Aircraft taxiing time prediction method based on double-scale attention fusion mechanism

Abstract

The invention relates to the technical field of air transportation management, in particular to an aircraft taxiing time prediction method based on a dual-scale attention fusion mechanism, which introduces dual mechanisms of time attention and space attention, adaptively models feature weights from a time scale and a space scale, and predicts taxiing time of an aircraft. And the capturing capability and prediction precision of the model on the sliding key factors are improved. Specifically, a time attention mechanism is used for dynamically modeling a key time slice in the taxiing process, and redundant feature interference is restrained; meanwhile, a space attention mechanism is introduced, feature weights of different sliding areas are distributed in a self-adaptive mode, influences of bottleneck nodes and abnormal areas are highlighted, and finally high-precision prediction of the sliding time is achieved. According to the method, the problem that an existing model is insufficient in spatial-temporal feature extraction in a complex taxiing environment is effectively solved, technical support is provided for airport ground operation optimization and intelligent decision making, and the modeling capacity and robustness of the model for uncertain factors in the complex taxiing environment are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of air transport management, and in particular to a method for predicting aircraft taxiing time based on a dual-scale attention fusion mechanism. Background Art

[0002] Aircraft taxiing is a crucial component of airport operations management, directly impacting flight takeoff and landing efficiency and operational safety. Taxiing time is subject to multiple factors, including the complex spatial environment, dynamic traffic flow, and real-time control strategies. It exhibits strong spatiotemporal coupling and significant uncertainty. Accurately predicting aircraft taxiing time is not only an important means of improving airport operational efficiency and reducing ground delays, but also a key technical support for achieving air-ground collaborative decision-making and intelligent airport management.

[0003] Existing taxi time prediction methods primarily rely on temporal or spatial features for modeling. Typical approaches include statistical regression models, machine learning models, and deep learning methods based on graph neural networks (GCNs) and recurrent neural networks (RNNs). Although deep models have made some progress in extracting spatiotemporal features, relying solely on time series features or spatial topology often fails to fully capture the complex dynamics of taxiing. This is particularly true in complex operational scenarios such as dense flight density and frequent bottleneck area conflicts, resulting in insufficient prediction accuracy and robustness. Summary of the Invention

[0004] The purpose of the present invention is to provide an aircraft taxiing time prediction method based on a dual-scale attention fusion mechanism, which introduces the dual mechanisms of temporal attention and spatial attention, adaptively models feature weights from the temporal scale and spatial scale, and improves the model's ability to capture key taxiing factors and prediction accuracy.

[0005] To achieve the above objectives, the present invention provides an aircraft taxiing time prediction method based on a dual-scale attention fusion mechanism, comprising the following steps:

[0006] Step 1: Conduct correlation analysis on factors affecting aircraft taxiing time;

[0007] Step 2: Based on the spatiotemporal graph neural network model, spatial and temporal features of the data are extracted to obtain the spatial relationship features between nodes and the time series features of the sliding process respectively;

[0008] Step 3: Construct a dual-scale attention mechanism to enhance the spatial and temporal features of the feature extraction results;

[0009] Step 4: Based on the spatial weight matrix generated by the spatial attention mechanism, the dependencies and influence strengths between nodes in the sliding network are dynamically characterized. Combined with the temporal feature matrix extracted by the temporal attention mechanism, the importance of each time step in the sliding process is measured. By weighted multiplication and fusion of spatial and temporal features, spatiotemporal information complementation and coupling are achieved.

[0010] Step 5: Output of taxi time prediction. By systematically evaluating the prediction results and combining multiple error analysis indicators, the accuracy and reliability of the model prediction are verified.

[0011] Optionally, the influencing factors in step 1 include parking stand location, aircraft type, pushback period, queue length, and surface accumulated flow.

[0012] Optionally, in the process of spatial feature extraction in step 2, the airport taxiway network is formally defined as an undirected graph G = (V, E, A), where V is the set of nodes in the graph, E is the set of edges in the graph, indicating the connectivity between nodes, and A is the adjacency matrix of the graph G. Each node v i ∈V represents a taxiway location, edge (v i ,v j )∈E represents the connection between two taxiway positions;

[0013] When extracting temporal features, a gated recurrent unit is used to process the features of each time point during the taxiing process of the aircraft.

[0014] Optionally, the calculation formula used in the spatial feature extraction module includes:

[0015] H l+1 =f(H (l) ,A)

[0016] f(H (l) ,A)=σ(AH (l) W (l) )

[0017]

[0018] Among them, H (l) is the node feature matrix of the lth layer, H (0) is the input node matrix, i.e., the initial features of each taxiway location; represents the normalized adjacency matrix, where A is the original adjacency matrix, D is the degree matrix; W (l) is the trainable weight matrix of the lth layer, σ is the activation function ReLU=max(0,x), f θ is a mapping function learned by a neural network, x i ,x jare the characteristics of glide paths i and j respectively, and t is the current time node.

[0019] Optionally, the calculation formula used in the time feature extraction module includes:

[0020]

[0021] R t =W R ·x t +B R

[0022]

[0023] Among them, z is the update gate, r is the reset gate, For temporary memory, h t is the hidden state, R is the residual term, and the input and output are directly connected through linear transformation; t 、r t 、 and represent the update gate, reset gate, temporary memory and hidden state at time t respectively; x t is the matrix input at time t, σ and tanh are activation functions; w z 、w r 、w h They are update gate, reset gate and calculation temporary memory respectively When the hidden state h t-1 The weight matrix of b z 、b r 、b h For updating gates, resetting gates, and calculating temporary memory The offset vector when .

[0024] Optionally, the execution process of step 3 includes the following steps:

[0025] Step 3.1: Use a multi-layer perceptron to model the extracted spatial features and calculate the relevance score of any node to the target node, thereby generating a spatial attention matrix with dimension N×N;

[0026] Step 3.2: Fuse the spatial attention matrix with the original adjacency matrix to obtain the weighted spatial feature propagation matrix;

[0027] Step 3.3: Based on the different degrees of influence of historical time series data on the predicted taxiing moments, a multi-layer perceptron is used as a scoring function to perform feature mapping on the hidden states at each moment generated by the GRU network. The attention mechanism is used to calculate the probability distribution weights in the time dimension, enabling the model to adaptively focus on key time nodes.

[0028] Step 3.4: Linearly weighted fusion is performed on the obtained attention weights and the corresponding GRU hidden states to generate a comprehensive representation vector that contains global temporal dependencies. This vector explicitly extracts the time-varying trend characteristics of flight delays by capturing the dynamic changes in the historical sequence. Ultimately, it is input into the downstream prediction module as a temporal feature expression for flight delay prediction.

[0029] Optionally, the calculation formula used by the spatial attention mechanism in step 3 includes:

[0030] e i =σ(ο·(w1 (1) X i-1 +b1 (1) ) T (w2 (1) X i-1 +b2 (1) ))

[0031]

[0032] A att =[α i,j ] N×N

[0033] After obtaining the attention weight matrix, the spatial feature extraction module of GCN is combined to propagate and update the spatial features. The final spatial feature propagation formula is as follows:

[0034]

[0035] Among them, e i is an intermediate variable, indicating the mutual influence of node keys; X i-1 is the input matrix of the previous moment; w1, w2, b1, b2 are trainable parameters, and σ is the activation function.

[0036] Optionally, the calculation formula used by the temporal attention mechanism in step 3 includes:

[0037] L t =w4tanh(w3h i-1 +b3)

[0038]

[0039] Among them L t is the attention score at each time step; W3, W4, b3 are all learnable matrices; h i-1 is the hidden layer state at the previous moment; β t is the attention weight of time step t, T is the length of the time series; H time is the fused temporal feature representation.

[0040] Optionally, in step 4, the attention weight matrix obtained by the spatial attention mechanism is fused with the temporal feature matrix obtained by the temporal attention mechanism to obtain the prediction result. The calculation formula used includes:

[0041]

[0042] Among them, e t is the attention weight matrix, H time is the time feature matrix, and b is the learned weight matrix.

[0043] Optionally, the multiple error analysis indicators in step 5 include mean square error, mean absolute error, mean absolute percentage error, root mean square error, ±3min accuracy and ±5min accuracy. The mean square error (MSE) reflects the overall size of the prediction error and is suitable for measuring the overall error fluctuation of the model; the mean absolute error (MAE) measures the average absolute size of the deviation between the predicted value and the actual value; the mean absolute percentage error (MAPE) is used to measure the relative size of the prediction error; the root mean square error (RMSE) indicator is used to comprehensively evaluate the overall error of the model; ±3min accuracy Refers to the proportion of predictions where the absolute error between the model prediction value and the actual value is within 3 minutes; ±5min accuracy Refers to the proportion of predictions where the absolute error between the model prediction value and the actual value is within 5 minutes;

[0044] The specific calculation formula includes:

[0045]

[0046]

[0047] Among them, y i is the actual taxiing time of the i-th time, is the i-th predicted taxiing time, and n is the number of prediction samples.

[0048] The present invention provides an aircraft taxiing time prediction method based on a dual-scale attention fusion mechanism, which introduces the dual mechanisms of time attention and space attention, adaptively models feature weights from the time scale and space scale, and improves the model's ability to capture key taxiing factors and prediction accuracy. Specifically, the time attention mechanism is used to dynamically model key time segments in the taxiing process to suppress redundant feature interference; at the same time, the spatial attention mechanism is introduced to adaptively allocate feature weights for different taxiing areas, highlighting the impact of bottleneck nodes and abnormal areas, and ultimately achieving high-precision prediction of taxiing time. The present invention effectively solves the problem of insufficient spatiotemporal feature extraction of existing models in complex taxiing environments, provides technical support for airport ground operation optimization and intelligent decision-making, and improves the modeling capability and robustness of the model for uncertain factors in complex taxiing environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0050] Figure 1 The present invention is a flowchart of the steps of an aircraft taxiing time prediction method based on a dual-scale attention fusion mechanism.

[0051] Figure 2 It is a schematic diagram of factors affecting the sliding time in the present invention.

[0052] Figure 3 It is a schematic diagram of the overall framework of the dual-scale attention fusion mechanism of the present invention.

[0053] Figure 4 It is a schematic diagram of the overall framework of the glide time prediction model of the present invention.

[0054] Figure 5 It is a schematic diagram comparing the prediction results and the true values at different time slices in a specific embodiment of the present invention. DETAILED DESCRIPTION

[0055] The following describes embodiments of the present invention in detail, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and are not to be construed as limiting the present invention.

[0056] The present invention provides an aircraft taxiing time prediction method based on a dual-scale attention fusion mechanism, comprising the following steps:

[0057] Step 1: Conduct correlation analysis on factors affecting aircraft taxiing time;

[0058] Step 2: Based on the spatiotemporal graph neural network model, spatial and temporal features of the data are extracted to obtain the spatial relationship features between nodes and the time series features of the sliding process respectively;

[0059] Step 3: Construct a dual-scale attention mechanism to enhance the spatial and temporal features of the feature extraction results;

[0060] Step 4: Based on the spatial weight matrix generated by the spatial attention mechanism, the dependencies and influence strengths between nodes in the sliding network are dynamically characterized. Combined with the temporal feature matrix extracted by the temporal attention mechanism, the importance of each time step in the sliding process is measured. By weighted multiplication and fusion of spatial and temporal features, spatiotemporal information complementation and coupling are achieved.

[0061] Step 5: Output of taxi time prediction. By systematically evaluating the prediction results and combining multiple error analysis indicators, the accuracy and reliability of the model prediction are verified.

[0062] The specific execution process steps are as follows Figure 1 As shown, the following steps are further explained:

[0063] Step 1: Conduct a correlation analysis on the factors that affect the aircraft taxiing time. The specific contents of each factor are as follows:

[0064] (1.1) Parking position. When an aircraft is pushed back from different parking positions, the pushback restrictions, taxiing distances, and the interaction between arriving and departing aircraft during taxiing all have certain differences, which have a significant impact on taxiing time.

[0065] (1.2) Aircraft type. Different aircraft types have different taxiing speeds and taxiing spacing requirements, which affect taxiing time. Therefore, the average taxiing time of different aircraft types at the airport varies.

[0066] (1.3) Pushback period. The pushback period has a certain impact on the average departure taxi time. The greater the average number of pushbacks at an airport during the same time period, the longer the average taxi time used.

[0067] (1.4) Queue length. The longer the queue length, the more severe the queue pressure during the departure phase will be, leading to a significant increase in taxiing time.

[0068] (1.5) Surface cumulative traffic: Surface cumulative traffic refers to the number of other flights taxiing on the ground at the same time as an aircraft is taxiing. It can reflect the busyness of the airport taxiway.

[0069] See also Figure 2 , shows a schematic diagram of factors affecting the taxiing time of an aircraft in the present invention, the specific contents are as follows;

[0070] As shown in Figure 2(a), when an aircraft is pushed back from different parking positions, the impact on taxi time varies significantly due to differences in pushback restrictions, taxi distances, and the interaction between arriving and departing aircraft during taxi. In actual operations, the choice of parking position directly affects the taxi time of an aircraft. Differences in pushback restrictions, taxi distances, and the interaction between arriving and departing aircraft at parking positions in different areas lead to significant variations in taxi time.

[0071] As shown in Figure 2(b), the average taxi times for different aircraft types at an airport vary due to differences in taxi speeds and taxi spacing requirements. The impact of aircraft type on taxi time is primarily reflected in taxi speed and spacing requirements. Differences in taxi speeds arise from differences in aircraft structure, engine performance, and takeoff and landing weights.

[0072] As shown in Figure 2(c), the pushback period has a certain impact on average departure taxi time. The greater the average number of pushbacks at an airport during the same time period, the longer the average taxi time. However, some time periods exhibit inconsistent trends. For example, during the nighttime hours of 00:00-05:00, flights at the airport operate on a single runway and cannot take off nearby, resulting in an overall longer taxi time.

[0073] As shown in Figure 2(d), when the queue length is high, the taxi time of aircraft usually increases significantly. For departing aircraft, a higher queue length means greater runway dispatch pressure, and aircraft may need to wait for a long time on the taxiway or at the runway threshold, resulting in longer taxi time.

[0074] Figure 2(e) shows the trend of departure taxi time as a function of cumulative surface traffic. Cumulative surface traffic represents the number of other flights taxiing simultaneously with a given aircraft, reflecting the airport's taxiway traffic level. An analysis of historical airport operational data revealed a Pearson correlation coefficient of 0.53 between cumulative surface traffic and taxi time.

[0075] In step 2, spatial and temporal features are extracted from the data based on the spatiotemporal graph neural network model to obtain the spatial relationship features between nodes and the time series features during the sliding process, respectively. The specific contents are as follows:

[0076] (2.1) The spatial feature extraction module utilizes the significant spatial feature extraction advantages of GCN to learn the airport structure information under each slice. The airport taxiway network can be formally defined as an undirected graph G = (V, E, A), where V is the set of nodes in the graph, E is the set of edges in the graph, indicating the connectivity between nodes, and A represents the adjacency matrix of the graph G. Each node vi ∈V represents a taxiway location, edge (v i ,v j )∈E represents the connection between two taxiway positions, and the calculation formula used by the spatial feature extraction module includes:

[0077] H l+1 =f(H (l) ,A)

[0078] f(H (l) ,A)=σ(AH (l) W (l) )

[0079]

[0080] Among them, H (l) is the node feature matrix of the lth layer, H (0) is the input node matrix, i.e., the initial features of each taxiway location; represents the normalized adjacency matrix, where A is the original adjacency matrix, D is the degree matrix; W (l) is the trainable weight matrix of the lth layer, and σ is the activation function ReLU=max(0,x). θ is a mapping function learned by a neural network, x i ,x j are the characteristics of glide paths i and j respectively, and t is the current time node;

[0081] (2.2) Using the temporal feature extraction module, a gated recurrent unit (GRU) is used to process the features of each time point during the taxiing process of the aircraft. The calculation formula used in the module includes:

[0082]

[0083] R t =W R ·x t +B R

[0084]

[0085] Among them, z is the update gate, r is the reset gate, For temporary memory, h t is the hidden state, R is the residual term, and the input and output are directly connected through linear transformation; t 、r t 、 and represent the update gate, reset gate, temporary memory and hidden state at time t respectively; x t is the matrix input at time t, σ and tanh are activation functions; wz 、w r 、w h They are update gate, reset gate and calculation temporary memory respectively When the hidden state h t-1 The weight matrix of b z 、b r 、b h For updating gates, resetting gates, and calculating temporary memory The offset vector when .

[0086] The execution process of step 3 includes the following steps:

[0087] (3.1) Use a multi-layer perceptron to model the extracted spatial features and calculate the relevance score of any node to the target node, thereby generating a spatial attention matrix with a dimension of N×N;

[0088] (3.2) The spatial attention matrix is fused with the original adjacency matrix to obtain a weighted spatial feature propagation matrix, which provides more accurate structural information for subsequent spatial feature extraction. The calculation formula used in the spatial attention mechanism includes:

[0089] e i =σ(ο·(w1 (1) X i-1 +b1 (1) ) T (w2 (1) X i-1 +b2 (1) ))

[0090]

[0091] A att =[α i,j ] N×N

[0092] After obtaining the attention weight matrix, the spatial feature extraction module of GCN is combined to propagate and update the spatial features. The final spatial feature propagation formula is as follows:

[0093]

[0094] Among them, e i is an intermediate variable, indicating the mutual influence of node keys; X i-1 is the input matrix of the previous moment; w1, w2, b1, b2 are trainable parameters, and σ is the activation function.

[0095] (3.3) Based on the varying degrees of influence of historical time series data on the predicted taxiing moments, a multi-layer perceptron is used as a scoring function to perform feature mapping on the hidden states generated by the GRU network at each moment. The attention mechanism is used to calculate the probability distribution weights in the time dimension. These weights reflect the importance of the features of different historical moments to the current prediction task, enabling the model to adaptively focus on key time nodes.

[0096] (3.4) The obtained attention weights are linearly weighted and fused with the corresponding GRU hidden states to generate a comprehensive representation vector that embodies the global temporal dependency. This vector explicitly extracts the time-varying trend characteristics of flight delays by capturing the dynamic changes in the historical sequence. Ultimately, it is input into the downstream prediction module as a temporal feature expression for flight delay prediction. The calculation formula used in the temporal attention mechanism includes:

[0097] L t =w4tanh(w3h i-1 +b3)

[0098]

[0099] Among them L t is the attention score at each time step; W3, W4, b3 are all learnable matrices; h i-1 is the hidden layer state at the previous moment; β t is the attention weight of time step t, T is the length of the time series; H time is the fused temporal feature representation.

[0100] For further information, see Figure 3 , Figure 3 The overall framework of the dual-scale attention fusion mechanism in this invention is shown. The specific steps are as follows:

[0101] Step (1), spatial feature input. Obtain the spatial feature matrix extracted by the spatial convolution module as the input for spatial feature attention calculation.

[0102] Step (2), spatial attention weight calculation. The spatial feature matrix is input into the spatial attention neural network, and the spatial attention weight coefficient is calculated by combining the hidden layer nonlinear transformation. The spatial attention neural network can automatically identify and enhance the attention to the key spatial areas during the sliding process.

[0103] Step (3), spatial feature weighting. The calculated spatial attention weights are weighted element-by-element with the original spatial feature matrix to generate a weighted spatial feature matrix, thereby highlighting the spatial region features that have a significant impact on the gliding time.

[0104] Step (4), temporal feature input. Obtain the temporal feature matrix extracted by the temporal convolution module as the input for temporal feature attention calculation.

[0105] Step (5) calculates the time attention weight. The time feature matrix is input into the time attention neural network, and the time attention weight coefficient is calculated by combining it with the hidden layer feature extraction. The time attention neural network can identify key time segments and enhance the time sensitivity modeling during the sliding process.

[0106] Step (6), time feature weighting. The calculated time attention weight is weighted element by element with the time feature matrix to generate a weighted time feature matrix, thereby effectively highlighting the key timing features in the sliding process.

[0107] Step (7), dual-scale feature fusion. The weighted spatial feature matrix and the temporal feature matrix are subjected to feature weighted multiplication and fusion processing to fully integrate the spatial and temporal dual-scale feature information.

[0108] Step (8), fusion feature output. Output the prediction result after fusion of spatiotemporal feature matrix.

[0109] Step 4: Based on the spatial weight matrix generated by the spatial attention mechanism, the dependency and influence strength between nodes in the sliding network are dynamically characterized. Combined with the temporal feature matrix extracted by the temporal attention mechanism, the importance of each time step in the sliding process is measured. By weighted multiplication, spatial and temporal features are fused to achieve spatiotemporal information complementarity and coupling. The calculation formula used by the fusion mechanism includes

[0110]

[0111] Among them, e t is the attention weight matrix, H time is the time feature matrix, and b is the learned weight matrix.

[0112] Step 5: By systematically evaluating the prediction results and combining multiple error analysis indicators, the accuracy and reliability of the model prediction are verified to ensure its applicability and effectiveness in practical application scenarios. The specific evaluation indicators include: mean square error (MSE), mean absolute error (MAE), mean absolute percentage error (MAPE), root mean square error (RMSE), ±3min prediction accuracy (±3min accuracy ) and ±5min prediction accuracy (±5min accuracy ).

[0113] The specific formula of the error analysis index is as follows:

[0114]

[0115]

[0116] Among them, y i is the actual taxiing time of the i-th time, is the i-th predicted taxiing time, and n is the number of prediction samples.

[0117] MAE measures the average absolute size of the deviation between the predicted value and the actual value. Compared with MSE, MAE is less sensitive to outliers and can more directly reflect the average level of prediction error.

[0118] MAPE is used to measure the relative size of the forecast error and can intuitively reflect the error level in the form of a percentage. However, it is more sensitive to data points with small actual values. Therefore, in practical applications, the balance of the forecast results must be comprehensively considered.

[0119] RMSE is similar to MSE in that it can amplify the impact of larger errors while maintaining the same dimension as the data, making it easier to interpret. This indicator is used to comprehensively evaluate the overall error of the model.

[0120] ±3min accuracy Refers to the proportion of predictions where the absolute error between the model prediction value and the actual value is within 3 minutes, which can reflect the model's prediction ability under high precision requirements;

[0121] ±5min accuracy It refers to the proportion of predictions where the absolute error between the model prediction value and the actual value is within 5 minutes, reflecting the prediction stability and reliability of the model within the acceptable error range in practical applications.

[0122] In summary, the overall framework of the aircraft taxiing time prediction model in the present invention is as follows: Figure 4 The specific steps are as follows:

[0123] Step 1: Feature data input. Collect and input multi-dimensional feature data including the airport taxi network structure and aircraft operation status as input features of the model.

[0124] Step 2: Spatial Feature Extraction. The taxi network features are fed into the spatial convolutional layer, which constructs a graph structure based on the nodes and connectivity relationships within the airport taxi area. Node features are extracted through the graph convolutional network, dynamically capturing the spatial dependencies between taxi areas. A dynamic weight matrix is generated that includes the influence strength of each node, enhancing the representation of key taxi area features.

[0125] Step 3: Temporal feature extraction. The temporal feature sequence is extracted and input into the temporal convolutional layer. A gated recurrent unit is used to model the temporal features of the gliding process, automatically extracting the temporal dependency information. A temporal residual structure is then used to deeply optimize the temporal features. Combined with convolution operations, this enhances feature expression, mitigates information loss during temporal feature extraction, and improves temporal feature modeling.

[0126] Step 4: Dual-Scale Attention Fusion Mechanism. The dynamic weight matrix generated by the spatial convolution layer and the temporal features extracted by the temporal convolution layer are input into the dual-scale attention mechanism module. This module integrates spatial and temporal attention to adaptively identify key sliding areas and key time steps in the sliding network, dynamically adjust feature weights, and achieve deep fusion and coupling of spatial and temporal features.

[0127] Step 5: Taxi time prediction output. The fused spatiotemporal features are input into the prediction layer to generate taxi time prediction results. The model comprehensively considers both spatial structure and temporal dynamics to accurately predict taxi time for aircraft taxiing at the airport.

[0128] Furthermore, the present invention also proposes a specific embodiment, please refer to Figure 5 , Figure 5 The following is a comparison diagram of the predicted results and the true values at different time slices of the present invention.

[0129] The 30-minute slice length demonstrated the best forecasting performance, achieving the lowest error and highest accuracy. This indicates that a 30-minute slice length provides the optimal balance of temporal granularity: retaining sufficient time series information to capture data trends and cyclical characteristics while not introducing excessive data noise and complexity. This moderate slice length enables the model to more effectively learn the inherent patterns of the data, thereby improving forecast accuracy.

[0130] In comparison, the 15-minute slice length performed poorly, as this length can lead to excessive data fragmentation, making it difficult for the model to capture long-term trends and patterns in the data. While short time slices can provide more granular information, they also increase data noise and volatility, which affects prediction accuracy. As the slice length increases to 45 minutes, prediction performance begins to decline significantly. This decline in performance may be due to the fact that longer time slices lead to excessive data aggregation, losing short-term fluctuations and important detailed features. Prediction performance at a 60-minute slice length further deteriorates, presenting the worst performance of all tested lengths. Such a long slice length can lead to over-smoothing, making it impossible for the model to distinguish data characteristics and patterns of change across different time periods.

[0131] The above disclosure is merely one or more preferred embodiments of the present invention, and certainly cannot be used to limit the scope of the present invention. A person skilled in the art can understand that all or part of the processes of the above embodiments and equivalent changes made in accordance with the claims of the present invention still fall within the scope of the invention.

Claims

1. A method for predicting aircraft taxiing time based on a dual-scale attention fusion mechanism, characterized in that: The following steps are involved: Step 1: Conduct correlation analysis on factors affecting aircraft taxiing time; Step 2: Based on the spatiotemporal graph neural network model, spatial and temporal features of the data are extracted to obtain the spatial relationship features between nodes and the time series features of the sliding process respectively; Step 3: Construct a dual-scale attention mechanism to enhance the spatial and temporal features of the feature extraction results; Step 4: Based on the spatial weight matrix generated by the spatial attention mechanism, the dependencies and influence strengths between nodes in the sliding network are dynamically characterized. Combined with the temporal feature matrix extracted by the temporal attention mechanism, the importance of each time step in the sliding process is measured. By weighted multiplication and fusion of spatial and temporal features, spatiotemporal information complementation and coupling are achieved. Step 5: Output of taxi time prediction. By systematically evaluating the prediction results and combining multiple error analysis indicators, the accuracy and reliability of the model prediction are verified.

2. The aircraft taxiing time prediction method based on the dual-scale attention fusion mechanism according to claim 1, characterized in that: The influencing factors in step 1 include parking stand location, aircraft type, pushback period, queue length, and surface accumulated flow.

3. The aircraft taxiing time prediction method based on the dual-scale attention fusion mechanism according to claim 2, characterized in that: In the process of spatial feature extraction in step 2, the airport taxiway network is formally defined as an undirected graph G = (V, E, A), where V is the set of nodes in the graph, E is the set of edges in the graph, indicating the connectivity between nodes, and A is the adjacency matrix of the graph G; each node v i ∈V represents a taxiway location, edge (v i ,v j )∈E represents the connection between two taxiway positions; When extracting temporal features, a gated recurrent unit is used to process the features of each time point during the taxiing process of the aircraft.

4. The aircraft taxiing time prediction method based on the dual-scale attention fusion mechanism according to claim 3, characterized in that: The calculation formula used in the spatial feature extraction module includes: H l+1 =f(H (l) ,A) f(H (l) ,A)=σ(AH (l) W (l) ) Among them, H (l) is the node feature matrix of the lth layer, H (0) is the input node matrix, i.e., the initial features of each taxiway location; represents the normalized adjacency matrix, where A is the original adjacency matrix, D is the degree matrix; W (l) is the trainable weight matrix of the lth layer, σ is the activation function ReLU=max(0,x), f θ is a mapping function learned by a neural network, x i ,x j are the characteristics of glide paths i and j respectively, and t is the current time node.

5. The aircraft taxiing time prediction method based on the dual-scale attention fusion mechanism according to claim 3, characterized in that: The calculation formula used in the time feature extraction module includes: z t =σ(W z ·[h t-1 ,x t ]+b t z ) r t =σ(W r ·[h t-1 ,x t ]+b t r ) R t =W R ·x t +B R Among them, z is the update gate, r is the reset gate, For temporary memory, h t is the hidden state, R is the residual term, and the input and output are directly connected through linear transformation; t 、r t 、 and represent the update gate, reset gate, temporary memory and hidden state at time t respectively; x t is the matrix input at time t, σ and tanh are activation functions; w z 、w r 、w h They are update gate, reset gate and calculation temporary memory respectively When the hidden state h t-1 The weight matrix of b z 、b r 、b h For updating gates, resetting gates, and calculating temporary memory The offset vector when .

6. The aircraft taxiing time prediction method based on the dual-scale attention fusion mechanism according to claim 1, characterized in that: The execution process of step 3 includes the following steps: Step 3.1: Use a multi-layer perceptron to model the extracted spatial features and calculate the relevance score of any node to the target node, thereby generating a spatial attention matrix with dimension N×N; Step 3.2: Fuse the spatial attention matrix with the original adjacency matrix to obtain the weighted spatial feature propagation matrix; Step 3.3: Based on the different degrees of influence of historical time series data on the predicted taxiing moments, a multi-layer perceptron is used as a scoring function to perform feature mapping on the hidden states at each moment generated by the GRU network. The attention mechanism is used to calculate the probability distribution weights in the time dimension, enabling the model to adaptively focus on key time nodes. Step 3.4: Linearly weighted fusion is performed on the obtained attention weights and the corresponding GRU hidden states to generate a comprehensive representation vector that contains global temporal dependencies. This vector explicitly extracts the time-varying trend characteristics of flight delays by capturing the dynamic changes in the historical sequence. Ultimately, it is input into the downstream prediction module as a temporal feature expression for flight delay prediction.

7. The aircraft taxiing time prediction method based on the dual-scale attention fusion mechanism according to claim 6, characterized in that: The calculation formula used by the spatial attention mechanism in step 3 is include: and i =σ(ο·(w1 (1) X i-1 +b1 (1) ) T (w2 (1) X i-1 +b2 (1) )) A att ＝[α i,j ] N×N After obtaining the attention weight matrix, the spatial feature extraction module of GCN is combined to propagate and update the spatial features. The final spatial feature propagation formula is as follows: Among them, e i is an intermediate variable, indicating the mutual influence of node keys; X i-1 is the input matrix of the previous moment; w1, w2, b1, b2 are trainable parameters, and σ is the activation function.

8. The aircraft taxiing time prediction method based on the dual-scale attention fusion mechanism according to claim 6, characterized in that: The calculation formula used by the temporal attention mechanism in step 3 is include: <h2 style=";text-align:left;direction:ltr">L<h2 style=";text-align:left;direction:ltr"> t <h2 style=";text-align:left;direction:ltr"> =w4tanh(w3h<h2 style=";text-align:left;direction:ltr"> i-1 <h2 style=";text-align:left;direction:ltr"> +b3) Among them L t is the attention score at each time step; W3, W4, b3 are all learnable matrices; h i-1 is the hidden layer state at the previous moment; β t is the attention weight of time step t, T is the length of the time series; H time is the fused temporal feature representation.

9. The aircraft taxiing time prediction method based on the dual-scale attention fusion mechanism according to claim 6, characterized in that: In step 4, the attention weight matrix obtained by the spatial attention mechanism is fused with the time feature matrix obtained by the time attention mechanism to obtain the prediction result. The calculation formula used is as follows: Among them, e t is the attention weight matrix, H time is the time feature matrix, and b is the learned weight matrix.

10. The aircraft taxiing time prediction method based on the dual-scale attention fusion mechanism according to claim 6, characterized in that: The multiple error analysis indicators in step 5 include mean square error, mean absolute error, mean absolute percentage error, root mean square error, ±3min accuracy and ±5min accuracy. The mean square error (MSE) reflects the overall size of the prediction error and is suitable for measuring the overall error fluctuation of the model; the mean absolute error (MAE) measures the average absolute size of the deviation between the predicted value and the actual value; the mean absolute percentage error (MAPE) is used to measure the relative size of the prediction error; the root mean square error (RMSE) indicator is used to comprehensively evaluate the overall error of this model; ±3min accuracy Refers to the proportion of predictions where the absolute error between the model prediction value and the actual value is within 3 minutes; ±5min accuracy Refers to the proportion of predictions where the absolute error between the model prediction value and the actual value is within 5 minutes; Specific calculation formula include: Among them, y i is the actual taxiing time of the i-th time, is the i-th predicted taxiing time, and n is the number of prediction samples.

Images