Flow refinement prediction method driven by air traffic flow pattern recognition

Abstract

The embodiment of the invention discloses a flow refined prediction method driven by air traffic flow pattern recognition, and relates to the technical field of air traffic situation awareness and flow management. According to the method, spatial-temporal characteristics in a flight path are efficiently extracted, a local flow mode is accurately identified, a graph model is constructed based on a representative flow mode, and fine-grained prediction of traffic flow in an airspace is realized, so that the fineness and dynamic response capability of local region regulation and control are greatly improved. The method is especially suitable for deep analysis of flight path data in an airspace sector, and provides accurate auxiliary decision support for air traffic control, route planning and airspace design.

Description

Technical Field

[0001] The present invention relates to the technical field of air traffic situation awareness and flow management, and in particular to a flow refinement prediction method driven by air traffic flow pattern recognition. Background Art

[0002] With the continuous growth of global air traffic and the increasing flight density within airspaces, terminal area flight trajectory clustering and flow prediction technologies have, to a certain extent, met the needs of flight scheduling and safety regulation. However, existing flight trajectory clustering algorithms primarily target relatively fixed and well-defined flight trajectories within the terminal area. These algorithms are typically based on pre-set morphological assumptions and fixed feature extraction methods, such as distance metrics, density estimation, and hierarchical clustering. Because flight trajectories in the terminal area often follow pre-planned routes, their operational status and spatial layout are relatively stable. Therefore, these traditional clustering methods are able to effectively extract the main flow patterns within the trajectory. However, these algorithms have significant shortcomings when dealing with flight trajectories within airspace sectors. Flight trajectories within a sector are influenced by a mix of multiple phases, such as takeoff, approach, and cruise, and are highly diverse and spatiotemporally coherent. Traditional methods tend to overlook subtle differences in flow patterns, resulting in overly coarse clustering results that fail to reveal the complex flow structure within the local airspace, thus providing limited information for local airspace regulation and fine-grained management.

[0003] Meanwhile, existing air traffic flow forecasting methods focus on overall flow trends and macro-level statistics and predictions. These forecasting models typically perform statistical and time series analysis based on holistic phenomena, employing models such as the Autoregressive Integrated Moving Average (ARIMA) model, Support Vector Regression (SVR), and early deep learning models such as Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTMs) for overall flow modeling. Their primary goal is to capture overall airspace flow trends. However, these deep learning-based forecasting models fail to consider the impact of airspace structure on flow forecasting. Currently, the latest forecasting methods are mostly based on Graph Neural Networks (GNNs). By defining nodes in the airspace and the relationships between them to form a graph, GNNs can effectively capture the complex topological structure and spatial dependencies of the airspace. This approach can reflect the complexity of the actual airspace and improve forecasting accuracy. However, in these GNN-based forecasting models, nodes are typically defined as sectors or large airspaces, ignoring the spatial distribution and flow direction of individual flow patterns within the sector. Consequently, they cannot accurately reveal the specific distribution and dynamic characteristics of individual flow patterns at a fine-grained level.

[0004] In summary, while nearly all traffic flow prediction methods, such as the aforementioned existing solutions, can effectively reflect overall traffic conditions, their prediction data often lacks sufficient detailed information for local airspace management, especially when dynamic regulation is implemented at the sector level. This hinders precise air traffic management and route optimization. Furthermore, existing technologies often treat flow pattern identification and flow flow prediction as two separate research tasks. Flow pattern identification focuses on analyzing the internal structure of flight trajectories, extracting representative tracks through clustering methods, and revealing the patterns of aircraft movement within the airspace. In contrast, flow flow prediction focuses on estimating the number of future flights and their flow trends within the airspace using statistical models or deep learning methods. While the two overlap in theory and method, since flow pattern identification reveals subtle, diverse, and directional flow patterns within a local airspace, while flow flow prediction models tend to capture macro-statistical trends, effective synergy between the two approaches is difficult in terms of data representation and feature fusion, limiting the effectiveness of fine-grained airspace flow prediction.

[0005] Therefore, how to further improve the in-depth analysis scheme of flight trajectory data within airspace sectors, thereby improving the precision and dynamic response capabilities of local area control, has become a topic that needs further research. Summary of the Invention

[0006] The embodiments of the present invention provide a method for fine-grained traffic flow prediction driven by air traffic flow pattern recognition, which further improves the deep analysis scheme of flight trajectory data within airspace sectors, and can improve the fineness and dynamic response capability of local area control.

[0007] To achieve the above objectives, the embodiments of the present invention adopt the following technical solutions:

[0008] A refined air traffic flow prediction method driven by air traffic flow pattern recognition includes:

[0009] Flight trajectory data is obtained from the Automatic Dependent Surveillance-Broadcast (ADS-B) system and preprocessed. During the preprocessing process, data is separated according to the spatiotemporal characteristics of the flight trajectory data. Specific sector range data can be obtained from the Air Traffic Management System (ATM).

[0010] Low-dimensional features are extracted from the separated flight trajectory data, where the low-dimensional features of temporal attributes and spatial attributes are fused into joint features;

[0011] A graph structure model is constructed, in which representative trajectories reflecting the characteristics of each flow pattern are used as nodes in the graph structure, and the connection weights of the edges in the graph structure are associated with the flight volume between adjacent nodes. Specifically, using the obtained representative flow pattern, representative flight trajectories within each airspace sector are used as nodes, and a graph model is constructed based on historical trajectory data. The edge weights between nodes are determined by the flight volume and flow direction between adjacent nodes, intuitively reflecting the distribution and direction of traffic flow within the airspace.

[0012] Through the established parallel fusion prediction model, the air traffic flow prediction results within the future predetermined time step are obtained.

[0013] In this embodiment, the preprocessing includes: summarizing the trajectories corresponding to the flight trajectory data into various sectors; wherein the flight trajectory data within a specified airspace is intercepted, for example, one month's flight trajectory data of all sectors in the Shanghai airspace is intercepted, and the trajectories are summarized into various sectors;

[0014] The flight trajectory data is segmented according to sampling intervals and then normalized. The collected raw trajectory data can be segmented according to sampling intervals (e.g., 10 seconds), and minimum-maximum normalization is applied to each data attribute, mapping each attribute value to the [0, 1] interval. For example, cubic spline interpolation can be used to uniformly sample different trajectories, ensuring that each trajectory has a fixed number of sampling points to ensure consistent data length for the deep learning framework. Raw flight trajectory data is obtained from the Automatic Dependent Surveillance-Broadcast (ADS-B) system. This data includes timestamps, call signs, ground speed, heading angle, vertical speed, location information (latitude, longitude, and altitude), flight number, and takeoff and landing airports. The data is sampled and normalized, and the flight trajectory data is separated into temporal and spatial attribute data based on its spatiotemporal characteristics. The collected flight trajectory data is preprocessed according to predetermined rules, and cubic spline interpolation is used to uniformly sample each trajectory, ensuring that all trajectories have the same number of sampling points. The trajectory is further separated into temporal attribute features and spatial attribute features, and Min-Max Normalization is used to map each attribute value to a unified interval.

[0015] In this embodiment, the process of extracting low-dimensional features includes: using a deep autoencoder constructed based on a temporal convolutional network (TCN) and a deep autoencoder constructed based on a convolutional neural network (CNN) to extract low-dimensional features from the separated temporal attribute data and spatial attribute data, respectively, to obtain the temporal attribute low-dimensional features and spatial attribute low-dimensional features. In the latent space represented by the joint features, a clustering algorithm is adopted, and a clustering allocation reinforcement strategy is introduced to self-tune the clustering results to achieve accurate division of different air flow patterns. Specifically, for temporal attribute data, a deep autoencoder (DAE) based on a temporal convolutional network (TCN) is adopted to efficiently capture the long-term time-dependent features in the flight trajectory; for spatial attribute data, a deep autoencoder based on CNN is adopted to fully extract the local spatial layout and geometric features of the flight trajectory; the low-dimensional feature representations with temporal features and spatial features output by the above two autoencoders are respectively recorded as and , and fuse them to generate a comprehensive spatiotemporal feature vector In the latent representation zz output by the encoder, a clustering algorithm is used to cluster the low-dimensional representation of the flight trajectory. At the same time, a clustering allocation enhancement strategy based on Student's t distribution is introduced. The clustering results are optimized by the KL divergence between the auxiliary target distribution to achieve a detailed division of different flow patterns in the airspace. Based on the clustering results, the representative flight trajectory of each cluster is extracted as the key information to characterize each local flow pattern.

[0016] Low-dimensional feature extraction is performed on the separated temporal attribute data. This involves modeling the dynamic characteristics of the trajectory at successive moments, constructing a deep autoencoder model based on the TCN, and then training the autoencoder model. Specifically, the dynamic characteristics of the trajectory at successive moments include ground speed, heading angle, vertical speed, and position information. Dilated convolutions and residual connections are used in the TCN architecture to fully capture long-term dependencies in the flight trajectory.

[0017] In this embodiment, low-dimensional features are extracted from the separated spatial attribute data, including: establishing a deep autoencoder model based on CNN, and using it to extract low-dimensional features of spatial attribute data, wherein the deep autoencoder model is used to extract low-dimensional features of spatial attribute data such as trajectory length, azimuth, flight status and airport information hidden in the flight trajectory. For non-numerical airport information, Embedding technology is used to convert it into a vector representation of fixed dimensions. By minimizing the reconstruction error, the low-dimensional representation output by the spatial autoencoder contains sufficient regional features, which is conducive to the subsequent distinction of flow patterns. The fused joint features are soft-clustered using a Gaussian mixture model; the probability distribution constructed based on the Student-t distribution is used to fit the cluster allocation, wherein the loss function in the cluster allocation process is optimized by KL divergence to obtain a more discriminative clustering result. The compactness of the trajectories of similar flow patterns in the cluster is enhanced through cluster allocation reinforcement, and the separation between different flow patterns is improved.

[0018] In this embodiment, the process of constructing a graph-structured model involves: using the representative trajectory of each flow pattern in the clustering results as a graph node, where each node reflects the primary flow direction within the corresponding airspace sector; counting the number of flight connections between different nodes based on historical trajectory data; constructing an adjacency matrix based on the statistical results; and deriving edge weights based on the flight traffic between nodes. This graph-structured model accurately reflects the spatial distribution and flow direction of each flow pattern within the airspace, facilitating subsequent refined traffic forecasting.

[0019] In this parallel fusion prediction model, a graph convolutional network (GCN) is used to extract spatial topological dependencies within the graph structure, and a long short-term memory (LSTM) network is used to capture the temporal dynamics of node traffic data within the graph structure. For example, a parallel architecture is used to perform multimodal feature modeling on the constructed graph model using a graph convolutional network (GCN) and a long short-term memory (LSTM) network, respectively. The spatial topological structure features and temporal dynamic features are extracted, followed by feature fusion. Finally, the parallel fusion prediction model outputs air traffic flow forecasts for multiple predetermined time steps in the future (e.g., 30 to 150 minutes), thereby supporting fine-grained regulation of air traffic control.

[0020] The parallel fusion prediction model includes: one branch uses a graph convolutional network (GCN) to perform spatial feature modeling on the constructed flight trajectory graph structure to extract the topological structure information between each node in the airspace; the other branch uses a long short-term memory network (LSTM) to perform time series modeling on node traffic data to capture the dynamic change trend of air traffic flow; the outputs of the two branches are integrated through splicing or other fusion strategies, and the traffic prediction results for the future preset time step are output through a fully connected layer.

[0021] The training process of the parallel fusion prediction model includes: a) data reconstruction stage: the TCN and CNN autoencoders are trained separately, low-dimensional representations in the time and spatial domains are obtained by minimizing the reconstruction error, and the pre-trained model is optimized using a joint loss function; b) clustering stage: the flight trajectory data is carefully clustered in the fused latent space through a joint clustering allocation reinforcement strategy and a Gaussian mixture model, and the clustering results are converted into an analytical representation of air traffic flow using the constructed graph model; c) traffic prediction stage: the constructed graph structure is trained as a whole using the parallel GCN and LSTM models, parameter optimization is performed by defining theoretical loss functions such as mean squared error, and air traffic flow prediction results for multiple future time steps are output, thereby theoretically achieving a detailed prediction of airspace flow status and supporting decisions such as air traffic management and route planning. The outputs of the GCN and LSTM branches are fully fused at the feature layer, and the fused features are processed through a set of fully connected layers. Finally, the air traffic flow forecast results for multiple time periods in the future (such as 30 minutes, 60 minutes, 90 minutes, 120 minutes, and 150 minutes) are output, thereby achieving high-precision prediction of fine-grained traffic flow within the sector.

[0022] The present invention provides a refined air traffic flow prediction method driven by flow pattern recognition. This method efficiently extracts spatiotemporal features from flight trajectories, accurately identifies local flow patterns, and constructs a graphical model based on representative flow patterns. This method enables fine-grained prediction of airspace traffic flow, significantly improving the precision and dynamic responsiveness of localized regional control. This method is particularly suitable for in-depth analysis of flight trajectory data within airspace sectors, providing accurate decision support for air traffic control, route planning, and airspace design. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0024] Figure 1 A flowchart of flight trajectory data preprocessing provided by an embodiment of the present invention;

[0025] Figure 2 A flow pattern recognition framework diagram provided by an embodiment of the present invention;

[0026] Figure 3 A flow prediction framework diagram provided by an embodiment of the present invention;

[0027] Figure 4A schematic diagram of the system framework provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0028] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. The embodiments of the present invention will be described in detail below, with examples of the embodiments illustrated in the accompanying drawings. Throughout, identical or similar reference numerals represent identical or similar elements or elements having identical or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and intended only to explain the present invention and are not to be construed as limiting the present invention. Those skilled in the art will appreciate that, unless otherwise stated, the singular forms "a," "an," "said," and "the" used herein may also include the plural forms. It should be further understood that the term "comprising" as used in the description of the present invention refers to the presence of the stated features, integers, steps, operations, elements, and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when an element is referred to as being "connected" or "coupled" to another element, it may be directly connected or coupled to the other element, or intervening elements may be present. Furthermore, "connected" or "coupled" as used herein may include wireless connections or couplings. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items. It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by those skilled in the art in the art to which this invention belongs. It should also be understood that terms such as those defined in common dictionaries should be understood to have meanings consistent with their meanings in the context of the prior art, and unless defined as such, will not be interpreted in an idealized or overly formal sense.

[0029] The design concept of this embodiment is to optimize and develop a new theoretical and technical framework that can achieve in-depth recognition of various flow patterns in flight trajectories at the airspace sector level and directly incorporate this detailed flow pattern information into the traffic prediction process. By adopting advanced feature extraction methods (such as a deep spatiotemporal fusion model based on dual autoencoders), a joint clustering allocation reinforcement strategy, and a traffic prediction method based on graph structure modeling, it can more accurately capture the spatiotemporal characteristics of various flow movements within the sector, achieve refined prediction of traffic flow within the local airspace, and provide a theoretical basis and technical support for dynamic regulation and refined airspace management. This solution is mainly used to provide more accurate decision-making basis for functional subsystems of the air traffic control system, such as airspace regulation, route planning, and flight safety management.

[0030] To better illustrate the practical application of this embodiment, real navigation data can be used, such as the trajectory data and sector range data of all aircraft in all adjacent sectors of the airspace of a certain eastern municipality from October 1 to October 31, 2021. The method for fine-grained air traffic flow prediction based on flow pattern recognition provided by the embodiment of the present invention specifically includes the following main steps: 1. Data preprocessing, 2. Intra-sector flow pattern recognition, and 3. Fine-grained prediction of sector flow.

[0031] Step 1: Data preprocessing. The process of data preprocessing is as follows: Figure 1 As shown, the specific implementation steps are divided into S1-S6.

[0032] S1. Summarize the ADS-B data of all aircraft in the airspace above a certain eastern municipality within a certain time period into sectors and obtain all the tracks in each sector.

[0033] S2. Resample the ADS-B data of each track within the sector into information points evenly distributed in time (e.g., one sampling point every 10 seconds).

[0034] S3. All information points of each trajectory in the sector (such as latitude and longitude, altitude, heading angle, vertical speed, and ground speed) are merged horizontally into a row. The trajectory data is represented as features with the same time interval between each column, and each row represents all the features of a trajectory. The trajectory data obtained by this processing method is time series data. .

[0035] S4. Separating features with spatial attributes from ADS-B , whose features include:

[0036] Track length: The number of position points in each original track; Azimuth: The heading from the start point to the end point of the track is calculated to obtain the path azimuth; Climb or Descent: If the track is climbing within the sector, it is 1; if it is descending, it is -1; Airport Information: The departure and landing airport information of each track.

[0037] S5. The airport information attribute values in [ ] are strings and need to be converted into numerical data. To avoid dimensionality explosion and sparse matrix, word embedding technology is used to encode the airport information.

[0038] S6. Separately and The minimum-maximum normalization method is used to map the values of all attributes to between [0,1]. The formula is as follows:

[0039]

[0040] in, The maximum value in each attribute column, The minimum value in each attribute column.

[0041] Step 2: Identification of flow patterns within sectors

[0042] The intra-sector flow pattern recognition framework proposed in this paper consists of two deep autoencoders, one of which is a TCN-based deep autoencoder for temporal attribute data of intra-sector trajectories. For reconstruction, another CNN-based deep autoencoder is used to reconstruct spatial attribute data Reconstruction, the specific framework is as follows Figure 2 As shown in the figure, during training, the two autoencoders are trained independently. After training, the latent representations output by the two encoders are fused and further clustered using a clustering algorithm. The entire process is divided into three parts: temporal attribute data reconstruction, spatial attribute data reconstruction, and cluster assignment reinforcement.

[0043] (1) Reconstruction of time attribute data

[0044] Given that the autoencoder based on the fully connected layer is more effective in reconstructing temporal features, It is difficult to capture the temporal dependencies, resulting in poor low-dimensional clustering of features. In order to better capture the temporal changes and dynamic characteristics of the flight trajectory, the present invention uses TCN to construct an autoencoder. It is well known that TCN can better capture the long-term dependencies in time series than the fully connected layer. TCN uses dilated convolution to give the network a larger receptive field and can obtain The long-term historical information of is calculated as follows:

[0045]

[0046] Among them, k is the size of the convolution kernel, d is the expansion coefficient, is the filter, T is the transpose operation, Represents the previous time-frequency information index, Indicates that in the sequence x T The input vector corresponding to the nth time step in , i represents the index of the convolution kernel, j represents the index of the time step, represents the convolution kernel i-th filter weight vector, Represents the data of the jth time step in the input sequence The output after temporal convolution.

[0047] The entire TCN network consists of multiple residual blocks. Residual connections are used for skip connections to alleviate the gradient vanishing problem and improve the stability and performance of the model. and dilated convolution operations , the output of the residual block is defined as: , where o represents the output of the trajectory after being processed by the residual block, represents the activation function. In addition, to prevent overfitting, an average pooling layer is added after the TCN layer.

[0048] Depend on Figure 1 It can be seen that the present invention uses multiple TCN layers to design a set of symmetrical encoders and decoders, aiming to achieve There are two stages of mapping and reconstruction of data into latent space, and the formula is:

[0049]

[0050] in, are the description matrices of the input trajectory and the reconstructed trajectory respectively, N is the number of trajectories, and K is the dimension of the trajectory. and are the parameters of the encoder and decoder respectively, 、 Represent the decoder and encoder for processing time attribute data respectively. During the training process, the objective function is minimized so that and As similar as possible, the objective function is defined as follows:

[0051]

[0052] Among them, the model parameters , It is the L2 norm. The model parameters can be continuously optimized and updated by back propagation during the training process, and finally a low-dimensional representation is obtained from the encoder. , the parameters before the semicolon in the formula are used as variables of the function, and the parameters after the semicolon are used as parameters of the autoencoder.

[0053] (2) Spatial attribute data reconstruction

[0054] Spatial attribute data It consists of four parts: trajectory length, azimuth, climb or descent, and airport information. The attribute value of airport information is a string, which needs to be converted into numerical data. A common method is one-hot encoding. However, there are 117 types of airport information attribute values. Using one-hot encoding will cause the matrix to be too sparse and cause dimensionality explosion. Therefore, the embedding technology is used to encode the airport information. For example: There is no relationship between each column of data, so it presents a high spatial characteristic. Therefore, CNN is used to build an autoencoder Reconstruct and obtain its low-dimensional feature representation CNN has significant advantages in processing spatial attribute features due to its local connection and weight sharing characteristics.

[0055] The autoencoder model parameters for reconstructing spatial attribute data can be obtained through the following two formulas and the objective function .

[0056]

[0057]

[0058] in, represents the reconstructed spatial attribute data, Represents the decoder parameters for processing spatial attribute data, represents the encoder parameters for processing spatial attribute data, Represents a decoder for processing spatial attribute data, Represents an encoder for processing spatial attribute data.

[0059] (3) Cluster allocation enhancement

[0060] According to the above content, we can get and The low-dimensional representation of the latent space and , further merge them to obtain fusion features Although clustering algorithms such as k-means or DBSCAN can be used to cluster the fused features directly, this paper introduces cluster allocation enhancement to further enhance clustering, by fine-tuning z to make it more suitable for clustering. Specifically, two distributions, Q and P, are introduced to enhance the clustering ability of z. Q is the student-t distribution, which is used to calculate the probability that trajectory i is clustered into cluster j. , P is the target distribution, which is used to guide the model to optimize the clustering distribution, so as to better fit the actual distribution of trajectory data and continuously improve the clustering effect during the iteration process. The calculation formula is:

[0061]

[0062]

[0063] in, is the centroid of j, the initial cluster centroid It is estimated on z by using the k-means algorithm. represents the target probability that i is assigned to j, is the embedding representation of the i-th track after passing through the autoencoder, Represents the cluster center number corresponding to the current data point. By using KL divergence to calculate the distance between P and Q, the clustering loss function can be defined:

[0064]

[0065] By minimizing the clustering loss, trajectories can be more accurately assigned to appropriate cluster centers, improving clustering results. This is a self-training strategy.

[0066] Since the present invention uses two autoencoders to process spatial features and temporal features respectively, and then merges their results for clustering assignment reinforcement, and optimizes the parameters of each part simultaneously with the help of fine-tuning process, the entire traffic pattern recognition model can be regarded as an end-to-end structure. Considering the reconstruction loss of the autoencoder used to process temporal attribute data, , autoencoder reconstruction loss for processing spatial attribute data and clustering loss , the objective function of the traffic pattern recognition model is defined as follows:

[0067]

[0068] Where β is the weight coefficient. After fine-tuning, trajectory i will obtain a cluster label, such as , k represents the number of clusters. It is worth mentioning that the present invention can adopt a variety of clustering algorithms in the cluster allocation enhancement process, such as K-means, DBSCAN and GMM.

[0069] Step 3: Refined prediction of sector traffic

[0070] The proposed framework for refined intra-sector traffic prediction is composed of GCN and LSTM, where GCN is used to learn the spatial features of the constructed air traffic topology, and LSTM is used to learn the traffic time series features of each node. Figure 3 The entire process is divided into two parts: the refined traffic prediction and composition strategy and the parallel fusion prediction model.

[0071] (1) Traffic Refinement Prediction and Mapping Strategy

[0072] As can be seen from the background technology, the existing GNNs-based air traffic flow prediction methods are mainly limited to the macro-prediction level, and ATC currently still manages flights in sectors. The knowledge and information provided to ATC by previous prediction methods are relatively limited, which limits the diversity and accuracy of ATFM decisions. Therefore, this section proposes a composition strategy based on sector traffic pattern recognition, which aims to predict the flow of each traffic pattern within the sector. According to the sector flow pattern recognition in Section 2, the clustering results of the internal trajectories of each sector can be obtained, and the representative trajectory of each cluster can be further calculated, and these representative trajectories are regarded as node compositions. It is worth mentioning that since the representative trajectory is directional, it is not only possible to predict the flow of each traffic pattern within the sector, but also to know the flow direction of the traffic pattern, thereby realizing refined prediction of sector traffic. By providing ATC with richer traffic prediction information within the sector, it helps to improve the accuracy and diversity of its decision-making.

[0073] Given that the trajectories between adjacent sector nodes are connected, the present invention constructs an adjacency graph based on historical trajectories. Specifically, let A and B be two nodes (representing trajectories) in two adjacent sectors, intercept all complete trajectories contained in the adjacent sectors, and assume that there are Tracks , the i-th trajectory Expressed as , and Represent the starting point and end point of the complete trajectory in the adjacent sector respectively. Now we need to obtain the weight of the connection from A to B, which is calculated as follows:

[0074]

[0075]

[0076] in, Represents the set of all trajectories of the cluster to which A belongs , Represents the set of all trajectories of the cluster to which B belongs , express arrive The number of flights, represents the connection weight from A to B, and K represents the clustering of representative trajectories in all adjacent sectors. The number of flights, Indicates the last track point of each track within the sector.

[0077] (2) Parallel fusion prediction model

[0078] Given that the air traffic flow prediction model of GCN and LSTM in series mode will weaken some spatial and temporal features during the information transmission process, the present invention adopts a parallel mode to avoid this problem. The input of GCN includes the traffic characteristics of each node and the adjacency matrix A. Let It is GCN Layer input information (traffic samples), is the input information of the l+1 layer and is calculated as follows:

[0079]

[0080] in, is the activation function, is the weight matrix, . is the degree matrix, , I represents the identity matrix.

[0081] LSTM is used to fully learn the time dependency of air traffic flow. Its calculation process can be referred to, and the present invention will not repeat it here. and They are the outputs of GCN and LSTM, and the features are further fused by splicing: In order to further predict the traffic flow, three fully connected layers are used to process the fusion features in sequence. By setting the dimension of the last layer to 5, the traffic flow prediction of 5 time steps is achieved, such as [30min, 60min, 90min, 120min, 150min].

[0082] During the training process of the parallel fusion prediction model, the parameters of GCN and LSTM Both are achieved by minimizing the average mean square error (AMSE) loss function Update, calculated as follows:

[0083]

[0084] Among them, Y is the true value matrix of traffic, is the traffic prediction value matrix, L is the total number of samples, T is the output time step, which is 5 here, and Respectively represent The true value and predicted value of a traffic sample at the tth time step.

[0085] The method for fine-grained air traffic flow prediction driven by flow pattern recognition, provided by an embodiment of the present invention, obtains raw flight trajectory data from the Automatic Dependent Surveillance-Broadcast (ADS-B) system and uses advanced data preprocessing techniques to split the collected trajectory data into temporal attribute data and spatial attribute data with distinct spatiotemporal characteristics. Secondly, a deep autoencoder based on a temporal convolutional network (TCN) is used to effectively model the temporal features in the flight trajectory data, while an autoencoder constructed using a convolutional neural network (CNN) is used to deeply extract spatial features, thereby obtaining a low-dimensional feature representation of the flight trajectory. Thirdly, based on these low-dimensional features, the present invention introduces a joint clustering allocation reinforcement strategy that uses a clustering algorithm to accurately identify various flow patterns within the airspace, especially capturing representative flow patterns with obvious directionality. Finally, a graph structure model is constructed based on the representative trajectories of the identified flow patterns. A multimodal prediction architecture, combining a graph convolutional network (GCN) and a long short-term memory network (LSTM), is used to perform directional and fine-grained predictions on each flow pattern, theoretically achieving high-precision predictions of traffic flow distribution and its dynamic trends within airspace sectors. This paper clearly proposes a theoretical method based on flow pattern recognition, which realizes accurate mapping of flow types through extracting spatiotemporal features through dual-path autoencoders and joint clustering allocation reinforcement. It then combines graph representation technology with parallel network prediction strategy, providing an efficient and advanced technical path for airspace design optimization, air traffic control and flight safety management.

[0086] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited to this. Any changes or replacements that can be easily thought of by any technician familiar with this technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A method for refined air traffic flow prediction driven by air traffic flow pattern recognition, characterized in that: include: Acquiring flight trajectory data from an Automatic Dependent Surveillance-Broadcast (ADS-B) system and performing preprocessing, wherein the preprocessing step includes performing data separation according to spatiotemporal characteristics of the flight trajectory data; Low-dimensional features are extracted from the separated flight trajectory data, where the low-dimensional features of temporal attributes and spatial attributes are fused into joint features; Construct a graph structure model, where the nodes in the graph are representative trajectories reflecting the characteristics of each flow pattern, and the connection weights of the edges in the graph structure are associated with the flight flow between adjacent nodes; Through the established parallel fusion prediction model, the air traffic flow prediction results within the future predetermined time step are obtained.

2. The method according to claim 1, characterized in that The pretreatment includes: Summarize the trajectories corresponding to the flight trajectory data into various sectors; The flight trajectory data is segmented according to the sampling interval and then normalized.

3. The method according to claim 1, characterized in that The process of low-dimensional feature extraction includes: A deep autoencoder constructed based on a time-domain convolutional network (TCN) and a deep autoencoder constructed based on a convolutional neural network (CNN) are used to perform low-dimensional feature extraction on the separated time attribute data and spatial attribute data, respectively, to obtain the low-dimensional features of the time attributes and the low-dimensional features of the spatial attributes.

4. The method according to claim 1 or 3, characterized in that Perform low-dimensional feature extraction on the separated time attribute data, including: The dynamic change characteristics of continuous moments in the track are modeled, and a deep autoencoder model based on TCN is constructed, and then the autoencoder model is trained.

5. The method according to claim 1 or 3, characterized in that Perform low-dimensional feature extraction on the separated spatial attribute data, including: A CNN-based deep autoencoder model was established to extract low-dimensional features of spatial attribute data, including trajectory length, azimuth, flight status, and airport information hidden in the flight trajectory. For non-numerical airport information, Embedding technology is used to convert it into a fixed-dimensional vector representation.

6. The method according to claim 5, characterized in that Also includes: Use Gaussian mixture model to perform soft clustering on the fused joint features; The cluster assignment is fitted using a probability distribution based on the Student-t distribution, where the loss function in the cluster assignment process is optimized by KL divergence.

7. The method according to claim 1, characterized in that The process of building a graph structure model includes: The representative trajectory of each flow pattern in the clustering result is used as a node of the graph, where each node is used to reflect the main flow direction in the corresponding airspace sector; The number of flight connections between different nodes is counted based on historical trajectory data, an adjacency matrix is constructed based on the statistical results, and the edge weights are obtained based on the flight traffic between nodes.

8. The method according to claim 1, characterized in that In the parallel fusion prediction model, a graph convolutional network (GCN) is used to extract the spatial topological dependencies in the graph structure, and a long short-term memory network (LSTM) is used to capture the temporal dynamic changes of node traffic data in the graph structure.

9. The method according to claim 1 or 8, characterized in that The parallel fusion prediction model includes: One branch uses a graph convolutional network (GCN) to model the spatial features of the constructed flight trajectory graph structure to extract the topological structure information between nodes in the airspace; Another branch uses the long short-term memory network (LSTM) to perform time series modeling on node traffic data to capture the dynamic trend of air traffic flow; The outputs of the two branches are feature-integrated, and the traffic prediction results for the future preset time step are output through the fully connected layer.

Images