Clear sky bump probability prediction method based on NGboost algorithm

Through the NGboost probability prediction method based on the NGboost algorithm, the problem of insufficient accuracy and deterministic forecasting of clear sky turbulence prediction in the prior art is solved, and higher accuracy and reliability of bump intensity prediction are achieved.

CN119066327APending Publication Date: 2024-12-03NANJING UNIV OF AERONAUTICS & ASTRONAUTICS
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410936553.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-12
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

The existing methods for predicting clear sky turbulence have limitations, and it is difficult to accurately predict clear sky bumps under different conditions. The traditional methods are mainly deterministic forecasts, which cannot effectively capture the spatial and temporal position probability of turbulence, and the radar and satellite detection capabilities are limited.

Method used

The NGboost algorithm is used to predict the probability of the sunny sky bumps based on the NGboost algorithm. By acquiring meteorological data and bump observation data, data preprocessing, oversampling, normalization and parameter tuning are performed, and the NGboost model is used to output the probability of the bump intensity prediction of the future hour by hour.

Benefits of technology

The accuracy and reliability of clear sky turbulence are improved, and factors such as wind shear and temperature gradient are comprehensively considered, the model complexity and training time are reduced, and the accuracy and reliability of bump prediction are enhanced.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119066327A_ABST
    Figure CN119066327A_ABST
Patent Text Reader

Abstract

The invention discloses a clear sky bumping probability prediction method based on an NGboost algorithm, and aims to improve the capability of capturing clear sky turbulence (CAT) of a model through the advantage of probability prediction of the NGboost algorithm. According to the method, a CAT diagnosis index is used as a feature vector, and a clear sky bumping strength probability prediction model based on natural gradient lifting is provided. A regular expression and an SMOT algorithm are used to carry out data quality control, sample category imbalance problem and other preprocessing on a data set combining meteorological data and observation data; then feature vectors are selected, data normalization is carried out, data set proportions are divided, and model optimization is carried out through parameter optimization such as five-fold cross validation; inputting a test set, and carrying out performance verification and sensitivity test on the model; and finally, the NGboost algorithm is used for the multi-dimensional data of the target space-time to obtain the hour-level moderate or larger CAT occurrence probability.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of aviation meteorology, in particular to the field of forecasting and detecting clear-air turbulence, and specifically relates to a clear-air turbulence probability prediction method based on an NGboost algorithm. Background Art

[0002] With the intensification of climate change, the probability of encountering clear-air turbulence (CAT) in the future is increasing. CAT refers to turbulence that occurs at an altitude of more than 6,000 meters and is unrelated to convective clouds. It is composed of vortex airflows of different scales that change with time, and the direction of turbulence is from large scale to small scale dissipation. CAT is also one of the dangerous meteorological phenomena in aviation and has an important impact on flight safety. According to relevant research statistics, CAT events account for 24% of flight accidents caused by weather, while turbulence is more common, reaching 71%. Civilian airliners encounter turbulence and irregular movement during the cruising phase (above 8km), which is called "clear-air turbulence", which is most likely to injure passengers and crew members because they will unfasten their seat belts. CAT often occurs near high-altitude jet streams and frontal systems. The physical mechanism of most CAT events is due to Kelvin–Helmholtz instability caused by strong vertical wind shear (VWS). This instability is caused by wind shear producing areas with low Richardson numbers (Ri<0.25). Although many CAT diagnostic indices have been developed so far, since different indices focus on different factors, the forecast effects of clear-air turbulence in different regions, seasons and causes will also be different. Therefore, it is very important to improve the index forecast algorithm to make it more suitable for predicting clear-air turbulence under different conditions.

[0003] The traditional CAT prediction method has certain limitations, and a more accurate and reliable prediction model is needed to improve the accuracy of the forecast. At present, the main prediction method of CAT is the operational numerical weather prediction (NWP), which is much larger than the CAT length scale, which means that the real-time NWP mode will not be able to completely solve the turbulence prediction problem. Due to the complex mechanism of clear air turbulence, the atmospheric parameters fluctuate and the parameters affect each other. The detection ability of remote sensing technologies such as radar and satellites for clear air turbulence is also limited. At present, domestic patents on turbulence prediction are roughly divided into two categories. One is a method based on the transmission detection of airborne devices, and the other is a turbulence prediction model based on the idea of ​​ensemble prediction of turbulence index. However, the above methods are all deterministic predictions, which is different from the clear air turbulence probability prediction method based on the NGboost algorithm proposed in this invention. The model will eventually output the probability of MOG turbulence at the target time and space position through training and optimization, rather than whether MOG turbulence will occur at the time and space position. With the development and application of machine learning technology, new possibilities are provided for the prediction and detection of clear air turbulence. In order to reduce flight accidents and economic losses caused by turbulence such as CAT and ensure civil aviation operations and passenger safety, we urgently need to implement a clear-air turbulence prediction method based on the NGBoost algorithm. Summary of the invention

[0004] To solve the above problems, the present application provides a clear-air turbulence probability prediction method based on the NGboost algorithm, which improves the accuracy of the model in capturing clear-air turbulence through probabilistic forecasting.

[0005] A clear-air turbulence probability prediction method based on NGboost algorithm, comprising:

[0006] Step 1: Obtain meteorological data and turbulence observation data of a certain area with FL≥200, calculate the CAT diagnostic value and observation data through meteorological data to form a total data set, and preprocess the data;

[0007] Step 2: Based on the SMOTE algorithm, the data obtained in step 1 are processed by oversampling technology. First, the turbulence observation data are converted into digital representation intensity using regular expressions. Then, several CAT diagnostic values ​​at the time and space scales and longitude and latitude points reported by pilots are obtained using interpolation methods. The total data set and feature matrix are obtained as model input.

[0008] Step 3: Normalize the total data set, divide the data set ratio, perform 5-fold cross validation and optimal parameter search, and complete parameter tuning to assist in optimizing the NGboost model;

[0009] Step 4: Use the NGboost model with optimized parameters to train and validate the data set, and output the predicted probability of clear-air turbulence intensity at the hourly level in the future.

[0010] Preferably, in step 1, data preprocessing includes:

[0011] Step 1.1, input meteorological factors to calculate and output different CAT diagnostic feature sequences, and use bilinear interpolation and vertical interpolation methods to unify the resolution of climate model and observation data;

[0012] Step 1.2, perform data quality control on the total data set, delete duplicate values ​​and missing values; use regular expressions to convert the turbulence intensity characters recorded in the observation data into digital labels;

[0013] Step 1.3, annotate the digital labels to obtain labeled data, which is simplified into a two-classification problem including no turbulence events and moderate or greater turbulence events.

[0014] Preferably, in step 2, data oversampling is to use the SMOTE algorithm to perform oversampling processing on the total data set, including:

[0015] Step 2.1: For each sample x in the minority class in the binary classification problem, calculate the distance d between the point and other sample points in the minority class based on the Euclidean formula to obtain K nearest neighbors.

[0016]

[0017] where F xj represents the value of the target sample on the jth feature, F yj Represents the value of other samples y on the jth feature;

[0018] Step 2.2, set the sampling ratio according to the sample imbalance ratio to determine the sampling rate. For each minority sample x, randomly select several samples from K nearest neighbors;

[0019] Step 2.3: Combine each randomly selected neighbor x* with the original sample x to construct a new sample x new , new sample x new Together with the original sample x, it forms the total data set:

[0020] x new =x+rand(0,1)*(x*-x) (2)

[0021] where x new For the position of the new sample, rand(0,1) generates a random number between 0 and 1, which determines the proportion of the step in the direction of the target position x*;

[0022] Step 2.4, separate the training set and the test set, arrange them in chronological order, select feature vectors from the data set, that is, calculate the output CAT diagnostic value, and form a feature matrix.

[0023] Preferably, in step 3, parameter tuning includes:

[0024] Step 3.1, data normalization: normalize only the training set, and apply the normalized transformation to both the training set and the test set;

[0025] Step 3.2, data set division: set the ratio of training set to test set to 8:2, and set the random seed number to ensure data consistency;

[0026] Step 3.3, cross-validation and parameter optimization: Use 5-fold cross-validation to divide the original data set into five subsets of equal size, four of which are used to train the model and one subset is used to test the model; repeat this process five times, each time selecting a different subset as the test set and the rest as the training set, to establish a clear air turbulence prediction model based on the NGboost algorithm;

[0027] A list of parameter values ​​is preset, and a random search algorithm is used to search for optimal parameters of the clear air turbulence prediction model, and the parameters that make the model perform best are found through network search.

[0028] Preferably, step 4 comprises:

[0029] Step 4.1, taking the predicted probability distribution P and an observation y as input, identify each distribution using parameters θ based on an appropriate scoring rule S and assign a score S(P, y);

[0030] Step 4.2, use decision tree as the base learner and negative log likelihood as the loss function L of the NGboost model;

[0031] Step 4.3, use NGboost to train the turbulence data and predict the probability y of the possible clear sky turbulence intensity in the next day.

[0032] Preferably, in step 4.1, the appropriate scoring rule S satisfies the following formula:

[0033]

[0034] Where Q is the true distribution of the output variable y, and P is any other distribution different from Q;

[0035] Preferably, in step 4.2, the negative log-likelihood L is used as the scoring rule. When it is minimized, the scoring calculation formula is as follows:

[0036] L(θ,y)=-log P θ (y) (5)

[0037] In the formula, the parameters θ, P θ(y) are all probability distribution parameters.

[0038] Preferably, step 4.3 comprises:

[0039] Step 4.31, loop through the preprocessed CAT diagnostic data, i.e., feature matrix and label data, and calculate the initial parameter θ (0) :

[0040]

[0041] N base learner outputs and initial parameters θ (0) The sum of θ and θ is used to obtain the parameter θ;

[0042] Step 4.32, the NGboost model calculates the loss function L(θ (n-1) ,y)’s natural gradient Denoted as G i (n) , expressed as:

[0043]

[0044] In the formula, F(θ (n-1) ) is the Fisher information matrix;

[0045] Step 4.33, after the NGboost model obtains the natural gradient of the current iteration, the decision tree is trained, which is expressed as:

[0046]

[0047] Where fit represents the training process of the shallow decision tree, ρ (n) is the scaling factor;

[0048] Step 4.34, use decision trees to build base learners. In each iteration n, the data fitting of the current learner is based on the natural gradient G of the prediction parameters of the previous set of base learners. i (n) ;

[0049] Step 4.35, determine the scaling factor by iteration, and then add it to each θ i (n-1) Add the negative scaled prediction gradient ρ*f of the i-th sample (n) (x i ) Update the prediction parameters θ i (n) :

[0050]

[0051] Where η is the learning rate.

[0052] Preferably, the NGboost model is evaluated, and the detection probability PODy of moderate or greater turbulence and the turbulence threshold of the detection probability PODn of no turbulence are statistically analyzed. A ROC curve is formed with PODn as the x-axis and PODy as the y-axis. The closer the curve is to 1, the more perfect the prediction effect is. The AUC value of the ROC curve is used as the performance skill score of turbulence prediction.

[0053] Preferably, for each probability threshold, PODy and PODn are defined as:

[0054]

[0055] Where a: the number of samples predicted to be positive and observed to be positive, b: the number of samples predicted to be positive but observed to be negative, c: the number of samples predicted to be negative but observed to be positive, d: the number of samples predicted to be negative and observed to be negative.

[0056] Beneficial Effects

[0057] (1) This application comprehensively considers the impact of factors such as horizontal and vertical wind shear, horizontal and vertical temperature gradients, and local temperature changes on CAT intensity, thereby increasing the accuracy of turbulence prediction;

[0058] (2) This application uses the NGboost algorithm with optimized parameters for model training, which reduces the model complexity and training time, and can obtain the predicted probability of turbulence intensity more quickly;

[0059] (3) The present invention proposes an oversampling technology based on the K-nearest neighbor algorithm to process the data set, thereby avoiding the impact of the imbalance of data sample categories on the performance of the model and improving the accuracy and reliability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 is an overall flow chart of an embodiment of the present invention;

[0061] Figure 2 It is a ROC curve of the NGboost model after 200 iterations of an embodiment of the invention;

[0062] Figure 3 It is an annual ROC curve based on the NGboost model of an embodiment of the invention;

[0063] Figure 4 It is a quarterly ROC curve based on the NGboost model of an embodiment of the invention;

[0064] Figure 5 The invention is an embodiment of the invention of the category prediction probability distribution of 500 random samples based on the NGboost model. DETAILED DESCRIPTION

[0065] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0066] The present invention proposes a clear air turbulence probability prediction method based on the NGboost algorithm. The method first uses regular expressions to convert the turbulence text reported by pilots into digital representations of intensity, and deletes duplicate values ​​and missing values. Secondly, the CAT diagnostic data is calculated, and the CAT diagnostic values ​​at the time and space scales and longitude and latitude points of the pilot report are obtained using a 4-dimensional interpolation (time, altitude, longitude, latitude) method to obtain a feature matrix as a model input; then data oversampling, data normalization, data set partitioning, 5-fold cross validation, learning rate, number of weak classifiers and other parameters are tuned to assist in optimizing the model; finally, the NGboost algorithm is used to train the multidimensional feature data to obtain the hourly MOG turbulence prediction probability.

[0067] The embodiment of the present invention comprises the following steps:

[0068] Step 1: Obtain turbulence observation data for a certain area for consecutive years, such as pilot reports or aircraft onboard data, to form operational data for turbulence intensity prediction.

[0069] Meteorological data with the same spatiotemporal characteristics were screened and downloaded, mainly including: u wind speed, v wind speed, vertical wind speed, temperature, gravity potential, etc., and then the CAT diagnostic values ​​were calculated using the above meteorological factors, including: eddy dissipation rate (EDR), Richardson number (Ri), Ellrod turbulence index (Ellrod1), relative vorticity squared (VOSQ), nonlinear equilibrium equation residual (UBF), vertical wind shear (VWS), horizontal temperature gradient (TEMPG) and Brown energy dissipation index (Brown2). Table 1 below provides the calculation methods of the above 8 indices:

[0070] Table 1 8 CAT diagnoses and their calculation methods

[0071]

[0072] Note: 1) C in EDR index kp is the Kolmogorov-Prandtl constant, 2) U is the full wind speed, N is the Brunt-Vaisala frequency, is the Laplace operator, Φ is the potential, J() is the Jacob operator, f and β are the Coriolis parameters and their latitudinal derivatives, ζ a is the absolute vorticity.

[0073] The above two parts of data are unified by bilinear interpolation and vertical interpolation methods to form a total data set S after unifying the spatiotemporal characteristics (unifying the resolution of climate models and observation data). The above calculation can be achieved using the 'RegularGridInterpolator' function in python. S is a two-dimensional matrix containing 13 columns of data, that is, the set consisting of time, altitude, longitude and latitude, CAT diagnosis, and turbulence intensity fields obtained by the above method. The feature vector is Fx = [Diy, i = 1, 2, ..., 8], where the value corresponding to the jth feature in the i-th row is S ij , let β = S 总行数 -1, n is the total number of samples, then 0≤i≤β and 0≤j≤8.

[0074] The quality control of the total data set was performed by using regular expressions to convert characters into numbers. Specifically, the characters recording the turbulence intensity levels in the pilot report data were converted into digital labels using the python're' library, such as: 'Null'→0, 'LGT'→1, 'LGT-MOD'→2, 'MOD'→3, 'MOD-SEV'→4, 'SEV'→5, 'EXT'→6, where 'Null', 'LGT', 'MOD', 'SEV', 'EXT' represent no, mild, moderate, severe, and extreme turbulence, respectively. Duplicate and missing values ​​were then deleted to avoid affecting the prediction performance of the model.

[0075] The digital labels are annotated and simplified into two categories: 0 represents no turbulence event (i.e., the above ‘Null’→0), and 1 represents a moderate or greater level of turbulence event (i.e., the above ‘MOD’→3, ‘MOD-SEV’→4, ‘SEV’→5, ‘EXT’→6).

[0076] Step 2: Use the SMOTE algorithm to oversample the total data set.

[0077] For each sample x in the minority class in the binary classification problem, the distance d between the point and other sample points in the minority class is calculated based on the Euclidean formula to obtain K nearest neighbors. The formula is as follows:

[0078]

[0079] where F xj represents the value of the target sample on the jth feature, F yj Represents the value of other samples y on the jth feature, and x and y represent samples at different spatiotemporal positions.

[0080] A sampling ratio is set according to the sample imbalance ratio to determine the sampling rate. For each minority sample x, several samples are randomly selected from the K neighbors. Assume that the selected neighbor is x*. For each randomly selected neighbor x*, the following formula is used to construct a new sample with the original sample:

[0081] x new =x+rand(0,1)*(x*-x) (2)

[0082] where x new For the position of the new sample, rand(0,1) generates a random number between 0 and 1, which determines the proportion of the step in the direction of the target position x*. new Together with the original sample x, it forms the total data set. The data processing using the SMOTE algorithm does not change the characteristics of the original data feature vector, thus ensuring the data characteristics of the original data.

[0083] The data set is divided into a training set and a test set, and arranged in chronological order. The feature vector Fx = [Diy, i = 1, 2, ..., 8] is selected from the data set, that is, the output CAT diagnostic value is calculated and the feature matrix is ​​formed for the initial parameter θ in the subsequent steps. (0) calculate.

[0084] Step 3: Normalize the preprocessed CAT diagnostic data using linear normalization, also known as the deviation standard method or min-max normalization. Only the training set is normalized, and the normalized transformation is applied to the training set and the test set to avoid directly normalizing the total data set and thus causing data information leakage; the purpose of normalization is to eliminate the dimensional differences between different feature vectors, reduce the redundant noise of the model input, scale the data to the mapping interval [0,1], and use the following formula for data maximum-minimum normalization:

[0085] First, calculate the normalized X of the original data std :

[0086]

[0087] Where X is the original data, X.min(axis=0) and X.max(axis=0) are the minimum and maximum values ​​of the data on each feature, respectively.

[0088] The normalized data is then scaled to the specified range [0,1]:

[0089] x scaled =X std ·(max-min)+min (4)

[0090] Where Xscaled is the scaled value. Set the ratio of training set to test set to 8:2, and set the random seed number to ensure data consistency.

[0091] Using 5-fold cross validation, the training set in the original dataset is divided into five subsets of equal size, four of which are used to train the model and one is used to test the model. This process is repeated five times, each time a different subset is selected as the test set and the rest are used as the training set.

[0092] Establish a clear-air turbulence prediction model based on the NGboost algorithm. Input the training set data and perform model training. Figure 1 This is the training and verification process of the clear air turbulence prediction model based on the NGboost algorithm. At the same time, the random search algorithm is used to find the optimal parameters. Based on the principle of random sampling, within a given variable change range, points are randomly generated and the values ​​of the constraint function and objective function are calculated to find the approximate solution of the optimal solution. Preset parameter value list, such as learning rate "learning_rate": [0.01, 0.05, 0.1], set step size to 0.01; number of weak classifiers "n_estimators": [100, 500, 1000]; minimum number of samples required for node splitting "min_samples_split": [2, 10, 1]; minimum number of samples required for each leaf node "min_samples_leaf": [1, 10, 1]; minibatch fraction used for sample sampling and obtaining a small sample set "minibatch_frac": [0.5, 1.0]; before splitting each node in the same layer, randomly select a part of the feature column subsampling "col_sample": [0.5, 1.0], etc. Find the parameters that can make the model perform best through network search. In this embodiment, n_estimators=500, learning_rate=0.01, min_samples_split=5, min_samples_le af=5, minibatch_frac=1.0, col_sample=0.5 are selected. At this time, the prediction effect of the model is better.

[0093] The base learner is the core component of the NGboost model. In each iteration, NGboost will use the residuals estimated in the previous iteration to train multiple base learners and add the trained base learners to the overall model, so that the updated overall model has a smaller prediction error.

[0094] Step 4: Use the NGboost model with optimized parameters to train and validate the dataset.

[0095] Taking the predicted probability distribution P and an observed result y as input, we assign a score S(P, y) to the predicted result based on an appropriate scoring rule S so that the true distribution of the predicted result can obtain the expected optimal score. The appropriate scoring rule S satisfies:

[0096]

[0097] Where Q is the true distribution of the output variable y, and P is any other distribution different from Q. When each distribution is identified using parameters θ, the score is S(θ,y).

[0098] Using decision trees as base learners, it is crucial to select an appropriate loss function for the NGboost algorithm for probability prediction. The present invention uses negative log-likelihood as the loss function of the NGboost model, which can better describe the degree of fit between the predicted probability distribution and the true target variable. When it is minimized, also known as MLE, the score calculation formula is as follows:

[0099] L(θ,y)=-log P θ (y) (6)

[0100] Parameters θ, P θ (y) are all probability distribution parameters, that is, the probability distribution specified in advance in the model. If the specified probability distribution is a normal distribution, θ contains two parameters: mean μ and standard deviation σ; if it is specified as a Bernoulli distribution, θ is the probability p of success (data label is 1). This embodiment uses the Bernoulli distribution. Regardless of which probability distribution is selected, the probability distribution parameters are learned and continuously updated through training data.

[0101] Then, NGboost is used to train the turbulence data to predict the probability of clear sky turbulence intensity in the next day. First, the preprocessed CAT diagnosis data, i.e., the feature matrix and label data, are looped through to calculate the initial parameter θ (0) :

[0102]

[0103] where y i Belongs to the label data. The parameter θ is output by N base learners and the initial parameter θ (0) The sum of the negative log-likelihood loss function is obtained. Directly using the gradient information of the negative log-likelihood loss function to train the decision tree will result in poor prediction performance of NGboost. This is because the gradient descent in the probability distribution space is asymmetric with the gradient descent in the parameter space. Using the natural gradient instead can avoid the above problem. If the current iteration optimization number is n, NGboost will calculate the loss function L(θ in n-1 iterations for all training samples. (n-1) ,y)’s natural gradient Denoted as G i (n) , and its calculation formula is shown as follows:

[0104]

[0105] in Represents the Fisher information matrix. After obtaining the natural gradient of the current iteration, NGboost will train multiple decision trees as shown below:

[0106]

[0107] Where fit represents the training process of the shallow decision tree, ρ (n) is the scaling factor. Use decision trees to build base learners. In each iteration n, the data fitting of the current learner is based on the natural gradient G of the prediction parameters of the previous set of base learners. i (n) The output of the fitted base learner is the projection of the natural gradient on the range of the base learner set, and its projected gradient is scaled by the scaling factor.

[0108] The scaling factor is determined iteratively and then added to each θ i (n-1) Add the negative scaled prediction gradient ρ*f of the i-th sample (n) (x i ) Update the prediction parameters θ i (n) :

[0109]

[0110] Where η is the learning rate. The above gradient is further scaled by a smaller η (0.01 or 0.1).

[0111] The roc-auc is selected as the evaluation index of the clear air turbulence intensity classification prediction model. The receiver operating characteristic (ROC) curve consists of 1-PODn (x-axis) and PODy (y-axis). The detection probability (PODy) of moderate and above turbulence (ie, data label: 1) and the turbulence threshold of no turbulence (ie, data label: 0) are statistically calculated, and the ROC curve calculation is generated. As the prediction threshold changes, the ROC curve plots the relationship between PODy and 1-PODn. The closer the curve is to the upper left corner (the closer to 1, 0.5 is a random guess), the more perfect the prediction effect. The AUC value of the ROC curve is ultimately used as the performance skill score for turbulence forecasting. For each probability threshold, PODy and PODn can be defined as:

[0112]

[0113] Where a: the number of samples predicted to be positive and observed to be positive, b: the number of samples predicted to be positive but observed to be negative, c: the number of samples predicted to be negative but observed to be positive, d: the number of samples predicted to be negative and observed to be negative.

[0114] Input the test set data to verify the prediction hit rate of the model. Figure 2 This is a sensitivity test to verify the NGboost model. The ROC curve is plotted after 200 iterations of randomly selecting available turbulence observation data using an observation subset. The green and red lines show the two ROC curves with the maximum and minimum AUC values ​​among the 200 ROC curves. The results show that the median AUC ranges from -2.0% to +1.0% compared to the total observations, indicating that the performance skills of all evaluation tests are statistically stable.

[0115] Furthermore, the sensitivity of the model on annual and quarterly basis was verified. Figure 3 The ROC curves of the model in different years are shown. It can be observed from the figure that in these 7 ROC curves, the forecast skill score performance varies with the year. Except for the low AUC value (0.89) in 2016, the AUCs of the other 6 years are all above 0.91. Among them, the forecast performance in 2022 is the largest compared with 2016.

[0116] Figure 4 The seasonal ROC curve of the model for the entire turbulence event is shown. It can be seen that the prediction skill of the NGboost model is good for summer (JJA) and winter (DJF), with AUC values ​​of 0.94 and 0.93 respectively, while the prediction ability of the model is slightly weaker in spring (MAM), but its AUC value also reaches 0.90, indicating that the model has good overall performance in seasonal climate change.

[0117] Figure 5 The predicted probability distribution of 500 random samples output by the NGboost model proposed in the present invention is shown, with the number of samples on the horizontal axis and the predicted probability on the vertical axis, where "Null" represents the probability of no turbulence and "MOG" represents the probability of moderate or greater turbulence. It can be observed from the figure that in the 500 randomly selected samples, the probability distributions of the two categories of Null and MOG are basically similar, indicating that the model has a good ability to distinguish Null from MOG.

[0118] Furthermore, the maximum true skill statistics (TSS) score of each model was calculated, also known as the Hanssen–Kuipers (HK) discriminant, which is defined as TSS = PODy + PODn – 1. Its value range is -1 to 1, where 1 means that the diagnostic model has perfect skill in classification prediction, and values ​​less than 0 represent negative discrimination skill (0 means no skill).

[0119] Table 2. Area under the ROC curve (AUC) and true skill score (TSS) of different models predicting CAT

[0120]

[0121] Table 2 shows the prediction evaluation of the overall CAT events by each forecast model. W-KNN, RF, and CATavg are weighted K nearest neighbor, random forest, and ensemble forecast models, respectively. The Max TSS values ​​obtained from the forecast methods of different diagnostic models are compared. The Ellrod1 index and UBF index have the highest (0.272) and lowest (0.095) maxTSS values ​​in the single diagnosis, respectively. It is worth noting that the ensemble forecast has a higher TSS in turbulence prediction compared with the single diagnosis, because the ensemble forecast improves the statistical performance by combining more types of turbulence generation mechanisms. Compared with W-KNN, RF, and CATavg, the NGboost forecast model shows the highest TSS in diagnosing moderate or greater turbulence events and no turbulence events. TSS increases from 0.315 to 0.660 for the ensemble forecast. For the NGboost forecast model, this performance gain is achievable, with a smaller computational cost than the ensemble forecast, and the prediction runtime efficiency is improved by 2.4-2.6 times. Through experimental comparison, it is found that the computation time of the ensemble prediction method increases linearly with the number of individual diagnostics used.

[0122] In summary, the NGboost-based CAT prediction model proposed in this invention greatly improves the ability to capture CAT, providing safety protection for coping with the future intensification of clear-air turbulence. It can provide practical reference for pilots before takeoff, thereby improving aviation safety factor and reducing economic costs. This has important practical significance for ensuring the safety and efficiency of civil aviation transportation.

[0123] The present invention encompasses any substitution, modification, equivalent method and scheme made on the essence and scope of the present invention. In order to make the public have a thorough understanding of the present invention, specific details are described in detail in the following preferred embodiments of the present invention, and those skilled in the art can fully understand the present invention without the description of these details.

[0124] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A clear-air turbulence probability prediction method based on the NGboost algorithm, characterized in that: include: Step 1: Obtain meteorological data and turbulence observation data of a certain area with FL≥200, calculate the CAT diagnostic value and observation data through meteorological data to form a total data set, and preprocess the data; Step 2: Based on the SMOTE algorithm, the data obtained in step 1 are processed by oversampling technology. First, the turbulence observation data are converted into digital representation intensity using regular expressions. Then, several CAT diagnostic values ​​at the time and space scales and longitude and latitude points reported by pilots are obtained using interpolation methods. The total data set and feature matrix are obtained as model input. Step 3: Normalize the total data set, divide the data set ratio, perform 5-fold cross validation and optimal parameter search, and complete parameter tuning to assist in optimizing the NGboost model; Step 4: Use the NGboost model with optimized parameters to train and validate the data set, and output the predicted probability of clear-air turbulence intensity at the hourly level in the future.

2. The clear-air turbulence probability prediction method based on the NGboost algorithm according to claim 1 is characterized in that: In step 1, data preprocessing includes: Step 1.1, input meteorological factors to calculate and output different CAT diagnostic feature sequences, and use bilinear interpolation and vertical interpolation methods to unify the resolution of climate model and observation data; Step 1.2, perform data quality control on the total data set, delete duplicate values ​​and missing values; use regular expressions to convert the turbulence intensity characters recorded in the observation data into digital labels; Step 1.3, annotate the digital labels to obtain labeled data, which is simplified into a two-classification problem including no turbulence events and moderate or greater turbulence events.

3. The clear-air turbulence probability prediction method based on the NGboost algorithm according to claim 2 is characterized in that: In step 2, data oversampling is to use the SMOTE algorithm to oversample the total data set, including: Step 2.1: For each sample x in the minority class in the binary classification problem, calculate the distance d between the point and other sample points in the minority class based on the Euclidean formula to obtain K nearest neighbors. Among them, F xj represents the value of the target sample on the jth feature, F yj Represents the value of other samples y on the jth feature; Step 2.2, set the sampling ratio according to the sample imbalance ratio to determine the sampling rate. For each minority sample x, randomly select several samples from K nearest neighbors; Step 2.3: Combine each randomly selected neighbor x* with the original sample x to construct a new sample x new , new sample x new Together with the original sample x, it forms the total data set: x new =x+rand(0,1)*(x*-x) (2) where x new For the position of the new sample, rand(0,1) generates a random number between 0 and 1, which determines the proportion of the step in the direction of the target position x*; Step 2.4, separate the training set and the test set, arrange them in chronological order, select feature vectors from the data set, that is, calculate the output CAT diagnostic value, and form a feature matrix.

4. The clear-air turbulence probability prediction method based on the NGboost algorithm according to claim 1, characterized in that: In step 3, parameter tuning includes: Step 3.1, data normalization: normalize only the training set, and apply the normalized transformation to both the training set and the test set; Step 3.2, data set division: set the ratio of training set to test set to 8:2, and set the random seed number to ensure data consistency; Step 3.3, cross-validation and parameter optimization: Use 5-fold cross-validation to divide the original data set into five subsets of equal size, four of which are used to train the model and one subset is used to test the model; repeat this process five times, each time selecting a different subset as the test set and the rest as the training set, to establish a clear air turbulence prediction model based on the NGboost algorithm; A list of parameter values ​​is preset, and a random search algorithm is used to search for optimal parameters of the clear air turbulence prediction model, and the parameters that make the model perform best are found through network search.

5. The clear air turbulence prediction method based on NGboost algorithm according to claim 3, characterized in that: Step 4 includes: Step 4.1, taking the predicted probability distribution P and an observation y as input, identify each distribution using parameters θ based on an appropriate scoring rule S and assign a score S(P, y); Step 4.2, use decision tree as the base learner and negative log likelihood as the loss function L of the NGboost model; Step 4.3, use NGboost to train the turbulence data and predict the probability y of the possible clear sky turbulence intensity in the next day.

6. The clear air turbulence prediction method based on NGboost algorithm according to claim 1, characterized in that: In step 4.1, the appropriate scoring rule S satisfies the following formula: Where Q is the true distribution of the output variable y, and P is any other distribution different from Q.

7. The clear air turbulence prediction method based on NGboost algorithm according to claim 1, characterized in that: In step 4.2, the negative log-likelihood L is used as the scoring rule. When it is minimized, the scoring calculation formula is as follows: L(θ,y) -log P θ (y) (5) In the formula, the parameters θ, P θ (y) are all probability distribution parameters.

8. The clear air turbulence prediction method based on NGboost algorithm according to claim 1, characterized in that: Step 4.3 includes: Step 4.31, loop through the preprocessed CAT diagnostic data, i.e., feature matrix and label data, and calculate the initial parameter θ (0) : N base learner outputs and initial parameters θ (0) The sum of the parameters θ is obtained; Step 4.32, the NGboost model calculates the loss function L(θ (n-1) ,y)’s natural gradient Denoted as G i (n) , expressed as: In the formula, F(θ (n-1) ) is the Fisher information matrix; Step 4.33, after the NGboost model obtains the natural gradient of the current iteration, the decision tree is trained, which is expressed as: Where fit represents the training process of the shallow decision tree, ρ (n) is the scaling factor; Step 4.34, use decision trees to build base learners. In each iteration n, the data fitting of the current learner is based on the natural gradient G of the prediction parameters of the previous set of base learners. i (n) ; Step 4.35, determine the scaling factor by iteration, and then add it to each θ i (n-1) Add the negative scaled prediction gradient ρ*f of the i-th sample (n) (x i ) Update the prediction parameters θ i (n) : Where η is the learning rate.

9. The clear air turbulence prediction method based on NGboost algorithm according to claim 1, characterized in that: The NGboost model is evaluated, and the detection probability PODy of moderate or greater turbulence and the turbulence threshold of the detection probability PODn of no turbulence are statistically analyzed. The ROC curve is formed with PODn as the x-axis and PODy as the y-axis. The closer the curve is to 1, the more perfect the prediction effect is. The AUC value of the ROC curve is used as the performance skill score of turbulence prediction.

10. The clear air turbulence prediction method based on NGboost algorithm according to claim 9, characterized in that: For each probability threshold, PODy and PODn are defined as: Where a: the number of samples predicted to be positive and observed to be positive, b: the number of samples predicted to be positive but observed to be negative, c: the number of samples predicted to be negative but observed to be positive, d: the number of samples predicted to be negative and observed to be negative.