| Summary: | The rise in air travel has drawn significant attention to reducing congestion, mitigating environmental impact, and ensuring safety and efficiency. However, growing air traffic has led to an increase in flight delays, causing environmental, economic, operational, and passenger-related issues. Although COVID-19 temporarily reduced delays due to decreased traffic, many large hubs have now returned to pre-pandemic levels, with strong growth expected in the coming years.
Air traffic management divides the airspace into three sections: the airport control zone (taxiing, takeoff, landing), the terminal manoeuvring area (TMA, for climbs and descents), and the en-route airspace (cruise). The TMA, a critical link between en-route airspace and the airport, is a major source of airborne delays. These delays stem from variable factors (e.g., weather) and structural factors (e.g., airspace structure, traffic demand). The TMA’s rigid structure, characterized by standard departure and arrival routes and their corresponding constraints, and its heavy traffic flow during climbs and descents, lead to congestion. Therefore, efficient TMA management requires seamless integration with en-route and runway operations, as they are interdependent.
Despite significant advancements achieved through measures like Departure Manager (DMAN), Arrival Manager (AMAN), Ground Delay Program (GDP), Air Traffic Flow Management (ATFM), Reduced Vertical Separation Minima (RVSM), and Wake Turbulence Re-Categorization (RECAT), the challenge of managing escalating air traffic within TMA persists. These efforts focus on departures and en-route operations, overlooking TMA congestion and arrival stream management. As a result, arrival flights face inefficiencies, leading to prolonged holding, delays, and increased fuel use in the TMA. The Extended Arrival Manager (E-AMAN) aims to reduce delays and holding times by adjusting aircraft cruising speeds during en-route phases, ensuring efficient traffic flow through congested TMAs. However, challenges like limited operational horizons, delay prediction accuracy, and understanding runway processes persist. This thesis addresses these challenges using machine learning to optimize runway and TMA operations.
As such, this thesis aims to address the following research questions: (RQ1) How can runway efficiency be increased by better prediction and reduction of runway occupancy time? and (RQ2) How does the concept of Extended AMAN improve the efficiency of TMA processes?
The first research question is further divided into the following two sub-research questions: (RQ1.1) What is the causal effect of aircraft landing parameters (e.g. speed and glideslope angle) on runway occupancy time? and (RQ1.2) To what extent can an explainable prediction model explain how the model predicts ROT? How is the prediction accuracy affected by the model's explainability?
This thesis addresses two sub-research questions of RQ1: reducing Runway Occupancy Time (ROT) to improve runway efficiency and accurately predicting ROT for reliable arrival flow information. Using causal machine learning, the impact of landing parameters on ROT was analyzed, showing that reducing aircraft speed during the final approach led to shorter ROT, and flexible glide-slope angles improved performance. For ROT prediction, an explainable decision tree model, when combined with relevant features, performed as well as or better than more complex algorithms.
The second research question on TMA optimisation is further broken down into the following three sub-research questions: (RQ2.1) How does including holding time as a feature and increasing the prediction horizon of TMA delay prediction affect the prediction performance of the prediction model?, (RQ2.2) How can delay prediction models be dynamically integrated with the speed control strategy to accommodate the dynamic and uncertain nature of the air traffic environment and (RQ2.3) How much economic and environmental benefits in terms of reduction in fuel consumption and $CO_2$ emissions can be reaped from the E-AMAN concept?
For RQ2.1, the thesis develops a data-driven model using machine learning to forecast TMA delays at 500NM, 400NM, and 300NM from the airport. The model integrates a holding detection system with a baseline TMA transit time model and includes holding time in the delay prediction. Results confirm that incorporating holding time improves prediction accuracy, and the models show consistent performance across all prediction horizons without significant differences.
To address RQ2.2 and RQ2.3, the thesis proposes implementing an extended arrival manager (E-AMAN) using the TMA delay prediction model and a meta-heuristics optimization algorithm to transfer delays from the TMA to en-route by adjusting aircraft cruising speed at 500NM, 400NM, and 300NM from the airport. This adaptable speed control strategy is validated in Singapore airspace using historical air traffic data. Results show that $65\%$ of TMA delays can be shifted to the cruise phase, saving 1.5 tonnes of fuel and reducing carbon dioxide emissions by 4.8 tonnes for one day’s arrivals. The findings indicate that implementing an extended arrival manager in Singapore and Southeast Asia can effectively manage TMA congestion, reduce fuel consumption, and lower carbon dioxide emissions.
To further address RQ2.2 and enhance the adaptability of the speed control strategy, the thesis explores the application of reinforcement learning (RL) techniques. By framing the speed control problem as a Markov Decision Process (MDP), RL agents learn to make sequential speed decisions based on real-time environmental feedback. The agents are trained over 1 million steps using the Proximal Policy Optimization (PPO) algorithm and evaluated through 100 simulations against two baseline scenarios: no action and random action. Results show that the trained policy consistently outperforms both baselines in minimizing delays.
Overall, this thesis presents a comprehensive approach to optimising TMA and runway to reduce airborne delays in the TMA. The proposed solutions, including the data-driven ROT optimization and prediction models, the TMA delay prediction model with holding prediction integration, the optimisation framework for the speed control strategy of E-AMAN, and a Reinforcement Learning (RL) framework for adaptive speed control. Together, these approaches demonstrate the potential to significantly improve the efficiency and sustainability of the ATM system.
|
|---|