Machine Learning for Accurate Intent-Based Trajectory Prediction in Conflict Resolution

1 MACHINE LEARNING FOR ACCURATE INTENTBASED TRAJECTORY PREDICTION IN CONFLICT RESOLUTION Mustafa ÖZDEMIR1,2 and Eri ITOH3 1Research Associate, Research Center for Advanced Science and Technology, The University of Tokyo (4-6-1 Komaba, Meguro-ku, Tokyo 153-8904, Japan) 2Department of Air Traffic Control, Erzincan Binali Yıldırım University Erzincan, Turkey E-mail: [email protected] 3Professor, Research Center for Advanced Science and Technology, The University of Tokyo (4-6-1 Komaba, Meguro-ku, Tokyo 153-8904, Japan) E-mail: [email protected] (Corresponding Author) The demand for air transportation is projected to increase substantially in the coming decades, particularly in emerging regions such as the Asia-Pacific. According to the International Civil Aviation Organization, intra-Asia/Pacific passenger traffic is expected to grow at an average annual rate of 5% between 2022 and 2032, resulting in a 63% increase over 2022 levels by 2032. This growth poses significant challenges for current air traffic management systems, as higher traffic volumes increase airspace complexity and strain controller workload. Airspace complexity arises from multiple factors, typically categorized as aircraft-related, conflict-related, and airspace-related. Accurately predicting future complexity is therefore crucial for effective traffic management. In this study, we focus on forecasting airspace complexity based on predicted aircraft trajectories, explicitly incorporating aircraft intent as the anticipated maneuver an aircraft is expected to execute in the future. We investigate the use of machine learning models to predict airspace complexity based on relevant airspace data. The proposed approach aims to generate reliable forecasts of airspace complexity, which are closely linked to controller workload and sector capacity, thereby supporting improved planning and operational decision-making in air traffic management. Key Words : air traffic management, complexity, trajectory prediction, aircraft intent, machine learning. 1. INTRODUCTION The demand for air transportation is projected to rise in the coming decades, particularly in emerging regions such as the Asia-Pacific. The International Civil Aviation Organization (ICAO) estimates that Intra-Asia/Pacific passenger traffic will grow at an average annual rate of 5% from 2022 to 2032, resulting in a total increase of 63% over the 2022 level by 2032 1). This increase in air traffic volume will pose significant challenges to the current air traffic management system, leading to congestion and an increased workload for air traffic controllers, which remains a primary limitation on overall airspace capacity 2), 3). Air traffic controller workload is closely linked to airspace complexity 4); therefore, measuring sector complexity is essential. Moreover, complexity may play a more critical role than traffic volume in determining air traffic controllers’ task assignment strategies, as demonstrated by Jumad et al. 5) through human-in-the-loop (HITL) simulation experiments. This is because high traffic volume does not necessarily translate to high workload if aircraft trajectories involve few crossings or potential conflicts. Since the air traffic system is highly dynamic, estimating complexity solely from pre-flight plans may not provide accurate results, as aircraft interactions, numbers, and relative positions can vary significantly during operations. Predicting future airspace complexity based on projected aircraft positions can support more efficient resource allocation, sectorization, and operational planning. Therefore, it is crucial to have precise knowledge of an aircraft's trajectory during flight. However, this task is challenging due to the highly dynamic air traffic system, which is influenced by uncertainties from various factors, including environmental conditions, aircraft sensor sensitivity, and the human behavior of those involved 40-04 第72回土木計画学研究発表会・講演集 - 40-04 - 2 in air traffic systems, such as air traffic controllers and pilots 6). To achieve accurate trajectory predictions, machine learning models could enhance reliability and improve performance compared to using a static aircraft performance database such as the Base of Aircraft Data (BADA) 7). Utilizing high-fidelity historical flight trajectory data enables the development of machine-learning models that can accurately predict future aircraft trajectories. Recent studies have highlighted the pivotal role of machine learning algorithms in accurately predicting both aircraft trajectories 8), 9) and airspace complexity 10), 11). This research aims to develop robust machine learning models for accurately predicting future airspace complexity during flight, taking aircraft intent into account. In this study, the term intent refers to the planned maneuver of the aircraft, specified at the maneuver initiation time and characterized by the commanded changes in speed, heading, or altitude as provided by air traffic control instructions. We leverage historical flight trajectory data and validate the machine learning models through HITL simulations conducted in the Escape Light simulator developed by Eurocontrol 12). This research aligns very well with recent innovation trends in air traffic management within the scope of trajectory-based operations. Both the United States and Europe have launched research and development initiatives: the Next Generation Air Transportation System (NextGen) 13) and the Single European Sky ATM Research (SESAR) 14), respectively. These programs share a common vision centered on 4D-Trajectory-Based Operation, where aircraft follow optimized trajectories in space and time, mutually agreed upon by the aircraft, air traffic controllers, and operators 6), 15). 2. RESEARCH METHODOLOGY Our methodology follows a systematic process designed to predict and validate airspace complexity through a combination of data-driven modeling and HITL simulations. The process begins with data preparation, where flight, sector, and operational data are collected and preprocessed. Next, current complexity assessment is performed to compute baseline complexity metrics from the existing traffic situation. Building upon this, machine learning models are employed for trajectory prediction, enabling the estimation of future aircraft positions and movements. The predicted trajectories, along with aircraft intent, are used to forecast airspace complexity, providing an estimate of future complexity levels over a defined prediction horizon. This combined output is evaluated through HITL simulations conducted using the Escape Light simulator, which allows assessment of the operational feasibility and controller response to predicted complexity scenarios. Finally, the evaluation and analysis stage compares predicted and observed complexity measures. Figure 1 provides a summary of this process. Fig. 1 Overview of the methodology. Aircraft trajectories are represented as four-dimensional sequences, comprising latitude, longitude, altitude, and time. Trajectory prediction involves estimating an aircraft’s future positions over specified prediction horizons. (1) Problem formulation for trajectory prediction To formulate trajectory prediction with aircraft intent, we follow a similar approach to that presented in Tran et al. 16). The aircraft state at time t can be represented as follows: 𝑠𝑡= [𝑝𝑡,𝑣𝑡,𝜓𝑡,ℎ𝑡,𝑖𝑡] (1) Where 𝑝𝑡 is current position, 𝑣𝑡 is ground speed, 𝜓𝑡 is heading, and ℎ𝑡 is altitude. 𝑖𝑡 represents the aircraft intent, encompassing parameters such as target altitude, target speed, intended heading, and upcoming maneuvers. The latter may arise from conflict resolution actions, whose influence on airspace complexity over future trajectories is explicitly taken into account. The historical input to the model is the sequence of the last 𝑘 states prior to and including the current time (𝑡0). 𝑆𝑡0= [𝑠𝑡0−𝑘,…,𝑠𝑡0] (2) A model 𝑓𝜃 is trained to produce a sequence of future positions, where 𝑝 represents the predicted positions at different time intervals. The objective is to 40-04 第72回土木計画学研究発表会・講演集 - 40-04 - 3 minimize the error between predicted and true positions over the prediction horizon. 𝑇 𝑡0 𝑛= [𝑝𝑡0+1,…, 𝑝𝑡0+𝑛] (3) In this study, aircraft intent is incorporated not only for improving trajectory prediction but also to assess its influence on airspace complexity. Specifically, future maneuvers arising from conflict-resolution actions are modeled through intent variables, and their effects on predicted traffic interactions and resulting complexity measures are quantitatively evaluated. Our methodology for predicting potential conflict resolution strategies involves extrapolating shortterm trajectories from current states, identifying potential conflicts, and generating feasible resolution maneuvers (heading, speed, or altitude), from which the most likely or optimal maneuver is selected. This process is illustrated in Figure 2. Fig. 2 Flowchart of aircraft intent definition. (2) Dataset This research is intended to utilize on-board flight data for training machine learning algorithms and radar surveillance data from Japan's airspace for evaluating trajectory prediction performance. In this context we plan to use aircraft’s current positions (3) Data pre-processing One of the initial stages of the proposed methodology is to perform pre-processing on all flight profiles to ensure consistency and eliminate any erroneous data. In this stage, missing values in the datasets are also addressed. An algorithm will be developed to extract the necessary data from relevant sections of each flight profile. (4) Variable selection Aircraft trajectory is influenced by several factors, including inherent characteristics such as aircraft design and gross weight, as well as performance parameters like climb and descent profiles, cruise airspeed, and cruising altitude. Meteorological conditions, particularly wind speed, wind direction, and temperature, also play a significant role. Very few studies analyze some of these variables to develop trajectory prediction models. For example, Ref. 17) utilized flight plans, radar tracks, and localized weather data to predict aircraft trajectories for entire flights, while Ref. 18) incorporated factors such as departure and arrival airports, the distance between airports, aircraft type, airline operator, and day of the week. Figure 3 presents a portion of the flight profile during the cruise phase at 36,000 feet, illustrating variations in aircraft position in terms of latitude, longitude, altitude, and time. Although the variations in latitude and longitude appear minor, they can translate to significant distances when converted. Additionally, the altitude is not constant, fluctuating between 36,045 feet and 36,020 feet. Fig. 3 Fluctuations in aircraft position during the cruise phase for a randomly selected aircraft. (5) Complexity metrics Airspace complexity factors are generally categorized into three main groups: aircraft-related, conflict-related, and airspace-related factors 3). Aircraftrelated factors typically include the number of air40-04 第72回土木計画学研究発表会・講演集 - 40-04 - 4 craft within a sector and their operational characteristics, such as altitude, heading, speed, and relative spacing. Airspace-related factors describe the structural and procedural characteristics of the sector, including geometry, route configuration, and traffic flow organization. Conflict-related factors capture the dynamic interactions among aircraft, such as conflict density, convergence angles, temporal proximity of conflicts, and the predicted number of potential conflicts within a future time horizon (Figure 4). In this study, we further incorporate aircraft intent, defined as the planned conflict resolution maneuver or strategy to be executed in the aircraft’s future trajectory. Aircraft intent is treated as an independent complexity indicator under the conflict-related category, as it can substantially influence controller workload and traffic predictability depending on the nature and coordination of the applied resolution strategy. Fig. 4 Complexity metrics can be calculated from the current aircraft positions. Based on trajectory predictions, future complexity metrics can also be estimated. Adapted from 19). (5) Machine learning algorithms This research investigates the use of various machine learning algorithms, including Multiple Linear Regression, Random Forest, XGBoost, and Deep Neural Network, to estimate aircraft trajectories during conflict resolution maneuvers. The selection of these algorithms is based on the aim of exploring a diverse range of machine-learning methods with distinct strengths. Multiple Linear Regression is chosen for its simplicity, interpretability, and effectiveness in capturing linear relationships between features and the target variable. Random Forest is a robust and versatile machine learning algorithm capable of efficiently handling complex datasets, capturing non-linear relationships, and providing accurate predictions through the aggregation of multiple decision trees. XGBoost further refines predictive accuracy by iteratively refining errors from preceding models, thus capturing intricate data relationships. Lastly, Deep Neural Network is included for their proficiency in learning complex data relationships through interconnected neuron layers, excelling particularly in capturing non-linear and high-dimensional patterns. Among these algorithms, various types of artificial neural networks are commonly employed for trajectory prediction across different phases of flight 8), 20), 21). Other algorithms, such as Linear Regression 22) and Random Forest 23), are also utilized in different studies. In the scope of the proposed research, the models will be trained and evaluated with a randomized data split, allocating 75% for training and reserving the remaining 25% for model assessment. The hyperparameters of the models will be optimized through a combination of grid search and random search, exploring a defined set of hyperparameter configurations. To address overfitting and ensure optimal parameter selection, a 10-fold cross-validation approach will be employed on the training dataset. After the training phase on the dataset, the overall prediction performance of each machine learning model will be evaluated using various metrics, including Mean Absolute Error, Root Mean Squared Error, and Mean Absolute Percentage Error. (6) HITL simulations In this study, we aim to conduct HITL simulation experiments using the Escape Light simulator to validate complexity predictions. This simulator has been utilized in several previous research studies. For example, Sekine et al. 24) validated the operational feasibility and effectiveness of an en-route arrival manager through HITL simulation experiments conducted using the Escape Light simulator. Guleria et al. 25) conducted experiments using this simulator to predict air traffic controllers’ conflict resolution preferences through various machine learning models. Özdemir et al. 26) utilized the Escape Light simulator to evaluate a genetic algorithm-based decision support system for conflict resolution through HITL simulations. 3. EXPECTED OUTCOMES AND DISCUSSION The proposed machine learning models are expected to provide accurate and timely predictions of airspace complexity by effectively integrating aircraft intent information. Compared to conventional methods based solely on trajectory data, the inclusion of intent is anticipated to enhance prediction reliability, especially under dynamic traffic conditions. Reliable complexity forecasts could enable more proactive air traffic management, improved controller workload balancing, and optimized sector capacity 40-04 第72回土木計画学研究発表会・講演集 - 40-04 - 5 planning. Such predictions may also support decision-support tools for dynamic airspace configuration. The accuracy of the predictions will depend on the quality of intent data and the fidelity of the trajectory prediction models. Future work will focus on testing the framework across diverse traffic scenarios and validating its operational applicability through HITL simulations. Overall, the expected findings are anticipated to demonstrate the feasibility and potential of integrating aircraft intent into machine learning–based complexity prediction frameworks, paving the way for more adaptive and data-driven air traffic management systems. ACKNOWLEDGMENTS: The authors gratefully acknowledge the Matsumae International Foundation Fellowship Program for its support of this research and for funding the presentation of this work at the conference. REFERENCES 1) ICAO. : Report of the Asia/Pacific Area Traffic Forecasting Group. Technical Report, Montreal: ICAO, 2012. 2) Mogford, R. H., Guttman, J. A., Morrow, S. L., & Kopardekar, P. : The complexity construct in air traffic control: A review and synthesis of the literature (DOT/FAA/- CT TN95/22). Atlantic City, NJ: FAA, 1995. 3) Göppel, S. B., Schultz, M., & Itoh, E. : Complexity Metrics for Moving Sectors. IEEE Access, 2024. 4) Stein, E. S. : Air traffic controller workload: An examination of workload probe (No. DOT/FAA/CT-TN84/24). United States. Department of Transportation. Federal Aviation Administration. William J. Hughes Technical Center, 1985. 5) Jumad, A. S. B., Tominaga, K., Yi, C. X., Duong, V. N., Itoh, E., & Schultz, M. : Flow-centric air traffic control: Human in the loop simulation experiment. 2023 IEEE/AIAA 42nd Digital Avionics Systems Conference (DASC) (pp. 18). IEEE, 2023. 6) Itoh, E., Brown, M., Senoguchi, A., Wickramasinghe, N., Fukushima, S. : Future arrival management collaborating with trajectory-based operations, Air Traffic Management and Systems II: Selected papers of the 4th ENRI international workshop, 2015, pp. 137–156, 2017. 7) EUROCONTROL. : AI/ML Based Augmented 4D Trajectory. Executive Report, Issue 5.4, Edition 1.0, pp. 1–14, 2021. 8) Ma, L., Meng, X., Wu, Z. : Data-Driven 4D Trajectory Prediction Model Using Attention-TCN-GRU, Aerospace, Vol. 11, No. 4, pp. 313, 2024. 9) Singh, J., Goyal, G., Ali, F., Shah, B., Pack, S. : Estimating fuel-efficient airplane trajectories using machine learning, Computers, Materials and Continua, Vol. 70, No. 3, pp. 6189, 2022. 10) Li, B., Du, W., Zhang, Y., Chen, J., Tang, K., & Cao, X. : A deep unsupervised learning approach for airspace complexity evaluation. IEEE Transactions on Intelligent Transportation Systems, 23(8), 11739-11751, 2021. 11) Moreno, F. P., Rodríguez, F. I., Comendador, V. F. G., Jurado, R. D. A., Suárez, M. Z., & Valdés, R. M. A. : Prediction of air traffic complexity through a dynamic complexity indicator and machine learning models. Journal of Air Transport Management, 119, 102632, 2024. 12) EUROCONTROL. : EUROCONTROL simulation capabilities and platform for experimentation, 2024, https://www.eurocontrol.int/simulator/escape. 13) FAA. : NextGen Implementation Plan. Technical Report, Washington, DC: FAA, 2012. 14) EUROCONTROL. : European ATM Master Plan, 2022. 15) Qian, X., Mao, J., Chen, C.-H., Chen, S., Yang, C. : Coordinated multi-aircraft 4D trajectories planning considering buffer safety distance and fuel consumption optimization via pure-strategy game, Transportation Research Part C: Emerging Technologies, Vol. 81, pp. 18–35, 2017. 16) Tran, P. N., Nguyen, H. Q., Pham, D. T., & Alam, S. : Aircraft trajectory prediction with enriched intent using encoder-decoder architecture. IEEE Access, 10, 17881-17896, 2022. 17) Georgiou, H., Pelekis, N., Sideridis, S., Scarlatti, D., Theodoridis, Y. : Semantic-aware aircraft trajectory prediction using flight plans, International Journal of Data Science and Analytics, Vol. 9, No. 2, pp. 215–228, 2020. 18) Alligier, R. and Gianazza, D. : Learning aircraft operational factors to improve aircraft climb prediction: A large scale multi-airport study, Transportation Research Part C: Emerging Technologies, Vol. 96, pp. 72–95, 2018. 19) Schultz, M., Tominaga, K., Itoh, E., & Duong, V. N. : Introduction of moving sectors for flow-centric airspace management. SESAR Innovation Days (SIDs), 2023. 20) Hashemi, S. M., Botez, R. M., Grigorie, T. L. : New reliability studies of data-driven aircraft trajectory prediction, Aerospace, Vol. 7, No. 10, pp. 145, 2020. 21) Gallego, C. E. V., Comendador, V. F. G., Carmona, M. A. A., Valdés, R. M. A., Nieto, F. J. S., Martínez, M. G. : A machine learning approach to air traffic interdependency modelling and its application to trajectory prediction, Transportation Research Part C: Emerging Technologies, Vol. 107, pp. 356–386, 2019. 22) Alligier, R., Gianazza, D., Durand, N. : Machine learning and mass estimation methods for ground-based aircraft climb prediction, IEEE Transactions on Intelligent Transportation Systems, Vol. 16, No. 6, pp. 3138–3149, 2015. 23) Pham, D.-T. : Machine learning-based flight trajectories prediction and air traffic conflict resolution advisory, PhD Thesis, Université Paris Sciences et Lettres, 2019. 24) Sekine, K.; Iwata, D., Bouchaudon, P., Tatsukawa, T., Fujii, K., Tominaga, K., Itoh, E. : Validating Flow-Based Arrival Management for En Route Airspace: Human-In-The-Loop Simulation Experiment with ESCAPE Light Simulator, Aerospace 2024, 11, 866. https://doi.org/10.3390/aerospace11110866. 25) Guleria, Y., Pham, D. T., Alam, S., Tran, P. N., & Durand, N. : Towards conformal automation in air traffic control: Learning conflict resolution strategies through behavior cloning. Advanced Engineering Informatics, 59, 102273, 2024. 26) Özdemir, M., Usanmaz, Ö., Bouchaudon, P., & Gorgues, N. : A Decision Support System for Conflict Resolution: Interfacing a Genetic Algorithm with an Air Traffic Control Simulator. 2025 Integrated Communications, Navigation and Surveillance Conference (ICNS), IEEE, 2025. 40-04 第72回土木計画学研究発表会・講演集 - 40-04 -