A regional air traffic forecasting method based on Mamba-GCN

Abstract

This application provides a regional-level air traffic flow prediction method based on Mamba-GCN, belonging to the technical field of air traffic flow prediction. The method comprises: constructing a Mamba-GCN collaborative network model; collecting historical trajectory data of the target airspace, dividing the target airspace into spatial grids, counting the number of aircraft in each spatial grid, and encoding the temporal and spatial characteristics of each aircraft to construct a spatiotemporal tensor; constructing a dynamic adjacency matrix and a dynamic weight graph based on the 8-neighborhood topology of the spatial grid; inputting the spatiotemporal tensor and dynamic weight graph into the Mamba-GCN collaborative network model for training and optimizing model parameters; and preprocessing the real-time flight trajectory data of the target airspace before inputting it into the trained Mamba-GCN collaborative network model to obtain an air traffic flow prediction result. The present invention can capture the long-term dependencies of air traffic flow in the temporal dimension and the grid associations in the spatial dimension, and has high prediction efficiency.

Description

Technical Field

[0001] The present application belongs to the technical field of air traffic flow prediction, and specifically relates to a regional air traffic flow prediction method based on Mamba-GCN. Background Art

[0002] With the continuous growth of global air passenger traffic, accurate prediction of air traffic flow has become the key to optimizing airspace resource allocation and preventing air congestion. Traditional air traffic flow prediction mainly relies on statistical methods, traditional machine learning methods, and deep learning methods.

[0003] Statistical methods primarily employ time series analysis models, such as the autoregressive moving average (ARMA), the autoregressive integrated moving average (ARIMA), and the seasonal autoregressive integrated moving average (SARIMA). While these models can capture the periodicity of time series data, they ignore the spatial dependencies between adjacent regions and are unable to effectively capture the spatiotemporal coupling of flights within regional airspaces. With the development of artificial intelligence, traditional machine learning methods such as wavelet neural networks, support vector machines, and Bayesian networks have been introduced to air traffic flow forecasting. While these machine learning methods have improved forecasting capabilities to a certain extent, they struggle to reveal underlying spatiotemporal correlations in load when processing large-scale data, limiting their predictive performance. Deep learning methods have also been applied to traffic flow prediction. Models such as the spatiotemporal graph convolutional network (STGCN), GraphWaveNet, and the adaptive graph convolutional recurrent network (AGCRN) achieve joint modeling of the spatiotemporal characteristics of traffic data by combining graph convolutional networks (GCN) with temporal convolutional networks (TCN) or recurrent neural networks (RNN). However, these deep learning methods have the following problems: First, graph convolutional networks (GCN) rely on a fixed adjacency matrix and cannot capture the dynamic changes in spatiotemporal topology as flight flows change. Second, temporal convolutional networks (TCNs) have a limited receptive field and have difficulty capturing long-term dependencies. Although the Transformer can model long-term dependencies, its quadratic computational complexity limits its application in large-scale grid scenarios.

[0004] To address the above issues, existing studies have attempted to introduce dynamic graph generation mechanisms (such as DGCRN and D2STNN) or improve pre-training frameworks (such as STDMAE), but there are still problems of high computational complexity or loss of spatial relationship accuracy, which cannot meet the needs of spatial traffic flow prediction. Summary of the Invention

[0005] In a first aspect, an embodiment of the present application provides a regional air traffic forecasting method based on Mamba-GCN, comprising the following steps:

[0006] S1. Use the Mamba model and the GCN model to build a Mamba-GNC collaborative network model;

[0007] S2. Collect historical ADS-B track data for the target airspace, divide the target airspace into spatial grids according to a preset spatial resolution, count the number of aircraft in each spatial grid within a preset time granularity, and encode the temporal and spatial characteristics of each aircraft to construct a space-time tensor;

[0008] S3. Define the basic connectivity between grids based on the 8-neighborhood topology of the spatial grid. Use a Gaussian kernel function to map geographic distances to initial connection weights. Dynamically adjust these initial connection weights based on real-time traffic correlation to generate a dynamic adjacency matrix and dynamic weight graph.

[0009] S4. Input the spatiotemporal tensor and dynamic weight map into the Mamba-GCN collaborative network model for training and optimizing model parameters;

[0010] S5. After preprocessing the real-time flight trajectory data of the target airspace, the data is input into the trained Mamba-GCN collaborative network model to obtain the flight traffic forecast results within the set time period of the target airspace.

[0011] Furthermore, in step S1, the Mamba-GNC collaborative network model is constructed by interactively stacking the Mamba model and the GCN model and introducing an auxiliary layer;

[0012] The Mamba-GNC collaborative network model is provided with a first Mamba model layer, a GCN model layer, a first linear transformation layer, a Relu activation layer, a second Mamba model layer, a temporal convolution layer and a second linear transformation layer in sequence from the input side to the output side;

[0013] First Mamba model layer, GCN model layer, first linear transformation layer, Relu activation layer, second Mamba model layer, temporal convolution layer, and second linear transformation layer;

[0014] The first Mamba model layer is used to model the long-term dependency of the input spatiotemporal tensor in the time dimension through a selective state space model and output the first hidden state sequence;

[0015] The GCN model layer is used to aggregate spatial features of the first hidden state sequence based on the dynamic adjacency matrix, and uses Chebyshev polynomials to propagate neighborhood information and output spatially enhanced features;

[0016] The first linear transformation layer is used to adjust the dimension of the spatial enhancement features to match the ReLU activation layer;

[0017] The ReLU activation layer is used to enhance the expressive power of the spatial enhancement features after dimension adjustment using nonlinear transformation;

[0018] The second Mamba model layer is used to perform secondary temporal dependency modeling on the spatially enhanced features after the expressiveness is enhanced, and to extract high-order temporal features;

[0019] The temporal convolution layer is used to expand the temporal receptive field through one-dimensional dilated convolution, extract and fuse the local temporal dimension of the extracted high-order temporal features, and obtain local temporal features;

[0020] The second linear transformation layer is used to map the local temporal features to the output dimension.

[0021] Furthermore, in step S1, the first Mamba model layer and the second Mamba model layer use the selective state space model to construct the state transition equation and the output equation, and model the input as a time series as a hidden state; the specific process is as follows:

[0022] The first and second Mamba model layers use the selective state space model to construct the state transfer equation, introduce the state transfer matrix and input projection matrix, and convert the spatiotemporal tensor into a hidden state;

[0023] The first and second Mamba model layers use the selective state space model to construct the output equation, converting the spatiotemporal tensor and hidden state into output;

[0024] Reshape the space-time tensor to match the requirements of the selective state-space model;

[0025] The forward selective state space model is used to capture historical time node information from the reshaped spatiotemporal tensor to obtain the forward hidden state sequence;

[0026] The subsequent selective state space model captures the future time node information from the reshaped spatiotemporal tensor to obtain the backward hidden state sequence;

[0027] Linearly concatenate the forward hidden state sequence and the backward hidden state sequence in the feature dimension to obtain a bidirectional concatenated hidden state. Then, use a dimensionality-reduced linear transformation operation to map the bidirectional concatenated hidden state to a space that matches the output dimension to obtain the output hidden state.

[0028] The predicted flow value is extracted from the spatiotemporal tensor and the output hidden state using the output equation.

[0029] Furthermore, in step S1, the GCN model layer aggregates the spatial features of the first hidden state sequence based on the dynamic adjacency matrix and uses Chebyshev polynomials to propagate neighborhood information. The specific process of outputting spatial enhanced features is as follows:

[0030] Introducing the identity matrix to the dynamic adjacency matrix and adding self-loops;

[0031] Calculate the degree of each grid node in the dynamic adjacency matrix after adding the self-loop to obtain the degree matrix, and then calculate the dynamic square root inverse of the degree matrix;

[0032] Construct a dynamic normalized aggregation matrix using the dynamic adjacency matrix after adding self-loops and the dynamic square root inverse of the degree matrix;

[0033] Obtain the node features of each time step from the spatiotemporal tensor to obtain the node feature matrix of each time step;

[0034] Combine the node feature matrix with the dynamic adjacency matrix and aggregate it using the dynamic normalized aggregation matrix to obtain the feature matrix and complete the spatial feature enhancement;

[0035] The weight matrix is ​​introduced into the first linear transformation layer and multiplied with the aggregated feature matrix to complete the linear transformation;

[0036] In the Relu activation layer, the activation function is used to perform nonlinear activation on the linearly transformed feature matrix to obtain the node feature matrix of the next time step.

[0037] Furthermore, the specific steps of step S2 are as follows:

[0038] S21. Determine the target spatial area, preset spatial resolution, and preset time granularity;

[0039] S22. Obtain the ADS-B historical track data of the target airspace and perform data cleaning to parse out the number of aircraft, latitude and longitude, and time;

[0040] S23. Divide the target spatial area into spatial grids according to a preset spatial resolution and using a two-dimensional plane projection method, and record the number of grids N;

[0041] S24. Sum the number of aircraft in each spatial grid at each moment according to the preset time granularity;

[0042] S25. Determine the position of each moment in the timeline of a day according to the time granularity as the moment-day feature, and encode it to obtain the moment-day feature code tod;

[0043] S26. Determine the position of the day to which each moment belongs in the timeline of a week according to the time granularity as a day-week feature, and encode it to obtain a day-week feature code dow;

[0044] S27. After normalizing the longitude and latitude of each grid at each moment in the ADS-B historical track data, obtain the longitude feature and the latitude feature, and encode them respectively to obtain the longitude encoding feature lon and the latitude encoding feature lat;

[0045] S28. The time-day feature code tod and the day-week feature code dow are used as time features, and the longitude feature code lon and the latitude feature code lat are used as spatial features;

[0046] S29. The number of aircraft in the target space area at each moment, the time-day feature tod encoding, the day-week feature dow encoding, the longitude feature lon encoding and the latitude feature lat encoding are used as a spatiotemporal tensor.

[0047] Furthermore, the specific steps of data cleaning in step S22 are as follows:

[0048] S221. Obtain the ADS-B historical track data of the target airspace, the ADS-B historical track data including the aircraft icao number, latitude and longitude, altitude, ground speed and time;

[0049] S222. Threshold constraints are set according to ICAO standards. Track data with acceleration exceeding the upper threshold is marked as abnormal track points and removed. Track data with vertical speed exceeding the upper threshold is marked as abnormal altitude change data and removed.

[0050] S223. Use cubic spline interpolation to resample the original second-level track data to minute granularity, complete the alignment of the track data in the time dimension, and obtain a standardized track data set.

[0051] Furthermore, the specific steps of step S3 are as follows:

[0052] S31. Define the connection relationship of each grid in the target airspace according to the 8-neighborhood rule to obtain an initial topology;

[0053] S32. Based on the initial topology, define the graph edge structure with each grid as a node and the actual geographical distance between grids as an edge;

[0054] S33. For the associated grid connected to each grid, the initial connection weight is calculated by the Gaussian kernel function, and the initial connection weight is added to the edge of the graph edge structure;

[0055] S34. The number of aircraft in the spatiotemporal tensor of the target airspace is collected as historical traffic data, and the traffic time series correlation of the rows and columns of the associated grid is calculated based on the historical traffic data;

[0056] S35. Dynamically adjust the connection weights on the edges of the graph edge structure based on the actual geographical distance and traffic time series correlation between each grid and the corresponding associated grid to obtain a dynamic adjacency matrix, and generate an N×N dimensional dynamic weight graph for each time step; where N represents the number of grids in the target airspace.

[0057] Furthermore, the specific steps of step S4 are as follows:

[0058] S41. Construct a loss function based on the difference between the predicted air traffic volume and the actual air traffic volume of the Mamba-GCN collaborative network model;

[0059] S42. Input the spatiotemporal tensor and dynamic weight map of the target spatial domain into the Mamba-GCN collaborative network model for training, and optimize the parameters of the Mamba-GCN collaborative network model with the goal of minimizing the loss function during the training process;

[0060] S43. The training is completed until the loss function converges or the preset number of iterations is reached.

[0061] Furthermore, the specific steps of step S5 are as follows:

[0062] S51. The generation process of the spatiotemporal tensor and the generation process of the dynamic weight map are used as input to the network;

[0063] S52. Provide the real-time acquired target airspace aviation trajectory data to the input network for preprocessing to obtain a spatiotemporal tensor and a dynamic weight map;

[0064] S53. Input the spatiotemporal tensor and dynamic weight map into the trained Mamba-GCN collaborative network model to predict the air traffic flow within the target airspace within the set time period;

[0065] S54. Convert the predicted air traffic into a predicted traffic heat map using the output network map.

[0066] Furthermore, the prediction result of the air traffic in step S53 is represented in the form of a spatiotemporal tensor and a dynamic weight map;

[0067] In step S54, after the spatial features in the spatiotemporal tensor of the predicted aviation traffic are denormalized, spatial coordinate mapping is performed, and a heat map of aircraft traffic at each predicted time step is generated using a visual output method.

[0068] It can be seen from the above technical solutions that this application has the following advantages:

[0069] The regional air traffic forecasting method based on Mamba-GCN provided in this application solves the problems of insufficient dynamic spatial topology modeling and low efficiency in capturing long-term dependencies in regional air traffic forecasting by constructing a Mamba-GCN collaborative network model and integrating spatiotemporal tensors and dynamic weight graphs. This method can predict air traffic flow and provide support for optimizing airspace resource allocation and preventing air congestion. Through the synergy of the Mamba model and the GCN model, the long-term dependencies in the time dimension and the dynamic topological structure in the spatial dimension are considered at the same time, thereby improving the accuracy and efficiency of the forecast. The use of ADS-B historical track data for model training can fully utilize existing actual flight data and enhance the reliability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] In order to more clearly illustrate the technical solution of the present application, the following is a brief introduction to the drawings required for the description. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0071] Figure 1 Schematic diagram of the flow of the regional air traffic forecasting method based on Mamba-GCN of the present invention.

[0072] Figure 2 Schematic diagram of the network structure of the Mamba-GNC collaborative network model of the present invention.

[0073] Figure 3 Schematic diagram of grid coding rules in an embodiment of the present invention.

[0074] Figure 4 This is a heat map of aircraft traffic at a moment in the data set in an embodiment of the present invention.

[0075] Figure 5 Schematic diagram of aircraft traffic statistics at each time step in a day according to the present invention.

[0076] Figure 6 This is a schematic diagram of aircraft traffic on each day of the week at the present invention point.

[0077] Figure 7 A schematic diagram is generated for the graph topology structure of the present invention; wherein (a) represents a spatial grid schematic diagram, (b) represents a schematic diagram for defining nodes and edges, and (c) represents a topology structure schematic diagram.

[0078] Figure 8 Schematic diagram of the input network of the present invention.

[0079] Figure 9 Schematic diagram of the output network of the present invention.

[0080] Figure 10 Schematic diagram of the overall processing framework of the present invention. DETAILED DESCRIPTION

[0081] The various embodiments of the present disclosure will be described more fully below in detail in the specific steps of the regional air traffic forecasting method based on Mamba-GCN. The present disclosure is capable of various embodiments, and modifications and variations therein are possible. However, it should be understood that there is no intention to limit the various embodiments of the present disclosure to the specific embodiments disclosed herein. Instead, the present disclosure is to be construed as encompassing all modifications, equivalents, and / or alternatives falling within the spirit and scope of the various embodiments of the present disclosure.

[0082] For example, with the continuous increase in global air passenger volume, accurately predicting air traffic flow has become increasingly critical for optimizing airspace resource allocation and preventing air congestion. Traditionally, air traffic flow prediction methods mainly rely on statistical methods, traditional machine learning methods, and deep learning methods.

[0083] Statistical methods often rely on time series analysis models, such as the autoregressive moving average (ARMA), autoregressive integrated moving average (ARIMA), and seasonal autoregressive integrated moving average (SARIMA). While these models can capture the periodic characteristics of time series data, they are inadequate for handling spatiotemporal coupling within regional airspaces because they ignore the spatial dependencies between adjacent regions. With advances in artificial intelligence (AI), traditional machine learning methods such as wavelet neural networks, support vector machines, and Bayesian networks have also been applied to air traffic flow forecasting. While these methods have improved predictive capabilities to some extent, they are insufficient in uncovering underlying spatiotemporal correlations when faced with large-scale data, thus limiting their performance. Deep learning methods have also been widely used in the field of traffic flow forecasting. Models such as the spatiotemporal graph convolutional network (STGCN), GraphWaveNet, and the adaptive graph convolutional recurrent network (AGCRN) combine graph convolutional networks (GCNs) with temporal convolutional networks (TCNs) or recurrent neural networks (RNNs) to jointly model the spatiotemporal characteristics of traffic data. However, these deep learning methods also have some problems: on the one hand, graph convolutional networks (GCNs) rely on a fixed adjacency matrix, which makes them unable to capture the spatiotemporal topological structure that changes dynamically with flight flows; on the other hand, the temporal convolutional network (TCN) has a limited receptive field and has difficulty capturing long-term dependencies. Although the Transformer model can model long-term dependencies, its quadratic computational complexity limits its application in large-scale grid scenarios.

[0084] To address these issues, existing studies have attempted to introduce dynamic graph generation mechanisms (e.g., DGCRN and D2STNN) or improve pre-training frameworks such as STDMAE. However, these methods still have some drawbacks, such as high computational complexity or loss of spatial relationship accuracy, which makes it difficult to meet the high-precision requirements of air traffic flow forecasting.

[0085] To address the above issues, this embodiment provides a regional air traffic forecasting method based on Mamba-GCN. By constructing a Mamba-GCN collaborative network model and integrating spatiotemporal tensors and dynamic weight graphs, it effectively addresses the problems of insufficient dynamic spatial topology modeling and low efficiency in capturing long-term dependencies in regional air traffic forecasting.

[0086] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0087] See also Figure 1 FIG. 1 is a flow chart of a method for regional air traffic forecasting based on Mamba-GCN in a specific embodiment. The method includes the following steps:

[0088] S1. Use the Mamba model and the GCN model to build a Mamba-GNC collaborative network model;

[0089] It should be noted that the Mamba-GCN collaborative network model integrates the advantages of time series modeling and spatial feature aggregation. By using the Mamba model and the GCN model, it can simultaneously capture long-term dependencies in the temporal dimension and complex topological structures in the spatial dimension, improving the model's ability to model the spatiotemporal coupling relationships in air traffic data and providing a foundation for subsequent predictions.

[0090] S2. Collect historical ADS-B track data for the target airspace, divide the target airspace into spatial grids according to a preset spatial resolution, count the number of aircraft in each spatial grid within a preset time granularity, and encode the temporal and spatial characteristics of each aircraft to construct a space-time tensor;

[0091] It should be noted that the collection of ADS-B historical track data and the construction of a space-time tensor capture the spatiotemporal characteristics of the data; the division of the target airspace into spatial grids and the counting of aircraft, while encoding the temporal and spatial characteristics, can comprehensively describe the spatiotemporal distribution of air traffic, providing high-quality input data for model training and improving prediction accuracy;

[0092] S3. Define the basic connectivity between grids based on the 8-neighborhood topology of the spatial grid. Use a Gaussian kernel function to map geographic distances to initial connection weights. Dynamically adjust these initial connection weights based on real-time traffic correlation to generate a dynamic adjacency matrix and dynamic weight graph.

[0093] It should be noted that a dynamic adjacency matrix and a dynamic weight graph are generated to capture the dynamic correlation of traffic between grids. The basic connection relationship is defined based on the 8-neighborhood topology of the spatial grid, and the weights are dynamically adjusted based on geographic distance and real-time traffic correlation. This can timely reflect the changing trends and spatial correlations of flight traffic in the airspace, enabling the model to adapt to the dynamic air traffic environment.

[0094] S4. Input the spatiotemporal tensor and dynamic weight map into the Mamba-GCN collaborative network model for training and optimizing model parameters;

[0095] It should be noted that training the Mamba-GCN collaborative network model ensures the optimization of model parameters and the improvement of prediction performance. By inputting the spatiotemporal tensor and dynamic weight map into the model for training, and optimizing the model parameters with the goal of minimizing the loss function, the model can fully learn the spatiotemporal characteristics and traffic patterns in the data, thereby making the prediction results accurate and reliable.

[0096] S5. After preprocessing the real-time target airspace flight trajectory data, input it into the trained Mamba-GCN collaborative network model to obtain the air traffic forecast results for the target airspace within the set time period;

[0097] It should be noted that real-time prediction of air traffic in the target airspace achieves accurate prediction of air traffic flow with high prediction efficiency; pre-processing the real-time aviation trajectory data and inputting it into the trained model can quickly output the air traffic forecast results within the target time period, providing timely decision-making support for aviation management departments, facilitating the optimization of airspace resource allocation and the prevention of air congestion.

[0098] This example builds a Mamba-GNC collaborative network model, leveraging the time series modeling capabilities of the Mamba model and the spatial feature aggregation capabilities of the GCN model to capture the spatiotemporal characteristics of air traffic flow. Combining ADS-B data processing, spatiotemporal tensor construction, and dynamic adjacency matrix generation, this model transforms actual air traffic data into effective information suitable for model training, enabling the model to learn the complex patterns of air traffic flow in real scenarios and providing a foundation for prediction.

[0099] Furthermore, as a refinement and extension of the above-mentioned embodiment, and to fully illustrate the specific implementation process of this embodiment, another regional air traffic forecasting method based on Mamba-GCN is provided. The method includes the following steps:

[0100] S1. Use the Mamba model and the GCN model to build a Mamba-GNC collaborative network model; in step S1, the Mamba model and the GCN model are interactively stacked and an auxiliary layer is introduced to build the Mamba-GNC collaborative network model;

[0101] The Mamba-GNC collaborative network model is provided with a first Mamba model layer, a GCN model layer, a first linear transformation layer, a Relu activation layer, a second Mamba model layer, a temporal convolution layer and a second linear transformation layer in sequence from the input side to the output side;

[0102] First Mamba model layer, GCN model layer, first linear transformation layer, Relu activation layer, second Mamba model layer, temporal convolution layer, and second linear transformation layer;

[0103] It should be noted that the Mamba model layer is good at capturing temporal dimension dependencies, while the GCN model layer focuses on spatial dimension modeling. The Mamba-GCN collaborative network adopts the strategy of interactive stacking of Mamba and GCN to deeply integrate spatiotemporal features. The network structure of the Mamba-GCN collaborative network model is as follows: Figure 2 As shown;

[0104] S2. Collect historical ADS-B track data for the target airspace, divide the target airspace into spatial grids according to a preset spatial resolution, count the number of aircraft in each spatial grid within a preset time granularity, and encode the temporal and spatial characteristics of each aircraft to construct a space-time tensor. The specific steps of step S2 are as follows:

[0105] S21. Determine the target spatial area, preset spatial resolution, and preset time granularity;

[0106] S22. Obtain the historical ADS-B track data of the target airspace and perform data cleaning to parse out the number of aircraft, latitude and longitude, and time. The specific steps for data cleaning in step S22 are as follows:

[0107] S221. Obtain the ADS-B historical track data of the target airspace, the ADS-B historical track data including the aircraft icao number, latitude and longitude, altitude, ground speed and time;

[0108] It should be noted that the icao number is the four-letter identification code for the airport and airline;

[0109] S222. Threshold constraints are set according to ICAO standards. Track data with acceleration exceeding the upper threshold is marked as abnormal track points and removed. Track data with vertical speed exceeding the upper threshold is marked as abnormal altitude change data and removed.

[0110] Specifically, threshold constraints are set according to ICAO standards, eliminating abnormal track points with acceleration exceeding 4.9m / s² and removing abnormal altitude change data with vertical rate exceeding 500ft / min, thereby ensuring data accuracy and reliability.

[0111] S223. Using cubic spline interpolation, resample the original second-level track data to minute granularity, align the track data in the time dimension, and obtain a standardized track data set;

[0112] For example, in terms of spatial dimension, the target airspace (22°-32°N, 112°-122°E) is divided into Divide the grid cells (about 27.5km×27.5km) and generate a unique spatial identifier for each grid according to the Earth Space Grid Coding Rules to achieve multi-level scalable indexing under the global spatial reference system. The grid coding rules are as follows: Figure 3 As shown;

[0113] In the time dimension, with 1 minute as the statistical window, the number of ICAO numbers in each grid is calculated as the flow characterization value to form a spatiotemporal flow matrix; in order to adapt to the input requirements of the traffic prediction model, the data storage adopts the PEMs data set standard format to form an aircraft flow data file and an adjacency matrix definition file respectively; the aircraft flow data file stores the aircraft flow data of all grids at all time steps, and the data dimension is (10080, 1600, 1), where 10080 represents the total time step of the data, 1600 represents the total number of grids, and 1 represents one channel feature of each grid at each time step, i.e., aircraft flow; the aircraft flow heat map corresponding to a certain moment in the data set is visualized as shown in the figure. Figure 4 As shown;

[0114] It should be noted that the spline interpolation method is used to overcome the discreteness of the track data in the time series. Each record contains time information, ICAO number, longitude and latitude, and ground speed information, and the time resolution is unified to 1 minute.

[0115] The track data after cleaning is shown in Table 1:

[0116] Table 1: Track data table

[0117]

[0118] S23. Divide the target spatial area into spatial grids according to a preset spatial resolution and using a two-dimensional plane projection method, and record the number of grids N;

[0119] S24. Sum the number of aircraft in each spatial grid at each moment according to the preset time granularity;

[0120] S25. Determine the position of each moment in the timeline of a day according to the time granularity as the moment-day feature, and encode it to obtain the moment-day feature code tod;

[0121] S26. Determine the position of the day to which each moment belongs in the timeline of a week according to the time granularity as a day-week feature, and encode it to obtain a day-week feature code dow;

[0122] S27. After normalizing the longitude and latitude of each grid at each moment in the ADS-B historical track data, obtain the longitude feature and the latitude feature, and encode them respectively to obtain the longitude encoding feature lon and the latitude encoding feature lat;

[0123] S28. The time-day feature code tod and the day-week feature code dow are used as time features, and the longitude feature code lon and the latitude feature code lat are used as spatial features;

[0124] S29. The number of aircraft in the target space area at each moment, the time-day feature tod code, the day-week feature dow code, the longitude feature lon code, and the latitude feature lat code are used as a spatiotemporal tensor;

[0125] For example, the target area is divided into spatial grids with a resolution of 0.25°×0.25°, and the number of aircraft in each grid at each moment is counted at a time granularity of 1 minute. In this way, the ADS-B track data is converted into a 3D tensor of shape T×N×D, where T represents the historical time step, N represents the number of grids in the target area, and D represents the number of features of each grid, i.e., aircraft traffic.

[0126] After being converted into spatiotemporal data format, spatiotemporal feature encoding is performed according to the spatiotemporal characteristics of ADS-B data. In terms of time characteristics, firstly, based on the periodicity of track data, such as Figure 5 and Figure 6 As shown in the figure, the position of each time step in the timeline of a day is calculated according to the time granularity, and then the position of the day corresponding to each time step in the week is calculated according to the time granularity, and represented by tod (time of day) and dow (day of week), respectively. Tod and dow both belong to [0,1]. Tod and dow are used as the second and third dimensional features of each grid respectively. In terms of spatial features, the longitude and latitude of the center point of each grid are normalized and represented by lon and lat respectively. Both lon and lat belong to [0,1]. lon and lat are used as the fourth and fifth dimensional features of each grid respectively.

[0127] S3. Define the basic connection relationship between grids based on the 8-neighborhood topology of the spatial grid. Use the Gaussian kernel function to map geographic distances to initial connection weights. Dynamically adjust the initial connection weights based on real-time traffic correlation to generate a dynamic adjacency matrix and dynamic weight graph. The specific steps of step S3 are as follows:

[0128] S31. Define the connection relationship of each grid in the target airspace according to the 8-neighborhood rule to obtain an initial topology;

[0129] S32. Based on the initial topology, define the graph edge structure with each grid as a node and the actual geographical distance between grids as an edge;

[0130] S33. For the associated grid connected to each grid, the initial connection weight is calculated by the Gaussian kernel function, and the initial connection weight is added to the edge of the graph edge structure;

[0131] Specifically, the basic connection relationship is predefined according to the grid 8 neighborhood topology, such as Figure 7 As shown in the figure, the value of the edge between nodes represents the actual geographical distance between the grid center points. When dynamic weight adjustment is performed, the connection strength of the edge is adjusted without changing the connection relationship between nodes. The weight value is usually inversely proportional to the distance. The closer the distance, the greater the weight. Therefore, the Gaussian kernel function is first used to map the geographical distance to the weight. The specific mapping formula is as follows:

[0132]

[0133] in, Representation Grid and Geographic Euclidean distance of the center point; Represents a learnable bandwidth parameter that controls the geographic attenuation rate. Represents the initial connection weight, which is a static value;

[0134] S34. The number of aircraft in the spatiotemporal tensor of the target airspace is collected as historical traffic data, and the traffic time series correlation of the rows and columns of the associated grid is calculated based on the historical traffic data;

[0135] Specifically, the traffic time series correlation calculation formula is as follows:

[0136]

[0137] in, It represents the correlation coefficient of the flow time series between grid i and grid j at time step t, and characterizes the correlation of flow time series between grids; Representation grid i At time step τ The flow observation value, Representation grid i In the time window arrive t The mean of the flow observations within ; Representation Grid j At time step τ The flow observation value, Representation grid j In the time window arrive t The mean of the flow observations within ; τ represents the time step variable, which is used to traverse from time steps to t; t represents the current time step; T represents the time window length, which is used to determine the number of historical time steps considered when calculating the correlation;

[0138] S35. Dynamically adjust the connection weights on the edges of the graph edge structure based on the actual geographic distance and traffic time series correlation between each grid and the corresponding associated grid, obtain a dynamic adjacency matrix, and generate an N×N dimensional dynamic weight graph for each time step; where N represents the number of grids in the target airspace;

[0139] Specifically, the formula for dynamically adjusting the connection weight is as follows:

[0140]

[0141] in, Indicates the connection weight after dynamic adjustment, which is a dynamic value; is the learnable weight coefficient, satisfying ;

[0142] S4. Input the spatiotemporal tensor and dynamic weight map into the Mamba-GCN collaborative network model for training and optimizing model parameters; the specific steps of step S4 are as follows:

[0143] S41. Construct a loss function based on the difference between the predicted air traffic volume and the actual air traffic volume of the Mamba-GCN collaborative network model;

[0144] For example, the target airspace is divided into As an example, each grid has a time step The flow observation value is ; For a given history The observation sequence of time steps ,in ,The goal of the traffic prediction task is to learn the mapping function , predicting the future The flow sequence of time steps:

[0145]

[0146] in, Indicates the The predicted flow of all grids at the time step;

[0147] The model is optimized by minimizing the loss between the predicted value and the true value. The loss function is defined as follows:

[0148]

[0149] S42. Input the spatiotemporal tensor and dynamic weight map of the target spatial domain into the Mamba-GCN collaborative network model for training, and optimize the parameters of the Mamba-GCN collaborative network model with the goal of minimizing the loss function during the training process;

[0150] S43. Training is completed until the loss function converges or the preset number of iterations is reached;

[0151] S5. After preprocessing the real-time target airspace aviation trajectory data, input the trained Mamba-GCN collaborative network model to obtain the target airspace air traffic forecast results within the set time period; the specific steps of step S5 are as follows:

[0152] S51. The generation process of the spatiotemporal tensor and the generation process of the dynamic weight map are used as input to the network;

[0153] For example, the input network is Figure 8 As shown;

[0154] S52. Provide the real-time acquired target airspace aviation trajectory data to the input network for preprocessing to obtain a spatiotemporal tensor and a dynamic weight map;

[0155] S53. Input the spatiotemporal tensor and dynamic weight map into the trained Mamba-GCN collaborative network model to predict the air traffic flow within the target airspace within the set time period;

[0156] S54. Convert the predicted air traffic flow using the output network map into a predicted traffic heat map;

[0157] The prediction result of the air traffic in step S53 is represented in the form of a spatiotemporal tensor and a dynamic weight map;

[0158] In step S54, after the spatial features in the spatiotemporal tensor of the predicted air traffic are denormalized, spatial coordinate mapping is performed, and a heat map of the aircraft traffic at each predicted time step is generated using a visual output method;

[0159] For example, the output network is as follows Figure 9 As shown;

[0160] Combining the input network, Mamba-GNC collaborative network model and output network, the overall processing framework diagram is as follows Figure 10 As shown;

[0161] The first Mamba model layer in step S1 is used to model the long-term dependency of the input spatiotemporal tensor in the time dimension through a selective state space model and output a first hidden state sequence;

[0162] The GCN model layer is used to aggregate spatial features of the first hidden state sequence based on the dynamic adjacency matrix, and uses Chebyshev polynomials to propagate neighborhood information and output spatially enhanced features;

[0163] The first linear transformation layer is used to adjust the dimension of the spatial enhancement features to match the ReLU activation layer;

[0164] The ReLU activation layer is used to enhance the expressive power of the spatial enhancement features after dimension adjustment using nonlinear transformation;

[0165] The second Mamba model layer is used to perform secondary temporal dependency modeling on the spatially enhanced features after the expressiveness is enhanced, and to extract high-order temporal features;

[0166] The temporal convolution layer is used to expand the temporal receptive field through one-dimensional dilated convolution, extract and fuse the local temporal dimension of the extracted high-order temporal features, and obtain local temporal features;

[0167] The second linear transformation layer is used to map the local temporal features to the output dimension;

[0168] In step S1, the first and second Mamba model layers use the selective state space model to construct the state transition equation and output equation, and model the input as a time series as a hidden state. The specific process is as follows:

[0169] The first and second Mamba model layers use the selective state space model to construct the state transfer equation, introduce the state transfer matrix and input projection matrix, and convert the spatiotemporal tensor into a hidden state;

[0170] The first and second Mamba model layers use a selective state space model and introduce a state projection matrix and an output projection matrix to construct an output equation, converting the spatiotemporal tensor and hidden state into output;

[0171] Reshape the space-time tensor to match the requirements of the selective state-space model;

[0172] The forward selective state space model is used to capture historical time node information from the reshaped spatiotemporal tensor to obtain the forward hidden state sequence;

[0173] The subsequent selective state space model captures the future time node information from the reshaped spatiotemporal tensor to obtain the backward hidden state sequence;

[0174] Linearly concatenate the forward hidden state sequence and the backward hidden state sequence in the feature dimension to obtain a bidirectional concatenated hidden state. Then, use a dimensionality-reduced linear transformation operation to map the bidirectional concatenated hidden state to a space that matches the output dimension to obtain the output hidden state.

[0175] Extract the predicted flow value from the spatiotemporal tensor and the output hidden state using the output equation;

[0176] Specifically, the Mamba model layer captures the long-range dependency of time series through the selective state space model SSM. Compared with the local convolution of TCN, it can model the global long-range dependency. Compared with Transformer, it reduces the time complexity from down to , with higher computational efficiency, effectively improving the model's training and reasoning capabilities;

[0177] The state space model SSM models the time series as a hidden state through the state equation. The state transition equation is as follows:

[0178]

[0179] The output equation is as follows:

[0180]

[0181] in, is the state transition matrix, 、 、 are the input projection matrix, state projection matrix and output projection matrix respectively, represents the space-time tensor at the t-th time step, represents the hidden state at the tth time step;

[0182] When the Mamba module models time series, it first needs to convert the spatiotemporal tensor Reshape into , and then the forward and backward SSM are used to capture the historical and future time node information respectively to obtain the bidirectional hidden state. Finally, the bidirectional hidden state is linearly spliced ​​and the dimension is reduced to output as shown below;

[0183] The forward SSM is expressed as follows:

[0184]

[0185] The backward SSM is expressed as follows:

[0186]

[0187] The bidirectional hidden state linear concatenation is expressed as follows:

[0188]

[0189] in, Dynamic parameters that are input dependent;

[0190] In step S1, the GCN model layer aggregates the spatial features of the first hidden state sequence based on the dynamic adjacency matrix and uses Chebyshev polynomials to propagate neighborhood information. The specific process of outputting spatial enhanced features is as follows:

[0191] Introducing the identity matrix to the dynamic adjacency matrix and adding self-loops;

[0192] By self-looping, it ensures that all nodes include their own characteristics in the aggregation range;

[0193] Calculate the degree of each grid node in the dynamic adjacency matrix after adding the self-loop to obtain the degree matrix, and then calculate the dynamic square root inverse of the degree matrix;

[0194] Construct a dynamic normalized aggregation matrix using the dynamic adjacency matrix after adding self-loops and the dynamic square root inverse of the degree matrix;

[0195] Obtain the node features of each time step from the spatiotemporal tensor to obtain the node feature matrix of each time step;

[0196] Combine the node feature matrix with the dynamic adjacency matrix and aggregate it using the dynamic normalized aggregation matrix to obtain the feature matrix and complete the spatial feature enhancement;

[0197] The weight matrix is ​​introduced into the first linear transformation layer and multiplied with the aggregated feature matrix to complete the linear transformation;

[0198] In the Relu activation layer, the activation function is used to perform nonlinear activation on the linearly transformed feature matrix to obtain the node feature matrix of the next time step;

[0199] Specifically, after calculating the dynamic weights, Chebyshev polynomials are used to achieve efficient spatial feature aggregation. The specific formula is as follows:

[0200]

[0201] in, represents the node features at the t-th time step, , is the number of nodes, is the characteristic dimension of the node; represents the degree matrix of the node, , is a diagonal matrix, where the elements on the diagonal line represent the degree of each node after adding self-loops (i.e., the number of connected edges, including self-loops); , is the dynamic-weight adjacency matrix after adding the self-loop; represents the weight matrix of the t-th time step; represents the activation function; represents the dynamic square root inverse of the degree matrix, D Take the reciprocal of the square root of each diagonal element of ; represents the dynamic normalized aggregation matrix;

[0202] It should be noted that when the Mamba-GNC collaborative network model is constructed, the parameters of the first Mamba model layer and the second Mamba model layer are initialized respectively, including the state transfer matrix, input projection matrix, state projection matrix, and output projection matrix; the graph convolution kernel weight matrix of the GCN model layer is initialized; the weights and biases of the first linear transformation layer and the second linear transformation layer are initialized; and the convolution kernel weights of the temporal convolution layer are initialized.

[0203] The output of the first Mamba model layer is transformed into a channel number through the first linear transformation layer and then input into the ReLU layer for nonlinear activation to obtain an activated feature map. The activated feature map is input into the GCN model layer, and spatial features are aggregated using the predefined initial adjacency matrix to obtain a spatially aggregated feature map. The spatially aggregated feature map is input into the second Mamba model layer to capture the temporal dimension dependency and output a time series feature map. The time series feature map is input into the temporal convolution layer to extract local features in the temporal dimension and obtain the prediction result.

[0204] A first linear transformation layer is introduced between the first Mamba model layer and the GCN model layer to map the high-dimensional time series features output by the Mamba model to a feature space that matches the input requirements of the GCN model, aligning the temporal and spatial features in terms of dimension and ensuring smooth data flow between different model layers. The ReLU activation layer receives the output of the first linear transformation layer and performs nonlinear activation on the data to enhance the nonlinear expression capability of the model. Its output serves as the input of the GCN model layer. A temporal convolution layer is introduced after the second Mamba model layer to extract and fuse features in the local time dimension of the time series features output by the Mamba model, capturing local correlations and trends in the time series. Its output is mapped to the dimensional space of the prediction target through the second linear transformation layer to obtain the final air traffic forecast result.

[0205] The regional air traffic forecasting method based on Mamba-GCN provided in the embodiments of the present application can be applied to electronic devices. Those skilled in the art will understand that the electronic device structure involved in the embodiments of the present invention does not constitute a limitation of the electronic device. The electronic device may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently. In the embodiments of the present invention, the electronic device includes but is not limited to laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smart phones, wearable devices and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the embodiments of the present application described and / or required herein.

[0206] The electronic device may include a processor, an external memory interface, an internal memory, a universal serial bus (USB) interface, a charging management module, a power management module, a battery, a wireless communication module, an audio module, a speaker, a microphone, a sensor module, a button, a camera, a display, and a SIM card interface, etc.

[0207] It is understood that the structures illustrated in the embodiments of the present application do not constitute specific limitations on the electronic device. In other embodiments of the present application, the electronic device may include more or fewer components than shown, or combine or separate certain components, or arrange the components differently. The illustrated components may be implemented in hardware, software, or a combination of software and hardware.

[0208] A processor may include one or more processing units, such as a central processing unit (CPU), an application processor (AP), a modem processor, a graphics processing unit (GPU), an image signal processor (ISP), a controller, a memory, a video codec, a digital signal processor (DSP), a baseband processor, and / or a neural-network processing unit (NPU). Different processing units may be independent devices or integrated into one or more processors.

[0209] The processor can be the nerve center and command center of the electronic device. The controller can generate operation control signals based on the instruction opcode and timing signal to complete the control of instruction fetching and execution.

[0210] The processor may also include a memory for storing instructions and data. In some embodiments, the memory in the processor is a cache memory. This memory can store instructions or data that the processor has just used or is reusing. If the processor needs to use the instruction or data again, it can directly call it from the memory. This avoids repeated accesses, reduces processor latency, and thus improves system efficiency.

[0211] The above-mentioned electronic device realizes the regional-level aviation traffic prediction method based on Mamba-GCN in this application, uses the Mamba model and the GCN model to construct a Mamba-GNC collaborative network model; collects ADS-B historical track data of the target airspace, divides the target airspace into spatial grids according to a preset spatial resolution, counts the number of aircraft in each spatial grid within a preset time granularity, and encodes the time characteristics and spatial characteristics of each aircraft to construct a time-space tensor; defines the basic connection relationship between grids based on the 8-neighborhood topology of the spatial grid, maps the geographic distance to the initial connection weight through the Gaussian kernel function, and then dynamically adjusts the initial connection weight in combination with the real-time traffic correlation to generate a dynamic adjacency matrix and a dynamic weight graph; inputs the time-space tensor and the dynamic weight graph into the Mamba-GCN collaborative network model for training and optimizes the model parameters; pre-processes the aviation trajectory data of the target airspace obtained in real time, and inputs the trained Mamba-GCN collaborative network model to obtain the aviation traffic prediction result within the set time period of the target airspace, thereby achieving the technical solution of integrating Mamba The model builds a collaborative network with the GCN model, combines data processing and dynamic weight adjustment, and solves the problems of complex spatiotemporal characteristics and dynamic data changes in air traffic forecasting; it can more accurately capture the long-term dependence of air traffic in the time dimension and the grid correlation characteristics in the spatial dimension, thereby improving the accuracy and timeliness of the forecast; at the same time, the generation of dynamic adjacency matrix and weight graph enables the model to adapt to the beneficial effects of real-time changes in air traffic data.

[0212] The storage medium provided in this application stores a program product that can implement a regional-level aviation traffic prediction method based on Mamba-GCN.

[0213] The regional air traffic forecasting method based on Mamba-GCN includes: constructing a Mamba-GNC collaborative network model using the Mamba model and the GCN model; collecting historical ADS-B track data of the target airspace, dividing the target airspace into spatial grids according to a preset spatial resolution, counting the number of aircraft in each spatial grid within a preset time granularity, and encoding the temporal and spatial characteristics of each aircraft to construct a spatiotemporal tensor; defining the basic connection relationship between grids based on the 8-neighborhood topology of the spatial grid, mapping the geographic distance to the initial connection weight using a Gaussian kernel function, and dynamically adjusting the initial connection weight based on real-time traffic correlation to generate a dynamic adjacency matrix and a dynamic weight graph; inputting the spatiotemporal tensor and dynamic weight graph into the Mamba-GCN collaborative network model for training and optimizing model parameters; and preprocessing the real-time air trajectory data of the target airspace before inputting it into the trained Mamba-GCN collaborative network model to obtain the air traffic forecast results for the target airspace within a set time period.

[0214] In some possible implementations, the Mamba-GCN-based regional air traffic prediction method disclosed herein may be implemented in the form of a program product, which includes program code. When the program product is run on a terminal device, the program code is used to enable the terminal device to execute the steps of various exemplary implementations of the present disclosure described in the above “Exemplary Method” section of this specification.

[0215] The storage medium of the present disclosure can adopt any combination of one or more readable media. The readable medium can be a readable signal medium or a readable storage medium. The readable storage medium can be, for example, but not limited to, a system, device or component of electricity, magnetism, light, electromagnetic, infrared, or semiconductor, or any combination thereof. More specific examples (non-exhaustive list) of readable storage media include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof.

[0216] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A regional air traffic forecasting method based on Mamba-GCN, characterized by: The steps include: S1. Use the Mamba model and the GCN model to build a Mamba-GNC collaborative network model; In step S1, the Mamba model and the GCN model are interactively stacked and an auxiliary layer is introduced to construct a Mamba-GNC collaborative network model; The Mamba-GNC collaborative network model is provided with a first Mamba model layer, a GCN model layer, a first linear transformation layer, a Relu activation layer, a second Mamba model layer, a temporal convolution layer and a second linear transformation layer in sequence from the input side to the output side; First Mamba model layer, GCN model layer, first linear transformation layer, Relu activation layer, second Mamba model layer, temporal convolution layer, and second linear transformation layer; The first Mamba model layer is used to model the long-term dependency of the input spatiotemporal tensor in the time dimension through a selective state space model and output the first hidden state sequence; The GCN model layer is used to aggregate spatial features of the first hidden state sequence based on the dynamic adjacency matrix, and uses Chebyshev polynomials to propagate neighborhood information and output spatially enhanced features; The first linear transformation layer is used to adjust the dimension of the spatial enhancement features to match the ReLU activation layer; The ReLU activation layer is used to enhance the expressive power of the spatial enhancement features after dimension adjustment using nonlinear transformation; The second Mamba model layer is used to perform secondary temporal dependency modeling on the spatially enhanced features after the expressiveness is enhanced, and to extract high-order temporal features; The temporal convolution layer is used to expand the temporal receptive field through one-dimensional dilated convolution, extract and fuse the local temporal dimension of the extracted high-order temporal features, and obtain local temporal features; The second linear transformation layer is used to map the local temporal features to the output dimension; S2. Collect historical ADS-B track data for the target airspace, divide the target airspace into spatial grids according to a preset spatial resolution, count the number of aircraft in each spatial grid within a preset time granularity, and encode the temporal and spatial characteristics of each aircraft to construct a space-time tensor; S3. Define the basic connectivity between grids based on the 8-neighborhood topology of the spatial grid. Use a Gaussian kernel function to map geographic distances to initial connection weights. Dynamically adjust these initial connection weights based on real-time traffic correlation to generate a dynamic adjacency matrix and dynamic weight graph. S4. Input the spatiotemporal tensor and dynamic weight map into the Mamba-GCN collaborative network model for training and optimizing model parameters; S5. After preprocessing the real-time flight trajectory data of the target airspace, the data is input into the trained Mamba-GCN collaborative network model to obtain the flight traffic forecast results within the set time period of the target airspace.

2. The regional air traffic forecasting method based on Mamba-GCN according to claim 1, characterized in that: In step S1, the first and second Mamba model layers use the selective state space model to construct the state transition equation and output equation, and model the input as a time series as a hidden state. The specific process is as follows: The first and second Mamba model layers use the selective state space model to construct the state transfer equation, introduce the state transfer matrix and input projection matrix, and convert the spatiotemporal tensor into a hidden state; The first and second Mamba model layers use the selective state space model to construct the output equation, converting the spatiotemporal tensor and hidden state into output; Reshape the space-time tensor to match the requirements of the selective state-space model; The forward selective state space model is used to capture historical time node information from the reshaped spatiotemporal tensor to obtain the forward hidden state sequence; The subsequent selective state space model captures the future time node information from the reshaped spatiotemporal tensor to obtain the backward hidden state sequence; Linearly concatenate the forward hidden state sequence and the backward hidden state sequence in the feature dimension to obtain a bidirectional concatenated hidden state. Then, use a dimensionality-reduced linear transformation operation to map the bidirectional concatenated hidden state to a space that matches the output dimension to obtain the output hidden state. The predicted flow value is extracted from the spatiotemporal tensor and the output hidden state using the output equation.

3. The regional air traffic forecasting method based on Mamba-GCN according to claim 2, characterized in that: In step S1, the GCN model layer aggregates the spatial features of the first hidden state sequence based on the dynamic adjacency matrix and uses Chebyshev polynomials to propagate neighborhood information. The specific process of outputting spatial enhanced features is as follows: Introducing the identity matrix to the dynamic adjacency matrix and adding self-loops; Calculate the degree of each grid node in the dynamic adjacency matrix after adding the self-loop to obtain the degree matrix, and then calculate the dynamic square root inverse of the degree matrix; Construct a dynamic normalized aggregation matrix using the dynamic adjacency matrix after adding self-loops and the dynamic square root inverse of the degree matrix; Obtain the node features of each time step from the spatiotemporal tensor to obtain the node feature matrix of each time step; Combine the node feature matrix with the dynamic adjacency matrix and aggregate it using the dynamic normalized aggregation matrix to obtain the feature matrix and complete the spatial feature enhancement; The weight matrix is ​​introduced into the first linear transformation layer and multiplied with the aggregated feature matrix to complete the linear transformation; In the Relu activation layer, the activation function is used to perform nonlinear activation on the linearly transformed feature matrix to obtain the node feature matrix of the next time step.

4. The regional air traffic forecasting method based on Mamba-GCN according to claim 1, characterized in that: The specific steps of step S2 are as follows: S21. Determine the target spatial area, preset spatial resolution, and preset time granularity; S22. Obtain the ADS-B historical track data of the target airspace and perform data cleaning to parse out the number of aircraft, latitude and longitude, and time; S23. Divide the target spatial area into spatial grids according to a preset spatial resolution and using a two-dimensional plane projection method, and record the number of grids N; S24. Sum the number of aircraft in each spatial grid at each moment according to the preset time granularity; S25. Determine the position of each moment in the timeline of a day according to the time granularity as the moment-day feature, and encode it to obtain the moment-day feature code tod; S26. Determine the position of the day to which each moment belongs in the timeline of a week according to the time granularity as a day-week feature, and encode it to obtain a day-week feature code dow; S27. After normalizing the longitude and latitude of each grid at each moment in the ADS-B historical track data, obtain the longitude feature and the latitude feature, and encode them respectively to obtain the longitude encoding feature lon and the latitude encoding feature lat; S28. The time-day feature code tod and the day-week feature code dow are used as time features, and the longitude feature code lon and the latitude feature code lat are used as spatial features; S29. The number of aircraft in the target space area at each moment, the time-day feature tod encoding, the day-week feature dow encoding, the longitude feature lon encoding and the latitude feature lat encoding are used as a spatiotemporal tensor.

5. The regional air traffic forecasting method based on Mamba-GCN according to claim 4, characterized in that: The specific steps of data cleaning in step S22 are as follows: S221. Obtain the ADS-B historical track data of the target airspace, the ADS-B historical track data including the aircraft icao number, latitude and longitude, altitude, ground speed and time; S222. Threshold constraints are set according to ICAO standards. Track data with acceleration exceeding the upper threshold is marked as abnormal track points and removed. Track data with vertical speed exceeding the upper threshold is marked as abnormal altitude change data and removed. S223. Use cubic spline interpolation to resample the original second-level track data to minute granularity, complete the alignment of the track data in the time dimension, and obtain a standardized track data set.

6. The regional air traffic forecasting method based on Mamba-GCN according to claim 4, characterized in that: The specific steps of step S3 are as follows: S31. Define the connection relationship of each grid in the target airspace according to the 8-neighborhood rule to obtain an initial topology; S32. Based on the initial topology, define the graph edge structure with each grid as a node and the actual geographical distance between grids as an edge; S33. For the associated grid connected to each grid, the initial connection weight is calculated by the Gaussian kernel function, and the initial connection weight is added to the edge of the graph edge structure; S34. The number of aircraft in the spatiotemporal tensor of the target airspace is collected as historical traffic data, and the traffic time series correlation of the rows and columns of the associated grid is calculated based on the historical traffic data; S35. Dynamically adjust the connection weights on the edges of the graph edge structure based on the actual geographical distance and traffic time series correlation between each grid and the corresponding associated grid to obtain a dynamic adjacency matrix, and generate an N×N dimensional dynamic weight graph for each time step; where N represents the number of grids in the target airspace.

7. The regional air traffic forecasting method based on Mamba-GCN according to claim 6, characterized in that: The specific steps of step S4 are as follows: S41. Construct a loss function based on the difference between the predicted air traffic volume and the actual air traffic volume of the Mamba-GCN collaborative network model; S42. Input the spatiotemporal tensor and dynamic weight map of the target spatial domain into the Mamba-GCN collaborative network model for training, and optimize the parameters of the Mamba-GCN collaborative network model with the goal of minimizing the loss function during the training process; S43. The training is completed until the loss function converges or the preset number of iterations is reached.

8. The regional air traffic forecasting method based on Mamba-GCN according to claim 7, characterized in that: The specific steps of step S5 are as follows: S51. The generation process of the spatiotemporal tensor and the generation process of the dynamic weight map are used as input to the network; S52. Provide the real-time acquired target airspace aviation trajectory data to the input network for preprocessing to obtain a spatiotemporal tensor and a dynamic weight map; S53. Input the spatiotemporal tensor and dynamic weight map into the trained Mamba-GCN collaborative network model to predict the air traffic flow within the target airspace within the set time period; S54. Convert the predicted air traffic into a predicted traffic heat map using the output network map.

9. The regional air traffic forecasting method based on Mamba-GCN according to claim 8, characterized in that: The prediction result of the air traffic in step S53 is represented in the form of a spatiotemporal tensor and a dynamic weight map; In step S54, after the spatial features in the spatiotemporal tensor of the predicted aviation traffic are denormalized, spatial coordinate mapping is performed, and a heat map of aircraft traffic at each predicted time step is generated using a visual output method.

Images