A method for estimating variable taxiing time of aircraft at airports based on space-time graph network

Abstract

The present invention relates to the field of airport intelligent operation of civil aviation information engineering, and specifically to a method for estimating the variable taxiing time of airport aircraft based on a space-time graph network, and the specific steps include: S1, acquisition and preprocessing of multi-source heterogeneous operation data of the airport scene to establish a complete taxiing process; S2, construction of a topological map of the airport scene traffic situation; S3, prediction of section speed based on a space-time graph neural network, and obtaining a section speed feature table predicted by a section-time window according to the distribution of the space-time section interval; S4, estimation of the variable taxiing time of the aircraft by integrating the section speed. The present invention performs graph learning on the airport scene traffic situation, predicts the movement speed of the aircraft in different sections based on the space-time graph neural network, and then converts it through speed synthesis and machine learning, estimates the variable taxiing time of the aircraft, provides guarantee for subsequent time node prediction and aircraft take-off and landing planning, and enables intelligent and efficient operation of the airport.

Description

Technical Field

[0001] The present invention relates to the field of airport intelligent operation of civil aviation information engineering, and in particular to a method for estimating variable taxiing time of airport aircraft based on a space-time graph network. Background Art

[0002] Variable taxi time refers to the estimated time it takes for an aircraft to taxi from a departure parking position to a takeoff runway, or from a landing runway to an arrival parking position, accurately calculated according to the specific conditions of each airport. It is an important indicator for evaluating the characteristics of airport traffic flow and operational efficiency. Variable taxi time involves ground control milestones such as wheel block entry, wheel block exit, and runway entry. Accurately predicting variable taxi time can help better plan the order of aircraft pushback, takeoff, and landing, as well as the order of runway use in air traffic control, thereby reducing congestion and delays, and improving the accuracy of predictions for various time nodes during flight. Therefore, accurately predicting the variable taxi time of aircraft at the airport is a key link in ensuring intelligent airport operations and collaborative decision-making.

[0003] At present, there are still the following problems in accurately predicting the variable taxi time of aircraft in large airports: first, the static structure of large airports is complex, and the location information such as the road section structure, parking positions and taxiway distribution, and taxi combination path is complex, which is manifested as a topological network with multiple nodes, multiple edges, and multiple road sections; second, there are many dynamic factors in the airport scene, such as the scene traffic conditions, operating time periods, dynamic flow of sections, meteorological characteristics, etc., and they are interconnected through the airport network topology structure, and the traffic interaction has a serious impact. Therefore, it is urgent to construct an accurate and effective method for estimating the variable taxi time of aircraft at the airport scene, so as to improve the utilization efficiency of airport runway resources, better balance airport capacity and demand, and ensure the safe, efficient, and sustainable development of the aviation industry. Summary of the invention

[0004] In order to overcome the defects of the existing technology, aimed at the current situation of increasing airport aircraft traffic flow, taking into account the difficulties such as the complex static structure and numerous dynamic factors of the airport scene, a variable taxiing time estimation method for aircraft at the airport scene based on speed conversion is proposed. Through the fusion prediction of spatiotemporal graph neural network and speed conversion, it aims to improve the prediction accuracy of variable taxiing time, ensure the safe and efficient operation of the airport, and promote the sustainable development of the aviation transportation industry.

[0005] The technical solutions adopted are as follows:

[0006] A method for estimating variable taxiing time of aircraft at an airport based on a spatiotemporal graph network, the specific steps are as follows:

[0007] S1: Acquisition and preprocessing of multi-source heterogeneous operation data at airports

[0008] Acquire multi-source heterogeneous original surface taxiing data sets of airport surface aircraft operations, and pre-process the original surface taxiing data sets; complete and correct the aircraft data by constructing a detection area to establish a complete taxiing process;

[0009] S2: Construction of airport traffic situation topology map

[0010] Based on the geographic information data in the original taxiing data set, the airport road network topology is optimized, and the time consumption of each aircraft section and the corresponding section is obtained through map matching to establish a taxiing topology map of the traffic on the ground;

[0011] S3: Road speed prediction based on spatiotemporal graph neural network

[0012] Construct a set of average section speed features for each section in different time windows, use a spatiotemporal graph neural network based on graph attention convolution and gated recurrent units to learn the scene traffic situation, perform time series prediction through sliding time windows, and obtain the speed prediction results of each section based on the spatiotemporal section interval distribution, that is, the section-time window prediction section speed feature table;

[0013] S4: Estimation of aircraft variable taxi time with integrated segment speed

[0014] By matching the spatiotemporal information of the section speed feature table predicted by the section-time window, speed synthesis is performed to obtain the initial predicted variable taxi time of the variable taxi time of different aircraft; the original scene taxi data set is mined, the global features affecting the variable taxi time are calculated, the initial predicted variable taxi time is used as a feature to be fused with the global features, and a random forest model is trained to obtain the final predicted variable taxi time.

[0015] Furthermore, S1 specifically includes the following steps:

[0016] S11: Acquisition and preprocessing of the original taxiing dataset:

[0017] The original taxiing data set is obtained from the air traffic control automation system and the surface radar, including aircraft operation data in csv or json format, runway usage data, weather message data, field monitoring trajectory data, and airport road network data in shp format. The aircraft operation data includes aircraft attributes, parking positions, runways, pushback time, taxi-in time, take-off and landing time, and original taxiing time information data; the field monitoring trajectory data includes the geographical, time, and motion status information data of the aircraft taxiing trajectory;

[0018] The preprocessing of the original taxiing data set includes: abnormal data removal, duplicate data screening and integration, and data format alignment and matching; after the preprocessing operation, the unique identifier of each aircraft is obtained. , complete data tables of aircraft attributes, position information, milestone times, taxi times, taxi trajectories and weather information;

[0019] S12: Establish a complete taxiing process: In order to address the problem of missing data on the aircraft's parking position, take-off and landing runways, push-out time and taxi-in time, and the problem of redundant runway occupancy time in the original taxiing time, the missing data is supplemented by demarcating the inspection area, the original taxiing time is corrected, and the complete taxiing process is established.

[0020] Furthermore, S11 specifically includes the following steps:

[0021] S111: Abnormal data removal: For abnormal records where the take-off time is earlier than the pushback time or the off-block time is earlier than the landing time, the corresponding aircraft operation data is removed; for the problem of excessive taxiing time due to confusion of connecting aircraft numbers, a time threshold is set, and the aircraft operation data whose taxiing time exceeds the time threshold is removed;

[0022] S112: Duplicate data screening and integration: For aircraft operation data from different sources, identify the aircraft number and its corresponding taxi start time, and concatenate the 8-digit date string corresponding to the taxi start time including the year, month and day with the aircraft number to construct a unique identifier ; Afterwards, based on the unique identifier To ensure consistency, various information data in aircraft operation data are screened and integrated, repeated fields are retained, and unique aircraft operation data from different sources are merged;

[0023] S113: Data format alignment and matching: For file formats, convert json files to csv files and uniformly process them according to the two-dimensional table data structure to facilitate subsequent unified matching, splicing, and filtering operations; for geographic information data, store them in longitude and latitude format, and uniformly convert the longitude and latitude information in the field monitoring trajectory data and road network geographic information data into decimal degrees, retaining two decimal places; for various time data, uniformly convert them to the East 8 time zone and store them in timestamp format to facilitate subsequent time comparison and time window matching operations; sort, filter, match, and splice various csv files based on different conditions.

[0024] Furthermore, the specific steps of completing the missing data by defining the detection area in S12 are as follows:

[0025] For the J parking stands set at the airport , define the corresponding detection area set The detection area is a circle with the parking space as the center and a radius of Circular area of

[0026] Complete the missing data of the aircraft's parking position: If the aircraft's parking position information data is missing, extract I continuous trajectory points of the aircraft , where 1≤i≤I, i is a natural number, Represents coordinate values, calculates or The distance between each parking stand in the parking stand set, the nearest parking stand is regarded as the corresponding missing parking stand, and the aircraft operation data is completed;

[0027] The missing data of pushback time of completed aircraft: For each aircraft, according to the aircraft taxi track and parking stand set The corresponding detection area set The relationship between , the specific judgment logic is as follows:

[0028]

[0029]

[0030]

[0031] in, represents the distance between a trajectory point and its adjacent trajectory points, represents the velocity of the trajectory point. When the trajectory point satisfies equations (1), (2) and (3) at the earliest, the aircraft is considered to have been pushed out at that trajectory point, and the corresponding time is taken as the push-out time of the missing aircraft. To complete;

[0032] The specific steps for correcting the original taxi time by defining a detection area are as follows: Define a rectangular detection area of ​​a preset size at the entrance of the airport runway , from I consecutive track points of each aircraft Select the first one to enter the corresponding rectangular detection area The time corresponding to the trajectory point is used as the runway usage time , the judgment logic is as follows:

[0033]

[0034] in, express The corresponding timestamp;

[0035] In order to avoid the impact of the aircraft runway redundancy time included in the original taxi time on the estimate of the variable taxi time, the original take-off time is replaced by the runway use time. For aircraft taxiing out of the airport, the updated aircraft actual variable taxi time , calculated by the following formula:

[0036] .

[0037] Furthermore, S2 specifically includes the following steps:

[0038] S21: Airport road network topology simplification: Define the airport road network as a directed graph ,in Represents a node set, which is the embodiment of intersections and connections in the scene road topology, including number and degree attributes; It is a set of directed edges, representing the road segments in the scene road topology, including the segment number, segment length, segment category, and start and end point attributes. For the original road network data, it is geographic information data stored in the shp format, where the derived segment attributes include the segment number, segment category, segment length, segment start node number, and segment end node number. The identified derived segment attributes include the number of vertices that make up the segment. and the geographical longitude and latitude of each vertex, that is, the total number of road network segments is N;

[0039] S22: Extraction of time-consuming sequence of road sections based on graph association: The original field monitoring trajectory data includes the timestamp, longitude and latitude, ground speed, and turning angle information corresponding to a series of continuous trajectory points during the taxiing process of the aircraft. The continuous trajectory points are graph-associated with the airport road network through a map matching algorithm, thereby outputting the maximum probability road section sequence of the aircraft, and combining the sequence of the continuous trajectory points to determine the road section corresponding to each trajectory point;

[0040] Calculation of time consumption for a route segment: traverse the continuous trajectory points of the aircraft one by one, and the continuous trajectory points of the aircraft The corresponding timestamp sequence is , the time difference between adjacent trajectory points is ;

[0041] (1) When the trajectory point and trajectory points There is only one crossing section between them, and the crossing section is not a trajectory point and trajectory points The time difference is Assigned to the crossing section, and mark the starting time of the crossing section as , the end time of the crossing section is ;

[0042] (2) When the trajectory point and trajectory points There is only one crossing section between them, and the trajectory point If the corresponding road section is the same as the crossing road section, the track point is skipped. Continue to determine the trajectory point Whether the corresponding road section is the same as the crossing road section, until the corresponding road section and the crossing road section are different trajectory points are obtained , take the trajectory point With trajectory points The time difference is assigned to the crossing section, and the starting time of the crossing section is marked as , the end time of the crossing section is ;

[0043] (3) When the trajectory point and trajectory points When there are multiple crossing sections, the time difference is allocated according to the ratio relationship of the multiple crossing sections. The specific steps are as follows:

[0044] Apply Haversine formula to calculate trajectory points With trajectory points The total distance between , the total distance is the sum of the distances of the trajectory points on all the crossing sections; the statistical trajectory points With trajectory points Between The distance across the road , q is the point on the trajectory With trajectory points The number of random crossing sections between ;

[0045] According to the track point With trajectory points The ratio of the distance of each crossing section to the total distance is calculated by Assign to track point With trajectory points On each crossing section between The time taken to cross a road section is:

[0046]

[0047] Mark The start time of the crossing section , No. End time of the crossing section ;

[0048] Initialize the taxi path information table to store the list of segment numbers of each aircraft during the taxiing process and the segment time, start time, and end time of each corresponding segment.

[0049] Furthermore, the airport road network topology simplification in S21 specifically includes:

[0050] S211: For the road segment sequence in the road network , and the node numbers of the corresponding road section starting points are respectively represented as , the node numbers of the end points of the road sections are represented as , the length of the road section is expressed as , the number of control points in the section interval passed by the road section is expressed as ; Set the length threshold ,for Less than Section To delete; The adjacent road segments are numbered as , where 1≤p≤N, p is a natural number. The starting point of the segment is the segment The end point of the road segment is The starting point of the segment is changed to the segment The starting point of the road segment; if the road segment The end point of the segment is the segment The starting point of the road segment is The end point of the segment is changed to the segment The starting point of the road section is determined by the node, thereby ensuring the connectivity of the road network, that is:

[0051] ;

[0052] S212: Filling section The starting point and end point of the segment, through the segment The number of control points in the included road section Determine the line shape and fill it. If the number of control points is less than 4, it is determined to be a straight line, and the road sections are connected with a straight line. The starting point and end point of the road segment; if the number of control points is greater than 4, it is determined to be a curve, and the interpolation density of different segments is controlled. The cubic spline interpolation method is used to generate a smooth curve to connect the road segments. The starting point and end point of the road segment;

[0053] S213: recalculating the length of each road section in the road network after the airport road network topology is simplified, and storing the updated road section length, the node number of the road section starting point, the node number of the road section end point, and the geographic information data of the road section starting point and the road section end point.

[0054] Furthermore, S3 specifically includes the following steps:

[0055] S31: embedding the road segment speed: based on the glide path information table obtained in S22, obtaining a road segment-time window matrix, and constructing a road segment-time window road segment speed feature table;

[0056] S311: Generate an empty matrix with the horizontal axis being the road section number and the vertical axis being the time window. Fill in the horizontal axis from small to large; for the time range covered by the original scene taxiing data set, take 5 minutes as the window size, generate a time window sequence, and fill in the vertical axis; for each position in the matrix, initialize a speed array;

[0057] S312: For road sections , identify the section passing through the glide path information table Aircraft operation data;

[0058] S313: Based on the recognition result of S312, extract the time sequence of the section obtained in S22 and the time of the aircraft on the section Start time , End time , Road section time ;

[0059] S314: For the extracted start time and end time, traverse the vertical axis time window. If the start time or end time is within a certain time window, calculate the aircraft's time in the route section. The road speed :

[0060]

[0061] The road speed Fill in the corresponding positions in the matrix generated by S311;

[0062] S315: Repeat S312-S314 until the road segment sequence is traversed , get the road section-time window matrix;

[0063] S316: extract the segment speed array of each position in the segment-time window matrix, and calculate the average segment speed in the segment speed array; for the segment or time window not covered by the original scene taxiing data set, fill the corresponding position with 0; in the segment-time window matrix, if there is a segment whose segment speed is 0 in all time windows, delete this segment; for the position of 0 in the remaining segment, fill the average segment speed of this segment in the same time window or the six time windows before and after on other dates, and finally obtain the segment-time window segment speed feature table distributed according to the time-space information;

[0064] S32: Construction of spatiotemporal graph neural network: The spatiotemporal graph neural network model adopted consists of a spatiotemporal graph convolution module composed of a spatial block based on graph attention convolution and a temporal block based on gated recurrent unit GRU, which is finally output through an activation function and a linear prediction layer;

[0065] In the spatial block, a graph attention network that introduces the attention mechanism into the graph structure is used as the convolution layer. The input of the convolution layer is a graph structure node feature set containing the road speed of each road section. Each road section in the road network is used as a node. The node features are updated through the attention mechanism, and a multi-head attention mechanism is introduced at the same time.

[0066] In the time block, GRU is used as the basic unit for time window sequence processing;

[0067] Finally, the prediction output is achieved by introducing a ReLU function with nonlinear and sparse activation and a linear prediction layer for dimensionality reduction and linear transformation;

[0068] S33: training the spatiotemporal graph neural network and predicting the speed of each coasting section according to the spatiotemporal section interval distribution.

[0069] Furthermore, in S32, the graph attention convolutional network completes its functions through two functions: the node message function and the message update function, where the node message function calculates the unnormalized attention coefficient between each pair of nodes and their neighboring nodes. , and apply Function activation; the message update function uses softmax to normalize the attention coefficient to obtain the attention weight of the graph node corresponding to each road section.

[0070] Furthermore, in S32, the time block in the spatiotemporal graph convolution module also serves as a spatiotemporal graph learning module, which is composed of a two-layer GRU encoder and a one-layer GRU decoder, wherein the GRU encoder iteratively processes the input road section speed sequence, calculates the update gate, reset gate and candidate hidden state, and encodes the road section speed sequence into a fixed-size context vector, i.e., the current hidden state; in the two-layer GRU encoder, the result of the first layer encoding is used as the input of the second layer encoder, so as to extract high-dimensional features and encode them, perform feature time series modeling, capture the long-distance time dependency between features, consider the time correlation of the road section speed in different time windows, and finally generate a fixed-length vector as the feature representation; the GRU decoder is composed of a GRU layer, which uses the last hidden state of the GRU encoder as the initial hidden state at each time step to generate a predicted output.

[0071] Further, S4 specifically includes the following steps:

[0072] S41: Speed-time conversion: synthesizing the predicted variable coasting time based on the speed characteristic table of the segment-time window prediction segment obtained in S33:

[0073] S411: Determine the time window to which the aircraft to be predicted belongs: For the aircraft to be predicted, use the unique identifier Match the corresponding aircraft operation data and extract the aircraft's estimated taxi start time and the expected segment sequence , M refers to the total number of sections that the aircraft is expected to pass through, and determines The time window to which it belongs;

[0074] S412: Extract the historical section speed as input, use the trained graph neural network model to predict the speed of each section within the taxiing time coverage of the aircraft to be predicted; S413: Calculate the estimated section time of each section based on the estimated taxiing path of the aircraft to be predicted: Read the estimated section sequence of the aircraft to be predicted And the length sequence corresponding to each section ; For the first section passed during the taxiing , select the predicted section speed corresponding to the starting time window, and calculate the estimated section time of the section ; For the next section , first calculate the aircraft's expected entry section Start time ,judge The time window to which it belongs, and then pass through the road section within this time window Corresponding predicted road speed, calculate the estimated road time , determine the time window to which it belongs, until the traversal of the expected route sequence of the aircraft to be predicted is completed, and the expected route time of each route is obtained;

[0075] S414: Obtaining an initial predicted variable taxiing time by accumulating the estimated segment time: For the aircraft to be predicted, read the estimated segment time of each segment to be passed during taxiing, and obtain the initial predicted variable taxiing time by accumulating ;

[0076] S415: According to S411-S414, traverse each aircraft to be predicted in the set of aircraft to be predicted to obtain an initial predicted variable taxiing time of each aircraft to be predicted;

[0077] S42: The final prediction result of the variable taxiing time is obtained based on the variable taxiing time fusion and random forest model training;

[0078] S421: Based on the complete data table obtained in S11, feature mining is performed to select some motion features, meteorological features and historical spatiotemporal features that are most strongly correlated with the taxiing time as the global feature set; wherein the mined motion features include the total taxiing turning angle and the total taxiing distance obtained by accumulating the field monitoring trajectory data; the meteorological features include visibility; the historical spatiotemporal features include the average taxiing time of the aircraft within 15 minutes before the aircraft starts taxiing, the total number of taxiing aircraft within the first 15 minutes, the parking stand corresponding to the parking stand, and the time period number of the taxiing start time within one day; the parking stand group and the time period number of the taxiing start time within one day are uniquely encoded;

[0079] S422: For the set of aircraft to be predicted, obtain the global feature set corresponding to all aircraft to be predicted , and associate the global features with the initial predicted variable taxi time through a unique identifier Match, concatenate and fuse global features to obtain the total sliding feature set after encoding representation ; The actual variable taxiing time of all aircraft to be predicted obtained after S12 processing is used as the target true value set ;

[0080] S423: The sliding total feature set The total sliding feature set is divided into test set, validation set and training set according to 3:1:1, and the random forest model is used for prediction. And the target truth value set The random forest model is input and the final predicted variable taxiing time is obtained by training the random forest model.

[0081] The advantages and positive effects of the variable taxiing time prediction method provided by the present invention are:

[0082] 1. Taking into account the static topological correlation and dynamic traffic interconnectivity of the airport surface road network, the spatiotemporal graph network and graph attention convolution, which are relatively rare in airport surface traffic prediction, are selected to complete the road section speed prediction task. In the subsequent speed conversion into time estimation process, machine learning methods are used for fusion prediction, thereby improving the prediction accuracy compared with traditional queuing theory or machine learning methods.

[0083] 2. Taking into account the diversity of airport surface traffic prediction problems, the prediction of variable taxi time is divided into two parts: the prediction of section speed in different time periods and sections, and the calculation of variable taxi time based on speed conversion. This can not only complete the section speed prediction task, but also estimate the variable taxi time; not only can the entire taxi time be predicted, but also segmented prediction can be performed, which broadens the research depth and breadth of the scope of application. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0085] Figure 1 The present invention is a flow chart of the method for estimating variable taxiing time of aircraft at an airport based on a space-time graph network.

[0086] Figure 2 This is a schematic diagram of the spatiotemporal graph neural network used in the present invention; DETAILED DESCRIPTION

[0087] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0088] The present invention provides a method for estimating the variable taxiing time of an airport aircraft based on a space-time graph network, the specific steps are as follows: Figure 1 As shown, including the following:

[0089] 1. Step 1: Acquisition and preprocessing of multi-source heterogeneous operation data at the airport surface to establish a complete taxiing process.

[0090] 1. Acquisition and preprocessing of the original scene taxiing dataset.

[0091] The original taxiing data set is obtained from sources such as the air traffic control automation system and the radar. Aircraft operation data in csv or json format, runway usage data, weather message data, and field monitoring trajectory data; airport road network data in shp format, etc., require a lot of format conversion, cleaning, matching and other operations. Aircraft operation data includes aircraft attributes, parking positions, runways, pushback time, taxi-in time, take-off and landing time, and original taxiing time; field monitoring trajectory data includes information data such as the geography, time, and motion status of the aircraft taxiing trajectory.

[0092] The preprocessing of the original taxiing dataset includes:

[0093] (1) Abnormal data elimination: For abnormal records where the take-off time is earlier than the pushback time or the off-block time is earlier than the landing time, the corresponding aircraft operation data shall be eliminated. For taxi time, the taxi time may be too long due to confusion of connecting aircraft numbers. A time threshold (for example, 40 minutes) shall be set to eliminate the aircraft operation data whose taxi time exceeds the time threshold.

[0094] (2) Duplicate data screening and integration: Construct unique identifiers for aircraft operation data from different sources. , identify the aircraft number and its corresponding taxi start time, and concatenate the 8-digit date string corresponding to the taxi start time including the year, month and day with the aircraft number, that is, ; Afterwards, based on the unique identifier In order to ensure consistency, various information data in the aircraft operation data are screened and integrated, repeated field columns are retained, and unique aircraft operation data from different sources are merged. Unique aircraft operation data refers to the "unique" features contained in each aircraft operation data. For example, if the two feature information of parking positions and runways only appear in the first aircraft operation data, they will be merged according to the unique identifier id.

[0095] (3) Data format alignment and matching: For file format, convert json files to csv files and uniformly process them according to the two-dimensional table data structure to facilitate subsequent unified matching, splicing, and filtering operations; for geographic information data, store them in longitude and latitude format. For the longitude and latitude information in the field monitoring trajectory data and the road network geographic information data, it is necessary to complete the unification of the degree, minute, and second longitude and latitude format and the decimal longitude and latitude format, and uniformly convert them into decimal degrees, retaining two decimal places; for time data, it is necessary to complete time zone conversion, timestamp format conversion and other operations, and uniformly convert them to the East 8 time zone, and store them in the timestamp format: YYYY-MM-DD HH:MI:SS, so as to facilitate subsequent time comparison, time window matching and other operations; finally, various csv files need to be sorted, filtered, matched, spliced ​​and other operations based on different conditions, such as dividing aircraft taxiing information into outbound aircraft and inbound aircraft according to operation; according to the unique identifier Sort and divide the aircraft into different time periods; perform data splicing and quantity statistics according to the nature of aircraft arrival and departure, taxi start and end time periods, and taxi time duration.

[0096] After preprocessing, the unique identifier of each aircraft within a certain time range is obtained. A complete data table of its corresponding aircraft attributes, location information, milestone time points, taxi time, taxi trajectory, and weather information.

[0097] 2. Construction of taxiing process based on detection area identification. In view of the missing data of aircraft parking positions, runways, push-out time and taxi-in time, and the redundant runway occupancy time in the original taxiing time, the complete taxiing process is constructed by demarcating the detection area, and some missing data are supplemented and the taxiing time is corrected.

[0098] For J airport parking stands , and its corresponding detection area set The parking space is the center of the circle and the radius is circular area.

[0099] To solve the problem of missing data of aircraft parking positions, the specific method is as follows:

[0100] First, the set of aircraft to be predicted is traversed one by one. If the parking position information data of an aircraft is missing, I continuous trajectory points of the aircraft are extracted. , where 1≤i≤I, i is a natural number, calculate or The distance between each parking stand in the parking stand set, the nearest parking stand is regarded as the corresponding parking stand, and the aircraft operation data is completed.

[0101] To complete the launch time of missing aircraft For each aircraft, according to the aircraft taxiing trajectory and parking space set The corresponding detection area set The relationship between , the specific judgment logic is as follows:

[0102]

[0103]

[0104]

[0105] in, represents the distance between a trajectory point and its adjacent trajectory points, Represents the ground speed of the track point. That is, for the earliest simultaneous meeting of the following three conditions: a. the aircraft track point is within the detection area; b. there is a deviation between the positions of two consecutive track points of the aircraft; c. the speeds of the two consecutive track points are both higher than 0.5m / s, the position of the aircraft at the track point is considered to have been pushed out, and its corresponding time is regarded as the missing aircraft push-out time to be completed.

[0106] To correct the original taxi time of the aircraft, the specific method is as follows:

[0107] A rectangular detection area Z of appropriate size is defined at the entrance of the airport runway, and the detection area Z is obtained from I consecutive trajectory points of each aircraft. Select the first one to enter the corresponding rectangular detection area The time corresponding to the trajectory point is used as the runway usage time , the judgment logic is as follows:

[0108]

[0109] in, express The corresponding timestamp;

[0110] In order to avoid the impact of the aircraft runway redundant time included in the original data taxi time on the variable taxi estimate, the runway time is used to replace the original takeoff time. Therefore, for the aircraft taxiing out of the airport, the updated aircraft actual variable taxi time is recorded as , calculated by the following formula:

[0111]

[0112] in, , They are respectively the aircraft's runway usage time and pushback time. The difference between the two represents the aircraft's variable taxiing time within the airport surface.

[0113] 2. Step 2: Construction of Airport Traffic Situation Topology Map

[0114] 1. Simplify the airport road network topology.

[0115] Define the airport road network as a directed graph .in Represents a set of nodes, which are the embodiment of intersections and connections in the scene road topology, and contain attributes such as number and degree; It is a set of directed edges, representing the directed road segments in the scene road topology, including the segment number, segment length, segment type, start and end points, etc. For the original road network data, it is stored in the shp geographic information data format, where the exportable segment attributes include the segment number, segment type, segment length, segment start node number, segment end node number, and the identifiable derived attributes include the number of vertices that make up the segment. And the geographical longitude and latitude of each vertex, that is, the total number of road sections in the road network is N.

[0116] To address the problem of too many fragmented sections in the airport road network, the road network is simplified and corrected. , the node numbers corresponding to the starting points of the road sections are represented as , the node numbers of the end points of the road sections are represented as , the length of the road section is expressed as , the number of control points in the section interval passed by the road section is expressed as . Set the length threshold ,for Less than Section , delete it. The number of the adjacent road segment is expressed as ,like The starting point of the road segment is The end point of the road segment is The starting point of the segment is changed to The starting point of the road section; if The end point of the road segment is The starting point of the road segment is The end point of the segment is changed to The starting point of the road section is determined to ensure the connectivity of the road network, where 1≤n≤N, 1≤p≤N, n, p are natural numbers, that is:

[0117] Editing a section After the start and end points of the segment are set, the shape of the changed segment needs to be filled. The number of control points in the included road section Determine the line shape and fill it. If the number of control points is less than 4, it can be determined as a straight line, and the road segments are connected with a straight line. The starting point and end point of the road segment; if the number of control points is greater than 4, it can be determined as a curve, and the interpolation density of different segments is controlled. The cubic spline interpolation method is used to generate a smooth curve to connect the road segments. The starting point and end point of the road segment;

[0118] Finally, the length of each road section in the simplified road network topology is recalculated, and the updated section length, node number of the section start point, node number of the section end point, and geographic information data of the section start point and the section end point are stored. 2. Extraction of time-consuming sequence of sections based on graph association.

[0119] The original field monitoring trajectory data includes the timestamp, longitude and latitude, ground speed, turning angle and other information corresponding to a series of continuous trajectory points during the taxiing process of the aircraft. Through the map matching algorithm, the continuous trajectory points can be associated with the airport road network, thereby outputting the maximum probability section sequence of the aircraft. The present invention adopts the hidden Markov model matching algorithm, defines the state space and observation space, calculates the initial probability, state transition probability, and observation probability, and uses the Viterbi algorithm to obtain the maximum probability section sequence, that is, the most likely taxiing path series. On this basis, the sequence of continuous trajectory points and the maximum probability section sequence are combined to determine the section corresponding to each trajectory point.

[0120] Then the time consumption of the road section is calculated, and the trajectory points are traversed one by one. , the corresponding timestamp sequence is , the time difference between adjacent trajectory points is .

[0121] When the trajectory point and trajectory points There is only one crossing section between them, and the crossing section is not a trajectory point and trajectory points The corresponding section directly converts the time difference Assigned to the segment and mark the start time of crossing the segment as , the end time of the crossing section is When the trajectory point and trajectory points There is only one crossing section between them, and the trajectory point If the corresponding road section is the same as the crossing road section, the track point is skipped. Continue to determine the trajectory point Whether the corresponding road section is the same as the crossing road section. And so on, until the corresponding road section is different from the crossing road section. , take the trajectory point With trajectory points The time difference is assigned to the crossing section, and the starting time of the crossing section is marked as , the end time of the crossing section is ; If the trajectory point and trajectory points When there are multiple crossing sections, the time difference is allocated according to the ratio relationship of the multiple crossing sections. The specific steps are as follows:

[0122] Apply Haversine formula to calculate trajectory points With trajectory points The total distance between , the total distance is the sum of the distances of the trajectory points on all crossing sections.

[0123] Statistical track points With trajectory points The distance between the qth crossing sections ,in Represents trajectory points With trajectory points The first A crossing section.

[0124] According to the ratio of the distance of each crossing section to the total distance, the time difference Specifically, for the first The crossing section has a section time of:

[0125]

[0126] The starting time for entering the qth crossing section is , the end time of the qth crossing section .

[0127] Initialize the taxi path information table to store the list of segment numbers of each aircraft during taxiing and the segment time, start time, and end time of each segment.

[0128] Step 3: Road speed prediction based on spatiotemporal graph learning neural network

[0129] 1. Section speed embedding. Based on the obtained sliding path information table, obtain the section-time window matrix and construct the section-time window section speed feature. The steps are as follows:

[0130] (1) Generate an empty matrix with the horizontal axis as the road segment number and the vertical axis as the time window. Fill in the horizontal axis from small to large; for the time range covered by the original scene taxiing data set, use 5 minutes as the window size to generate a time window sequence and fill in the vertical axis. For each position in the matrix, initialize a road segment speed array.

[0131] (2) For road sections , identify the section passing through this route in the taxi path information table aircraft operation data.

[0132] (3) Based on the recognition results of the previous step, extract the aircraft’s The start time of the segment , End time , Road section time .

[0133] (4) For the extracted start time and end time, traverse the vertical axis time window. If the start time or end time is within a certain time window, calculate the aircraft's time in that section. The road speed :

[0134]

[0135] The road speed Fill in the corresponding positions in the previous matrix.

[0136] Repeat the above steps until the road segment sequence is traversed Get the road section-time window matrix;

[0137] Then, for the segment-time window matrix, the segment speed array of each position is extracted, and the average segment speed is calculated as the new value to fill the matrix; for the segments or time windows not covered by the original scene taxiing data set, the corresponding positions are first filled with 0; in the segment-time window matrix, if there is a segment whose segment speed is 0 in all time windows, it indicates that the possibility of the taxiing path passing through this segment is very low, and this segment can be deleted; for the positions with 0 in the remaining segments, the average segment speed of this segment in the same time window on other dates or in the six time windows before and after is filled, and finally a segment-time window segment speed feature table distributed according to the spatiotemporal information is obtained.

[0138] 2. Construction of spatiotemporal graph neural network, such as Figure 2 shown.

[0139] The spatiotemporal graph neural network can capture the data similarity of similar time windows in the airport road network and the traffic interaction of each road section, and can more accurately reflect the traffic conditions of a specific road section corresponding to a specific time, and obtain a more accurate road section speed prediction result. The spatiotemporal graph neural network used in the present invention is composed of a spatiotemporal graph convolution module consisting of a spatial block based on graph attention convolution and a temporal block based on a gated recurrent unit GRU (Gated Recurrent Unit), and is finally output through an activation function and a linear prediction layer.

[0140] In the spatial block, a graph attention network that introduces the attention mechanism into the graph structure is used as the convolution layer. The input of the convolution layer is a graph structure node feature set containing the segment speed of each segment, where each segment in the road network is used as a node, and the node features are updated through the attention mechanism. At the same time, a multi-head attention mechanism is introduced to help the model understand the topological structure and relationship between nodes, so that the degree of interaction between each segment in the airport road network is taken into account during the learning process, and the impact of the geographical structure of the road network on the taxiing time of aircraft is better captured.

[0141] In the graph attention convolutional network adopted by the present invention, the functions are completed by two functions: the node message function and the message update function. Among them, the node message function calculates the unnormalized attention coefficient between each pair of nodes and their neighboring nodes. , and apply Function activation. The message update function uses softmax to normalize the attention coefficient and obtain the attention weight of the graph node corresponding to each road segment.

[0142] In the specific implementation, the GAT layer (GATConv) in PyTorch Geometric is used to complete the graph attention convolution operation. From the perspective of graph structure data, the dynamic correlation of the traffic situation of relevant sections in the airport road network is captured through the multi-head attention mechanism. The attention coefficient is calculated to measure the degree of interaction between each section, especially focusing on the interaction between each section and its most relevant section, so as to obtain the section speed characteristics that take into account the static network structure and dynamic historical information.

[0143] The GAT layer (GATConv) in PyTorch Geometric is used to complete the graph attention convolution operation. From a graph structure data perspective, the dynamic correlation of the traffic situation of relevant sections in the airport road network is captured through a multi-head attention mechanism. The attention coefficient is calculated to measure the degree of interaction between each section, especially focusing on the interaction between each section and its most relevant section, so as to obtain the section speed characteristics that consider the static network structure and dynamic historical information.

[0144] In the time block, GRU is used as the basic unit for time window sequence processing. Since GRU has a strong memory for sequence data, it processes data in different time steps separately and can capture long-distance temporal correlations between data. For airport traffic conditions, the historical information of adjacent time steps is very useful for reference. Using GRU can more effectively capture the correlation of traffic conditions in similar time windows.

[0145] The GRU time block in the spatiotemporal graph convolution module constructed by the present invention also serves as a spatiotemporal graph learning module, which is composed of a double-layer GRU encoder and a layer of GRU decoder. The GRU encoder iteratively processes the input road section speed sequence, and encodes the road section speed sequence into a fixed-size context vector, i.e., the current hidden state, by calculating the update gate, the reset gate, and the candidate hidden state. In the double-layer GRU encoder, the result after encoding in the first layer is used as the input of the second layer encoder, so as to extract higher-dimensional features and encode them, perform feature time series modeling, capture the long-distance time dependency between features, consider the time correlation of the road section speed in different time windows, and finally generate a fixed-length vector as the feature representation. The GRU decoder consists of a GRU layer, but usually uses the last hidden state of the GRU encoder as the initial hidden state at each time step to generate the prediction output.

[0146] Finally, the prediction output is obtained by introducing the ReLU function with nonlinear and sparse activation and the linear prediction layer with dimensionality reduction and linear transformation to obtain the final output result.

[0147] 3. Road section speed prediction training process.

[0148] Based on the airport road network directed graph G(V,E) and the section speed feature table of the section-time window, a spatiotemporal graph neural network is used to traverse all time windows. The training process includes four steps: forward propagation, loss function calculation, back propagation, and network parameter update. The final output is the aircraft section speed prediction result under different section-time window conditions with the minimum loss, that is, the future section-time window predicted section speed table.

[0149] First, define the forward propagation function to complete the graph convolution, vector dimension matching, encoding, decoding, prediction, dimension sorting and data conversion based on the graph attention layer. In the forward propagation process, traverse all the time steps in the road section-time window road section speed feature table, and the window sliding step of each time window is ,in, The number of historical time windows required for prediction is equal to The number of time windows that the prediction can cover The sum of , and generate the corresponding graph structure data; in this process, the entire time series data is divided into multiple sliding windows, each sliding window contains Step historical data and The prediction data of the first step is then used for subsequent graph convolution and encoding and decoding operations to achieve time series prediction.

[0150] After that, the loss function of the model is calculated, and the gradient of each network parameter with respect to the loss is calculated through back propagation.

[0151] Finally, the Adam optimizer is used to update and optimize parameters by adjusting and correcting the adaptive learning rate.

[0152] During the training process, the initial parameters such as the hidden state value of the GRU in the model, the learning rate of the Adam optimizer, and the number of training cycles epochs are first defined in the main function, and the section-time window section speed feature table is divided into a training set, a test set, and a validation set according to the proportion. After that, in each iteration, the model performs a parameter update through the above steps. Selecting an appropriate number of training cycles can make the final prediction result close to the optimal convergence. Through model training and optimization, a model with better performance can be finally obtained, that is, a two-dimensional table of aircraft section speed prediction results with the smallest loss.

[0153] Step 4: Estimation of aircraft variable taxi time with integrated segment speed

[0154] 1. Speed-time conversion. Since the airport road network has a relatively complex topological connection relationship, and for aircraft, before pushout or taxiing, the parking positions and runways have been deployed, and the pushout and taxiing paths have been determined, the predicted variable taxiing time can be synthesized based on the road section speed predicted by the spatiotemporal graph neural network.

[0155] Based on the future section-time window predicted section speed table, the steps for synthesizing and converting the section time prediction results are as follows:

[0156] (1) Determine the time window to which the aircraft to be predicted belongs.

[0157] For the set of aircraft to be predicted, a unique identifier is Match the corresponding aircraft operation data and extract the aircraft's estimated taxi start time and the expected segment sequence , M refers to the total number of sections that the aircraft is expected to pass through, and determines The time window.

[0158] (2) Extract historical section speeds as input, and use the trained graph neural network to predict the speed of each section within the taxiing time coverage of the aircraft to be predicted. The specific method is as follows: for the set of aircraft to be predicted, identify the time window in which the taxiing start time of each aircraft is located, and use the historical section speeds of the 20 time windows before this time window as input. Through the trained graph neural network, obtain the predicted section speed of each section in the subsequent 6 time windows of the first 20 time windows.

[0159] (3) Based on the estimated taxi path of the aircraft to be predicted, calculate the estimated time for each segment.

[0160] For the aircraft to be predicted, read its expected route sequence And the length sequence corresponding to each section For the first section of the road you pass during the taxi ride , select the predicted section speed of the section corresponding to the starting time window, and calculate the estimated time consumption of the section ; For the next section , first calculate the aircraft's expected entry section Start time ,judge The time window to which it belongs, and then pass through the road section within this time window Corresponding predicted road speed, calculate the estimated time , determine the time window to which it belongs; and so on, until the traversal of the expected route sequence of the aircraft to be predicted is completed, and the expected time consumption of each route is obtained.

[0161] (4) The initial predicted variable taxiing time is obtained by accumulating the estimated time.

[0162] For the predicted aircraft, read the estimated time of each section to be passed during taxiing, and get the initial predicted variable taxiing time by accumulation .

[0163] (5) According to the above steps (1) to (4), each aircraft in the set of aircraft to be predicted is traversed to obtain the initial predicted variable taxi time of each aircraft to be predicted.

[0164] For the initial predicted variable taxiing time, it is not directly compared with the actual taxiing time due to reasons such as the possible missing of road sections in map matching and the error caused by mean filling. Instead, it is used as a feature reflecting the spatiotemporal situation of the scene traffic to be integrated with machine learning to obtain the final predicted variable taxiing time.

[0165] 2. Feature fusion and training

[0166] Based on the complete data table, feature mining is performed to select some motion features, meteorological features, historical spatiotemporal features, etc. that are most strongly correlated with taxiing time as the global feature set. Among them, the mined motion features include the total taxiing angle and total taxiing distance obtained by accumulating trajectory data; meteorological data include visibility; historical spatiotemporal features include the average taxiing time of aircraft within 15 minutes before the aircraft starts taxiing, the total number of taxiing aircraft within the first 15 minutes, the corresponding parking stand group, and the time period number of the taxiing start time within a day.

[0167] For the stand group and taxi start time within a day, the row number is one-hot encoded. Then, for the set of aircraft to be predicted, the global feature set corresponding to all aircraft is calculated , and associate the global features with the initial predicted variable taxi time through a unique identifier Match, concatenate and fuse global features to obtain the total sliding feature set after encoding representation At the same time, the actual variable taxi time of all aircraft to be predicted is taken as the target true value set .

[0168] The total sliding feature set The data was divided into test set, validation set and training set according to the ratio of 3:1:1, and the random forest model was selected for further prediction. The random forest model is an ensemble learning algorithm often used for classification and regression problems. It consists of multiple decision trees generated by learning random subsamples of the training data, including the steps of constructing random samples, constructing decision trees, and ensemble prediction. The mean squared error (MSE) was used as the splitting criterion in the decision tree construction, and the splitting point that minimizes the mean squared error was selected.

[0169] Based on the random forest model, the initial predicted variable taxi time of the aircraft to be predicted is combined with the global feature set The total sliding feature set obtained by splicing and the target truth set As input, the final predicted variable taxiing time is obtained through model training. During the training process, Bayesian optimization can be used to tune parameters, such as the number of trees, the maximum depth of the tree, and the maximum number of features considered when splitting.

[0170] The estimated time consumption during taxiing obtained by the present invention through space-time graph prediction, speed conversion, and fusion prediction integrates the spatiotemporal dependence of variable taxiing time and dynamic and static characteristics of aircraft, airport scene environment, taxiing time history information and other factors captured by the space-time graph neural network. It has high accuracy and stability, which helps to reduce congestion and delays, optimize aircraft take-off and landing sequence planning, and ensure efficient and intelligent operation of large airports.

[0171] Finally, it should be noted that the above is only a specific implementation of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can still modify the technical solutions recorded in the above embodiments within the technical scope disclosed by the present invention, or replace some of the technical features therein with equivalents; and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A method for estimating variable taxiing time of aircraft at an airport based on a spatiotemporal graph network, characterized in that: The specific steps are as follows: S1: Acquisition and preprocessing of multi-source heterogeneous operation data at airports Acquire multi-source heterogeneous original surface taxiing data sets of airport surface aircraft operations, and pre-process the original surface taxiing data sets; complete and correct the aircraft data by constructing a detection area to establish a complete taxiing process; S2: Construction of airport traffic situation topology map Based on the geographic information data in the original taxiing data set, the airport road network topology is optimized, and the time consumption of each aircraft section and the corresponding section is obtained through map matching to establish a taxiing topology map of the traffic on the ground; S3: Road speed prediction based on spatiotemporal graph neural network Construct a set of average section speed features for each section in different time windows, use a spatiotemporal graph neural network based on graph attention convolution and gated recurrent units to learn the scene traffic situation, perform time series prediction through sliding time windows, and obtain the speed prediction results of each section based on the spatiotemporal section interval distribution, that is, the section-time window prediction section speed feature table; S4: Estimation of aircraft variable taxi time with integrated segment speed By matching the spatiotemporal information of the section speed feature table predicted by the section-time window, speed synthesis is performed to obtain the initial predicted variable taxi time of the variable taxi time of different aircraft; the original scene taxi data set is mined, the global features affecting the variable taxi time are calculated, the initial predicted variable taxi time is used as a feature to be fused with the global features, and a random forest model is trained to obtain the final predicted variable taxi time.

2. The method for estimating variable taxiing time of aircraft at an airport based on a spatiotemporal graph network according to claim 1, characterized in that: S1 specifically includes the following steps: S11: Acquisition and preprocessing of the original taxiing dataset: The original taxiing data set is obtained from the air traffic control automation system and the surface radar, including aircraft operation data in csv or json format, runway usage data, weather message data, field monitoring trajectory data, and airport road network data in shp format. The aircraft operation data includes aircraft attributes, parking positions, runways, pushback time, taxi-in time, take-off and landing time, and original taxiing time information data; the field monitoring trajectory data includes the geographical, time, and motion status information data of the aircraft taxiing trajectory; The preprocessing of the original taxiing data set includes: abnormal data removal, duplicate data screening and integration, and data format alignment and matching; after the preprocessing operation, the unique identifier of each aircraft is obtained. , complete data tables of aircraft attributes, position information, milestone times, taxi times, taxi trajectories and weather information; S12: Establish a complete taxiing process: In order to address the problem of missing data on the aircraft's parking position, take-off and landing runways, push-out time and taxi-in time, and the problem of redundant runway occupancy time in the original taxiing time, the missing data is supplemented by demarcating the inspection area, the original taxiing time is corrected, and the complete taxiing process is established.

3. The method for estimating variable taxiing time of aircraft at an airport based on a spatiotemporal graph network according to claim 2, characterized in that: S11 specifically includes the following steps: S111: Abnormal data removal: For abnormal records where the take-off time is earlier than the pushback time or the off-block time is earlier than the landing time, the corresponding aircraft operation data is removed; for the problem of excessive taxiing time due to confusion of connecting aircraft numbers, a time threshold is set, and the aircraft operation data whose taxiing time exceeds the time threshold is removed; S112: Duplicate data screening and integration: For aircraft operation data from different sources, identify the aircraft number and its corresponding taxi start time, and concatenate the 8-digit date string corresponding to the taxi start time including the year, month and day with the aircraft number to construct a unique identifier ; Afterwards, based on the unique identifier To ensure consistency, various information data in aircraft operation data are screened and integrated, repeated fields are retained, and unique aircraft operation data from different sources are merged; S113: Data format alignment and matching: For file formats, convert json files to csv files and uniformly process them according to the two-dimensional table data structure to facilitate subsequent unified matching, splicing, and filtering operations; for geographic information data, store them in longitude and latitude format, and uniformly convert the longitude and latitude information in the field monitoring trajectory data and road network geographic information data into decimal degrees, retaining two decimal places; for various time data, uniformly convert them to the East 8 time zone and store them in timestamp format to facilitate subsequent time comparison and time window matching operations; sort, filter, match, and splice various csv files based on different conditions.

4. The method for estimating variable taxiing time of aircraft at an airport based on a spatiotemporal graph network according to claim 3, characterized in that: The specific steps for completing missing data by defining the detection area in S12 are as follows: For the J parking stands set at the airport , define the corresponding detection area set The detection area is a circle with the parking space as the center and a radius of Circular area of Complete the missing data of the aircraft's parking position: If the aircraft's parking position information data is missing, extract I continuous trajectory points of the aircraft , where 1≤i≤I, i is a natural number, Represents coordinate values, calculates or The distance between each parking stand in the parking stand set, the nearest parking stand is regarded as the corresponding missing parking stand, and the aircraft operation data is completed; The missing data of pushback time of completed aircraft: For each aircraft, according to the aircraft taxi track and parking stand set The corresponding detection area set The relationship between , the specific judgment logic is as follows: in, represents the distance between a trajectory point and its adjacent trajectory points, represents the velocity of the trajectory point. When the trajectory point satisfies equations (1), (2) and (3) at the earliest, the aircraft is considered to have been pushed out at that trajectory point, and the corresponding time is taken as the push-out time of the missing aircraft. To complete; The specific steps for correcting the original taxi time by defining a detection area are as follows: Define a rectangular detection area of ​​a preset size at the entrance of the airport runway , from I consecutive track points of each aircraft Select the first one to enter the corresponding rectangular detection area The time corresponding to the trajectory point is used as the runway usage time , the judgment logic is as follows: in, express The corresponding timestamp; In order to avoid the impact of the aircraft runway redundancy time included in the original taxi time on the estimate of the variable taxi time, the original take-off time is replaced by the runway use time. For aircraft taxiing out of the airport, the updated aircraft actual variable taxi time , calculated by the following formula: 。 5. The method for estimating variable taxiing time of aircraft at an airport based on a spatiotemporal graph network according to claim 4, characterized in that: S2 specifically includes the following steps: S21: Airport road network topology simplification: Define the airport road network as a directed graph ,in Represents a node set, which is the embodiment of intersections and connections in the scene road topology, including number and degree attributes; It is a set of directed edges, representing the road segments in the scene road topology, including the segment number, segment length, segment category, and start and end point attributes. For the original road network data, it is geographic information data stored in the shp format, where the derived segment attributes include the segment number, segment category, segment length, segment start node number, and segment end node number. The identified derived segment attributes include the number of vertices that make up the segment. and the geographical longitude and latitude of each vertex, that is, the total number of road network segments is N; S22: Extraction of time-consuming sequence of road sections based on graph association: The original field monitoring trajectory data includes the timestamp, longitude and latitude, ground speed, and turning angle information corresponding to a series of continuous trajectory points during the taxiing process of the aircraft. The continuous trajectory points are graph-associated with the airport road network through a map matching algorithm, thereby outputting the maximum probability road section sequence of the aircraft, and combining the sequence of the continuous trajectory points to determine the road section corresponding to each trajectory point; Calculation of time consumption for a route segment: traverse the continuous trajectory points of the aircraft one by one, and the continuous trajectory points of the aircraft The corresponding timestamp sequence is , the time difference between adjacent trajectory points is ; (1) When the trajectory point and trajectory points There is only one crossing section between them, and the crossing section is not a trajectory point and trajectory points The time difference is Assigned to the crossing section, and mark the starting time of the crossing section as , the end time of the crossing section is ; (2) When the trajectory point and trajectory points There is only one crossing section between them, and the trajectory point If the corresponding road section is the same as the crossing road section, the track point is skipped. Continue to determine the trajectory point Whether the corresponding road section is the same as the crossing road section, until the corresponding road section and the crossing road section are different trajectory points are obtained , take the trajectory point With trajectory points The time difference is assigned to the crossing section, and the starting time of the crossing section is marked as , the end time of the crossing section is ; (3) When the trajectory point and trajectory points When there are multiple crossing sections, the time difference is allocated according to the ratio relationship of the multiple crossing sections. The specific steps are as follows: Apply Haversine formula to calculate trajectory points With trajectory points The total distance between , the total distance is the sum of the distances of the trajectory points on all the crossing sections; the statistical trajectory points With trajectory points Between The distance across the road , q is the point on the trajectory With trajectory points The number of random crossing sections between ; According to the track point With trajectory points The ratio of the distance of each crossing section to the total distance is calculated by Assign to track point With trajectory points On each crossing section between The time taken to cross a road section is: Mark The start time of the crossing section , No. End time of the crossing section ; Initialize the taxi path information table to store the list of segment numbers of each aircraft during the taxiing process and the segment time, start time, and end time of each corresponding segment.

6. The method for estimating variable taxiing time of aircraft at an airport based on a spatiotemporal graph network according to claim 5, characterized in that: The airport road network topology simplification in S21 specifically includes: S211: For the road segment sequence in the road network , and the node numbers of the corresponding road section starting points are respectively represented as , the node numbers of the end points of the road sections are represented as , the length of the road section is expressed as , the number of control points in the section interval passed by the road section is expressed as ; Set the length threshold ,for Less than Section To delete; The adjacent road segments are numbered as , where 1≤p≤N, p is a natural number. The starting point of the segment is the segment The end point of the road segment is The starting point of the segment is changed to the segment The starting point of the road segment; if the road segment The end point of the segment is the segment The starting point of the road segment is The end point of the segment is changed to the segment The starting point of the road section is determined by the node, thereby ensuring the connectivity of the road network, that is: ； S212: Filling section The starting point and end point of the segment, through the segment The number of control points in the included road section Determine the line shape and fill it. If the number of control points is less than 4, it is determined to be a straight line, and the road sections are connected with a straight line. The starting point and end point of the road segment; if the number of control points is greater than 4, it is determined to be a curve, and the interpolation density of different segments is controlled. The cubic spline interpolation method is used to generate a smooth curve to connect the road segments. The starting point and end point of the road segment; S213: recalculating the length of each road section in the road network after the airport road network topology is simplified, and storing the updated road section length, the node number of the road section starting point, the node number of the road section end point, and the geographic information data of the road section starting point and the road section end point.

7. The method for estimating variable taxiing time of aircraft at an airport based on a spatiotemporal graph network according to claim 6, characterized in that: S3 specifically includes the following steps: S31: embedding the road segment speed: based on the glide path information table obtained in S22, obtaining a road segment-time window matrix, and constructing a road segment-time window road segment speed feature table; S311: Generate an empty matrix with the horizontal axis being the road section number and the vertical axis being the time window. Fill in the horizontal axis from small to large; for the time range covered by the original scene taxiing data set, take 5 minutes as the window size, generate a time window sequence, and fill in the vertical axis; for each position in the matrix, initialize a speed array; S312: For road sections , identify the section passing through the glide path information table Aircraft operation data; S313: Based on the recognition result of S312, extract the time sequence of the section obtained in S22 and the time of the aircraft on the section Start time , End time , Road section time ; S314: For the extracted start time and end time, traverse the vertical axis time window. If the start time or end time is within a certain time window, calculate the aircraft's time in the route section. The road speed : The road speed Fill in the corresponding positions in the matrix generated by S311; S315: Repeat S312-S314 until the road segment sequence is traversed , get the road section-time window matrix; S316: extract the segment speed array of each position in the segment-time window matrix, and calculate the average segment speed in the segment speed array; for the segment or time window not covered by the original scene taxiing data set, fill the corresponding position with 0; in the segment-time window matrix, if there is a segment whose segment speed is 0 in all time windows, delete this segment; for the position of 0 in the remaining segment, fill the average segment speed of this segment in the same time window or the six time windows before and after on other dates, and finally obtain the segment-time window segment speed feature table distributed according to the time-space information; S32: Construction of spatiotemporal graph neural network: The spatiotemporal graph neural network model adopted consists of a spatiotemporal graph convolution module composed of a spatial block based on graph attention convolution and a temporal block based on gated recurrent unit GRU, which is finally output through an activation function and a linear prediction layer; In the spatial block, a graph attention network that introduces the attention mechanism into the graph structure is used as the convolution layer. The input of the convolution layer is a graph structure node feature set containing the road speed of each road section. Each road section in the road network is used as a node. The node features are updated through the attention mechanism, and a multi-head attention mechanism is introduced at the same time. In the time block, GRU is used as the basic unit for time window sequence processing; Finally, the prediction output is achieved by introducing a ReLU function with nonlinear and sparse activation and a linear prediction layer for dimensionality reduction and linear transformation; S33: training the spatiotemporal graph neural network and predicting the speed of each coasting section according to the spatiotemporal section interval distribution.

8. The method for estimating variable taxiing time of aircraft at an airport based on a spatiotemporal graph network according to claim 7, characterized in that: In S32, the graph attention convolutional network completes its functions through two functions: node message function and message update function. The node message function calculates the unnormalized attention coefficient between each pair of nodes and their neighboring nodes. , and apply Function activation; the message update function uses softmax to normalize the attention coefficient to obtain the attention weight of the graph node corresponding to each road section.

9. The method for estimating variable taxiing time of aircraft at an airport based on a space-time graph network according to claim 8, characterized in that: In S32, the time block in the spatiotemporal graph convolution module also serves as a spatiotemporal graph learning module, which consists of a two-layer GRU encoder and a one-layer GRU decoder, wherein the GRU encoder iteratively processes the input road section speed sequence, calculates the update gate, reset gate and candidate hidden state, and encodes the road section speed sequence into a fixed-size context vector, i.e., the current hidden state; in the two-layer GRU encoder, the result of the first layer encoding is used as the input of the second layer encoder, so as to extract high-dimensional features and encode them, perform feature time series modeling, capture the long-distance time dependency between features, consider the time correlation of the road section speed in different time windows, and finally generate a fixed-length vector as the feature representation; the GRU decoder consists of a GRU layer, which uses the last hidden state of the GRU encoder as the initial hidden state at each time step to generate a predicted output.

10. The method for estimating variable taxiing time of aircraft at an airport based on a space-time graph network according to claim 9, characterized in that: S4 specifically includes the following steps: S41: Speed-time conversion: synthesizing the predicted variable coasting time based on the speed characteristic table of the segment-time window prediction segment obtained in S33: S411: Determine the time window to which the aircraft to be predicted belongs: For the aircraft to be predicted, use the unique identifier Match the corresponding aircraft operation data and extract the aircraft's estimated taxi start time and the expected segment sequence , M refers to the total number of sections that the aircraft is expected to pass through, and determines The time window to which it belongs; S412: Extract the historical section speed as input, use the trained graph neural network model to predict the speed of each section within the taxiing time coverage of the aircraft to be predicted; S413: Calculate the estimated section time of each section based on the estimated taxiing path of the aircraft to be predicted: Read the estimated section sequence of the aircraft to be predicted And the length sequence corresponding to each section ; For the first section passed during the slide , select the predicted section speed corresponding to the starting time window, and calculate the estimated section time of the section ; For the next section , first calculate the aircraft's expected entry section Start time ,judge The time window to which it belongs, and then pass through the road section within this time window Corresponding predicted road speed, calculate the estimated road time , determine the time window to which it belongs, until the traversal of the expected route sequence of the aircraft to be predicted is completed, and the expected route time of each route is obtained; S414: Obtaining an initial predicted variable taxiing time by accumulating the estimated segment time: For the aircraft to be predicted, read the estimated segment time of each segment to be passed during taxiing, and obtain the initial predicted variable taxiing time by accumulating ; S415: According to S411-S414, traverse each aircraft to be predicted in the set of aircraft to be predicted to obtain an initial predicted variable taxiing time of each aircraft to be predicted; S42: The final prediction result of the variable taxiing time is obtained based on the variable taxiing time fusion and random forest model training; S421: Based on the complete data table obtained in S11, feature mining is performed to select some motion features, meteorological features and historical spatiotemporal features that are most strongly correlated with the taxiing time as the global feature set; wherein the mined motion features include the total taxiing turning angle and the total taxiing distance obtained by accumulating the field monitoring trajectory data; the meteorological features include visibility; the historical spatiotemporal features include the average taxiing time of the aircraft within 15 minutes before the aircraft starts taxiing, the total number of taxiing aircraft within the first 15 minutes, the parking stand corresponding to the parking stand, and the time period number of the taxiing start time within one day; the parking stand group and the time period number of the taxiing start time within one day are uniquely encoded; S422: For the set of aircraft to be predicted, obtain the global feature set corresponding to all aircraft to be predicted , and associate the global features with the initial predicted variable taxi time through a unique identifier Match, concatenate and fuse global features to obtain the total sliding feature set after encoding representation ; The actual variable taxiing time of all aircraft to be predicted obtained after S12 processing is used as the target true value set ; S423: The sliding total feature set The total sliding feature set is divided into test set, validation set and training set according to 3:1:1, and the random forest model is used for prediction. And the target truth value set The random forest model is input and the final predicted variable taxiing time is obtained by training the random forest model.

Images