The Research On Air Traffic Flow Spatial-Temporal Modeling And Prediction Based On Graph Neural Network

With the advent of the post-pandemic era globally,the demand for air travel in our country has surged significantly,leading to increasingly complex air traffic flow distribution and imbalances during peak periods.To promote the high-quality development of the civil aviation industry,the construction of efficient and intelligent traffic flow management is crucial.Air traffic flow prediction is the foundation and key to flow management.Therefore,how to enhance the perception capability of air traffic flow and achieve globally optimal and accurate air traffic flow prediction methods has become an urgent issue to address.The air traffic flow prediction can essentially be attributed to a spatial-temporal network prediction problem.However,most existing air traffic flow prediction methods focus on mining the temporal variation patterns of single airspace traffic flows,neglecting the impact of airport networks on the spatialtemporal distribution of air traffic flow in different management phases,resulting in the prediction accuracy that fails to meet the requirements for intelligent flow management construction.Therefore,this paper,focusing on airport traffic flow data and the demand for network-level airport arrival flow predictions,based on the graph neural network,proposes a comprehensive technique system for strategic,pre-tactical,and tactical phase air traffic flow predictions,covering multiple airports,perspectives,and scales.It emphasizes modeling the spatial-temporal operational characters of multi-phase air traffic operations.The primary innovative research in this paper can be concluded as follows:(1)To address the deficiency in medium-to-long-term,fine-grained air situation awareness capability during the strategic phase,we propose a Multi-view Attention Spatial-temporal Network model,MASTNet.By modeling the spatial-temporal correlation,spatial heterogeneity,and planned nature of air traffic flow,it achieves accurate strategic air traffic flow prediction.Considering the influence of flight planning and airport location distribution on the variation of air traffic flow during the strategic phase,we introduce the airport schedule graph and airport duration graph to construct a static airport network,representing the air traffic operating environment in the strategic phase.For the heterogeneous airport network,we propose a multiview attention mechanism in the air traffic flow prediction domain for the first time,quantifying the contribution strength of different airport network relationships in the prediction task from both macro and micro perspectives,achieving effective integration of heterogeneous airport networks.Based on the integration results of the airport network,we apply graph convolution models and gating mechanisms to mine the spatial-temporal correlation of air traffic flow.We introduce a time embedding module and global traffic features to enhance time features and traffic information,improving MASTNet’s medium-to-long-term prediction capability.Experimental results on a real domestic airport traffic flow dataset showed that MASTNet has excellent prediction performance.Compared to the sub-optimal T-MGCN,it achieved a relative improvement of 7%,7%,and 19% in MAE,MAPE,and RMSE metrics,respectively.(2)To address the inadequacy in modeling the dynamic air traffic operating environment during the pre-tactical phase,we propose a Dynamic Spatial-temporal Graph Neural Network model,DSTGNN.By exploring the spatial-temporal dynamics,local tendency,and periodicity of air traffic flow,it achieves accurate pre-tactical air traffic flow prediction.Considering the influence of dynamic flight demands during the pre-tactical phase on the evolution patterns of air traffic flow,we introduce a discrete dynamic airport network based on flight plans and prior air traffic control element information,representing the highly variable air traffic operating environment.Spatially,we design a dynamic multi-graph neural network based on the self-attention mechanism to adaptively model the spatial dynamics and heterogeneity inherent in air traffic flow data.Temporally,we proposed a dual-path temporal awareness attention module that utilizes convolutional attention and differential operations to extract local change trends and dynamic temporal dependencies from time series data.To extract the periodicity and planned nature of traffic flow data,we construct a prior-guided recalibration fusion module to learn an effective representation of prior global traffic features.Experimental results on a real domestic airport traffic flow dataset showed that DSTGNN has excellent situational-level traffic flow prediction performance.Compared to the sub-optimal MASTNet model,the prediction performance was significantly improved,with an average relative increase of 10%,9%,and 21% in MAE,MAPE,and RMSE metrics,respectively.(3)To address the problem of insufficient representation capability for multi-source,heterogeneous flight operation dynamic information during the tactical phase,we propose a Traffic Situation Awareness Network(TSAN)based on event-driven mechanisms.This network uncovers the potential correlation between micro flight operation dynamics and macro air traffic flow change patterns,achieving accurate tactical air traffic flow prediction.Considering the influence of real-time external driving factors on the evolution direction of air traffic flow during the tactical phase,we introduce a continuous dynamic airport network based on flight dynamic events and prior air traffic control information,representing the real-time changes in the air traffic operating environment on the day of flight operation.Taking into account the macro airport operation status,we designed a spatial-temporal embedding module for modeling the flight operation context.To learn the complex,heterogeneous flight dynamics,we construct a flight dynamic event embedding module with spatial-temporal perception capability to extract effective flight dynamic event representations.Combining the aforementioned embedding information,we introduce the temporal dynamic graph neural network method to the field of air traffic flow prediction,learning the evolution process of air traffic situations under mixed en-route and planned flight operations,and extracting rich semantic air traffic situation graph embedding representations.Considering the complex flight operation process during the tactical phase,we propose an airport traffic situation multi-task decoder module to explore the dynamic coupling relationship between micro flight dynamics,air traffic situations,and macro traffic flows.Experimental results on a real domestic airport traffic flow dataset verified the effectiveness of the TSAN model.Compared to the sub-optimal TGN model,there was an average relative improvement of 9%,8%,and 10% in the MAE,MAPE,and RMSE metrics,respectively.In summary,this paper considers the demands of air traffic flow management and the controlling characteristics of air traffic across multiple phases,which proposes a series of spatialtemporal network data prediction methods based on graph neural networks.It establishes a comprehensive air traffic flow intelligent prediction technology system,significantly enhancing the prediction accuracy of air traffic flow across various periods.The research results of this paper can effectively assist the development of flow management toward intelligence,offering valuable insights and holding substantial application potential.