initial assessment showed that the predicted polygon on the windward side expanded outward, with an intersection-over-union ratio (IoU) of 0.42, a peak time difference of 6 minutes, and a systematic upwind shift in geometric bias. The system calculated the coverage difference and provided a regularization term, indicating a high synthetic loss. A first amplitude-limited update was performed: the surface roughness coefficient in the similarity mapping decreased slightly, the stability correlation coefficient increased slightly, the weight of the slow branch of the hysteresis kernel was increased, the time scale was slightly relaxed, and the generation variance was moderately increased in the radar obstruction zone. The next window assessment showed that the IoU increased to 0.56, the peak time difference decreased to 3 minutes, the geometric bias was halved, the coverage reached the target level, and the parameter changes were all within the upper limit of the amplitude, meeting the stability requirements.
[0171] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects.
[0172] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.
Claims
1. A method for predicting extreme wind weather based on a large-scale meteorological forecasting model, characterized in that, include: The system acquires three-dimensional physical fields for numerical weather prediction, weather radar, meteorological satellites and ground meteorological observations, unifies the grid and assimilates the data through observation operators to form a consistent three-dimensional physical field and atmospheric stability and low-level wind shear. Time progression is performed and vertical acceleration is calculated based on a micro-downburst calculation program. The sinking potential energy is obtained by integration, the depth of the cold pool and the dummy temperature difference are diagnosed, and the location and probability of occurrence of the cold pool front are constructed. Based on the sinking potential energy, cold pool depth and illusory temperature difference, coarse resolution gusts are obtained through boundary layer similarity mapping and hysteresis kernel. Optimal transmission downsampling with advection constraints is adopted and the cold pool front is used as an insurmountable boundary. Combined with the conditional diffusion probability generation model, the pixel gust distribution and over-threshold probability are output and the wind damage polygon is extracted. Aligning radar lower-level divergence, satellite brightness temperature cooling rate, lightning density variation, and ground gusts, applying monotonically normalized flow for probabilistic calibration, and using crossover ratio and conformal coverage for event consistency verification, the boundary layer similarity mapping parameters, hysteresis kernel, and generation variance are updated under amplitude-limited conditions.
2. The method according to claim 1, characterized in that, When acquiring and assimilating data from numerical weather prediction's three-dimensional physical field, weather radar, meteorological satellites, and ground meteorological observations, the observation operators generate simulated quantities of radar reflectivity and radial velocity, satellite brightness temperature and cloud top height, and ground wind speed and ground pressure, respectively. In a unified grid and terrain-following vertical coordinate system, variational assimilation weighted by observation errors generates a consistent three-dimensional physical field, atmospheric stability, and low-level wind shear.
3. The method according to claim 1, characterized in that, When advancing the time, a neural operator is used to update the three-dimensional physical field, and the update results are corrected by using the residuals of the continuity equation and the total water conservation residuals as regularization terms.
4. The method according to claim 1, characterized in that, When calculating vertical acceleration based on the micro-downburst calculation program, the vertical acceleration is obtained by adding the evaporative cooling buoyancy and the precipitation load. The sinking potential energy is obtained by integrating the negative value of the vertical acceleration along the height. The depth of the cold pool and the virtual temperature difference are determined based on the vertical distribution of the virtual temperature difference. The horizontal set function of the location and probability of occurrence of the cold pool front is constructed by the virtual temperature difference gradient and the low-level divergence.
5. The method according to claim 1, characterized in that, When obtaining coarse-resolution gusts based on sinking potential energy, cold pool depth, and virtual temperature difference through boundary layer similarity mapping, the boundary layer similarity mapping takes frictional velocity, stability parameters, surface roughness, turbulent kinetic energy, and near-surface wind shear as inputs, and applies monotonicity constraints to surface roughness and stability parameters.
6. The method according to claim 1, characterized in that, When obtaining coarse-resolution gusts based on sinking potential energy, cold pool depth, and illusory temperature difference through a hysteresis kernel, the hysteresis kernel is composed of a time convolution kernel formed by adding two exponential kernels, which is used to align the time series of coarse-resolution gusts with the time scale of cold pool evolution.
7. The method according to claim 1, characterized in that, When using optimal transport downsampling with advection constraints and taking the cold pool front as an insurmountable boundary, the near-surface average wind field is used as the advection constraint. Boundary conditions of zero normal flux and tangential energy redistribution are applied at the location of the cold pool front to obtain the fine-scale field of gusts.
8. The method according to claim 1, characterized in that, When combining the conditional diffusion probability generation model to output the cell gust distribution and over-threshold probability, the fine-scale field of gusts, surface roughness, terrain data, building height, low-level wind shear, turbulent kinetic energy, cold pool front occurrence probability, low-level divergence and ensemble confidence weight are used as the condition set, and energy consistency regularization, topological persistence filtering and extreme value distribution embedding are applied to the cell gust distribution.
9. The method according to claim 1, characterized in that, After aligning the radar lower-level divergence, satellite brightness temperature cooling rate, lightning density variation, and ground gusts, a monotonically normalized flow is applied to perform probability calibration on the over-threshold probability. The monotonically normalized flow adopts a piecewise affine transformation while maintaining monotonicity and endpoint invariance, and the consistency between the Brill score and the reliability curve within the sliding time window is used as the optimization objective.
10. The method according to claim 1, characterized in that, When performing event consistency verification using cross-union ratio and conformal coverage, a synthetic loss is formed based on cross-union ratio, peak occurrence time difference, and cold pool front geometric deviation. Under amplitude-limited conditions, the boundary layer similarity mapping parameters, hysteresis kernel, and generation variance are updated.
Citation Information
Patent Citations
Method and device for estimating thunderstorm gale extreme wind speed based on dual-polarization weather radar
CN118731948A
Low altitude wind shear detection with airport surveillance radars
US5093662A
