Airport surface taxiing time prediction method and system based on interpretable similarity

Abstract

The present invention discloses a method and system for predicting airport scene taxi time based on interpretable similarity. The method includes the following steps: Step 1: Collecting airport scene data from the A-CDM system and airport weather data from the aviation meteorological department, and constructing a candidate scene database; Step 2: Extracting features related to taxi time, dividing the results into taxi scene static feature similarity and taxi scene periodic feature similarity, and performing data construction for each; Step 3: Calculating taxi time based on dynamic interpretable similarity, calculating the scene similarity between the target scene and candidate similar scenes based on static features and periodic features, and performing weighted summation to obtain an integrated scene similarity. The taxi time of all candidate scenes is weighted according to the obtained scene similarity to linearly generate a taxi time prediction result for the target scene. The present invention achieves linear generation of taxi time, improving the prediction accuracy of the model.

Description

Technical Field

[0001] The present invention belongs to the field of airport scene key indicator prediction, and in particular relates to a taxiing time prediction method applicable to flight dynamic attribute characteristics with some static attribute characteristics that do not change with time and periodic characteristics. Background Art

[0002] With the continuous increase in flight volume, improving the operational efficiency of air transport airports requires refined surface dispatching. At large transport airports, surface structures are interconnected and complex, with dense traffic and a volatile environment. This creates challenges for controllers in actual operations, especially during peak hours. Currently, Airport Collaborative Decision Making (A-CDM) systems are widely used in surface operations management. This system encourages multi-party collaboration to coordinate airport resources, ultimately improving the operational efficiency of the airport network. As of 2023, A-CDM systems have been implemented and are in use at 33 European airports. Variable taxi time (VTT) is a key metric in surface dispatching using A-CDM systems. Failure to accurately estimate taxi time can lead to waste of surface resources and environmental pollution.

[0003] As prediction models become increasingly complex, taxi time prediction accuracy is also improving. However, in actual surface control operations, the opacity of the taxi time prediction process prevents air traffic controllers from fully understanding the model's workings. This leads them to rely more on their own experience to ensure operational safety, hindering the widespread adoption of related technologies. To enhance air traffic controllers' trust in taxi time prediction models, a feasible approach is to employ a prediction model based on similar scenarios. By comparing the actual operational information differences between historical and current scenarios, this model provides historical statistical results for predicting indicators in the target scenario. In the field of taxi time prediction, this approach uses multiple historical taxi data sets that are most similar to the current scenario to improve the accuracy of taxi time predictions for departing aircraft in the target scenario. Given that the scenarios obtained by this method are based on actual historical operational data and are fully visible, the prediction results are more easily accepted by frontline operators, and therefore the process is considered explainable. Therefore, a taxi time prediction method that can provide more historical reference information is urgently needed. Summary of the Invention

[0004] The purpose of the present invention is to provide an airport scene taxiing time prediction method and system based on interpretable similarity, which performs multi-time scale comparison according to heterogeneous feature types, calculates scene interpretable similarity and linearly generates taxiing time prediction results to achieve precise scene control.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] A method for predicting airport surface taxiing time based on interpretable similarity includes the following steps:

[0007] Step 1: Processing raw scene data: Collect airport scene data from the A-CDM system and airport weather data from the aviation meteorological department, pre-process the data, obtain a complete data list that conforms to actual operation conditions and does not contain any missing data, and construct a candidate scene database;

[0008] Step 2: Build a taxiing scene feature system. Starting with flight model data, flight schedule data, scene situation data, and airport environment data, extract features related to taxiing time. Classify the data based on whether it has periodic properties. The results are divided into taxiing scene static feature similarity and taxiing scene periodic feature similarity, and data is constructed separately.

[0009] Step 3: Dynamic similarity calculation and glide time prediction. The glide time is calculated based on the dynamic explainable similarity. The scene similarity between the target scene and the candidate similar scenes is calculated based on the static features and periodic features, and the weighted sum is used to obtain the integrated scene similarity. The glide time of all candidate scenes is linearly weighted according to the obtained scene similarity to generate the glide time prediction result under the target scene.

[0010] The step 1 specifically includes:

[0011] (1.1) Collect collaborative decision-making data from the airport's A-CDM system, flight plan data from airlines, and weather message data from the airport's meteorological department, and match the data involved in all flight scenarios;

[0012] (1.2) Preprocess the matched data to remove all erroneous data, complete missing data, filter out all abnormal data, and build the original scene database.

[0013] The second step specifically includes:

[0014] (2.1) Based on the original taxiing scene dataset generated by screening, the static attribute features of the taxiing scene are extracted. The static attribute features of the taxiing scene include eight common features: flight number, airline, runway and apron group, parking position, aircraft type, destination airport, hour and minute. The taxiing trajectory is reflected by pairing the runway and the parking position. The above eight features are all categorical features, so the entity embedding method is introduced for re-encoding. The taxiing time under different scenarios is used as the supervision condition to construct a neural network. An additional embedding layer is added to the network to embed the categorical features of all samples into a space with half the original dimension.

[0015] (2.2) The periodic attribute characteristics of the taxiing scenario reflect the possible mutual influence between the departure flight and other departure or arrival flights, as well as the weather. Four categories and eight surface traffic characteristics based on the spatiotemporal network topology are used, which comprehensively consider the possible relationship between departure flights and arrival flights. These include the departure / arrival surface instantaneous flow index (SIFI), which represents the number of arriving or departing aircraft taxiing on the surface when the target aircraft is pushed out; the departure / arrival surface cumulative flow index (SCFI), which represents the number of departure and arrival aircraft taxiing when the target aircraft is taxiing; the aircraft queue length index (AQLI), which represents the total number of aircraft taking off or landing during the taxiing process of the target aircraft; and the slot resource demand index (SRDI), which represents the total number of aircraft pushing out or landing within 15 minutes before and after the push-out time of the target aircraft.

[0016] (2.3) Construct the data structure required for the input network. According to the characteristics of the two types of data, in the construction of the static attribute feature data structure of the sliding scene, the vector for data embedding is called the static attribute vector θ i , all sample target scenes and their respective candidate similar scene sets are stacked into an input structure, and the final input data format is s×n×2×α cat , where s represents the number of sample scenes, α cat is the dimension of the processed static attribute vector, and n is the number of candidate scenarios. Under this construction method, the static attributes of all sample target scenarios are compared with those of each candidate similar scenario. These attributes are only related to the flight plan or airport and weather data and have nothing to do with the actual status of the airport surface.

[0017] (2.4) In terms of the periodic attribute characteristics of the taxiing scenario, considering that the flight schedule may have periodic characteristics, the environmental dynamic attribute characteristics of all scenarios at the same time in the previous 1 day, 7 days, and 28 days are spliced ​​to form a multi-time scale environmental dynamic attribute input vector for one day, called the dynamic attribute vector θ e , process the vectors of all scene data according to the above steps; combine all scenes with their respective candidate similar scene sets to finally form the input data.

[0018] In step (2.2), the estimated pushback time and planned takeoff time are used when calculating these characteristics, taking into account the late and tactical operations.

[0019] The step three specifically includes:

[0020] (3.1) Scenario index decomposition: For the gliding time T(n) under the nth candidate similar scenario, the gliding time T of the target scenario ξ is used. ξ The deviation δ between the nth candidate similar scene and the target scene n Therefore, the following method is used to model and analyze the taxi time composition of candidate similar scenarios:

[0021] T(n)=T ξ +δ n

[0022] Various factors and uncertainties involved in the scene will cause the taxiing time of the current scene to be shortened or prolonged compared with the target scene. Therefore, for the convenience of analysis, the taxiing time deviation in the nth candidate scene is assumed to obey a normal distribution with a mean of 0, so that in The variance is generated under the condition of this scene due to the difference in key features between it and the target scene;

[0023] (3.2) Assume that there are only two different departure flight operation scenarios i and j in a scenario class, and rewrite the taxi time as Similarly, rewrite another sample as

[0024] Combine multiple samples with the same mean but different variances and apply weights to each sample to minimize the overall variance and obtain a more accurate T ξ Value:

[0025]

[0026] Among them, ω is the scene similarity, is the estimated value of the sliding time of the target scene, and the sum of the restricted similarities ω i +ω j =1;

[0027] (3.3) Through statistical knowledge The variance value of

[0028]

[0029] Derivative the similarity in the formula:

[0030]

[0031] Obviously, the second-order derivative is greater than zero, so the minimum value of the variance is obtained, and the corresponding similarity values ​​are:

[0032]

[0033] If the corresponding similarity of multiple scenes is determined, the expression is as follows:

[0034]

[0035] Where S is a scene set, which contains n different scenes, so the following results are obtained:

[0036]

[0037] The sum of all similarities is 1, that is,

[0038] (3.4) Taking into account the minimum value of the sample variance and combining the final prediction results, the glide time of the target scene is approximately considered to be the weighted sum of the samples, which can be expressed as:

[0039]

[0040] The same similarity sum is 1, that is Each weight ω i That is the explainable similarity of the scene;

[0041] Based on the characteristics of interpretable similarity, the two types of processed data are respectively input into two neural networks NET1 and NET2. The calculation process is as follows:

[0042] (3.5) Data downsampling and visualization: downsample the data using a 1×2 convolution kernel; expand the simplified data to form a two-dimensional tensor, whose size and shape are similar to common images;

[0043] (3.6) Image convolution: The convolution pooling process is performed on the tensor using the LeNet-based network framework. The fully connected layer is connected to the end of the two neural networks to obtain their respective similarity vectors. The weights are combined to obtain an n×1 similarity fitting result, the number of which is consistent with the number of candidate similar scenes. The loss function is calculated as follows:

[0044]

[0045] Where ω is the similarity set of candidate similar scenes under all target scenes, ω mn is the similarity of the nth scene under the mth target scene, The meaning of is the mth target scene feature, x nThe meaning is the nth historical scene feature, α is the target scene set, β is the candidate similar scene set, λ is the deviation importance parameter, K is the number of target operation scenes of all departing aircraft during the training process, T(x n ) is the sliding time of the nth candidate similar scene, For the target scene The loss function consists of two parts: minimizing the taxi time prediction error on the training set while reducing the similarity value of high-deviation scenarios. The actual role of the neural network in this process is to calculate the similarity between different scenarios and the target scenario and use it to linearly generate the taxi time of the departing flight under the target scenario.

[0046] (3.7) Priori processing, in order to ensure that the sum of similarities is 1, the following conditions are required: Based on this constraint, the similarities output by the two neural networks need to be processed a priori: first, all negative similarities need to be mapped to 0, that is, an additional ReLU layer is added to the end of the two neural networks. Second, all weights need to be normalized, that is, a normalization layer is added after the ReLU layer of the two neural networks. The similarities processed above are combined and weighted to produce the overall similarity between the target scene and the candidate similar scenes, which is calculated as follows:

[0047]

[0048] Where μ is the combination similarity distribution coefficient, and are the static and dynamic vector similarities of the candidate similar scenes output by NET1 and NET2 under the nth sample target scene;

[0049] (3.8) Iterative training and similarity extraction: glide time is linearly generated, and the predicted combined similarity is linearly weighted summed with the historical glide time in the candidate similar scenes. Iterative training is performed based on the loss function to reduce the error between the two.

[0050] (3.9) Based on the scenario similarity calculation results, the taxiing time of the corresponding flights in each candidate similar scenario is weightedly summed and the result is used to represent the taxiing time result in the target scenario.

[0051] An airport surface taxiing time prediction system based on interpretable similarity, comprising:

[0052] Airport surface flight taxiing scene data processing module, used to collect data from multiple data sources and perform pre-processing;

[0053] The flight taxi scene feature extraction and grouping module extracts features based on whether the data has periodic characteristics, and divides the results into static attribute features of the taxi scene and periodic attribute features of the taxi scene;

[0054] The scene dynamic interpretable similarity calculation and taxiing time prediction module is used to integrate input data and calculate the dynamic interpretable similarity between the target scene and the historical operation scene.

[0055] Beneficial effects: The airport surface taxiing time prediction method and system based on interpretable similarity of the present invention designs a taxiing time prediction process and system based on interpretable similarity, compares static features and data with periodic features according to the characteristics of feature data, and provides a new method for achieving accurate taxiing time prediction. It can provide strong support for subsequent decision-making content such as surface aircraft launch control, and fills the technical gap in the dynamic interpretable similarity calculation method of airport surface operations and taxiing time prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 This is a hierarchical diagram of the airport surface taxi time prediction method based on interpretable similarity;

[0057] Figure 2 This is the core flow chart of the airport surface taxi time prediction method based on interpretable similarity;

[0058] Figure 3 It is the static attribute feature data structure of taxiing scene for the airport taxiing time prediction method based on interpretable similarity;

[0059] Figure 4 The data structure of taxiing scene periodic attribute characteristics for the airport taxiing time prediction method based on interpretable similarity;

[0060] Figure 5 This is a diagram of the neural network structure in the airport surface taxiing time prediction method based on interpretable similarity;

[0061] Figure 6 This is a comparison chart of the prediction accuracy of the airport surface taxi time prediction method based on interpretable similarity;

[0062] Figure 7 This is the weight distribution diagram obtained by the airport surface taxiing time prediction method based on interpretable similarity. DETAILED DESCRIPTION

[0063] To facilitate understanding of the design objectives, improvements, and combined advantages of this invention, the following seven diagrams and implementation examples are used to further explain this method. It should also be noted that the scene classification samples used in this section are only used to illustrate the feasibility of this classification method and are not solely applicable to the processes included in this invention.

[0064] Figure 1 This is a hierarchical diagram of the airport surface taxi time prediction method based on interpretable similarity, which explains the entire process from data processing to data construction to taxi time prediction in the present invention and summarizes the overall implementation plan of the present invention.

[0065] Figure 2 This is the core flow chart of the airport surface taxiing time prediction method based on interpretable similarity, and the implementation sequence and specific execution method of each step in the present invention are explained in detail.

[0066] The airport surface taxiing time prediction method based on interpretable similarity includes the following steps:

[0067] Step 1: Scene data collection and preprocessing, which is essentially data cleaning and integration. Collect and match relevant data in the scene and process any erroneous or missing data. This includes the following steps:

[0068] (1.1) Collect collaborative decision-making data from the airport's A-CDM system, flight plan data from airlines, and weather message data from the airport's meteorological department. Match and process the data involved in all flight scenarios;

[0069] (1.2) Preprocess the matched data to remove all erroneous data, complete missing data, and filter out all abnormal data to build the original scene database;

[0070] Collect various scenario data within the target timeframe at the busy airport, including but not limited to airport A-CDM data, airline-submitted flight plan data, and airport weather report times. Erroneous data is directly deleted, and missing data is supplemented using difference filling or upward filling. The cleaned raw data is integrated to form a database of original departure flight scenarios.

[0071] Step 2: Feature processing and data construction. The essence is to extract features from the processed data based on the data characteristics. The features are divided into static attribute features of the taxiing scene and periodic attribute features of the taxiing scene, and input data is constructed based on similar scene characteristics. The specific steps include the following:

[0072] (2.1) Based on the original data set of taxiing scenes generated by screening, the static attribute features of the taxiing scenes are extracted, which refer to the features determined by the flight model and the location of the flight, including relevant information such as flight number, airline, model and parking position. The present invention introduces a total of 8 common features, namely flight number, airline, runway apron group, parking position, model, destination airport, hour and minute. The present invention uses the pairing of runway and parking position to reflect the taxiing trajectory. The above 8 features are all categorical features, so the present invention introduces the entity embedding method for re-encoding. The present invention constructs a neural network with the taxiing time under different scenarios as the supervision condition, adds an additional embedding layer in the network, and embeds the classification features in all samples into a space with half the original dimension. The present invention can significantly reduce the cost required for the prediction model.

[0073] (2.2) The main content reflected by the periodic attribute characteristics of the taxiing scenario is the possible mutual influence between the departing flight and other departing or arriving flights, as well as the weather. Four categories and eight surface traffic characteristics based on the spatiotemporal network topology are used, among which the possible relationship between departing flights and arriving flights is comprehensively considered. These include the departure / arrival surface instantaneous flow index (SIFI), which represents the number of arriving or departing aircraft taxiing on the surface when the target aircraft is pushed out; the departure / arrival surface cumulative flow index (SCFI), which represents the number of departing and arriving aircraft taxiing when the target aircraft is taxiing; the aircraft queue length index (AQLI), which represents the total number of aircraft taking off or landing during the taxiing process of the target aircraft; and the slot resource demand index (SRDI), which represents the total number of aircraft pushing out or landing within 15 minutes before and after the push-out time of the target aircraft. It is important to note that the estimated time of launch and planned takeoff are used when calculating these characteristics, taking into account late-stage and tactical operations.

[0074] (2.3) Construct the data structure required for the input network. According to the characteristics of the two types of data, in the construction of the static attribute feature data structure of the sliding scene, the vector for data embedding is called the static attribute vector θ i , all sample target scenes and their respective candidate similar scene sets are stacked into an input structure, and the final input data format is s×n×2×α cat , where s represents the number of sample scenes, α catis the dimension of the processed static attribute vector, and n is the number of candidate scenarios. In this construction, the static flight attributes of all sample target scenarios are compared with those of each candidate similar scenario. These attributes are only related to flight schedules or airport and weather data, not the actual state of the airport surface.

[0075] (2.4) In terms of the periodic attribute characteristics of the taxiing scenario, considering that the flight schedule may have periodic characteristics, the environmental dynamic attribute characteristics of all scenarios at the same time in the previous 1 day, 7 days, and 28 days are spliced ​​to form a multi-time scale environmental dynamic attribute input vector for one day, called the dynamic attribute vector θ e , process the vectors of all scene data according to the above steps; respectively combine all scenes with their respective candidate similar scene sets to form the input data;

[0076] The features related to taxiing time are extracted and classified according to the static attribute features of the taxiing scene and the periodic attribute features of the taxiing scene. The static attribute features include flight number, aircraft model, etc., while the periodic attribute features include congestion evaluation index, airport weather conditions, etc. The categorical variables are embedded in the entity. The two features are mapped and stored separately. The result is as follows: Figure 3 and Figure 4 shown.

[0077] Step 3: Calculate the dynamic interpretable similarity of the scene and predict the taxiing time. This is essentially based on the similarity calculation results of the target scene and the candidate scene in the proposed neural network, and then perform similarity fusion to finally generate the scene interpretable similarity. Finally, the interpretable similarity of the candidate scene and the aircraft taxiing time in the scene are weighted summed to obtain the taxiing time prediction result for the target scene. The specific steps include:

[0078] (3.1) Scenario index decomposition. For the gliding time T(n) under the nth candidate similar scenario, this study uses the gliding time T of the target scenario ξ ξ The deviation δ between the nth candidate similar scene and the target scene n Therefore, the following method is used to model and analyze the departure taxi time composition of candidate similar scenarios:

[0079] T(n)=T ξ +δ n

[0080] Various factors and uncertainties involved in the scene will cause the taxiing time of the current scene to be shortened or prolonged compared with the target scene. Therefore, for the convenience of analysis, the taxiing time deviation in the nth candidate scene is assumed to obey a normal distribution with a mean of 0, so that in The variance is generated under the condition of this scene due to the difference in key features between it and the target scene.

[0081] (3.2) Assuming that there are only two different departure flight operation scenarios i and j in a scenario class, the taxi time is rewritten as Another example can be rewritten as Combine multiple samples with the same mean but different variances and apply weights to each sample to minimize the overall variance and obtain a more accurate T ξ Value:

[0082]

[0083] Among them, ω is the scene similarity, is the estimated value of the sliding time of the target scene, and the sum of the restricted similarities ω i +ω j =1.

[0084] (3.3) Through statistical knowledge, we can get The variance value of

[0085]

[0086] By taking the derivative of the similarity in the formula, we can get:

[0087]

[0088] Obviously, the second-order derivative is greater than zero, so the minimum value of the variance can be obtained, and the corresponding similarity values ​​are:

[0089]

[0090]

[0091] If the corresponding similarity of multiple scenes is determined, the expression is as follows:

[0092]

[0093] Where S is a scene set, which contains n different scenes. Therefore, we can get the following results:

[0094]

[0095] The sum of all similarities is 1, that is,

[0096] (3.4) Taking into account the minimum value of the sample variance and combining it with the final prediction result, we can approximate that the glide time of the target scene is the weighted sum of the samples. The experimental results show that this approximation method can meet the actual calculation requirements, and its expression is:

[0097]

[0098] The same similarity sum is 1, that is Each weight ω i This is the explainable similarity of the scene. Based on the characteristics of explainable similarity, the two types of processed data are respectively input into two neural networks NET1 and NET2.

[0099] (3.5) Data downsampling and visualization: downsample the data using a 1×2 convolution kernel; expand the simplified data to form a two-dimensional tensor, whose size and shape are similar to common images;

[0100] (3.6) Image convolution, using a LeNet-based network framework to perform convolution pooling on the tensor. At the end of the two neural networks, a fully connected layer is connected to obtain their respective similarity vectors, and the weights are combined to obtain an n×1 similarity fitting result, which is consistent with the number of candidate similar scenes used in this invention. The loss function is calculated as follows:

[0101]

[0102] Where ω is the similarity set of candidate similar scenes under all target scenes, ω mn is the similarity of the nth scene under the mth target scene, The meaning of is the mth target scene feature, x n The meaning is the nth historical scene feature, α is the target scene set, β is the candidate similar scene set, λ is the deviation importance parameter, K is the number of target operation scenes of all departing aircraft during the training process, T(x n ) is the sliding time of the nth candidate similar scene, For the target scene The loss function consists of two parts: minimizing the taxi time prediction error on the training set while reducing the similarity value of high-deviation scenarios. The actual role of the neural network in this process is to calculate the similarity between different scenarios and the target scenario and use it to linearly generate the taxi time of the departing flight under the target scenario.

[0103] (3.7) Priori processing, in order to ensure that the sum of similarities is 1, the following conditions are required: Based on this constraint, the similarities output by the two neural networks need to be processed a priori: first, all negative similarities need to be mapped to 0, that is, an additional ReLU layer is added to the end of the two neural networks. Second, all weights need to be normalized, that is, a normalization layer is added after the ReLU layer of the two neural networks. The combined weighted output of the similarities completed above can produce the overall similarity between the target scene and the candidate similar scenes, which is calculated as follows:

[0104]

[0105] Where μ is the combination similarity distribution coefficient, and The static and dynamic vector similarities of the candidate similar scenes output by NET1 and NET2 under the nth sample target scene are respectively. This method can significantly reduce the parameter search space and accelerate the convergence of the neural network.

[0106] (3.8) Iterative training and similarity extraction: glide time is linearly generated, and the predicted combined similarity is linearly weighted summed with the historical glide time in the candidate similar scenes. Iterative training is performed according to the loss function to reduce the error between the two.

[0107] (3.9) Based on the scenario similarity calculation results, the taxiing time of the corresponding flights in each candidate similar scenario is weightedly summed and the result is used to represent the taxiing time result in the target scenario.

[0108] The parameters of the glide time prediction network proposed by the present invention are trained using the scene data in step 2. The network structure is as follows: Figure 5 As shown in the figure, the scene similarity of the static attribute features of the taxiing scene and the periodic attribute features of the taxiing scene are compared respectively. Finally, the similarity between the target scene and all candidate similar scenes is evaluated based on the scene metric vector after the fully connected layer and prior processing. The obtained dynamic similarity is weighted with the taxiing time of the candidate taxiing scene to obtain the taxiing time prediction result of the target scene. The prediction accuracy is shown in the figure. Figure 6 As shown, the weight distribution is as Figure 7 shown.

[0109] This implementation example also provides an airport surface taxiing time prediction system based on explainable similarity, including:

[0110] Airport surface flight taxiing scene data processing module, used to collect data from multiple data sources and perform pre-processing;

[0111] The flight taxi scene feature extraction and grouping module extracts features based on whether the data has periodic characteristics, and divides the results into static attribute features of the taxi scene and periodic attribute features of the taxi scene;

[0112] The scene dynamic interpretable similarity calculation and taxiing time prediction module is used to integrate input data and calculate the dynamic interpretable similarity between the target scene and the historical operation scene.

[0113] The present invention provides a method and system for predicting airport taxi time based on interpretable similarity. While numerous methods and approaches exist for implementing this technical solution, the foregoing description represents only a preferred embodiment of the present invention. It should be noted that improvements and modifications could be made by those skilled in the art without departing from the principles of the present invention, and such improvements and modifications are also within the scope of protection of the present invention. Components not specified in this embodiment may be implemented using existing technologies.

Claims

1. A method for predicting airport surface taxiing time based on interpretable similarity, characterized by: The following steps are involved: Step 1: Processing raw scene data: Collect airport scene data from the A-CDM system and airport weather data from the aviation meteorological department, pre-process the data, obtain a complete data list that conforms to actual operation conditions and does not contain any missing data, and construct a candidate scene database; Step 2: Build a taxiing scene feature system. Starting with flight model data, flight schedule data, scene situation data, and airport environment data, extract features related to taxiing time. Classify the data based on whether it has periodic properties. The results are divided into taxiing scene static feature similarity and taxiing scene periodic feature similarity, and data is constructed separately. Step 3: Dynamic similarity calculation and glide time prediction. Glide time is calculated based on the dynamic explainable similarity. The scene similarity between the target scene and candidate similar scenes is calculated based on static features and periodic features. The weighted sum is then used to obtain the integrated scene similarity. The glide time of all candidate scenes is linearly weighted according to the obtained scene similarity to generate the glide time prediction result for the target scene. Step three specifically includes: (3.1) Scenario index decomposition: For the gliding time T(n) under the nth candidate similar scenario, the gliding time T of the target scenario ξ is used. ξ The deviation δ between the nth candidate similar scene and the target scene n Therefore, the following method is used to model and analyze the taxi time composition of candidate similar scenarios: T(n)=T ξ +δ n Various factors and uncertainties involved in the scene will cause the taxiing time of the current scene to be shortened or prolonged compared with the target scene. Therefore, for the convenience of analysis, the taxiing time deviation in the nth candidate scene is assumed to obey a normal distribution with a mean of 0, so that in The variance is generated under the condition of this scene due to the difference in key features between it and the target scene; (3.2) Assume that there are only two different departure flight operation scenarios i and j in a scenario class, and rewrite the taxi time as Similarly, rewrite another sample as Combine multiple samples with the same mean but different variances and apply weights to each sample to minimize the overall variance and obtain a more accurate T ξ Value: Among them, ω is the scene similarity, is the estimated value of the sliding time of the target scene, and the sum of the restricted similarities ω i +ω j =1; (3.3) Through statistical knowledge The variance value of Derivative the similarity in the formula: Obviously, the second-order derivative is greater than zero, so the minimum value of the variance is obtained, and the corresponding similarity values ​​are: If the corresponding similarity of multiple scenes is determined, the expression is as follows: Where S is a scene set, which contains n different scenes, so the following results are obtained: The sum of all similarities is 1, that is, (3.4) Taking into account the minimum value of the sample variance and combining the final prediction results, the glide time of the target scene is approximately considered to be the weighted sum of the samples, which can be expressed as: The same similarity sum is 1, that is Each weight ω i That is the explainable similarity of the scene; Based on the characteristics of interpretable similarity, the two types of processed data are respectively input into two neural networks NET1 and NET2. The calculation process is as follows: (3.5) Data downsampling and visualization: downsample the data using a 1×2 convolution kernel; expand the simplified data to form a two-dimensional tensor, whose size and shape are similar to common images; (3.6) Image convolution: The convolution pooling process is performed on the tensor using the LeNet-based network framework. The fully connected layer is connected to the end of the two neural networks to obtain their respective similarity vectors. The weights are combined to obtain an n×1 similarity fitting result, the number of which is consistent with the number of candidate similar scenes. The loss function is calculated as follows: Where ω is the similarity set of candidate similar scenes under all target scenes, ω mn is the similarity of the nth scene under the mth target scene, The meaning of is the mth target scene feature, x n The meaning is the nth historical scene feature, α is the target scene set, β is the candidate similar scene set, λ is the deviation importance parameter, K is the number of target operation scenes of all departing aircraft during the training process, T(x n ) is the sliding time of the nth candidate similar scene, For the target scene The loss function consists of two parts: minimizing the taxi time prediction error on the training set while reducing the similarity value of high-deviation scenarios. The actual role of the neural network in this process is to calculate the similarity between different scenarios and the target scenario and use it to linearly generate the taxi time of the departing flight under the target scenario. (3.7) Priori processing, in order to ensure that the sum of similarities is 1, the following conditions are required: Based on this constraint, the similarities output by the two neural networks need to be processed a priori: first, all negative similarities need to be mapped to 0, that is, an additional ReLU layer is added to the end of the two neural networks. Second, all weights need to be normalized, that is, a normalization layer is added after the ReLU layer of the two neural networks. The similarities processed above are combined and weighted to produce the overall similarity between the target scene and the candidate similar scenes, which is calculated as follows: Where μ is the combination similarity distribution coefficient, and are the static and dynamic vector similarities of the candidate similar scenes output by NET1 and NET2 under the nth sample target scene; (3.8) Iterative training and similarity extraction: glide time is linearly generated, and the predicted combined similarity is linearly weighted summed with the historical glide time in the candidate similar scenes. Iterative training is performed based on the loss function to reduce the error between the two. (3.9) Based on the scenario similarity calculation results, the taxiing time of the corresponding flights in each candidate similar scenario is weightedly summed and the result is used to represent the taxiing time result in the target scenario.

2. The method for predicting airport surface taxiing time based on interpretable similarity according to claim 1, characterized in that: The step 1 specifically includes: (1.1) Collect collaborative decision-making data from the airport's A-CDM system, flight plan data from airlines, and weather message data from the airport's meteorological department, and match the data involved in all flight scenarios; (1.2) Preprocess the matched data to remove all erroneous data, complete missing data, filter out all abnormal data, and build the original scene database.

3. The airport surface taxiing time prediction method based on interpretable similarity according to claim 1, characterized in that: The second step specifically includes: (2.1) Based on the original taxiing scene dataset generated by screening, the static attribute features of the taxiing scene are extracted. The static attribute features of the taxiing scene include eight common features: flight number, airline, runway and apron group, parking position, aircraft type, destination airport, hour and minute. The taxiing trajectory is reflected by pairing the runway and the parking position. The above eight features are all categorical features, so the entity embedding method is introduced for re-encoding. The taxiing time under different scenarios is used as the supervision condition to construct a neural network. An additional embedding layer is added to the network to embed the categorical features of all samples into a space with half the original dimension. (2.2) The periodic attribute characteristics of the taxiing scenario reflect the possible mutual influence between departing flights and other departing or arriving flights, as well as the weather. Four categories and eight surface traffic characteristics based on the spatiotemporal network topology are used, which comprehensively consider the possible relationship between departing and arriving flights. These include the departure / arrival surface instantaneous flow index (SIFI), which represents the number of arriving or departing aircraft taxiing on the surface when the target aircraft is pushed out; the departure / arrival surface cumulative flow index (SCFI), which represents the number of departing and arriving aircraft taxiing when the target aircraft is taxiing; the aircraft queue length index (AQLI), which represents the total number of aircraft taking off or landing during the taxiing process of the target aircraft; and the slot resource demand index (SRDI), which represents the total number of aircraft pushing out or landing within 15 minutes before and after the push-out time of the target aircraft. (2.3) Construct the data structure required for the input network. According to the characteristics of the two types of data, in the construction of the static attribute feature data structure of the sliding scene, the vector for data embedding is called the static attribute vector θ i , all sample target scenes and their respective candidate similar scene sets are stacked into an input structure, and the final input data format is s×n×2×α cat , where s represents the number of sample scenes, α cat is the dimension of the processed static attribute vector, and n is the number of candidate scenarios. Under this construction method, the static attributes of all sample target scenarios are compared with those of each candidate similar scenario. These attributes are only related to the flight plan or airport and weather data and have nothing to do with the actual status of the airport surface. (2.4) In terms of the periodic attribute characteristics of the taxiing scenario, considering that the flight schedule may have periodic characteristics, the environmental dynamic attribute characteristics of all scenarios at the same time in the previous 1 day, 7 days, and 28 days are spliced ​​to form a multi-time scale environmental dynamic attribute input vector for one day, called the dynamic attribute vector θ e , process the vectors of all scene data according to the above steps; combine all scenes with their respective candidate similar scene sets to finally form the input data.

4. The method for predicting airport surface taxiing time based on interpretable similarity according to claim 3, characterized in that: In step (2.2), the estimated pushback time and planned takeoff time are used when calculating these characteristics, taking into account the late and tactical operations.

5. An airport surface taxiing time prediction system based on interpretable similarity, used to implement the method according to any one of claims 1 to 4, characterized in that: include: Airport surface flight taxiing scene data processing module, used to collect data from multiple data sources and perform pre-processing; The flight taxi scene feature extraction and grouping module extracts features based on whether the data has periodic characteristics, and divides the results into static attribute features of the taxi scene and periodic attribute features of the taxi scene; The scene dynamic interpretable similarity calculation and taxiing time prediction module is used to integrate input data and calculate the dynamic interpretable similarity between the target scene and the historical operation scene.

Images