An airport runway wind condition forecasting method and short-term rolling forecasting device

Abstract

The application provides an airport runway wind condition prediction method and short-term rolling prediction device, comprising: extracting the review U and V component sequences of a target height layer from the review U and V component sequences respectively, and arranging the review U and V component sequences of the target height layer in time step order to form an input vector; inputting the input vector into the trained U component time sequence direct-pushing model and V component time sequence direct-pushing model to obtain the predicted U and V component sequences of the target height layer respectively; inverting the predicted U and V component sequences of the target height layer into a predicted wind speed sequence and a predicted wind direction sequence, and performing phase-neutral rolling smoothing processing on the predicted wind speed sequence to obtain a smoothed wind speed sequence. The application provides an airport runway wind condition prediction method and short-term rolling prediction device to solve the problems of insufficient runway-level refined wind condition prediction capability, low observation and prediction fusion degree and information dispersion in multiple independent systems in the prior art.

Description

Technical Field

[0001] This invention relates to the field of wind condition forecasting technology, and in particular to a method for forecasting wind conditions on airport runways and a short-term rolling forecasting device. Background Technology

[0002] Currently, runway wind condition monitoring and forecasting in airport operations mainly rely on the following methods: Firstly, traditional meteorological observation equipment, such as wind beacons and other conventional meteorological instruments, are usually deployed in the middle of the runway at airports to measure near-surface wind speed and direction in real time. These devices can only provide instantaneous wind observation values ​​at a single altitude level and cannot obtain wind profile information in the vertical direction. Secondly, numerical weather prediction models are currently used as a reference for airport operations. However, existing meteorological numerical prediction models or observation assimilation systems involve large computational loads and long update cycles. Although the spatial resolution has been refined to the kilometer or hundred-meter level, there is still a significant mismatch between the model output results and the actual runway operational requirements at the meter and minute levels in terms of space and time. In addition, model calculations rely on initial field and boundary conditions, and the forecast bias is large under specific runway environments or rapidly changing microscale weather conditions. Third, multiple systems operate in parallel. Currently, airport meteorological support usually relies on multiple independent observation and forecasting systems, including ground automatic weather stations, wind profiler radar, and airport terminal area weather forecasting systems. The data formats, update frequencies, and spatial coverage of each system are inconsistent. Operators need to switch between multiple information sources and make comprehensive judgments, which makes the operation process cumbersome and not conducive to rapid decision-making.

[0003] Meanwhile, although Doppler lidar has been widely used in areas such as airport wind shear detection, current applications mainly focus on wind shear warnings and real-time observation and display of wind profiles, without fully utilizing the high spatiotemporal resolution of lidar data for quantitative runway-level wind condition forecasting. The temporal evolution patterns inherent in the dense wind profile observation data collected by lidar have not yet been effectively explored and applied to the forecasting process.

[0004] In summary, the main shortcomings of the existing technology include: 1) The ability to provide detailed runway-level wind condition forecasts is insufficient. Existing forecasting methods are unable to directly provide quantitative forecasts of the parameters required for operation, such as wind speed, wind direction, headwind, crosswind, and gusts for specific runways. 2) The integration of observations and forecasts is low, and observational data has failed to effectively drive the continuous calibration and updating of forecast models; 3) Information is scattered across multiple independent systems, and there is a lack of an integrated solution that combines multi-layer wind profile observation information with runway operation parameter forecasts into a unified interface. As a result, operators have difficulty obtaining complete runway wind condition information on a single interface, leading to low decision-making efficiency. Summary of the Invention

[0005] This invention provides an airport runway wind condition forecasting method and a short-term rolling forecasting device to solve the problems of insufficient runway-level refined wind condition forecasting capability, low degree of integration of observation and forecasting, and information being scattered across multiple independent systems in the prior art.

[0006] On one hand, the present invention provides a method for forecasting airport runway wind conditions, including: The wind profile observation data at the current moment is preprocessed to obtain the U and V components at the current moment. The wind profile observation data includes horizontal wind speed, wind direction and vertical velocity at multiple height levels. The preprocessing includes missing value processing, physical boundary verification, spatiotemporal sorting and deduplication, missing value judgment and wind component decomposition. Based on the U and V components at the current time and the U and V components of the past N-1 consecutive time steps, a retrospective U and V component sequence is constructed, and the length of the retrospective U and V component sequence is N time steps. Extract the U and V component sequences of the target height layer from the U and V component sequences of the retrospective, and arrange the U and V component sequences of the target height layer in an interleaved manner according to the time step order as an input vector. The target height layer is a preset height layer or the height layer closest to the preset height layer. The input vector is input into the trained U-component temporal direct propagation model and V-component temporal direct propagation model to obtain the predicted U and V component sequences of the target height layer, respectively. The U-component temporal direct propagation model and V-component temporal direct propagation model are multi-output regression models containing M independent lightweight gradient boosting decision tree regressors, and the length of the predicted U and V component sequences of the target height layer is M time steps. The predicted U and V component sequences of the target height layer are inverted into predicted wind speed sequences and predicted wind direction sequences. The predicted wind speed sequences are then subjected to phase-neutral rolling smoothing to obtain smoothed wind speed sequences. Based on the azimuth of the in-use runway, the smoothed wind speed sequence and the forecast wind direction sequence are decomposed into forecast headwind, crosswind and gust component sequences at the target height level.

[0007] Optional, also includes: A multidimensional feature vector is constructed based on wind profile observation data from N consecutive time steps and the retrospective U and V component sequences. The wind profile observation data from the N consecutive time steps is constructed from the current wind profile observation data and the wind profile observation data from the past N-1 consecutive time steps. The multidimensional feature vector includes basic state features, time-coded features, multi-scale trend and acceleration features, vertical structure features, full profile statistical features, multi-scale rolling statistics and anomaly features, wind direction features, vertical velocity and turbulence features, and interaction and time-delay features. The basic state features include the U and V components and the composite wind speed at all observation heights at the current time. The time-coded features include intraday time features and annual cycle time features constructed based on sine-cosine coding. The multi-scale trend and acceleration features include the first-order variation trend and second-order acceleration of the U and V components. The vertical structure features... The system includes wind components of adjacent layers at the target height, vertical wind shear intensity, and wind shear trend. The full profile statistical features include the mean and standard deviation of the U and V components of the full profile, the mean, maximum, and standard deviation of the composite wind speed of the full profile, the full profile consistency index, and the deviation of the U and V components of the target height layer from the mean of the U and V components of the full profile. The multi-scale rolling statistics and anomaly features include the rolling mean, rolling standard deviation, and anomaly features of the U and V components and the composite wind speed. The wind direction features include the change in wind direction and the rolling standard deviation. The vertical velocity and turbulence features include the mean, standard deviation, maximum, and minimum of the vertical velocity of the full profile, the difference and time variation trend of the vertical velocity between adjacent height layers, and the approximate index of turbulent kinetic energy. The interaction and time lag features include the interaction features of wind speed and wind shear with hourly coding features, and the lag features of the U and V components at a preset historical time. The multidimensional feature vector is input into the trained U-component multidimensional feature model and V-component multidimensional feature model to obtain the enhanced forecast U and V components for a specific time period. The U-component multidimensional feature model and the V-component multidimensional feature model are single-output regression models containing an independent lightweight gradient boosting decision tree regressor. The enhanced forecast U and V components for a specific time period are inverted into enhanced forecast wind speed and direction for that specific time period.

[0008] On the other hand, the present invention provides a short-term rolling forecasting device for airport runway wind conditions, comprising a data acquisition module, a forecast scheduling module, a preprocessing module, a data integration module, a data extraction module, a time-series direct propagation module, a first inversion module, a runway decomposition module, a feature construction module, an enhanced forecast module, a second inversion module, a data storage module, and a comprehensive display module, wherein: The data acquisition module continuously receives wind profile observation data output by the laser wind radar and pushes it to the forecast scheduling module, the data storage module and the integrated display module. The wind profile observation data includes horizontal wind speed, wind direction and vertical velocity at multiple height levels. The forecasting and scheduling module triggers the preprocessing module when it receives the wind profile observation data at the current moment. The preprocessing module preprocesses the wind profile observation data at the current moment to obtain the U and V components at the current moment and pushes them to the data storage module. The preprocessing includes missing measurement processing, physical boundary verification, spatiotemporal sorting and deduplication, missing data judgment and wind component decomposition. The data integration module constructs a retrospective U and V component sequence based on the current U and V components and the U and V components of the past N-1 consecutive time steps, respectively. The length of the retrospective U and V component sequence is N time steps. The U and V components of the past N-1 consecutive time steps are obtained from the data storage module. The data extraction module extracts the U and V component sequences of the target height layer from the U and V component sequences of the retrospective, and arranges the U and V component sequences of the target height layer in an interleaved manner according to the time step order as an input vector. The target height layer is a preset height layer or the height layer closest to the preset height layer. The time-series inversion module inputs the input vector into the trained U-component time-series inversion model and V-component time-series inversion model to obtain the predicted U and V component sequences of the target height layer, respectively. The U-component time-series inversion model and the V-component time-series inversion model are multi-output regression models containing M independent lightweight gradient boosting decision tree regressors, and the length of the predicted U and V component sequences of the target height layer is M time steps. The first inversion module inverts the predicted U and V component sequences of the target height layer into predicted wind speed sequences and predicted wind direction sequences, performs phase-neutral rolling smoothing on the predicted wind speed sequences to obtain smoothed wind speed sequences, and pushes the smoothed wind speed sequences and the predicted wind direction sequences to the data storage module and the integrated display module. The runway decomposition module decomposes the smoothed wind speed sequence and the forecast wind direction sequence into forecast headwind, crosswind and gust component sequences at the target height level based on the azimuth of the in-use runway and pushes them to the integrated display module. The feature construction module constructs a multi-dimensional feature vector based on wind profile observation data from N consecutive time steps and the retrospective U and V component sequences, and pushes it to the data storage module. The wind profile observation data from the N consecutive time steps is constructed based on the wind profile observation data at the current moment and the wind profile observation data from the past N-1 consecutive time steps. The wind profile observation data from the past N-1 consecutive time steps is obtained from the data storage module. The multi-dimensional feature vector includes basic state features, time-coded features, multi-scale trend and acceleration features, vertical structure features, full profile statistical features, multi-scale rolling statistics and anomaly features, wind direction features, vertical velocity and turbulence features, and interaction and time-delay features. The basic state features include the U and V components and composite wind speed at all observation heights at the current moment. The time-coded features include intraday time features and annual cycle time features constructed based on sine-cosine coding. The multi-scale trend and acceleration features include... The vertical structure features include the wind components of adjacent layers at the target height, vertical wind shear intensity, and wind shear trend. The full profile statistical features include the mean and standard deviation of the U and V components of the full profile, the mean, maximum, and standard deviation of the composite wind speed of the full profile, the full profile consistency index, and the deviation between the U and V components of the target height and the mean of the U and V components of the full profile. The multi-scale rolling statistics and anomaly features include the rolling mean, rolling standard deviation, and anomaly features of the U and V components and the composite wind speed. The wind direction features include the change in wind direction and the rolling standard deviation. The vertical velocity and turbulence features include the mean, standard deviation, maximum, and minimum of the vertical velocity of the full profile, the difference and time variation trend of the vertical velocity between adjacent height layers, and the approximate index of turbulent kinetic energy. The interaction and time lag features include the interaction features of wind speed and wind shear with hourly coding features, and the lag features of the U and V components at a preset historical time. The enhanced forecast module inputs the multidimensional feature vector into the trained U-component multidimensional feature model and V-component multidimensional feature model to obtain the enhanced forecast U and V components for a specific time period. The U-component multidimensional feature model and the V-component multidimensional feature model are single-output regression models containing an independent lightweight gradient boosting decision tree regressor. The second inversion module inverts the enhanced forecast U and V components for a specific time period into enhanced forecast wind speed and wind direction for that specific time period and pushes them to the integrated display module; The data storage module stores the received data; The integrated display module displays the horizontal wind speed and direction at the target altitude level, as well as the smoothed wind speed sequence and the predicted wind direction sequence, in the wind profile observation data for N consecutive time steps. It also overlays the horizontal wind speed and direction at the target altitude level in the subsequently received wind profile observation data onto the corresponding forecast curve using a time nearest neighbor matching method. The module displays the predicted headwind, crosswind, and gust component sequences at the target altitude level in the form of time series curves. It displays the horizontal wind speed and direction in the wind profile observation data at the current moment, as well as the enhanced forecast wind speed and direction at a specific time, in the form of a compass. Finally, it displays the wind profile observation data at the current moment and the vertical wind shear intensity in the form of a profile.

[0009] The beneficial effects of this invention are as follows: 1) Data-driven runway-level refined forecasting The machine learning model is trained by directly using continuous wind profile observation data collected by laser wind radar deployed in the airport runway area to achieve wind condition forecasting for the runway area. Compared with numerical weather prediction models, it does not require complex downscaling and post-processing, and can directly output the wind speed, wind direction, headwind, crosswind, gust and other information required for operation. 2) Dual-model collaborative forecasting architecture A dual-model architecture combining a time-series direct propagation model and a multi-dimensional feature model was constructed: the time-series direct propagation model predicts the wind field evolution at multiple future time steps based on historical time window data; the multi-dimensional feature model utilizes wind field structure features to enhance forecasts for specific future time periods; the two types of models extract wind condition evolution patterns from different information dimensions, achieving complementary advantages. 3) Minute-level rolling forecasts and updates Based on the continuously updated detection data from the laser wind radar, the forecasting process is re-executed after acquiring new observation data, so that the forecast results continuously reflect the latest observation status and improve the ability to respond to rapidly changing wind conditions. 4) Integrated display of observation and forecasting Real-time observation data and forecast results are displayed together on the same timeline, and subsequent observation data are dynamically superimposed to verify the forecast results, enabling operators to intuitively assess the reliability of the forecast. 5) Comprehensive utilization of multi-layered wind profile information By leveraging the ability of laser wind radar to simultaneously acquire wind information at multiple altitude levels, it can simultaneously display complete wind profiles and vertical wind shear intensity, providing richer vertical dimension references for operational decisions. 6) Adapt to runway configuration and operating conditions The system automatically calculates the headwind and crosswind components of each runway based on their actual azimuth angle, and can adapt to changes in runway operating status, dynamically adjusting the output and display content. 7) Integrated operation support capability By integrating real-time wind profile observation, vertical wind shear intensity, and runway-level wind condition forecasting functions into a single system, it can serve as a comprehensive runway wind condition monitoring and early warning tool for airport operations and management personnel, reducing reliance on multiple independent systems, simplifying operational procedures, and improving decision-making efficiency. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating an airport runway wind condition forecasting method provided in Embodiment 1 of the present invention. Figure 2 This is a schematic diagram of the structure of an airport runway wind condition short-term rolling forecast device provided in Embodiment 2 of the present invention. Detailed Implementation

[0011] The present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. The advantages and features of the present application will become clearer from the following description and claims. It should be noted that the drawings are all in a very simplified form and are not to scale, and are only used to facilitate and clarify the illustration of the embodiments of the present application.

[0012] It should be noted that, in order to clearly illustrate the content of this application, several embodiments are provided to further explain the different implementations of this application. These embodiments are enumerated rather than exhaustive. Furthermore, for the sake of brevity, content mentioned in the preceding embodiments is often omitted in the following embodiments. Therefore, content not mentioned in the following embodiments can be referred to in the preceding embodiments.

[0013] Example 1 Please refer to Figure 1 , Figure 1 The image shows an airport runway wind condition forecasting method provided in this embodiment, including: S1. Preprocess the wind profile observation data at the current moment to obtain the U and V components at the current moment. The wind profile observation data includes horizontal wind speed, wind direction and vertical velocity at multiple height levels. The preprocessing includes missing measurement processing, physical boundary verification, spatiotemporal sorting and deduplication, missing data judgment and wind component decomposition. In this embodiment, the wind profile observation data at the current moment is output by the laser wind measurement radar.

[0014] Laser wind radar obtains radial wind speed information at different distance gates by performing Doppler frequency shift analysis on the backscattered signals of aerosol particles in the atmosphere. After combined processing by scanning strategies, it generates wind profile observation data covering multiple height layers.

[0015] In this embodiment, the time resolution of the laser wind-measuring radar is one minute.

[0016] In this embodiment, missing value processing includes identifying missing values ​​and replacing them with missing markers.

[0017] In laser wind radar detection, specific values ​​are used to mark invalid or missing detection results. Replacing such missing values ​​with missing markers (NaN) can prevent invalid data from entering subsequent calculations.

[0018] In this embodiment, physical boundary verification includes: Determine whether the wind profile observation data is within a reasonable range; If not present, replace with a missing marker.

[0019] Apply a reasonable boundary check to each physical quantity in the wind profile observation data, such as whether the horizontal wind speed is within a reasonable meteorological range (e.g., 0 to 100 m / s), whether the vertical speed is within a reasonable range (e.g., -50 to 50 m / s), and whether the wind direction is within the range of 0° and 360°.

[0020] In this embodiment, spatiotemporal sorting and deduplication includes: The wind profile observation data were sorted by time and altitude, and duplicate data at the same time and altitude were removed.

[0021] In this embodiment, the missing information determination includes: Determine whether the proportion of missing markers in the wind profile observation data is higher than a preset threshold. If so, skip this forecast. If not, use forward or backward filling to complete the missing markers. Missing markers are generated during missing measurement processing or physical boundary verification.

[0022] In this embodiment, the wind component is decomposed into U (zonal, east-west) and V (longitudinal, north-south) components, which are converted from the commonly used meteorological polar coordinate representation of wind speed and direction to a rectangular coordinate system. The conversion formula is as follows: in, For horizontal wind speed, The sign represents the wind direction (defined as the direction from which the wind is coming), and the negative sign is used to convert the wind direction (coming direction) into the direction of the velocity vector (going direction).

[0023] Wind component decomposition not only eliminates the discontinuity of wind direction at 0° and 360°, but also the U and V components obtained by decomposition are numerically continuous and have good algebraic properties, making them more suitable as input and output vectors for regression models.

[0024] S2. Construct retrospective U and V component sequences based on the current U and V components and the U and V components of the past N-1 consecutive time steps, respectively. The length of the retrospective U and V component sequences is N time steps. In this embodiment, one time step is one minute.

[0025] S3. Extract the U and V component sequences of the target height layer from the U and V component sequences of the retrospective, and arrange the U and V component sequences of the target height layer in the time step order as an input vector. The target height layer is the preset height layer or the height layer closest to the preset height layer. In this embodiment, the input vector takes the following form: The staggered arrangement of time steps ensures that the U and V components of each time step are arranged adjacently in the input vector. The U-component temporal propagation model and the V-component temporal propagation model can perceive the coupling relationship of the two-dimensional wind vector at the same moment while learning the temporal pattern. Compared with the U and V component separate input, the gradient path is shorter and the convergence is faster during the training initialization stage.

[0026] S4. Input the input vector into the trained U-component time-series direct propagation model and V-component time-series direct propagation model to obtain the predicted U and V component sequences of the target height layer, respectively. The U-component time-series direct propagation model and V-component time-series direct propagation model are multi-output regression models containing M independent lightweight gradient boosting decision tree regressors. The length of the predicted U and V component sequences of the target height layer is M time steps. Both the U-component and V-component time-series probabilistic models train an independent gradient boosting decision tree regressor for each prediction time step. During the inference phase, only forward computation is performed, and the model parameters are not updated. The core parameters of the regressor include: maximum number of iterations, learning rate, number of leaf nodes, maximum depth, minimum number of leaf samples, feature subsampling ratio, sample subsampling ratio, L1 regularization coefficient, and L2 regularization coefficient, all of which are determined through automated tuning.

[0027] In this embodiment, the training process of the U-component temporal transductive model and the V-component temporal transductive model includes: S41. Calculate the time interval between adjacent timestamps in the historical wind profile observation data, mark the positions where the time interval exceeds one time step as data gaps, and divide the historical wind profile observation data into several time-continuous data blocks with the data gaps as the boundary. S42. Within at least one data block, a sliding window is used to acquire several data segments. The input window of the sliding window has N time steps, the output window has M time steps, and the step size is one time step. S43. Preprocess several data segments to obtain several pairs of U and V component sample sequences respectively; S44. Extract several pairs of U and V component sample sequences of the target height layer from several pairs of U and V component sample sequences respectively; S45. For each pair of target height layer U and V component sample sequences, the U and V components of the first N time steps are arranged alternately according to the time step order and used as the input vectors of the U component time series direct push model and the V component time series direct push model. The U and V components of the next M time steps are used as the output vectors of the U component time series direct push model and the V component time series direct push model, respectively. S46. Automatic hyperparameter optimization is performed based on the Bayesian optimization method, and time series cross-validation is used to train and validate the U-component time series inverse model and the V-component time series inverse model. The objective function of hyperparameter optimization is the weighted root mean square error, and the weight values ​​decay exponentially with the increase of the prediction time step.

[0028] In this embodiment, the Bayesian optimization method is TPE.

[0029] In this embodiment, the weighted root mean square error (WRMSE) is: in, The total number of sample sequences, For sample index, Forecast value, For observations, weight values for: in, For forecast time step index, As the decay rate parameter, this weight value gives higher weight to recent forecasts during model optimization.

[0030] This weighted strategy focuses hyperparameter optimization on improving near-term forecast accuracy, which aligns with the actual need for airport operational decisions to be more sensitive to the accuracy of near-term forecasts.

[0031] In this embodiment, to accommodate the computational complexity of multi-output regression, the basic regressors corresponding to all M forecast time steps are not trained and validated during the hyperparameter optimization process. Instead, a number of representative forecast time steps are selected from the M forecast time steps for surrogate evaluation, and the objective function value is calculated based on the representative forecast time steps, thereby reducing the computational cost of a single hyperparameter search.

[0032] Time-series cross-validation was used for training. All sample sequences were divided into several folds in chronological order. Each fold used the preceding sample sequences for training and the following sample sequences for validation, ensuring that validation samples always came after the training samples to avoid future information leakage. The results of cross-validation were used to evaluate the model's generalization performance and the stability of hyperparameter selection. Finally, the model was retrained using all sample sequences.

[0033] In this embodiment, M is 120, which means the U and V components are predicted minute by minute for the next 2 hours.

[0034] S5. Invert the forecast U and V component sequences of the target height layer into forecast wind speed and forecast wind direction sequences, and perform phase-neutral rolling smoothing on the forecast wind speed sequence to obtain a smoothed wind speed sequence. The forecast U and V components are inverted into the forecast wind speed and forecast wind direction using the following formulas: in, , These are the forecast U and V components, respectively. , These are the forecast wind speed and forecast wind direction, respectively. It is a four-quadrant arctangent function, ensuring that the wind direction is within the range of 0° to 360°.

[0035] In this embodiment, the phase-neutral rolling smoothing process includes: A centered sliding window is used to take the average value within a symmetrical time range before and after the current forecast time step. The length of the centered sliding window is a configurable parameter.

[0036] The forecast wind speed sequence output step by step may have high-frequency fluctuations. By using a centered sliding window to take the average of the symmetrical time range before and after the current moment, the high-frequency noise can be eliminated. Furthermore, since the centered sliding window is symmetrical about the current moment, the smoothing operation does not change the time phase of the signal, so it does not introduce any time displacement. Ultimately, this makes the forecast curve smoother and easier for operators to interpret.

[0037] S6. Based on the azimuth of the in-use runway, the smoothed wind speed sequence and the forecast wind direction sequence are decomposed into forecast headwind, crosswind and gust component sequences at the target height level.

[0038] The formulas for calculating the forecast headwind and crosswind are as follows: in, , These represent the components of the forecast headwind (positive value indicates headwind) and crosswind (positive value indicates wind coming from the right). The azimuth angle of the runway in use is calculated using trigonometric functions with the angle difference in radians.

[0039] Based on the vertical distribution characteristics of the wind profile and the momentum transfer effect, the predicted gust component is estimated using the following formula: in, To forecast the strength of gusts, This is an empirical enhancement factor used to quantify the effect of wind speed difference between upper and target altitude layers on the transport of upper-level momentum to the near-surface layer. To predict wind speed from the target height level using power-law wind profile relationships The formula for calculating the upper-level wind speed by extrapolating upwards is as follows: in, The target height level (runway height). For high-rise buildings, This is the vertical distribution index of wind speed.

[0040] This embodiment also includes: S7. Construct a multidimensional feature vector based on wind profile observation data from N consecutive time steps and retrospective U and V component sequences. The wind profile observation data from N consecutive time steps is constructed based on the wind profile observation data at the current moment and the wind profile observation data from the past N-1 consecutive time steps. The multidimensional feature vector includes basic state features, time-coded features, multi-scale trend and acceleration features, vertical structure features, full profile statistical features, multi-scale rolling statistics and anomaly features, wind direction features, vertical velocity and turbulence features, and interaction and time-delay features. The basic state features include the U and V components and the composite wind speed at all observation heights at the current moment. The time-coded features include intraday time features and annual cycle time features constructed based on sine-cosine coding. The multi-scale trend and acceleration features include the first-order variation trend and second-order acceleration of the U and V components. The vertical structure features include... The data includes wind components of adjacent layers at the target height, vertical wind shear intensity and wind shear trend; full profile statistical features include the mean and standard deviation of the U and V components of the full profile, the mean, maximum and standard deviation of the composite wind speed of the full profile, the consistency index of the full profile and the deviation of the U and V components of the target height layer from the mean of the U and V components of the full profile; multi-scale rolling statistics and anomaly features include the rolling mean, rolling standard deviation and anomaly features of the U and V components and the composite wind speed; wind direction features include the change in wind direction and the rolling standard deviation; vertical velocity and turbulence features include the mean, standard deviation, maximum and minimum of the vertical velocity of the full profile, the difference and time variation trend of vertical velocity between adjacent height layers and the approximate index of turbulent kinetic energy; and interaction and time lag features include the interaction features of wind speed and wind shear with hourly coding features and the lag features of the U and V components at preset historical times. The multidimensional feature vectors total approximately 60 to 70 dimensions, as detailed below: 1) Basic state characteristics This includes the U and V components and the composite wind speed WS at all observation heights at the current moment.

[0041] 2) Time coding features Sine-cosine coding is used to map the time of day and the accumulated day of the year into a continuous and smooth two-dimensional representation, avoiding discontinuities at the boundary of the cycle (such as the jump between 23:00 and 0:00), while enabling the model to learn the impact of diurnal and seasonal variations on wind conditions.

[0042] The intraday time characteristics are defined as follows: in, , These represent the hour and minute corresponding to the current time, respectively. Define the annual cycle time characteristics as: in, It is accumulated over the years.

[0043] 3) Multi-scale trends and acceleration characteristics Within N consecutive time steps, the first-order variation trend and second-order acceleration characteristics of the U and V components of all observed height layers are explicitly constructed using sliding windows of different time scales, enabling the model to identify dynamic information such as whether the wind conditions are accelerating, decelerating, or stabilizing.

[0044] The first-order trend is defined as the difference between the value at the current time step and the values ​​at past time steps, reflecting the direction and rate of wind change. The calculation formula is as follows: Among them, wind component , This is the current end time step of the sliding window. This is the length of the sliding window.

[0045] Second-order acceleration is defined as the difference between the first-order trends over two consecutive time intervals, reflecting whether the rate of change of wind itself is increasing or decreasing. The calculation formula is as follows: 4) Vertical structural features The current moment The observed height is denoted as an ordered set. ,in, For a given target height layer Define its directly above adjacent height layer as The wind components are respectively denoted as and Define the directly adjacent height layer below as The wind components are respectively denoted as and If the target height layer is located at the profile boundary, the corresponding feature can be set to a missing value.

[0046] The vertical wind shear component, defined based on the wind components of adjacent layers, is: Vertical wind shear intensity (composite shear value) is defined as: The wind shear trend is calculated based on the vertical wind shear intensity, reflecting the change trend of vertical wind shear intensity over time, so as to capture the indication information of short-term forecasts such as the establishment of low-level jet or the enhancement of vertical mixing in the boundary layer.

[0047] Wind shear trend is defined as: in, The wind shear trend is the linear trend or rate of change of the vertical wind shear intensity within the sliding window. An increase in the wind shear trend usually indicates an intensified difference in wind speed between upper and lower layers, which may indicate the establishment of a low-level jet, enhanced vertical mixing in the boundary layer, or abrupt changes in the wind profile structure. It has indicative significance for short-term wind condition forecasting.

[0048] 5) Statistical characteristics of the entire profile The mean and standard deviation of the U and V components of the entire profile are calculated across all height layers. The mean, maximum, and standard deviation of the composite wind speed of the entire profile are also calculated. A profile consistency index (defined as the ratio of the standard deviation to the mean of the composite wind speed; this index measures the consistency of wind speed across different height layers, with a larger value indicating greater wind speed differences and stronger profile shear) is also calculated. Simultaneously, the deviation of the U and V components of the target height layer from the mean of the entire profile is calculated, reflecting the relative position of the wind speed at the target height layer within the entire vertical profile. Positive values ​​indicate that the wind speed at that layer is above the average level of the profile, while negative values ​​indicate that it is below the average level.

[0049] 6) Multi-scale rolling statistics and anomaly characteristics The rolling mean and rolling standard deviation of the U and V components and the composite wind speed WS are calculated within multiple sliding windows, and anomaly features are constructed based on the rolling mean. The anomaly feature is defined as the difference between the current value and the rolling mean, which measures the degree to which the current value deviates from the recent average level. A positive value indicates that the current wind component is higher than the recent average, which is helpful for the model to capture sudden changes in wind speed or direction.

[0050] 7) Wind direction characteristics After retrieving wind direction from the U and V components, the change in wind direction over a certain time interval is calculated. Circular interpolation is used to correctly handle cases spanning 0° and 360°. The formula is as follows: At the same time, the rolling standard deviation of wind direction is calculated to reflect the degree of wind direction change. The calculation formula is as follows: in, This is the current end time step of the sliding window. The length of the sliding window. For the time step index within the window, For a moment The radian of the wind direction angle.

[0051] 8) Vertical velocity and turbulence characteristics Construct statistics on the vertical velocity of the entire profile (mean, standard deviation, maximum, minimum), the differences in vertical velocity between adjacent height layers, and the trend of its temporal variation.

[0052] The difference in vertical velocity between adjacent height levels is defined as: The time-varying trend of vertical velocity is defined as: The approximate index of turbulent kinetic energy, constructed based on the three-dimensional wind component fluctuation statistics, is defined as: in, , and These represent the standard deviations of the U, V components and the vertical velocity across the entire profile, respectively. This index, through the comprehensive quantification of velocity fluctuations in all directions, provides an approximate characterization of the intensity of atmospheric boundary layer turbulence, offering auxiliary information for gust forecasting and early warning of sudden wind changes.

[0053] 9) Interaction and time delay characteristics We construct interaction features between wind speed and wind shear and hourly coding features to capture the nonlinear modulation effects of wind speed and wind shear on a diurnal scale.

[0054] The interaction feature between wind speed and hourly coding features is defined as follows: in, For the current moment The corresponding hour, such as 12:25 and 12:47, both correspond to the hour 12. and To Perform sine and cosine encoding.

[0055] The interaction feature between wind shear and hourly encoded features is defined as follows: Furthermore, the lag characteristics of the U and V components at preset historical moments are introduced as input variables to enhance the model's ability to represent time dependence and short-term evolution trends. The lag characteristics of the U and V components at preset historical moments are defined as follows: in, This indicates the time interval between the current moment and the preset historical moment.

[0056] S8. Input the multidimensional feature vector into the trained U-component multidimensional feature model and V-component multidimensional feature model to obtain the enhanced forecast U and V components for specific time periods. The U-component multidimensional feature model and V-component multidimensional feature model are single-output regression models containing an independent lightweight gradient boosting decision tree regressor. In this embodiment, the training process of the U-component multidimensional feature model and the V-component multidimensional feature model includes: S81. Calculate the time interval between adjacent timestamps in the historical wind profile observation data, mark the positions where the time interval exceeds one time step as data gaps, and divide the historical wind profile observation data into several time-continuous data blocks with the data gaps as the boundary. S82. Construct several matching data segments within at least one data block. The matching data segments include wind profile observation data with a length of N time steps and wind profile observation data with a specific time step thereafter. The length of the matching data segments is N+1 time steps. S83. Preprocess several matching data segments to obtain several pairs of U and V component matching sequences respectively; S84. For each matching data segment and its corresponding U and V component matching sequence, construct a multidimensional feature vector based on the wind profile observation data and U and V components of the first N time steps. Use the multidimensional feature vector as the input vector of the U component multidimensional feature model and the V component multidimensional feature model, and use the U and V components of the last time step as the output targets of the U component multidimensional feature model and the V component multidimensional feature model, respectively. S85. Automatic hyperparameter optimization is performed based on Bayesian optimization method, and the U-component multidimensional feature model and V-component multidimensional feature model are trained and validated by combining early stopping mechanism and time series cross-validation.

[0057] An early stopping mechanism is used during training. Training is automatically stopped when the prediction error on the validation set no longer decreases in several consecutive iterations to prevent overfitting and determine the optimal number of iterations. During the cross-validation phase, the optimal number of iterations for each fold is recorded. The final model uses the average of the optimal number of iterations for each fold as the total number of training rounds and is retrained on all data to obtain the final model.

[0058] It should be noted that different U-component multidimensional feature models and V-component multidimensional feature models need to be trained for different specific timeframes, such as the next 30 minutes or the next hour.

[0059] In the final model validation phase, the system will evaluate the model's predictive performance at each altitude level. The evaluation metrics include the root mean square error of the U and V components, as well as the root mean square error and mean absolute error of the composite wind speed.

[0060] S9. Invert the U and V components of the enhanced forecast for a specific time period into the enhanced forecast wind speed and direction for that specific time period.

[0061] The multidimensional feature model explicitly constructs multidimensional feature vectors with clear physical meanings, making full use of the vertical structure information and atmospheric dynamic characteristics in the multi-height wind profile observation data of laser wind radar, and providing additional forecast references at specific time points.

[0062] Example 2 Please refer to Figure 2 , Figure 2 The image shows a short-term rolling forecast device for airport runway wind conditions provided in this embodiment, including a data acquisition module, a forecast scheduling module, a preprocessing module, a data integration module, a data extraction module, a time series direct propagation module, a first inversion module, a runway decomposition module, a feature construction module, an enhanced forecast module, a second inversion module, a data storage module, and a comprehensive display module.

[0063] The data acquisition module continuously receives wind profile observation data output by the laser wind radar and pushes it to the forecasting and scheduling module, data storage module, and integrated display module. The wind profile observation data includes horizontal wind speed, wind direction, and vertical velocity at multiple height levels.

[0064] The forecasting and scheduling module triggers the preprocessing module when it receives the wind profile observation data at the current moment.

[0065] As the lidar wind profile observation data is updated every minute, a new round of preprocessing and forecasting is automatically triggered, continuously generating forecast curves based on the latest observation data. This rolling mechanism achieves continuous calibration of forecast results, enabling forecasts to respond quickly to changes in wind conditions.

[0066] The preprocessing module preprocesses the wind profile observation data at the current moment, obtains the U and V components at the current moment, and pushes them to the data storage module. The preprocessing includes missing measurement processing, physical boundary verification, spatiotemporal sorting and deduplication, missing data judgment, and wind component decomposition.

[0067] The data integration module constructs retrospective U and V component sequences based on the current U and V components and the U and V components of the past N-1 consecutive time steps. The length of the retrospective U and V component sequences is N time steps. The U and V components of the past N-1 consecutive time steps are obtained from the data storage module.

[0068] The data extraction module extracts the U and V component sequences of the target height layer from the U and V component sequences of the retrospective, and arranges the U and V component sequences of the target height layer in an interleaved manner according to the time step order as an input vector. The target height layer is the preset height layer or the height layer closest to the preset height layer.

[0069] The temporal inversion module inputs the input vector into the trained U-component temporal inversion model and V-component temporal inversion model to obtain the predicted U and V component sequences of the target height layer, respectively. The U-component temporal inversion model and the V-component temporal inversion model are multi-output regression models containing M independent lightweight gradient boosting decision tree regressors. The length of the predicted U and V component sequences of the target height layer is M time steps.

[0070] The first inversion module inverts the forecast U and V component sequences of the target height layer into forecast wind speed sequences and forecast wind direction sequences. It then performs phase-neutral rolling smoothing on the forecast wind speed sequences to obtain smoothed wind speed sequences, and pushes the smoothed wind speed sequences and forecast wind direction sequences to the data storage module and the integrated display module.

[0071] The runway decomposition module decomposes the smoothed wind speed sequence and forecast wind direction sequence into forecast headwind, crosswind and gust component sequences at the target height level based on the azimuth of the in-use runway and pushes them to the integrated display module.

[0072] It should be noted that when the operational status of an in-use runway changes, such as when switching runways, the runway decomposition module automatically updates the azimuth parameters, recalculates the forecast headwind, crosswind, and gust component sequences, and pushes them to the integrated display module to ensure that operators always see the forecast information for the currently in-use runway.

[0073] The feature construction module constructs a multidimensional feature vector based on wind profile observation data from N consecutive time steps and the retrospective U and V component sequences, and pushes it to the data storage module. The wind profile observation data from N consecutive time steps is constructed based on the wind profile observation data at the current moment and the wind profile observation data from the past N-1 consecutive time steps. The wind profile observation data from the past N-1 consecutive time steps is obtained from the data storage module. The multidimensional feature vector includes basic state features, time-coded features, multi-scale trend and acceleration features, vertical structure features, full profile statistical features, multi-scale rolling statistics and anomaly features, wind direction features, vertical velocity and turbulence features, and interaction and time-delay features. The basic state features include the U and V components and the composite wind speed at all observation heights at the current moment. The time-coded features include intraday time features and annual cycle time features constructed based on sine-cosine coding. The multi-scale trend and acceleration features include the U and V components. The first-order trend and second-order acceleration of the vertical structure features include the wind components of adjacent layers at the target height, the vertical wind shear intensity and the wind shear trend; the full profile statistical features include the mean and standard deviation of the U and V components of the full profile, the mean, maximum and standard deviation of the composite wind speed of the full profile, the full profile consistency index and the deviation of the U and V components of the target height layer from the mean of the U and V components of the full profile; the multi-scale rolling statistics and anomaly features include the rolling mean, rolling standard deviation and anomaly features of the U and V components and the composite wind speed; the wind direction features include the change in wind direction and the rolling standard deviation; the vertical velocity and turbulence features include the mean, standard deviation, maximum and minimum of the vertical velocity of the full profile, the difference and time trend of vertical velocity between adjacent height layers and the approximate index of turbulent kinetic energy; and the interaction and time lag features include the interaction features of wind speed and wind shear with the hourly coding features and the lag features of the U and V components at the preset historical time.

[0074] The enhanced forecast module inputs the multidimensional feature vectors into the trained U-component multidimensional feature model and V-component multidimensional feature model to obtain the enhanced forecast U and V components for a specific time period. The U-component multidimensional feature model and V-component multidimensional feature model are single-output regression models containing an independent lightweight gradient boosting decision tree regressor.

[0075] The second inversion module inverts the U and V components of the enhanced forecast for a specific time period into the enhanced forecast wind speed and direction for that specific time period and pushes them to the integrated display module.

[0076] The data storage module stores the received data.

[0077] The integrated display module presents the horizontal wind speed and direction at the target altitude level, as well as the smoothed wind speed sequence and forecast wind direction sequence, in the form of time series curves on the same time axis, connecting and displaying the horizontal wind speed and direction at the target altitude level in the wind profile observation data of N consecutive time steps. It also superimposes the horizontal wind speed and direction at the target altitude level in the wind profile observation data of subsequent received wind profile observation data onto the corresponding forecast curve using a time nearest neighbor matching method. It displays the forecast headwind, crosswind, and gust component sequences at the target altitude level in the form of time series curves. It displays the horizontal wind speed and direction in the wind profile observation data of the current moment, as well as the enhanced forecast wind speed and direction at a specific time, in the form of a compass. Finally, it displays the wind profile observation data and vertical wind shear intensity of the current moment in the form of a profile.

[0078] For each published forecast time step, the corresponding observation value is searched in the subsequent wind profile observation data using the nearest neighbor matching method, with a search tolerance of a configurable time threshold. The observation value is dynamically overlaid and displayed on the forecast curve that transitions from the forecast period to the historical period, forming a dynamic comparison sequence between the forecast and the observation. This enables rolling comparison and verification on the same time axis, supporting operators in real-time assessment of forecast accuracy.

[0079] Example 3 This embodiment provides a device for airport runway wind condition forecasting, which may specifically include: A processor; and a memory arranged to store computer-executable instructions, which, when executed, cause the processor to perform the steps as described in the above embodiments.

[0080] Example 4 This embodiment provides a storage medium for airport runway wind condition forecasts. Specifically, the storage medium may include: The airport runway wind condition forecast processing program is stored on the storage medium. When the airport runway wind condition forecast processing program is executed by the processor, it implements the steps in the above embodiments.

[0081] The above are merely embodiments of this application and are not intended to limit 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 principle of this application should be included within the scope of the claims of this application.

Claims

1. A method for forecasting wind conditions on airport runways, characterized in that, include: The wind profile observation data at the current moment is preprocessed to obtain the U and V components at the current moment. The wind profile observation data includes horizontal wind speed, wind direction and vertical velocity at multiple height levels. The preprocessing includes missing value processing, physical boundary verification, spatiotemporal sorting and deduplication, missing value judgment and wind component decomposition. Based on the U and V components at the current time and the U and V components of the past N-1 consecutive time steps, a retrospective U and V component sequence is constructed, and the length of the retrospective U and V component sequence is N time steps. Extract the U and V component sequences of the target height layer from the U and V component sequences of the retrospective, and arrange the U and V component sequences of the target height layer in an interleaved manner according to the time step order as an input vector. The target height layer is a preset height layer or the height layer closest to the preset height layer. The input vector is input into the trained U-component temporal direct propagation model and V-component temporal direct propagation model to obtain the predicted U and V component sequences of the target height layer, respectively. The U-component temporal direct propagation model and V-component temporal direct propagation model are multi-output regression models containing M independent lightweight gradient boosting decision tree regressors, and the length of the predicted U and V component sequences of the target height layer is M time steps. The predicted U and V component sequences of the target height layer are inverted into predicted wind speed sequences and predicted wind direction sequences. The predicted wind speed sequences are then subjected to phase-neutral rolling smoothing to obtain smoothed wind speed sequences. Based on the in-use runway azimuth, the smoothed wind speed sequence and the forecast wind direction sequence are decomposed into forecast headwind, crosswind and gust component sequences at the target height level; A multidimensional feature vector is constructed based on wind profile observation data from N consecutive time steps and the retrospective U and V component sequences. The wind profile observation data from the N consecutive time steps is constructed from the current wind profile observation data and the wind profile observation data from the past N-1 consecutive time steps. The multidimensional feature vector includes basic state features, time-coded features, multi-scale trend and acceleration features, vertical structure features, full profile statistical features, multi-scale rolling statistics and anomaly features, wind direction features, vertical velocity and turbulence features, and interaction and time-delay features. The basic state features include the U and V components and the composite wind speed at all observation heights at the current time. The time-coded features include intraday time features and annual cycle time features constructed based on sine-cosine coding. The multi-scale trend and acceleration features include the first-order variation trend and second-order acceleration of the U and V components. The vertical structure features... The system includes wind components of adjacent layers at the target height, vertical wind shear intensity, and wind shear trend. The full profile statistical features include the mean and standard deviation of the U and V components of the full profile, the mean, maximum, and standard deviation of the composite wind speed of the full profile, the full profile consistency index, and the deviation of the U and V components of the target height layer from the mean of the U and V components of the full profile. The multi-scale rolling statistics and anomaly features include the rolling mean, rolling standard deviation, and anomaly features of the U and V components and the composite wind speed. The wind direction features include the change in wind direction and the rolling standard deviation. The vertical velocity and turbulence features include the mean, standard deviation, maximum, and minimum of the vertical velocity of the full profile, the difference and time variation trend of the vertical velocity between adjacent height layers, and the approximate index of turbulent kinetic energy. The interaction and time lag features include the interaction features of wind speed and wind shear with hourly coding features, and the lag features of the U and V components at a preset historical time. The multidimensional feature vector is input into the trained U-component multidimensional feature model and V-component multidimensional feature model to obtain the enhanced forecast U and V components for a specific time period. The U-component multidimensional feature model and the V-component multidimensional feature model are single-output regression models containing an independent lightweight gradient boosting decision tree regressor. The enhanced forecast U and V components for a specific time period are inverted into enhanced forecast wind speed and direction for that specific time period.

2. The method according to claim 1, characterized in that, The missing value processing includes identifying missing values ​​and replacing them with missing markers.

3. The method according to claim 1, characterized in that, The physical boundary verification includes: Determine whether the wind profile observation data is within a reasonable range; If not present, replace with a missing marker.

4. The method according to claim 1, characterized in that, The missing information determination includes: Determine whether the proportion of missing markers in the wind profile observation data is higher than a preset threshold. If so, skip this forecast. If not, use forward or backward filling to complete the missing markers. The missing markers are generated in the missing measurement value processing or the physical boundary verification.

5. The method according to claim 1, characterized in that, The spatiotemporal sorting and deduplication includes: The wind profile observation data were sorted by time and altitude, and duplicate data at the same time and altitude were removed.

6. The method according to claim 1, characterized in that, The training process of the U-component temporal transductive model and the V-component temporal transductive model includes: Calculate the time interval between adjacent timestamps in the historical wind profile observation data, mark the positions where the time interval exceeds one time step as data gaps, and divide the historical wind profile observation data into several time-continuous data blocks with the data gaps as boundaries; Within at least one data block, a sliding window is used to acquire several data segments. The input window of the sliding window has N time steps, the output window has M time steps, and the step size is one time step. The aforementioned preprocessing is performed on the aforementioned data segments to obtain several pairs of U and V component sample sequences, respectively; Extract several pairs of U and V component sample sequences of the target height layer from the aforementioned pairs of U and V component sample sequences; For each pair of target height layer U and V component sample sequences, the U and V components of the first N time steps are arranged in an alternating order according to the time step and used as the input vectors of the U component time-series direct push model and the V component time-series direct push model. The U and V components of the next M time steps are used as the output vectors of the U component time-series direct push model and the V component time-series direct push model, respectively. Automatic hyperparameter optimization is performed based on the Bayesian optimization method, and the U-component time-series inductive model and the V-component time-series inductive model are trained and validated using time-series cross-validation. The objective function of the hyperparameter optimization is the weighted root mean square error, and the weight values ​​decay exponentially with the increase of the prediction time step.

7. The method according to claim 1, characterized in that, The phase-neutral rolling smoothing process includes: A centered sliding window is used to take the average value within a symmetrical time range before and after the current forecast time step. The length of the centered sliding window is a configurable parameter.

8. The method according to claim 1, characterized in that, The training process of the U-component multidimensional feature model and the V-component multidimensional feature model includes: Calculate the time interval between adjacent timestamps in the historical wind profile observation data, mark the positions where the time interval exceeds one time step as data gaps, and divide the historical wind profile observation data into several time-continuous data blocks with the data gaps as boundaries; Within at least one data block, several matching data segments are constructed. Each matching data segment includes wind profile observation data of length N time steps and subsequent wind profile observation data of a specific time period. The length of each matching data segment is N+1 time steps. The preprocessing is performed on the aforementioned matched data segments to obtain several pairs of U and V component matching sequences; For each matching data segment and its corresponding U and V component matching sequence, the multidimensional feature vector is constructed based on the wind profile observation data and U and V components of the first N time steps. The multidimensional feature vector is used as the input vector of the U component multidimensional feature model and the V component multidimensional feature model, and the U and V components of the last time step are used as the output targets of the U component multidimensional feature model and the V component multidimensional feature model, respectively. Automatic hyperparameter optimization is performed based on the Bayesian optimization method, and the U-component multidimensional feature model and the V-component multidimensional feature model are trained and validated by combining early stopping mechanism and time series cross-validation.

9. A short-term rolling forecasting device for airport runway wind conditions, characterized in that, It includes a data acquisition module, a forecasting and scheduling module, a preprocessing module, a data integration module, a data extraction module, a time-series direct propagation module, a first inversion module, a runway decomposition module, a feature construction module, an enhanced forecasting module, a second inversion module, a data storage module, and a comprehensive display module, among which: The data acquisition module continuously receives wind profile observation data output by the laser wind radar and pushes it to the forecast scheduling module, the data storage module and the integrated display module. The wind profile observation data includes horizontal wind speed, wind direction and vertical velocity at multiple height levels. The forecasting and scheduling module triggers the preprocessing module when it receives the wind profile observation data at the current moment. The preprocessing module preprocesses the wind profile observation data at the current moment to obtain the U and V components at the current moment and pushes them to the data storage module. The preprocessing includes missing measurement processing, physical boundary verification, spatiotemporal sorting and deduplication, missing data judgment and wind component decomposition. The data integration module constructs a retrospective U and V component sequence based on the current U and V components and the U and V components of the past N-1 consecutive time steps, respectively. The length of the retrospective U and V component sequence is N time steps. The U and V components of the past N-1 consecutive time steps are obtained from the data storage module. The data extraction module extracts the U and V component sequences of the target height layer from the U and V component sequences of the retrospective, and arranges the U and V component sequences of the target height layer in an interleaved manner according to the time step order as an input vector. The target height layer is a preset height layer or the height layer closest to the preset height layer. The time-series inversion module inputs the input vector into the trained U-component time-series inversion model and V-component time-series inversion model to obtain the predicted U and V component sequences of the target height layer, respectively. The U-component time-series inversion model and the V-component time-series inversion model are multi-output regression models containing M independent lightweight gradient boosting decision tree regressors, and the length of the predicted U and V component sequences of the target height layer is M time steps. The first inversion module inverts the predicted U and V component sequences of the target height layer into predicted wind speed sequences and predicted wind direction sequences, performs phase-neutral rolling smoothing on the predicted wind speed sequences to obtain smoothed wind speed sequences, and pushes the smoothed wind speed sequences and the predicted wind direction sequences to the data storage module and the integrated display module. The runway decomposition module decomposes the smoothed wind speed sequence and the forecast wind direction sequence into forecast headwind, crosswind and gust component sequences at the target height level based on the azimuth of the in-use runway and pushes them to the integrated display module. The feature construction module constructs a multi-dimensional feature vector based on wind profile observation data from N consecutive time steps and the retrospective U and V component sequences, and pushes it to the data storage module. The wind profile observation data from the N consecutive time steps is constructed based on the wind profile observation data at the current moment and the wind profile observation data from the past N-1 consecutive time steps. The wind profile observation data from the past N-1 consecutive time steps is obtained from the data storage module. The multi-dimensional feature vector includes basic state features, time-coded features, multi-scale trend and acceleration features, vertical structure features, full profile statistical features, multi-scale rolling statistics and anomaly features, wind direction features, vertical velocity and turbulence features, and interaction and time-delay features. The basic state features include the U and V components and composite wind speed at all observation heights at the current moment. The time-coded features include intraday time features and annual cycle time features constructed based on sine-cosine coding. The multi-scale trend and acceleration features include... The vertical structure features include the wind components of adjacent layers at the target height, vertical wind shear intensity, and wind shear trend. The full profile statistical features include the mean and standard deviation of the U and V components of the full profile, the mean, maximum, and standard deviation of the composite wind speed of the full profile, the full profile consistency index, and the deviation between the U and V components of the target height and the mean of the U and V components of the full profile. The multi-scale rolling statistics and anomaly features include the rolling mean, rolling standard deviation, and anomaly features of the U and V components and the composite wind speed. The wind direction features include the change in wind direction and the rolling standard deviation. The vertical velocity and turbulence features include the mean, standard deviation, maximum, and minimum of the vertical velocity of the full profile, the difference and time variation trend of the vertical velocity between adjacent height layers, and the approximate index of turbulent kinetic energy. The interaction and time lag features include the interaction features of wind speed and wind shear with hourly coding features, and the lag features of the U and V components at a preset historical time. The enhanced forecast module inputs the multidimensional feature vector into the trained U-component multidimensional feature model and V-component multidimensional feature model to obtain the enhanced forecast U and V components for a specific time period. The U-component multidimensional feature model and the V-component multidimensional feature model are single-output regression models containing an independent lightweight gradient boosting decision tree regressor. The second inversion module inverts the enhanced forecast U and V components for a specific time period into enhanced forecast wind speed and wind direction for that specific time period and pushes them to the integrated display module; The data storage module stores the received data; The integrated display module displays the horizontal wind speed and direction at the target altitude level, as well as the smoothed wind speed sequence and the predicted wind direction sequence, in the form of time-series curves on the same time axis from wind profile observation data of N consecutive time steps. It also superimposes the horizontal wind speed and direction at the target altitude level from subsequently received wind profile observation data onto the corresponding prediction curve using a time nearest neighbor matching method. The module displays the predicted headwind, crosswind, and gust component sequences at the target altitude level in the form of time-series curves. It displays the horizontal wind speed and direction at the current moment from the wind profile observation data, as well as the enhanced predicted wind speed and direction at a specific time, in the form of a compass. Finally, it displays the current moment from the wind profile observation data and the vertical wind shear intensity in the form of a profile.

Images