Airport flight delay prediction method based on multi-dimensional data fusion

Abstract

The invention relates to the technical field of airport flight delay prediction, and provides an airport flight delay prediction method based on multi-dimensional data fusion, which comprises the following steps: selecting multi-attribute data; converting the multi-attribute data into a spatial data variable, a spatio-temporal data variable and a time data variable; inputting the spatial data variables into a graph convolutional neural network, and extracting data spatial features by using the graph convolutional neural network on the basis of the congestion index of the window moment transport topology network; inputting the spatio-temporal data variables into a three-dimensional convolutional neural network, calculating features from time and space dimensions, and obtaining spatio-temporal features of dynamic data with space-time dependence; inputting the time data variable into a long short-term memory network to extract time characteristics of the data; and fusing the spatial features, the spatial-temporal features and the time features, and inputting the fused features into a long-short-term memory network for delay identification to obtain the delay time of the take-off flight. The airport delay can be effectively and accurately identified.

Description

Technical Field

[0001] The present invention relates to the technical field of airport flight delay prediction, and in particular to an airport flight delay prediction method based on multi-dimensional data fusion. Background Art

[0002] Airport flight delays are mainly caused by a variety of factors. In addition to common factors such as weather, air traffic control and airport ground operation management, there are also factors affecting the operation of aircraft at the airport, such as flight arrival and departure, ground taxiing, boarding and unloading. There are complex spatiotemporal correlations between these factors. Related research at home and abroad mainly uses classification models, traditional machine learning, deep neural networks and other methods to predict airport flight takeoff and landing delays, and has been well applied in the research of short-term delay prediction models. Based on the multi-dimensional attribute data of airport aircraft operations, on the basis of the existing airport delay prediction model, considering the time, space and spatiotemporal characteristics of the interaction between aircraft and other flights at the airport and different ground resources, a new airport delay prediction model is constructed to further improve the accuracy and efficiency of real-time prediction data of airport flight delays.

[0003] The increasing use of large-scale data sets and sensor technology has driven airport operations and management into a revolutionary information age. The applicability of large-scale data enables the Airport Operation Control Center (AOCC) to capture the potential characteristics of complex dynamic systems from the perspective of aviation big data, such as aircraft trajectories, weather, network systems, etc. The diversity of data provides more comprehensive and intelligent data support for accurately predicting airport flight delays.

[0004] However, dynamic system data has the characteristics of subject complexity and interdependence, dimensional diversity, and spatial-temporal dependency. The AOCC system includes subsystems such as people, equipment, and weather. These factors may affect flight delays independently or have a compound impact on flight delays. Due to the complexity and interdependence between subsystems, the extracted flight delay data will be affected by multiple factors, some of which are difficult to be directly used in functional models, such as human stress response and weather mutation. In the AOCC system, the research object of airport departure flight delays is the movement of aircraft, which leads to an increase in data dimensions, including spatial dimensions (i.e., aircraft operation status and location), time series dimensions (i.e., aircraft operation process, corresponding weather, delay time, crew, passengers and other related characteristics) and non-time series dimensions (i.e., aircraft equipment status and operators, etc.). The increase in flight dimensions reduces the accuracy and efficiency of existing flight delay prediction models. The dynamic movement of the aircraft means that the system state is continuous, and the time state and position state are simultaneously related, so there is a spatial and temporal dependency between the delay state and the aircraft's past operation state. Summary of the invention

[0005] The present invention aims to solve at least one of the technical problems existing in the related art. To this end, the present invention provides an airport flight delay prediction method based on multi-dimensional data fusion, combines a three-dimensional convolutional neural network, a graph convolutional neural network and a long short-term memory network to build a recognition model, generates input variables of spatial data variables, spatiotemporal data variables and time data variables from multi-attribute data, inputs the spatial data variables into the graph convolutional neural network to extract data spatial features; inputs the spatiotemporal data variables into the three-dimensional convolutional neural network to obtain the spatiotemporal features of dynamic data with space-time dependence; inputs the time data variables into the long short-term memory network to extract the time features of the data; the spatial features, spatiotemporal features and time features are fused and input into the long short-term memory network for delay identification, and the airport departure flight delay time is used as the output variable, which can effectively and accurately identify airport delays.

[0006] The present invention provides an airport flight delay prediction method based on multi-dimensional data fusion, comprising: S1: By collecting the time points of the aircraft in the door closing stage, wheel block removal stage, runway waiting stage, takeoff clearance stage and take-off stage, and calculating the aircraft congestion index, the multi-attribute data of the aircraft operation status including spatial characteristics, temporal characteristics and spatiotemporal characteristics are selected; S2: Convert multi-attribute data into spatial data variables, spatiotemporal data variables, and time data variables; S3: Input the spatial data variables into the graph convolutional neural network, and use the graph convolutional neural network to extract the spatial characteristics of the actual congestion of the target road section at each window time based on the congestion index of the transportation topology network at the window time; S4: Input the spatiotemporal data variables into the three-dimensional convolutional neural network, calculate the features from the time and space dimensions, and obtain the spatiotemporal features of dynamic data with space-time dependence; S5: Input the time data variable into the long short-term memory network to extract the time features of the data; S6: The spatial features, spatiotemporal features and time features are integrated and input into the long short-term memory network for delay identification to obtain the departure flight delay time.

[0007] According to the present invention, a method for predicting airport flight delays based on multidimensional data fusion is provided, which further includes estimating the airport congestion degree according to the ratio of the number of aircraft to the operating capacity within a fixed time interval in the step S1, and calculating the aircraft congestion index. The calculation expression of the aircraft congestion index is: in, is the aircraft congestion index, is the time interval, for The movement status value of the aircraft in the air or on the ground during the period, For the operating capacity, for During this period, the aircraft was in state The duration of sliding or moving, For flights in All aircraft in operation during the period, is the number of aircraft.

[0008] According to the method for predicting airport flight delays based on multidimensional data fusion provided by the present invention, the extraction of spatial data variables in the step S2 includes: S21: The network topology is constructed by taking the latitude and longitude coordinates of the airport parking apron, taxiway, runway and airway as points and the aircraft congestion status as the network formed by the connecting arcs. The network topology expression is: in, is the network topology, is a set of aerial channel nodes, , is the number of nodes, is a set of edges; S22: Define the adjacency matrix based on the network topology and the feature matrix ; Adjacency Matrix For the connection between air channels, , the adjacency matrix only contains elements of 0 and 1, 0 means there is no connection between the two channels, and 1 means there is a connection between the two channels; the congestion index information of each road section is used as the attribute characteristics of the network node to form a feature matrix , ,in is the number of node attribute features; S23: Based on network topology and the feature matrix The learning mapping function , calculate the next Congestion index at the moment: in, is the length of the time series to be predicted, for The congestion index of each channel at the moment; S24: Calculate the next Congestion index at the time; S25: Extract spatial data based on the structural layout of the airport and the congestion index of each road section of the airport.

[0009] According to the present invention, a method for predicting airport flight delays based on multidimensional data fusion is provided, which also includes converting the door closing stage, wheel block removal stage, runway waiting stage, takeoff clearance stage and liftoff stage of the aircraft operation into a transportation grid in the said S2 step, and adding the dynamic state of the aircraft on the basis of the transportation grid to obtain spatiotemporal data variables.

[0010] According to the present invention, a method for predicting airport flight delays based on multidimensional data fusion also includes converting the factors affecting each flight delay in crew information, airline information, aircraft information and part of the weather information into digital features through one-hot encoding in step S2 to obtain time data variables.

[0011] According to the method for predicting airport flight delays based on multidimensional data fusion provided by the present invention, the method further includes extracting spatial features through a graph convolutional neural network in the step S3, and the expression of the graph convolutional neural network is: in, For the The node feature matrix of the layer is a The matrix of is the number of nodes, is the dimension of node features, is the adjacency matrix, is the adjacency matrix The normalized version of is the identity matrix, for The degree matrix of For the The weight matrix of the layer is used to linearly transform the node features. Sigmoid function is the nonlinear model.

[0012] According to the present invention, a method for predicting airport flight delays based on multidimensional data fusion further includes the step S4 of performing three-dimensional convolution in the convolution stage of the two-dimensional convolutional neural network to calculate features from the time and space dimensions; connecting the feature map in the convolution layer to multiple adjacent frames in the previous layer to obtain airport aircraft sliding information, and extracting the time and space features through the three-dimensional convolutional neural network. The calculation expression is: in, For the Tier The feature map is in coordinates ( ), For the Tier The bias of the feature map, is the pair connected to the current feature map The feature map set in the layer is indexed, For the Tier Features connected to the Layer The kernel coordinates of the feature maps The value of For the The height of the layer core, For the The width of the layer core, For the The size of the 3D kernel of the layer along the time dimension, For the Tier The feature map is in coordinates ( ) value.

[0013] According to the method for predicting airport flight delays based on multidimensional data fusion provided by the present invention, the method further includes step S5, wherein the long short-term memory network structure includes an input layer, a stacked hidden layer and an output layer, and the gate mechanism of the input gate, the forget gate and the output gate as well as the cell state and the hidden state are used to extract the time feature; Input layer input Crew, aircraft and weather at the time and Momentary memory information, The memory information of the moment includes The hidden state at the moment and The cell state at a moment in time, The calculation expression for the retention degree of historical information controlled by the forget gate is: in, for The forget gate controls the degree of retention of historical information. for The weight matrix of for The paranoia, for Crew, aircraft and weather information at the moment, for The hidden state of the moment; Control the retention degree of the current input information. The calculation expression is: in, for The degree of retention of the gate control information input at all times, for The weight matrix of for of bigotry; Establish candidate vectors, the calculation expression of candidate vectors is: in, is the candidate vector, is the weight matrix of the Tanh function, is the bias of the Tanh function; merge Moment and The state at the moment is used to obtain the output gate state. The output gate state expression is: in, for Output door status at all times, for Input the state of the door at all times, is the Hadama product; The calculation expression of the retention degree of output gate control information is: in, for The degree of retention of the output gate control information at all times, for The weight matrix of for of bigotry; The hidden state calculation expression of the output gate is: in, for Output the hidden state of the gate at all times.

[0014] According to the present invention, a method for predicting airport flight delays based on multidimensional data fusion further includes using an L2 regularization method to calculate a loss function on a mean square error. The calculation expression of the loss function on the mean square error is: in, is the loss function on the mean square error, For the The actual value of the aircraft, For the The predicted value for each aircraft, is the number of aircraft, is the proportion of L2 regularization in the total loss, For the The trainable weights of the layer.

[0015] According to the present invention, a method for predicting airport flight delays based on multidimensional data fusion is provided, which also includes using mean absolute error, root mean square error and mean absolute percentage error to evaluate the performance of the method. The mean absolute error calculation expression is: in, is the mean absolute error, For the The actual value of the aircraft, For the The predicted value for each aircraft, is the number of aircraft; The root mean square error calculation expression is: in, is the root mean square error; The mean absolute percentage error calculation expression is: in, is the mean absolute percentage error.

[0016] The above one or more technical solutions in the embodiments of the invention have at least one of the following technical effects: The present invention provides an airport flight delay prediction method based on multidimensional data fusion, which combines a three-dimensional convolutional neural network, a graph convolutional neural network and a long short-term memory network to build a recognition model, generates input variables of spatial data variables, spatiotemporal data variables and time data variables from multi-attribute data, inputs the spatial data variables into the graph convolutional neural network to extract data spatial features; inputs the spatiotemporal data variables into the three-dimensional convolutional neural network to obtain the spatiotemporal features of dynamic data with space-time dependence; inputs the time data variables into the long short-term memory network to extract the time features of the data; and fuses the spatial features with the spatiotemporal features and the time features and inputs them into the long short-term memory network for delay recognition, and takes the airport take-off flight delay time as the output variable, so as to effectively and accurately identify airport delays.

[0017] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0019] Figure 1 It is a flow chart of an airport flight delay prediction method based on multi-dimensional data fusion provided by the present invention.

[0020] Figure 2 It is the loss result of six optimizers with a learning rate of 0.01 for an airport flight delay prediction method based on multidimensional data fusion provided by the present invention.

[0021] Figure 3 It is the loss result of six optimizers with a learning rate of 0.001 for an airport flight delay prediction method based on multidimensional data fusion provided by the present invention.

[0022] Figure 4 These are the loss results of different learning rates of the Adam optimizer of the airport flight delay prediction method based on multidimensional data fusion provided by the present invention.

[0023] Figure 5 These are the loss results of different learning rates of the Nadam optimizer of the airport flight delay prediction method based on multidimensional data fusion provided by the present invention.

[0024] Figure 6 The present invention provides an airport flight delay prediction method based on multi-dimensional data fusion and provides prediction results of each model for each month from January to June.

[0025] Figure 7 The present invention provides an airport flight delay prediction method based on multi-dimensional data fusion and provides prediction results of each model for each month from July to December.

[0026] Figure 8 These are the prediction results of each model in each quarter of the airport flight delay prediction method based on multi-dimensional data fusion provided by the present invention.

[0027] Fig. 9 The invention provides an airport flight delay prediction method based on multi-dimensional data fusion and provides prediction results of each model every half year and every year.

[0028] Fig.10 The invention provides a method for predicting airport flight delays based on multi-dimensional data fusion and prediction results of each model of a single aircraft runway.

[0029] Fig.11 The invention provides a method for predicting airport flight delays based on multi-dimensional data fusion and prediction results of dual-aircraft runway models. DETAILED DESCRIPTION

[0030] In order to make the purpose, technical scheme and advantages of the present invention clearer, the technical scheme in the present invention will be clearly and completely described below. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present invention. The following embodiments are used to illustrate the present invention, but cannot be used to limit the scope of the present invention.

[0031] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the embodiment of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0032] Combine the following Figures 1 to 11 The present invention describes an airport flight delay prediction method based on multi-dimensional data fusion.

[0033] like Figure 1 As shown, a method for predicting airport flight delays based on multidimensional data fusion includes: S1: By collecting the time points of the aircraft in the door closing stage, wheel block removal stage, runway waiting stage, takeoff clearance stage and take-off stage, and calculating the aircraft congestion index, the multi-attribute data of the aircraft operation status including spatial characteristics, temporal characteristics and spatiotemporal characteristics are selected; In this embodiment, flight data of Beijing Capital International Airport (PEK) was collected. The research period was from December 1, 2017 to November 30, 2018. The data sources available for analysis can be divided into two categories: flight data and meteorological information. The operation data of PEK Airport in 2018 was collected, with a total of 279,822 flight information records. For each record, specific flight information was extracted from the data set, including arrival and departure date (UTC time), flight ID, aircraft type, apron area, runway information, historical time of aircraft taxiing record points, arrival and departure points, arrival and departure directions, and delay duration. Airport meteorological information collected report data on the weather conditions of Beijing Capital International Airport (PEK) in 2018, including airport code, UTC time, hourly average temperature, wind direction, wind speed, precipitation, thunderstorms, fog, snowfall, haze and other weather phenomena.

[0034] Through the layout analysis of the runways, taxiways and parking stands of the Capital Airport, the parking stands are first divided into 14 parking stand areas according to the area names. On the other hand, 152 taxi structure data with network graph topology characteristics are generated at the same time. In the graph topology data, there are mainly two data subsets. The first data subset is the total time from the apron to takeoff or the total time from landing to the parking stand represented by each edge data. The second data subset is the spatial association matrix data with 0 and 1 characteristics. When the value is 1, it proves that the two edges have at least one taxi intersection. The selected data includes spatial features, temporal features and spatiotemporal features. As shown in Table 1, the spatial feature data values ​​are spatially variable, temporally static or relatively static data, including the spatial locations of departure points, arrival points, aprons, the moving distance between runway and taxi record points, and the aircraft queue congestion index within a fixed time interval. Temporal characteristic data are data that change over time and are static in space, including date, time, crew information, and weather conditions; spatiotemporal characteristic data refer to data that change in both time and space dimensions and have time and space dependencies and correlations. Spatiotemporal data include aircraft operating status data.

[0035] Table 1 Description of spatial features, temporal features and spatiotemporal features

[0036] There are five stages of aircraft operation at the airport, namely the door closing stage, the wheel block removal stage, the runway waiting stage, the takeoff clearance release stage and the liftoff stage. The time points of each stage of each flight are recorded through information collection. Among them, the door closing time point refers to the time when the aircraft door is closed after passengers, cargo and luggage are boarded; the wheel block removal time point is the operation time from the completion of engine start-up to taxiing preparation and the maintenance to remove the wheel blocks; the runway waiting time point refers to the time when the aircraft taxis on the ground to the runway waiting position; the takeoff clearance release time point is the time when the airport tower issues the release permit after receiving the captain's takeoff application; the liftoff time refers to the time when the aircraft taxis from the runway to the takeoff abandonment critical value and lifts the front wheels.

[0037] The airport congestion degree is estimated based on the ratio of the number of aircraft to the operating capacity within a fixed time interval, and the aircraft congestion index is calculated. The calculation expression of the aircraft congestion index is: in, is the aircraft congestion index, is the time interval, for The movement status value of the aircraft in the air or on the ground during the period, For the operating capacity, for During this period, the aircraft was in state The duration of sliding or moving, For flights in All aircraft in operation during the period, is the number of aircraft.

[0038] S2: Convert multi-attribute data into spatial data variables, spatiotemporal data variables, and time data variables; The extraction of spatial data variables includes: S21: The network topology is constructed by taking the latitude and longitude coordinates of the airport parking apron, taxiway, runway and airway as points and the aircraft congestion status as the network formed by the connecting arcs. The network topology expression is: in, is the network topology, is a set of aerial channel nodes, , is the number of nodes, is a set of edges; S22: Define the adjacency matrix based on the network topology Feature Matrix ; Adjacency Matrix For the connection between air channels, , the adjacency matrix only contains elements of 0 and 1, 0 means there is no connection between the two channels, and 1 means there is a connection between the two channels; the congestion index information of each road section is used as the attribute characteristics of the network node to form a feature matrix , ,in is the number of node attribute features; S23: Based on network topology and the feature matrix The learning mapping function , calculate the next Congestion index at the moment: in, is the length of the time series to be predicted, is the congestion index of each channel; S24: Calculate the next Congestion index at the time; S25: Extract spatial data based on the structural layout of the airport and the congestion index of each road section of the airport.

[0039] In this embodiment, the airport's structural layout and the congestion index of each section of the airport are data with spatial characteristics. The network formed by taking the longitude and latitude coordinates of the airport parking apron, taxiway, runway and airway as points and the aircraft congestion status as connecting arcs can effectively extract the spatial characteristics of the data and form the basis of the neural network topology.

[0040] The airport structure layout has 3 runways, 15 apron areas, 87 taxiways, and 59 airways. To construct the neural network input graph, the airport's ground structure, runway distribution, and airway structure graphs are converted into the network structure graph shown in the figure, and a topological structure graph is constructed based on the aircraft congestion status. When constructing the influencing factors of delay prediction, the congestion index data is the spatial feature data of each section, which describes the density of aircraft on the section within a fixed time interval. The connection characteristics of the arcs in the network are used to reflect the density. Therefore, when constructing the congestion index characteristics, the nodes in the topological network are defined as the connection arcs in the original network, and the connection lines between the nodes are the directions of aircraft operation. Node 1 represents the section information between apron PW1 and runway 36R. The edge between node 1 and node 65 represents the waiting point of runway 36R. In the final constructed topological graph, nodes 1-93 are the congestion of the ground channel, and nodes 94-152 are the congestion of the air channel.

[0041] The door closing phase, wheel block removal phase, runway waiting phase, takeoff clearance phase and liftoff phase of the aircraft operation are converted into transportation grids, and the dynamic state of the aircraft is added on the basis of the transportation grid to obtain spatiotemporal data variables.

[0042] The factors that affect each flight delay in crew information, airline information, aircraft information and some weather information are converted into digital features through one-hot encoding to obtain time data variables.

[0043] S3: Input the spatial data variables into the graph convolutional neural network, and use the graph convolutional neural network to extract the spatial characteristics of the actual congestion of the target road section at each window time based on the congestion index of the transportation topology network at the window time; Spatial feature extraction is performed through graph convolutional neural network, and the expression of graph convolutional neural network is: in, For the The node feature matrix of the layer is a The matrix of is the number of nodes, is the dimension of node features, is the adjacency matrix, is the adjacency matrix The normalized version of is the identity matrix, for The degree matrix of For the The weight matrix of the layer is used to linearly transform the node features. Sigmoid function is the nonlinear model.

[0044] In this embodiment, the congestion index of a road section is spatially affected by the congestion index of an adjacent road section. The congestion index will be relatively stable within a fixed time interval (within 1 to 10 minutes), but within the overall flight operation range, it will still be affected by the congestion conditions at historical times. Therefore, the feature extraction of the target road section congestion index needs to take into account its spatiotemporal characteristics. The transportation network established based on the airport distribution structure can well reflect the congested road section data source. Based on the congestion index of the window time transportation topology network, the GCN neural network is used to extract the spatial features of the actual congestion of the target road section at each window time, and the actual congestion status of the target road section at the output window time is mapped.

[0045] Given an adjacency matrix and the feature matrix , the GCN model constructs a filter in the Fourier domain. The filter acts on the nodes in the graph, captures the spatial features between nodes through its first-order neighborhood, and then builds the GCN model by stacking multiple convolutional layers.

[0046] S4: Input the spatiotemporal data variables into the three-dimensional convolutional neural network, calculate the features from the time and space dimensions, and obtain the spatiotemporal features of dynamic continuous mobile data with space-time dependence; In the convolution stage of the two-dimensional convolutional neural network (2DCNN), three-dimensional convolution is performed to calculate features from the time and space dimensions. The feature map in the convolution layer is connected to multiple adjacent frames in the previous layer to obtain the airport aircraft sliding information. The calculation expression for extracting spatiotemporal features through 3D-CNN is: in, For the Tier The feature map is in coordinates ( ), For the Tier The bias of the feature map, is the pair connected to the current feature map The feature map set in the layer is indexed, For the Tier Features connected to the Layer The kernel coordinates of the feature maps The value of For the The height of the layer core, For the The width of the layer core, For the The size of the 3D kernel of the layer along the time dimension, For the Tier The feature map is in coordinates ( ) value.

[0047] In this embodiment, three-dimensional convolution is performed in the convolution stage of a two-dimensional convolutional neural network (2DCNN) to calculate features from both time and space dimensions. The three-dimensional convolution is achieved by convolving a three-dimensional kernel into a cube formed by stacking multiple consecutive frames together. Through this construction, the feature map in the convolution layer is connected to multiple adjacent frames in the previous layer, thereby obtaining airport aircraft sliding information.

[0048] S5: Input the time data variable into the long short-term memory network to extract the time features of the data; Use the LSTM algorithm to extract time features, such as flight features, weather features, etc. The LSTM structure includes an input layer, a stacked hidden layer, and an output layer. It uses the gate mechanism of the input gate, forget gate, and output gate as well as the cell state and hidden state to extract time features; Input layer input Crew, aircraft and weather at the time and Momentary memory information, The memory information of the moment includes The hidden state at the moment and The cell state at a moment in time, The calculation expression for the retention degree of historical information controlled by the forget gate is: in, for The forget gate controls the degree of retention of historical information. for The weight matrix of for The paranoia, for Crew, aircraft and weather information at the moment, for The hidden state of the moment; Control the retention degree of the current input information. The calculation expression is: in, for The degree of retention of the gate control information input at all times, for The weight matrix of for of bigotry; Establish candidate vectors, the calculation expression of candidate vectors is: in, is the candidate vector, is the weight matrix of the Tanh function, is the bias of the Tanh function; merge Moment and The state at the moment is used to obtain the output gate state. The output gate state expression is: in, for Output door status at all times, for Input the state of the door at all times, is the Hadama product; The calculation expression of the retention degree of output gate control information is: in, for The degree of retention of the output gate control information at all times, for The weight matrix of for of bigotry; The hidden state calculation expression of the output gate is: in, for Output the hidden state of the gate at all times.

[0049] S6: The spatial features, spatiotemporal features and time features are integrated and input into the long short-term memory network for delay identification to obtain the departure flight delay time.

[0050] The L2 regularization method is used to calculate the loss function on the mean square error. The loss function calculation expression on the mean square error is: in, is the loss function on the mean square error, For the The actual value of the aircraft, For the The predicted value for each aircraft is is the number of aircraft, is the proportion of L2 regularization in the total loss, For the The trainable weights of the layer.

[0051] The performance of the method was evaluated using mean absolute error, root mean square error, and mean absolute percentage error. The mean absolute error calculation expression is: in, is the mean absolute error, For the The actual value of the aircraft, For the The predicted value for each aircraft is is the number of aircraft; The root mean square error calculation expression is: in, is the root mean square error; The mean absolute percentage error calculation expression is: in, is the mean absolute percentage error.

[0052] In this embodiment, the data of delays and key factors affecting delays at Beijing Capital International Airport (PEK) are obtained by sorting the data. In order to verify the performance of the proposed 3DF-DSCL model, the selected benchmark methods including LSTM, CNN, GCN, GCN+LSTM, 2D CNN+LSTM, 3D CNN+LSTM, and 2D+CNN+GCN+LSTM were tested and compared. All benchmark methods were trained and tested using the same input variables to ensure the comparability of the models. The present invention compares the model results using the Nadam optimizer and five widely used optimizers, namely Adam, Adagrad, SGD, Adamax, and Adadelta. First, the batch_size is set to 32. MSE is used as the loss function, and the loss function using L2 regularization is shown, and each epoch is set to 200. Five different optimizers under different learning rates (LR) are summarized. As Figure 2 The loss results of the six optimizers with a learning rate of 0.01 are shown below. Figure 3 The loss results of the six optimizers with a learning rate of 0.001 are shown below. Figure 4 The loss results of the Adam optimizer with different learning rates are shown in Figure 5 As shown, the loss results of Nadam optimizer with different learning rates are: Figures 2 to 5 It shows that the lossy convergence speed using SGD, Adagard and Adadelta optimizers is quite slow compared to the other three optimizers. For Adam, Adamax and Nadam, their convergence speed and training performance are similar. By comparing Nadam and Adam, we can determine a suitable value for the test performance at different learning rates. As can be seen from the figure, the performance of the two optimizers at different learning rates is similar. Therefore, in the following experiments, the Nadam optimizer was used.

[0053] In order to verify the performance of the proposed 3DF-DSCL model, several benchmark methods are also tested and compared. The selected benchmark methods include LSTM, ARIMA, CNN, GCN, GCN+LSTM, 3D CNN+LSTM, and 2D CNN+GCN+LSTM, which are commonly used in previous short-term time series data forecasting studies.

[0054] All benchmark methods are trained and tested with the same input variables to ensure model comparability. In order to verify the performance of the proposed 3DF-DSCL model, commonly used methods are also used as test cases as benchmarks. All benchmark methods are trained and tested with the same input variables to ensure model comparability.

[0055] As shown in Table 2, the MAE, MAPE and RMSE values ​​of each benchmark model are shown. Through analysis, among the LSTM, CNN, GCN, GCN+LSTM, 2D CNN+LSTM, 3D CNN+LSTM, 2D+CNN+GCN+LSTM and 3DF-DSCL methods, the proposed 3DF-DSCL method has achieved the most stable and satisfactory performance in terms of MAE, MAPE and RMSE. The performance of 2DCNN+GCN+LSTM, GCN+LSTM and 3DCNN+LSTM is far better than other methods because they integrate LSTM and GCN (CNN) structures and consider spatiotemporal features at the same time. In addition, 3DF-DSCL is more effective than 2DCNN+GCN+LSTM because it is more accurate in extracting dynamic data. Compared with the 2D CNN+GCN+LSTM model, 3DF-DSCL has the smallest MAPE, RMSE and MAE values.

[0056] Table 2 MAE, MAPE and RMSE values ​​of the benchmark model

[0057] Figures 6 to 11 The prediction results between 3DF-DSCL and 2DF-DSCL models under different data scales and data dimensions are shown. Figure 6 The prediction results of each model for each month from January to June are: Figure 7 The prediction results of each model for each month from July to December are: Figure 8 For each model prediction result for each quarter, Fig. 9 The prediction results of each model for each half year and each year are as follows: Fig.10 The prediction results of each model for a single runway are: Fig.11 The prediction results of each model for the dual runway are shown in Figure 2. In a specific prediction scenario, the MAE values ​​of the 3DF-DSCL method are all less than 0.6, which proves the robustness of the model to different data scales and dimensions. In addition, it can be observed that the prediction accuracy of large-scale datasets is relatively higher than that of small datasets, which shows that the proposed model is more suitable for large-scale datasets. Overall, in all test cases, 3DF-DSCL has the smallest MAPE, RMSE, and MAE values ​​compared to the 2DF-DSCL model.

[0058] The beneficial effects of the present invention are as follows: the present invention combines 3D CNN, GCN and LSTM to construct a recognition model, generates input variables of spatial data variables, spatiotemporal data variables and time data variables from multi-attribute data, inputs the spatial data variables into GCN to extract data spatial features; inputs the spatiotemporal data variables into 3D CNN to obtain spatiotemporal features of dynamic continuous mobile data with space-time dependency; inputs the time data variables into LSTM to extract the time features of the data; the spatial features, spatiotemporal features and time features are fused and input into LSTM for delay identification, and the airport departure flight delay time is used as the output variable, so as to effectively and accurately identify airport delays.

[0059] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting airport flight delays based on multidimensional data fusion, characterized in that: include: S1: By collecting the time points of the aircraft in the door closing stage, wheel block removal stage, runway waiting stage, takeoff clearance stage and take-off stage, and calculating the aircraft congestion index, the multi-attribute data of the aircraft operation status including spatial characteristics, temporal characteristics and spatiotemporal characteristics are selected; S2: Convert multi-attribute data into spatial data variables, spatiotemporal data variables, and time data variables; S3: Input the spatial data variables into the graph convolutional neural network, and use the graph convolutional neural network to extract the spatial characteristics of the actual congestion of the target road section at each window time based on the congestion index of the transportation topology network at the window time; S4: Input the spatiotemporal data variables into the three-dimensional convolutional neural network, calculate the features from the time and space dimensions, and obtain the spatiotemporal features of dynamic data with space-time dependence; S5: Input the time data variable into the long short-term memory network to extract the time features of the data; S6: The spatial features, spatiotemporal features and time features are integrated and input into the long short-term memory network for delay identification to obtain the departure flight delay time.

2. The method for predicting airport flight delays based on multidimensional data fusion according to claim 1, characterized in that: In the step S1, the airport congestion level is estimated based on the ratio of the number of aircraft to the operating capacity within a fixed time interval, and the aircraft congestion index is calculated. The calculation expression of the aircraft congestion index is: in, is the aircraft congestion index, is the time interval, for The movement status value of the aircraft in the air or on the ground during the period, For the operating capacity, for During this period, the aircraft was in state The duration of sliding or moving, For flights in All aircraft in operation during the period, is the number of aircraft.

3. The method for predicting airport flight delays based on multidimensional data fusion according to claim 1, characterized in that: In step S2, the extraction of spatial data variables includes: S21: The network topology is constructed by taking the latitude and longitude coordinates of the airport parking apron, taxiway, runway and airway as points and the aircraft congestion status as the network formed by the connecting arcs. The network topology expression is: in, is the network topology, is a set of aerial channel nodes, , is the number of nodes, is a set of edges; S22: Define the adjacency matrix based on the network topology and the feature matrix ; Adjacency Matrix For the connection between air channels, , the adjacency matrix only contains elements of 0 and 1, 0 means there is no connection between the two channels, and 1 means there is a connection between the two channels; the congestion index information of each road section is used as the attribute characteristics of the network node to form a feature matrix , ,in is the number of node attribute features; S23: Based on network topology and the feature matrix The learning mapping function , calculate the next Congestion index at the moment: in, is the length of the time series to be predicted, for The congestion index of each channel at the moment; S24: Calculate the next Congestion index at the time; S25: Extract spatial data based on the structural layout of the airport and the congestion index of each road section of the airport.

4. The method for predicting airport flight delays based on multidimensional data fusion according to claim 1, characterized in that: In the step S2, the door closing phase, wheel block removal phase, runway waiting phase, takeoff clearance phase and liftoff phase of the aircraft operation are converted into a transportation grid, and the dynamic state of the aircraft is added on the basis of the transportation grid to obtain the spatiotemporal data variables.

5. The method for predicting airport flight delays based on multidimensional data fusion according to claim 1, characterized in that: In the step S2, the factors affecting each flight delay in the crew information, airline information, aircraft information and part of the weather information are converted into digital features through one-hot encoding to obtain time data variables.

6. The method for predicting airport flight delays based on multidimensional data fusion according to claim 1, characterized in that: In the step S3, spatial features are extracted by using a graph convolutional neural network, and the expression of the graph convolutional neural network is: in, For the The node feature matrix of the layer is a The matrix of is the number of nodes, is the dimension of node features, is the adjacency matrix, is the adjacency matrix The normalized version of is the identity matrix, for The degree matrix of For the The weight matrix of the layer is used to linearly transform the node features. Sigmoid function is the nonlinear model.

7. The method for predicting airport flight delays based on multidimensional data fusion according to claim 1, characterized in that: In the step S4, three-dimensional convolution is performed in the convolution stage of the two-dimensional convolutional neural network to calculate features from the time and space dimensions; the feature map in the convolution layer is connected to multiple adjacent frames in the previous layer to obtain airport aircraft sliding information, and the calculation expression for extracting spatiotemporal features through the three-dimensional convolutional neural network is: in, For the Tier The feature map is in coordinates ( ), For the Tier The bias of the feature map, is the pair connected to the current feature map The feature map set in the layer is indexed, For the Tier Features connected to the Layer The kernel coordinates of the feature maps The value of For the The height of the layer core, For the The width of the layer core, For the The size of the 3D kernel of the layer along the time dimension, For the Tier The feature map is in coordinates ( ) value.

8. The method for predicting airport flight delays based on multidimensional data fusion according to claim 1, characterized in that: In the step S5, the long short-term memory network structure includes an input layer, a stacked hidden layer and an output layer, and uses the gate mechanism of the input gate, the forget gate and the output gate as well as the cell state and the hidden state to extract the time feature; Input layer input Crew, aircraft and weather at the time and Momentary memory information, The memory information of the moment includes The hidden state at the moment and The cell state at a moment in time, The calculation expression for the retention degree of historical information controlled by the forget gate is: in, for The forget gate controls the degree of retention of historical information. for The weight matrix of for The paranoia, for Crew, aircraft and weather information at the moment, for The hidden state of the moment; Control the retention degree of the current input information. The calculation expression is: in, for The degree of retention of the gate control information input at all times, for The weight matrix of for of bigotry; Establish candidate vectors, the calculation expression of candidate vectors is: in, is the candidate vector, is the weight matrix of the Tanh function, is the bias of the Tanh function; merge Moment and The state at the moment is used to obtain the output gate state. The output gate state expression is: in, for Output door status at all times, for Input the state of the door at all times, is the Hadama product; The calculation expression of the retention degree of output gate control information is: in, for The degree of retention of the output gate control information at all times, for The weight matrix of for of bigotry; The hidden state calculation expression of the output gate is: in, for Output the hidden state of the gate at all times.

9. The method for predicting airport flight delays based on multidimensional data fusion according to claim 1, characterized in that: The L2 regularization method is used to calculate the loss function on the mean square error. The loss function calculation expression on the mean square error is: in, is the loss function on the mean square error, For the The actual value of the aircraft, For the The predicted value for each aircraft, is the number of aircraft, is the proportion of L2 regularization in the total loss, For the The trainable weights of the layer.

10. The method for predicting airport flight delays based on multidimensional data fusion according to claim 1, characterized in that: The performance of the method was evaluated using mean absolute error, root mean square error, and mean absolute percentage error. The mean absolute error calculation expression is: in, is the mean absolute error, For the The actual value of the aircraft, For the The predicted value for each aircraft, is the number of aircraft; The root mean square error calculation expression is: in, is the root mean square error; The mean absolute percentage error calculation expression is: in, is the mean absolute percentage error.

Images