An aircraft-intention and data-driven short-term trajectory prediction method

Abstract

The application discloses a kind of short-term track prediction methods based on aircraft intention and data driving.First, a large number of track ADS-B feature information is obtained by selecting airport, and data is preprocessed using DBSCAN and Newton interpolation method.Second, based on the segmented track density, the track is assigned a weight, and the segmented nominal track is obtained.Third, based on the TBO four-dimensional track prediction strategy, an aircraft intention model for short-term prediction is constructed.Finally, the intention information and the processed ADS-B position information are input to train the Seq2Seq model, and the short-term prediction results of different step lengths are obtained.The segmented nominal track obtained by the application can accurately distinguish the characteristics of different flight segments from a macroscopic perspective, and the nominal track obtained based on the track density can perceive the main development trend of the track from a microscopic perspective.The aircraft intention information obtained from the time and space aspects can be integrated into the data-driven model, which can enhance the accuracy of short-term prediction.

Description

TECHNICAL FIELD

[0001] The application relates to an aircraft-intention and data-driven short-term trajectory prediction method and belongs to the field of trajectory prediction. BACKGROUND

[0002] The trajectory prediction problem involves meteorology, performance, intention and the like, currently used trajectory prediction methods are mainly divided into three categories, namely state estimation models, dynamic models and machine learning models, but are mostly used for strategic and pre-tactical stages and implement long-term prediction, and short-term prediction is less and prediction accuracy needs to be improved. Short-term trajectory prediction belongs to the tactical stage, and prediction accuracy greatly affects the probability of conflict existence and the accuracy of conflict detection, and is crucial to traffic management, arrival management, anomaly detection and the like. Therefore, the short-term online trajectory prediction problem needs to be solved. SUMMARY

[0003] The application provides an aircraft-intention and data-driven short-term trajectory prediction method, which realizes high-precision prediction in a short time and can be used for conflict detection and resolution. The steps are as follows:

[0004] Step 1: select ADS-B data of trajectories between two airports in one year, and perform data processing by using DBSCAN and Newton interpolation method, including completing missing data and removing abnormal data.

[0005] Step 2: segmentally extract nominal trajectories. A method based on density distribution weight is used to give different trajectories weight values, and a weighted sum is used to obtain the nominal trajectory of the current segment.

[0006] Step 3: based on the strategy of TBO four-dimensional trajectory prediction, the time change caused by the angle change and the arc distance difference is introduced into the trajectory intention, and the short-term trajectory intention information is enriched.

[0007] Step 4: input position information and intention information of different step lengths, train a Seq2Seq model with an embedded neural network, and obtain output contents of different step lengths.

[0008] Step 1 specifically includes:

[0009] Step 1.1: the longitude, latitude, heading and speed in the ADS-B data are subjected to MinMax standardization processing according to the following formula, so that the trajectory features are dimensionless and the subsequent model training speed is accelerated.

[0010]

[0011] Wherein, x i is a standard value, x i is an original value, x min is a minimum feature value, and x max is a maximum feature value.

[0012] Step 1.2: The DBSCAN method based on density clustering is used to cluster the track positions, the neighborhood radius Eps and the threshold value MinPts defining the core points are determined, the best cluster number can be automatically obtained, and abnormal tracks can be removed.

[0013] Step 1.3: The mean reconstruction is used to reconstruct the feature values corresponding to the time stamps by taking T as the time slice. Given the time stamps t1, t2 and the features x1, x2, if t2-t1

[0014]

[0015] Step 1.4: Since the single track has no periodicity and no specific rules, Newton interpolation method can be used to complete the missing values. The known track points are used to establish a suitable interpolation function f(x),

[0016]

[0017] where f[x0,…,x n ] is the n-order difference quotient, and the calculation method is as follows:

[0018]

[0019] The missing point p i The function value f(p i ) can be obtained from the interpolation function f(x), and the approximate value obtained is used as the missing value to complete the missing track features.

[0020] Step 2 specifically includes:

[0021] Step 2.1: According to the characteristics of different flight segments, m flight segments M={M1, M2, …, M m} are divided, for the current flight segment M i , the track with larger density is given larger weight, and the track with smaller density is given smaller weight, and the outermost and innermost tracks are selected as the frame range, when calculating the density of each track, the track is taken as the center, and the frame range is opened to both sides, the total track a ij in the calculation range is calculated, and the linear density is:

[0022]

[0023] Step 2.2: The linear density is z-score standardized,

[0024]

[0025] where μ i is the mean of the i-th flight segment track set, and σ iFor variance.

[0026] Step 2.3: Obtain the leg weight set ω i = {ω i1 , ω i2 , …, ω in}(i = 1, 2, …, m), and the nominal track of each leg is obtained by density weighted summation, M i The nominal track of the leg is:

[0027]

[0028] Where l j is the jth track.

[0029] Step 3 specifically includes:

[0030] Step 3.1: Based on the strategy of TBO four-dimensional track, the angle difference and the arc distance are introduced into the track intention to construct a short-term track intention vector.

[0031] P1 is the current nominal track point, P2 is the next nominal track point, P0 is the current recognized track point,

[0032] d0 is the arc distance of the concentric circles with P1 as the center, the radius of l0 and l, and a0 is the angle between l0 and l.

[0033] Step 3.2: Recognize the current position, use the Euclidean distance to obtain the nearest nominal track point, dynamically adjust the radius r of the circle with P1 as the center, and include P0 in the circle, but not in the circle with P2 as the center and radius r.

[0034] Step 3.3: a0 is the angle difference between l0 and l, which is used to describe the direction intention of the aircraft,

[0035] a0 = l0 - l

[0036] If l0 precedes l, a0 takes a positive value, and if l0 lags behind l, a0 takes a negative value.

[0037] d0 represents the controllable distance from the nominal track point to the next track point. The quotient of the current ground speed is obtained, and T0 is obtained, which can control the predicted time of arrival at the next nominal track point,

[0038]

[0039] If l > l0, T0 takes a positive value, and vice versa.

[0040] Therefore, the intention information I = [a0, T0] is constructed.

[0041] Step 4 specifically includes:

[0042] Step 4.1: The application uses a Seq2Seq model for prediction, and the encoding network and decoding network can be set to LSTM, GRU, etc. The input information is X t ={x t-q ,x t-q+1 ,…,x t} with a length of q, x i =[t i ,lon i ,lat i ,α 0i ,T 0i ], and the output information is Y t ={y t+1 ,y t+2 ,…,y t+p} with a length of p, y j =[t j ,lon i ,lat i ]. The track feature information is divided into position information P=[t,lon,lat] and intention information I=[α0,T0].

[0043] Step 4.2: The Seq2Seq encodes the input information X t , which is processed by the following formula.

[0044]

[0045] wherein, is the encoder learning parameter.

[0046] According to the above, the hidden sequence H t ={h t-q ,h t-q+1 ,…,h t} can be obtained.

[0047] Step 4.3: The decoder takes the output Z j of the encoder q(h1,h2,…,h q ), the previous time prediction value and the latent representation u j-1 as the input of the decoder, and after the decoder, the parameters in the network are learned, and the prediction value of the future p time can be obtained.

[0048] Step 4.4: The root mean square error (RMSE), mean absolute error (MAE), and mean relative error (MRE) calculation methods are used to train the model, so that the prediction value and the true value interpolation are minimized.

[0049]

[0050]

[0051]

[0052] wherein, p i is a real value, is a predicted value.

[0053] The beneficial effects achieved by the present application are: segmenting the nominal track can more accurately distinguish the characteristics of different segments from a macroscopic perspective, and segmenting the nominal track based on track density can perceive the main development trend of the track from a microscopic perspective; adding aircraft intention information that can control the arrival of the aircraft from both time and space aspects enhances the short-term accuracy of the prediction; different subjects have different needs for short-term tracks, so different time scale prediction information needs to be output, and the Seq2Seq structure can meet this requirement. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 is a flowchart of the method of the present application;

[0055] Figure 2 is a track intention diagram in a conventional operation mode;

[0056] Figure 3 is a Seq2Seq structure diagram;

[0057] Figure 4 is an ADS-B data diagram;

[0058] Figure 5 is a ZUUU-ZSHC track diagram. DETAILED EMBODIMENT

[0059] To make the purpose, technical scheme and points of the present application more clear, the track between Chengdu Shuangliu Airport (ZUUU) and Hangzhou Xiaoshan Airport (ZSHC) is selected to further illustrate the present application in combination with the drawings.

[0060] The present application proposes a short-term track prediction method based on aircraft intention and data driving, and the flowchart is as shown in Figure 1 The specific implementation steps are as follows:

[0061] Step 1: Obtain the ADS-B data of flight 3U8088 from ZUUU to ZSHC from July 1, 2022 to June 30, 2023 through python crawler variflight, as shown in Figure 4 DBSCAN and Newton interpolation method are used for data processing, including completing missing data, eliminating abnormal data, etc.

[0062] Step 1.1: MinMax standardization is performed on the longitude, latitude, heading, and speed in ADS-B data according to the following formula to make the track features dimensionless and accelerate the subsequent model training speed.

[0063]

[0064] Step 1.2: DBSCAN method based on density clustering is used to cluster the track positions, with a neighborhood radius Eps = 0.1 and a core point threshold MinPts = 300, to obtain the best clustering number of 2 classes, and to eliminate abnormal tracks, as shown in Figure 5

[0065] Step 1.3: In this example, 10 min is taken as the time slice, and the feature values corresponding to the time stamps are reconstructed by mean value. Given time stamps t1, t2 and features x1, x2, if t2-t1 < 10 min, then the reconstructed standard feature value is

[0066]

[0067] Step 1.4: Due to the characteristics of single track without periodicity and specific rules, Newton interpolation method is used to complete the missing values. An appropriate interpolation function f(x) is established using known track points,

[0068]

[0069] where f[x0,…,x n ] is the n-th difference quotient, and its calculation method is as follows:

[0070]

[0071] The missing point p i The function value f(p i ) can be obtained from the interpolation function f(x), and the approximate value obtained is used as the missing value to complete the missing track features.

[0072] Step 2: Extract the nominal track in segments. The method based on density distribution weight gives different weights to different tracks, and the weighted sum is obtained to get the nominal track of the current segment.

[0073] Step 2.1: According to the characteristics of different segments, m segments M = {M1, M2, …, M m} are divided, and for the current segment M i , the track with larger density is given larger weight, and the track with smaller density is given smaller weight. The outermost and innermost tracks are selected as the frame range, and when calculating the density of each track, the track is taken as the center, and the frame range is opened to both sides. The total track a ij in the calculation range is calculated, and the linear density is: ​

[0074]

[0075] Step 2.2: z-score standardization is performed on the linear density,

[0076]

[0077] where μ i is the mean of the i-th leg track set, and σ i is the variance.

[0078] Step 2.3: Obtain the leg weight set ω i = {ω i1 , ω i2 , …, ω in}(i = 1, 2, …, m), and the nominal track of each leg is obtained by density-weighted summation, M i The leg nominal track is:

[0079]

[0080] where l j is the j-th track.

[0081] Step 3: Based on the strategy of TBO four-dimensional track prediction, the time change caused by the angle change and the arc distance difference is introduced into the track intention, and the short-term track intention information is enriched.

[0082] Step 3.1: Based on the strategy of TBO four-dimensional track, the angle difference and the arc distance are introduced into the track intention, and the short-term track intention vector is constructed, as shown in the schematic diagram Figure 2 .

[0083] P1 is the current nominal track point, P2 is the next nominal track point, P0 is the current recognized track point,

[0084] d0 is the arc distance of the concentric circles with P1 as the center and the radii of l0 and l, and a0 is the included angle of l0 and l.

[0085] Step 3.2: The current position is recognized, the Euclidean distance is used to obtain the nearest nominal track point, and the radius r of the circle with P1 as the center is dynamically adjusted to include P0 in the circle, but not in the circle with P2 as the center and the radius r.

[0086] Step 3.3: a0 is the angle difference between l0 and l, which is used to describe the direction intention of the aircraft,

[0087] a0 = l0 - l

[0088] If l0 precedes l, a0 takes a positive value, and if l0 lags behind l, a0 takes a negative value.

[0089] d0 represents the controllable distance from the nominal track point and the current track point to the next track point. The quotient of the current ground speed is obtained, and T0 is obtained, which can control the predicted time of arrival at the next nominal track point,

[0090]

[0091] If l > l0, T0 takes a positive value, otherwise takes a negative value.

[0092] Therefore, the intention information I = [a0, T0] is constructed.

[0093] Step 4: input position information and intention information with different step lengths, and train the Seq2Seq model of the built-in LSTM network to obtain output content with different step lengths.

[0094] Step 4.1: the application uses a Seq2Seq model for prediction, as shown in Figure 3 The encoding network and the decoding network are set to the LSTM structure. The input information is X t ={x t-q ,x t-q+1 ,…,x t}, the length is q, x i =[t i ,lon i ,lat i ,α 0i ,T 0i ], and the output information is Y t ={y t+1 ,y t+2 ,…,y t+p}, the length is p, and y j =[t j ,lon i ,lat i ]. The track feature information is divided into position information P = [t, lon, lat] and intention information I = [a0, T0].

[0095] Step 4.2: the Seq2Seq encodes the input information X t , and processes the input information by the following formula.

[0096]

[0097] Where, is the encoder learning parameter.

[0098] According to the above content, the hidden sequence H t ={h t-q ,h t-q+1 ,…,h t} can be obtained.

[0099] Step 4.3: The decoder takes the output Z of the encoder as input, and the prediction value of the previous time and the latent representation u as input, and passes through the decoder to obtain the prediction value of the future p time. j q j-1

[0100] Step 4.4: Train the model by using the root mean square error calculation method to make the prediction value and the true value interpolation minimum.

[0101]

[0102] Wherein, p i is the true value, is the prediction value.

[0103] The above is only an embodiment of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application are included in the scope of the claims of the application to be approved.​​​​

Claims

1. A short-term trajectory prediction method based on aircraft intent and data-driven approach, characterized in that: Includes the following steps: Step 1: Select ADS-B data of flight tracks between two airports within one year, and use DBSCAN and Newton interpolation to process the data, including completing missing data and removing outlier data; Step 2: Extract nominal tracks in segments, assign weights to different tracks based on density-based weighting, and sum the weighted values ​​to obtain the nominal track of the current segment; Step 3: Based on the TBO four-dimensional trajectory prediction strategy, the time changes caused by angle changes and arc distance differences are introduced into the trajectory intent to enrich the short-term trajectory intent information; Step 4: Input the location and intent information of different time lengths, train the Seq2Seq model with built-in neural network, and obtain the output content of different time lengths; Step 2 specifically includes: Step 2.1: Divide the flight into m segments M = {M1, M2, ..., M} based on the characteristics of different segments. m For the current flight segment M i Tracks with higher line density are assigned higher weights, and tracks with lower line density are assigned lower weights. Using the outermost and innermost tracks as the selection area, the selection area is expanded to both sides of each track when calculating its density, and the total track density 'a' within that area is calculated. ij The linear density is: Step 2.2: Standardize the linear density using the z-score. Where μ i Let σ be the mean of the set of tracks for the i-th segment. i For variance; Step 2.3: Obtain the segment weight set ω i ={ω i1 ,ω i2 ,…,ω in }(i=1,2,…,m), the nominal track of each segment is obtained by density-weighted summation, M i The segment's nominal track is: Among them l j For the j-th track; Step 3 specifically includes: Step 3.1: Based on the TBO four-dimensional trajectory prediction strategy, the angle difference and arc distance are introduced into the trajectory intent to construct a short-term trajectory intent vector; P1 is the current nominal track point, P2 is the next nominal track point, and P0 is the currently identified track point. d0 is the arc distance of concentric circles with P1 as the center and radii of l0 and l, and α0 is the angle between l0 and l; Step 3.2: Identify the current position, use Euclidean distance to obtain the nearest nominal track point, and dynamically adjust the radius r of the circle centered at P1 to include P0 within the circle, but not within the circle centered at P2 with radius r. Step 3.3: α0 is the angular difference between l0 and l, used to describe the aircraft's directional intention. α0＝l0-l If l0 precedes l, then α0 takes a positive value; if l0 lags l, then it takes a negative value. d0 represents the controllable distance from the nominal track point and the current track point to the next track point; the quotient of the current ground speed is used to obtain T0, which controls the estimated time to reach the next nominal track point. If l > l0, then T0 takes a positive value; otherwise, it takes a negative value. Therefore, the intent information I = [α0, T0] is constructed.

2. The short-term trajectory prediction method based on aircraft intent and data-driven approach according to claim 1, characterized in that: Step 1 specifically includes: Select the air route between the two airports, import the ADS-B data for that air route and perform preprocessing; Step 1.1: Perform MinMax standardization on the longitude, latitude, heading, and speed in the ADS-B data according to the following formula to make the track features dimensionless and speed up the subsequent model training. Where, x′ i For the standard value, x i x is the original value. min For the eigenvalue minimum, x max It is the maximum value of the characteristic; Step 1.2: The DBSCAN method based on density clustering is used to cluster the track positions, determine the neighborhood radius Eps and define the threshold MinPts of the core point, automatically obtain the optimal number of clusters, and remove abnormal tracks; Step 1.3: Using time slice T, reconstruct the feature values ​​corresponding to the timestamps using the mean; given timestamps t1, t2, and features x1, x2, if t2 - t1 < T, then the reconstructed standard feature value is Step 1.4: Due to the non-periodic and irregular nature of single-track points, Newton's interpolation method is used to fill in the missing values; a suitable interpolation function f(x) is established using the known track points. Where, f[x0,…,x n The quotient is the nth-order difference, and it is calculated as follows: Missing point p i The function value f(p) is obtained from the interpolation function f(x). i The approximate values ​​obtained are used as missing values ​​to complete the missing trajectory features.

3. The short-term trajectory prediction method based on aircraft intent and data-driven approach according to claim 1, characterized in that: Step 4 specifically includes: Step 4.1: Use the Seq2Seq model for prediction, with the encoder and decoder networks configured as LSTM and GRU structures; the input information is X. t ={x t-q ,x t-q+1 ,…,x t }, with length q, x i =[t i ,lon i ,lat i ,α 0i ,T 0i The output information is Y. t ={y t+1 ,y t+2 ,…,y t+p }, with length p, y j =[t j ,lon i ,lat i The trajectory feature information is divided into position information P = [t, lon, lat] and intent information I = [α0, T0]. Step 4.2: Seq2Seq processes the input information X t Encode the input information and process it using the following formula; in, Learn parameters for the encoder; Based on the above, the hidden sequence H is obtained. t ={h t-q ,h t-q+1 ,…,h t }; Step 4.3: The decoder will convert the encoder's output Z... j =q(h1,h2,…,h q ), the predicted value y at the previous time step j-1 and potential representation u j-1 As input to the decoder, the parameters in the network are learned by the decoder to obtain the predicted value Y for the next p time steps. t ={y t+1 ,y t+2 ,…,y t+p }; Step 4.4: Train the model using the root mean square error (RMSE), mean absolute error (MAE), and mean relative error (MRE) calculation methods to minimize the interpolation between the predicted and actual values. Where, p i p is the true value i These are predicted values.

Images